In arid and semi-arid regions, pressure on groundwater resources has reached critical levels. Long-term over-pumping has depleted many aquifers, and climate change is intensifying this process. Rising temperatures increase evaporation from rivers and reservoirs, reducing the amount of surface water available for infiltration and natural recharge. Under these conditions, the use of surface water during periods of availability and its storage underground represents a key mechanism of managed aquifer recharge, effectively avoiding evaporation losses.

In this study, a practical framework is developed and tested to identify feasible ways to transfer accumulated surface water toward stressed aquifers. Rather than relying on complex ranking approaches, the locations of existing water infrastructure specifically wells and traditional khettara systems are used as reference points. These features indicate where aquifers are accessible and provide realistic spatial anchors for planning recharge at the regional scale.

The method combines satellite imagery to map surface water, geographic information systems (GIS) to identify cost-effective transfer pathways across the landscape, and multi-objective optimization to evaluate trade-offs between competing objectives. Feasibility is assessed through a cost function that accounts for terrain slope, elevation differences, transfer distance, pumping energy requirements, infrastructure costs, and potential water treatment needs.

The approach is applied to the Draa Oued Noun Basin in southern Morocco, a region strongly affected by water scarcity, high evaporation rates, and declining groundwater levels. Several surface water sources are examined, and feasible conveyance routes toward aquifers supplying key wells and khettara systems are identified.

The results show substantial variations in cost between water sources. Available water volume, transfer distance, and especially elevation lift emerge as the main cost drivers. Trade-off analysis helps identify the most cost-effective projects under limited budgets. The results also highlight opportunities for cost reduction: where gravity-driven transfer is possible, costs are significantly lower, and where pumping is required, solar energy offers a viable option for reducing long-term operational expenses.

Overall, this work provides a spatially explicit and realistic basis for planning artificial groundwater recharge, while respecting economic constraints and supporting sustainable groundwater management in highly water-stressed regions.

How to cite: Fri, R., Scozzari, A., Haida, S., Kili, M., Chao, J., Mridekh, A., and El Mansouri, B.: A Multi-Objective Cost Minimization Framework for Managed Aquifer Recharge Integrating Pareto Optimization and Least-Cost Path Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5340, https://doi.org/10.5194/egusphere-egu26-5340, 2026.

Modern Earth system models increasingly hit I/O limits—not only in performance, but also in reproducibility, maintainability, and developer productivity. As data volumes and workflows evolve, tightly coupled, file-centric I/O approaches can become hard to scale and hard to extend.

We present the design and lessons learned from introducing an asynchronous, modular I/O server concept in the Modular Earth Submodel System (MESSy). I/O operations were decoupled from the Fortran-based scientific core and implemented as separate Python services, where the communication between the two components was implemented using the Yet Another Coupler (YAC) library. This architecture was chosen to improve flexibility and long-term maintainability, while enabling heterogeneous workflows and evolving storage backends.

Using MESSy as a case study, we discuss practical decision criteria for selecting an I/O concept in large models (e.g., scaling behavior, accessibility for developers, testing and CI strategies, and reproducibility).  We conclude with lessons learned from bridging Fortran and Python communities and from lowering entry barriers for user-developers in a large modeling system.

How to cite: Mitic, A., Jöckel, P., Kerkweg, A., Hartung, K., Kern, B., and Hanke, M.: Choosing an I/O approach for Earth system models: lessons learned from a modular I/O server for MESSy, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11154, https://doi.org/10.5194/egusphere-egu26-11154, 2026.

Modern digital technologies and geoinformatics have experienced rapid growth, offering powerful tools to bridge the gap between scientific communities and society in landscape assessment and mapping. This research details the application of a crowdsourcing scheme that utilizes a dedicated mobile application to facilitate direct public participation in quantifying perceptions of urban landscapes and architecture. Initially developed as an educational tool, the methodology has been tested by university students across Italy, Greece, and France, providing a foundational phase for assessing landscape quality and urban typologies. Building upon these educational pilot studies, the work explores the evolution of this methodology into a broader, multicultural citizen science initiative designed to improve the quality and quantity of available landscape perception data.

A significant technical advancement in this research involves the integration of automated image analysis to process the novel data generated by participants from any location. The photographic material was examined using stochastic image analysis based on climacograms, in which images are treated as two-dimensional grayscale intensity fields and analyzed across multiple spatial scales. The method enables the comparison of image patterns based on the visual complexity of the uploaded photographs. A primary challenge addressed was the algorithm's performance when processing real-world, non-curated smartphone images. The analysis began an assessment on how the methodology handles environmental noise, such as sky, trees, and unconventional capture angles, which are inherent to bottom-up crowdsourcing schemes.

The early results indicate that the method can reveal group-level tendencies associated with differing architectural characteristics, particularly in relation to visual complexity, while not supporting reliable classification at the level of individual image. In detail, the findings indicate a trend towards two categorizations: firstly, between modernist-type movements, characterized by minimal elements, and secondly between eclectic or decorative movements, which exhibited higher measured complexity; however, this this behaviour was not observed universally on all analyzed movements The stochastic analysis also indicated theoretical overlaps between certain movements, such as Postmodernism and Eclecticism, based on shared decorative patterns. While the results highlight that environmental factors can influence the analysis of individual photographs, the method utilized presents potential for distinguishing movement trends with logical consistency even from unfiltered data.

Scientifically, this yield of quantitative data sets the groundwork for improved research in the humanities and culture, showing a strong correlation with established landscape quality indices. Socially, the project provides a scalable model for participatory mapping that fosters critical thinking about urban quality, creating new conditions for communication between universities and the broader public. Overall, the presented work reports on the early-stage results of this methodological exploration and aims to evaluate the combined use of participatory mobile data collection and exploratory image-based analysis for landscape and architectural studies, while identifying key challenges related to data quality, interpretation, and future methodological refinement.

How to cite: Kopelia, S., Tepetidis, N., Tzortzi, J. N., Sargentis, G.-F., and Ioannidis, R.: Integrating Participatory Perception-Mapping Data and Stochastic Image Analysis for Urban Landscape Assessment, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13270, https://doi.org/10.5194/egusphere-egu26-13270, 2026.

Accurate monitoring of land cover is essential for sustainable environmental management and urban planning in arid regions. However, rapid changes in land use often make it difficult to distinguish between different surface types, such as urban areas and bare soil, using standard satellite data alone. This research examines land-use changes in the Bahr Qarun district of Fayoum, Egypt, during 2019, 2021, and 2023. The study used Sentinel-2 and Landsat OLI 8 satellite images taken each April to ensure data consistency. We applied the Maximum Likelihood (ML) method to classify Sentinel-2 images. They used 30 training samples for each land category to guide the process. The results achieved a Kappa coefficient above 75%, indicating a reliable level of accuracy. We measured vegetation using the Normalized Difference Vegetation Index (NDVI) and urban areas using the Normalized Difference Built-up Index (NDBI). A comparative analysis revealed that NDVI results were closely aligned with those obtained from supervised classification, reflecting its strong capability in accurately identifying vegetated areas. In contrast, NDBI exhibited a tendency to overestimate urban extent, primarily due to spectral confusion between built-up surfaces and bare soil within individual pixels. The study concludes that NDVI is an effective tool for mapping the green cover in this area.

Keywords: Land Cover Change, Sentinel-2, Landsat OLI 8, Supervised Classification,  Spectral Indices (NDVI & NDBI), Bahr Qarun, Egypt.

How to cite: Elsehsah, A., Negm, A., Ashour, E., and Elsahabi, M.: Monitoring Land Cover Dynamics in Bahr Qarun District, Egypt, via Remote Sensing Data , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13783, https://doi.org/10.5194/egusphere-egu26-13783, 2026.

Satellite-Derived Bathymetry (SDB) offers a cost-effective alternative to traditional shipborne surveys for mapping large coastal areas. This technique utilizes optical remote sensing data from multispectral sensors to estimate water depth. The fundamental principle relies on the behavior of light as it travels through the water column; as depth increases, light intensity decreases due to absorption and scattering. Different wavelengths penetrate to varying degrees, with blue light reaching the greatest depths while red light is absorbed quickly. By analyzing these spectral features, researchers can calculate underwater topography. Currently, SDB techniques are categorized into two primary groups: physically based (analytical) models, which simulate light propagation without needing local in-situ depth calibration, and statistical (empirical) models, which correlate satellite data with known depth measurements from nautical charts, ship-based acoustic surveys or airborne LiDAR.

While both approaches provide extensive spatial coverage at a lower cost, they are generally limited to clear, shallow waters, typically reaching depths of less than 20 meters. Analytical models are highly accurate but complex and data-intensive, whereas empirical models are more accessible but rely heavily on the quality of reference data. Recent advancements in machine learning have significantly improved the automation and performance of these empirical methods. This study evaluates the core concepts, advantages, and limitations of various SDB approaches, with a focus on Landsat-8 and Sentinel-2 data. Furthermore, the research details essential processes for empirical model calibration, validation, and detecting model bias. The findings emphasize that rigorous evaluation and bias correction are critical for ensuring the reliability of depth data in diverse coastal environments.

Keywords: Satellite-Derived Bathymetry, Remote Sensing, Empirical Models, Stumpf Algorithm, Coastal Waters, Model Bias Detection and Correction.

How to cite: Abdalla, M. H., Elhalawany, H., Abdelrahman, S. M., Negm, A., and Scozzari, A.: Monitoring Shallow Water Depths: A Review of Satellite-Derived Bathymetry Methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13852, https://doi.org/10.5194/egusphere-egu26-13852, 2026.

Significant heterogeneity in metadata schemas, vocabularies, and ontologies hinders the discovery, reuse, and integration of European environmental data infrastructures across national and disciplinary boundaries. Recent initiatives have identified semantic interoperability as a vital enabler of FAIR data flows between infrastructures, paving the way for sophisticated, AI-driven, large-scale analyses.

Powered by OntoPortal technology, EarthPortal is a specialised catalogue of semantic resources (ontologies, thesauri and controlled vocabularies) for Earth and environmental sciences. It provides navigation, multi-ontology searching, mapping management, text annotation and recommendation services via web interfaces and REST APIs. These support data catalogues and repositories in an interoperable way.

EOSC LUMEN builds an interoperable discovery ecosystem across multiple domains (including Earth System Science, Social Sciences and Humanities, and Mathematics) to enable cross-platform search and meaningful reuse across communities. Rather than focusing only on metadata aggregation, LUMEN targets the practical enablers of interoperability that make resources discoverable and machine-actionable across infrastructures.

LUMIS (LUMEN Infrastructure for Semantics) is the shared semantic layer of LUMEN. It supports the end-to-end lifecycle of semantic artefacts (ontologies and controlled vocabularies, including SKOS resources) from scoping and requirements to implementation, publication and long-term maintenance. LUMIS focuses on governance, provenance, versioning and quality checks, while adopting an integration-first strategy: it connects and orchestrates established community tools (deployed services and/or API-based components) into coherent workflows, so that semantic resources can be created, aligned, validated and delivered in reusable forms for discovery platforms.

Integrating EarthPortal into LUMIS links a domain-specific semantic catalogue to a cross-domain discovery ecosystem. This enables repositories to annotate metadata using EarthPortal resources, while making use of LUMIS’s lifecycle-driven workflows and FAIR-aligned governance and quality checks.

In this presentation, we will demonstrate how integrating EarthPortal into the LUMIS platform supports more consistent semantic interoperability and FAIR-aligned practices across European Earth System Science infrastructures. We will showcase practical data workflows to enhance interdisciplinary research.

How to cite: Homo, J., Pierkot, C., Darty, K., and Allem, H.: Operationalising Semantic Interoperability for Cross-domain Discovery with LUMIS, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19784, https://doi.org/10.5194/egusphere-egu26-19784, 2026.

Modern Earth system research relies on integrating heterogeneous datasets such as reanalysis, satellite observations, in situ measurements, climate model ensembles, and reforecasts, yet these data are often stored in fragmented, inconsistent, and difficult to reuse forms. This limits reproducibility, slows modelling workflows, and constrains the development of operational digital twins for water and climate risk management.

This contribution presents a scalable, FAIR aligned data lake architecture implemented on the EVE high performance computing environment. The system transforms a large, unstructured source pool of more than two million files into a curated, duplication free, metadata rich repository designed for hydrological modelling, machine learning, and climate analytics. The architecture follows a four stage lifecycle: raw, curated, database ready, and ancillary GIS layers, reflecting data governance practices used by major climate centres.

A reproducible ingestion workflow classifies, deduplicates, and standardizes datasets from ERA5, ERA5 Land, MERRA 2, PRISM, E OBS, GPM IMERG, CMIP6, ISIMIP3, ECMWF reforecasts, MODIS, CHIRPS, GFED, GRDC, GSIM, and other sources. A Python based metadata extractor, built on CF convention standards, automatically captures variables, units, dimensions, spatial resolution, temporal coverage, coordinate reference systems, and checksums. Metadata are stored both as dataset level JSON and as a global inventory, enabling transparent provenance tracking and rapid dataset discovery.

The curated data hub is implemented under /data/db/earth_system and organized by scientific domain, temporal resolution, spatial extent, and processing stage. The system supports SLURM based workflows, HPC native processing, and cloud optimized formats such as Zarr.

This work demonstrates how a single researcher can design and operationalize a modern, HPC native data infrastructure that accelerates hydro climate research and forms the backbone of an emerging Digital Hydro Twin. The approach is transferable to institutions seeking to modernize their data ecosystems and improve reproducibility in environmental modelling.

How to cite: Amin, B., Zscheischler, J., Samaniego, L., Peng, J., García-García, A., and Harzendorf, T.: A Scalable, FAIR‑Aligned Data Lake Architecture for Earth System Modelling: From Heterogeneous Raw Archives to Curated, Metadata‑Rich, Analysis‑Ready Climate Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20391, https://doi.org/10.5194/egusphere-egu26-20391, 2026.

The Upper Senegal River Basin is a strategic water resource system supporting agriculture, hydropower generation, and essential ecosystem services in West Africa. However, a comprehensive understanding of its hydrological dynamics remains constrained by the limited availability of in situ hydroclimatic observations. This study applies the Soil and Water Assessment Tool (SWAT) to simulate hydrological processes in the basin, with a particular emphasis on the influence of precipitation data sources on model performance and uncertainty. Hydrological simulations were conducted at six representative gauging stations (Bakel, Kayes, Gourbassy, Oualia, Bafing Makana, and Daka Saidou) over the period 1983–2021, using a combination of ground-based observations, satellite precipitation products, and reanalysis datasets (ERA5, MERRA-2, PERSIANN, and CHIRPS). Model calibration demonstrated satisfactory performance, with Nash–Sutcliffe Efficiency (NSE) values reaching up to 0.74 at upstream stations, while reduced performance was observed downstream. Validation results showed a moderate decline in model efficiency, highlighting the sensitivity of SWAT outputs to precipitation inputs and data uncertainty. The comparative analysis of precipitation datasets reveals substantial variability in simulated streamflow and water balance components, underscoring the importance of precipitation data selection in data-scarce regions. These findings highlight the need for robust, multi-source hydroclimatic data integration to improve hydrological modelling reliability and support informed water resource management decisions.

Keywords: Upper Senegal River, SWAT, Hydrological modelling, Precipitation uncertainty; Satellite rainfall; Reanalysis data.

How to cite: Boussabou, S. M., Taia, S., El Mansouri, B., Kebd, A., Mohamedou Idriss, A., Legsabi, H., and Erraioui, L.: Hydrological Modelling of the Upper Senegal River Basin Using SWAT: Assessing the Impact of Multi-Source Precipitation Data on Model Performance, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21793, https://doi.org/10.5194/egusphere-egu26-21793, 2026.

The establishment of the European Open Science Cloud (EOSC) places renewed emphasis on the role of national e-infrastructures in enabling standards-based, interoperable, and reusable research workflows in the EU. Within the context of Ireland’s EOSC Node, there is particular interest in demonstrating how European-scale open-data services can be digested by national research clouds, transformed into analysis-ready assets, and made available for both open research and applied industry use-cases. Earth Observation (EO) provides a strong test case, given the volume and complexity of the data involved, and its growing role in scalable environments that support operational decision-making.

This pilot project, QuarryLink, presents a Phase-1 study focused on building a reproducible EO data ingestion workflow that connects the Copernicus Data Space Ecosystem with the HEAnet Research Cloud, operating on the SURF Research Cloud platform. Through a real-world quarry case-study in the Dublin region (Ireland), the work demonstrates how EOSC-aligned principles, including auditable machine-readable workflows, can be applied from the outset of the EO research process. We will demonstrate how precise spatial boundaries can be defined and validated; how modern OAuth-based authentication mechanisms can be integrated into research cloud workflows; and how Sentinel-2 Level-2A products can be programmatically discovered, retrieved, and prepared for downstream analysis using current Copernicus services.

By executing the ingestion workflow on the HEAnet Research Cloud using open-source geospatial tooling, the pilot aims to establish an analytics-ready foundation for working with Sentinel-2 data in a reproducible research cloud environment. The resulting data products are structured to support downstream analysis, with compute resources accessed dynamically through the HEAnet Research Cloud workspace as required. Building on this foundation, Phase 2 will focus on developing time-series analyses, EO data cubes, and derived environmental indicators to support both research-driven investigation and applied monitoring scenarios in European quarry environments.

More broadly, the pilot seeks to illustrate how EOSC-aligned integration across data ingestion and compute layers can support open research practices while enabling scalable, real-world EO-enabled industrial applications, providing a practical pathway for national EOSC Nodes to translate open data into shareable analytics and societal impact.

How to cite: Neff, F., Sabatino, R., Arreba, A., and Sweeney, J.: An EOSC Node Ireland Pilot Study: Bridging European and National e-Infrastructures for Reproducible Sentinel-2 Data Ingestion in Quarry Applications, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21965, https://doi.org/10.5194/egusphere-egu26-21965, 2026.

Groundwater in arid regions is highly sensitive to human activity, especially when untreated wastewater interacts with shallow aquifers. This study evaluates the hydrogeochemical response of the Kima aquifer in Aswan, Egypt, following the Kima Drain Covering Project. The research uses an integrated framework of field measurements, geospatial analysis, and multi-criteria decision-making. The team analyzed groundwater samples from 2020 and 2025. They tested eleven physicochemical parameters and six irrigation indices. Spatial interpolation through Inverse Distance Weighting (IDW) helped map temporal variations and identify contamination hotspots. To classify water suitability, the study standardized values according to WHO and Egyptian guidelines. The Analytical Hierarchy Process (AHP) was used to determine the importance of various drinking and irrigation indicators. Finally, a Weighted Linear Combination (WLC) generated composite Groundwater Quality Index (GWQI) maps. The results show a significant improvement in groundwater quality after the drain was covered. Levels of TDS, chloride, sulfate, sodium, and magnesium decreased substantially across the area. The ionic balance shifted toward a more favorable calcium-magnesium-bicarbonate facies. Irrigation indices also improved, with most parameters falling into safe or excellent ranges. The 2025 GWQI map reveals a transition from "good–permissible" to "excellent–safe" zones. This confirms that eliminating direct seepage from the drain had a positive environmental impact. This integrated AHP–GIS–IDW approach is an effective tool for monitoring groundwater changes. It provides a robust decision-support system for managing water resources in arid urban environments.

How to cite: Khairy, M., S. Nour-Eldeen, A., Hossen, H., Abd-Elaty, I., and Negm, A.: Monitoring Groundwater Quality and Improvement in the Kima Area, Aswan, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22084, https://doi.org/10.5194/egusphere-egu26-22084, 2026.

Hyperspectral image (HSI) classification often struggles with feature interference across different scales and the inherent challenges of data imbalance and sample scarcity. While deep learning models have significantly advanced the field, traditional single-branch architectures often suffer from scale-related noise, where features from different receptive fields interfere with one another. To address this, we propose the Multibranch Adaptive Feature Fusion Network (MBAFFN). Our approach utilizes three parallel branches to independently extract scale-specific features, effectively decoupling the multiscale information to prevent interference. This architecture is enhanced by two specialized modules: Global Detail Attention (GDA) for capturing broad contextual dependencies and Distance Suppression Attention (DSA) for refining local pixel-level discrimination. Furthermore, a pixel-wise adaptive fusion mechanism is introduced to dynamically weigh and integrate these features, prioritizing the most relevant scales for final classification. The performance of MBAFFN was validated on four benchmark datasets: Indian Pines (IP), Pavia University (PU), Longkou (LK), and Hanchuan (HC). Compared to current state-of-the-art methods, our model improved Overall Accuracy (OA) by 0.91%, 1.71%, 0.86%, and 3.16% on the IP, PU, LK, and HC datasets, respectively. The significant improvement on the HC and PU datasets underscores the model’s robustness in scenarios with limited training samples and complex class distributions. These results, supported by detailed ablation studies, demonstrate that adaptive fusion and scale-specific branching are effective strategies for mitigating feature interference in hyperspectral analysis.

How to cite: Li, C. and Du, B.: Multibranch Adaptive Feature Fusion for Hyperspectral Image Classification, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3080, https://doi.org/10.5194/egusphere-egu26-3080, 2026.

Characterizing the thermal state of urban surfaces is fundamental for mitigating the impacts of the Surface Urban Heat Island (SUHI) effect. This study presents an intensive in-situ thermal infrared monitoring campaign in the high-density urban core of Athens, Greece. Utilizing a calibrated handheld TIR sensor (7.5–14 μm), surface temperatures were recorded across strategic locations in the center of Athens during hot weather conditions. The methodology emphasizes the critical role of material-specific parameterization, where thermographic data were post-processed to account for emissivity (ε) variations and surface temperature, ensuring high-fidelity measurements.

Experimental results reveal extreme thermal stress, with maximum surface temperatures reaching 56.0°C on conventional paving materials, while the mean ambient air temperature was close to 35.0°C during peak solar hours (13:00–18:00LT). Spatial analysis and visualization of the results were performed using QGIS, correlating thermal signatures with urban geometry, shading conditions, and vegetation density. The aim of this study was to highlight the significant cooling potential of specific urban materials and nature-based solutions.

How to cite: Gkountaras, O., Georgakis, C., Velissaridis, T., and Assimakopoulos, M. N.: In-situ Thermal Infrared Monitoring in an Urban Area: A Case Study of Micro-scale Thermal Transitions during Hot Weather Conditions in Athens, Greece., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3363, https://doi.org/10.5194/egusphere-egu26-3363, 2026.

Landslides claim thousands of lives and cause billions in economic losses annually, with impacts disproportionately concentrated in developing regions across Asia, Africa, and Latin America. Paradoxically, the current trajectory of artificial intelligence in geohazard detection—characterized by billion-parameter foundation models requiring substantial computational infrastructure—risks widening, rather than closing, the gap between technological capability and operational deployment where it is needed most. We argue that this paradigm requires fundamental reconsideration, proposing domain adaptation on strategically curated geological datasets as a more equitable and effective path toward globally accessible landslide detection systems.

Foundation models like the Segment Anything Model (SAM), pre-trained on over one billion masks, demand computational resources—312 million parameters, 1,376 GFLOPs per inference, specialized GPU infrastructure—that remain inaccessible to disaster management agencies in resource-constrained regions. Beyond these practical constraints, we contend that the apparent generalization capabilities of such models reflect pattern coverage in training data rather than emergent understanding transferable to geological contexts. The SA-1B dataset, despite its scale, was not curated to systematically represent landslide morphological diversity, creating coverage gaps for rare failure types, unusual triggering mechanisms, and underrepresented terrain configurations precisely where robust detection is operationally critical.

Given these limitations, we propose that effective generalization for geological applications emerges not from architectural scale but from strategic coverage of domain-relevant pattern space. We developed and tested GeoNeXt, a lightweight architecture that exploits the hierarchical transferability of geological features through targeted domain adaptation. Low-level representations (edges, spectral gradients) transfer universally across sensors and terrain; mid-level patterns (drainage networks, slope morphology) require adaptation to local expressions; and high-level configurations (failure geometries, trigger signatures) demand targeted training. Our results showed that this approach outperformed SAM-based methods across three independent benchmarks while requiring 10× fewer parameters (32.2M versus 312.5M) and a 62% reduction in computational cost. Zero-shot transferability to geographically distinct test sites (74–78% F1 score) emerged from the training dataset's systematic morphological diversity rather than parameter count. Inference at 10.6 frames per second on standard hardware, versus 3.0 frames per second for foundation model alternatives, transforms theoretical capability into deployable technology for resource-constrained environments. These findings suggest that strategic domain adaptation, rather than architectural scale, offers the most viable path toward operational landslide detection in vulnerable regions.

How to cite: Uribe-Ventura, R., Viveen, W., Pineda-Ancco, F., and Beltrán-Castañon, C.: Democratizing landslide detection for vulnerable regions beyond resource-intensive foundation models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3619, https://doi.org/10.5194/egusphere-egu26-3619, 2026.

Numerical modeling is a fundamental tool for understanding physically driven processes in geosciences. In multiparametric settings, the Finite Element Method is widely used because it can accommodate irregular geometries and complex boundary conditions. However, this advantage critically depends on the quality of the computational mesh, which must faithfully represent geological features such as faults, stratigraphic interfaces, and wells. In practice, mesh generation remains a major bottleneck, requiring specialized expertise and significant manual effort. We present Geo2Gmsh, an automated, lightweight workflow built on Gmsh (Geuzaine & Remacle, 2009), that generates geological meshes directly from simple text‐based descriptions of topological elements, including surfaces, lines, and points. These elements correspond to geologically meaningful features, allowing users to define faults, horizons, wells, and domain boundaries in a transparent, reproducible, and solver‐independent way. The workflow is demonstrated using two contrasting case studies: (1) Ringvent, an active sill‐driven hydrothermal system in the Guaymas Basin, and (2) the Eastern Llanos Basin, a foreland basin in eastern Colombia. To evaluate solver compatibility, we solved the heat equation in SfePy (https://sfepy.org/doc-devel/index.html) using the Eastern Llanos Basin model as the computational domain. Although the simulation is illustrative and not calibrated to observations, it confirms that meshes produced by Geo2Gmsh can be readily incorporated into numerical solvers. By explicitly embedding wells, faults, and geological interfaces in the mesh, Geo2Gmsh enables boundary conditions to be applied directly to physically meaningful features and allows model outputs to be extracted along them, simplifying both model setup and post‐processing. Meshes can be exported in standard formats (e.g., VTK, MSH, and Exodus via meshio), ensuring broad interoperability. Overall, Geo2Gmsh provides a lightweight, scalable, and reproducible workflow that dramatically lowers the technical barrier to geological mesh generation. This contribution establishes a practical foundation for reproducible, open-source numerical modeling in geosciences, facilitating the integration of geological knowledge into high-fidelity computational simulations.

How to cite: Buitrago, H., Contreras, J., and Neumann, F.: Geo2Gmsh: A Scalable Workflow for Automated Mesh Generation of Geological Models Using Gmsh, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6022, https://doi.org/10.5194/egusphere-egu26-6022, 2026.

Many geoscientific datasets, such as those produced by climate and weather models, are stored in the NetCDF file format.  These datasets are typically very large and often strain institutional data storage resources. While lossy compression methods for scientific data have become more studied and adopted in recent years, most advanced lossy approaches do not work easily and/or transparently with NetCDF files. For example, they may require a file format conversion or they may not work correctly with “missing values” or “fill values” that are often present in model outputs.  While lossy quantization approaches such at BitRound and Granular BitRound have built-in support by NetCDF and are quite easy to use, such approaches are generally not able to reduce the data size as much as more advanced compressors (for a fixed error metric), like SPERR, ZFP, or SZ3.

We are particularly interested in reducing the data size of the CONUS404 dataset.  CONUS404 is a publicly available unique high-resolution hydro-climate dataset produced by Weather Research and Forecasting (WRF) Model simulations that cover the CONtiguous United States (CONUS) for 40 years at 4-km resolution (a collaboration between NSF National Center for Atmospheric Research the U.S. Geological Survey Water Mission Area). 

Here, we investigate one advanced lossy compressor, SPERR [1], together with its plugin for NetCDF files, H5Z-SPERR [2], in a Python-based workflow to compress and analyze CONUS404 data.  SPERR is attractive due to its support for quality control in terms of both maximum point-wise error (PWE) and peak signal-to-noise ratio (PSNR), enabling easy experimenting of storage-quality tradeoffs. Further, given a target quality metric, previous work has shown that SPERR likely produces the smallest compressed file size compared to other advanced compressors. It leverages the HDF5 dynamic plugin mechanism to enable users to stay in the NetCDF ecosystem with minimal to no change to existing analysis workflows, whenever a typical NetCDF file is able to be read. And, importantly for our work, the SPERR plugin supports efficient masking of “missing values,” which are common to climate and weather model output.  The support for missing values enables compression on many variables which are not naturally handled by other advanced compressors that rely on HDF5 plugins. Further, because H5Z-SPERR directly handles missing values, they can be stored in a much more compact format (and are restored during decompression), further improving compression efficiency. (Note that built-in NetCDF quantization approaches can work with missing values.) 

Our experimentation demonstrates the benefit of enabling advanced lossy (de)compression in the NetCDF ecosystem: adoption friction is kept at the minimum with little change to workflows, while storage requirements are greatly reduced.

[1] https://github.com/NCAR/SPERR

[2] https://github.com/NCAR/H5Z-SPERR

How to cite: Li, S., Baker, A., and Xue, L.: Application of advanced lossy compression in the NetCDF ecosystem for CONUS404 data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6232, https://doi.org/10.5194/egusphere-egu26-6232, 2026.

Investigations have been carried out into the initiation of the Pangu weather model, initiating the model with both ERA5 data (on which it was trained) and with the Met Office’s Global UM model data. There are many consistent local biases at ground level between these two sets of initial conditions. The geographically local biases are not dissipated by the Pangu model with timestep but instead remain geographically fixed and gradually decrease with lead time. Whilst the Pangu model initiated with UM initial conditions remains further from the ERA5 truth than the ERA5-initiated Pangu model at all timesteps, it initially moves towards the ERA5 truth with timestep, as the geographically static differences in initiation decrease, before moving further away from the ERA5 truth as differences in large-scale systems begin to dominate.

Also investigated was the difference between the Pangu model 24-hour timesteps and 6-hour timesteps; it was found that the 6-hour timesteps were better able to reduce the geographically static initial differences than the 24-hour timesteps.

If time permits, a similar analysis will be made of the FastNet and GraphCast models.

How to cite: Buttery, H.: Investigations into the Reaction of the Pangu ML Weather Model to Different Initial Conditions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7344, https://doi.org/10.5194/egusphere-egu26-7344, 2026.

Solar phenomena such as flares, coronal mass ejections (CMEs), and solar energetic particles (SEPs) are actively monitored and assessed for space weather hazards. In recent years, machine learning has demonstrated considerable success in solar flare forecasting. Accurate SEP forecasting remains challenging in space weather monitoring due to the complexity of SEP event origins and propagation. We introduce SEPNET, an innovative multi-task neural network that integrates forecasting of solar flares and CME summary statistics into the SEP prediction model, leveraging their shared dependence on space-weather HMI active region patches (SHARP) magnetic field parameters. SEPNET incorporates long short-term memory and transformer architectures to capture contextual dependencies. The performance of SEPNET is evaluated on the state-of-the-art SEPVAL SEP dataset and compared with classical machine learning methods and current state-of-the-art pre-eruptive SEP prediction models. The results show that SEPNET achieves higher detection rates and skill scores while being suitable for real-time space weather alert operations.

How to cite: Chen, Y., Yu, Y., Zhao, L., Whitman, K., Manchester, W., and Gombosi, T.: SEPNET: a multi-task deep learning framework for SEP forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11945, https://doi.org/10.5194/egusphere-egu26-11945, 2026.

Quantifying soil properties such as Soil Organic Carbon (SOC), texture, and Calcium Carbonate (CaCO₃) is essential for assessing soil health and ensuring food security. While Visible, Near Infrared, and Short Wave Infrared (VSWIR) remote sensing is a standard operational tool, the Longwave Infrared (LWIR, 8-14 μm) offer complementary information on mineralogy and moisture that are still not yet fully explored for this specific application. This study investigates the synergy between VSWIR and LWIR data that will be available with future hyperspectral satellite missions. Among them, the European Space Agency's Copernicus Expansion missions that will add to the EO capacity the Hyperspectral Imaging Mission for the Environment (CHIME) and Land Surface Temperature Monitoring (LSTM) mission. Alongside are the NASA's Surface Biology and Geology (SBG and SBG-TIR) missions.

The research focuses on Jolanda di Savoia (Italy), an agricultural landscape resulting from land reclamation projects in the late 19th century. Ground truth data were collected during a field campaign on June 22, 2023, providing 59 topsoil samples further analysed for SOC, texture, and CaCO₃. Field campaign was coincident with an airborne survey carried out with the LWIR Hyperspectral Thermal Emission Spectrometer (HyTES) sensor. HyTES captured data across 256 spectral bands from 7.5 to 11.5 μm, providing a pixel size of approximately 2.3 meters.

To evaluate the multi-frequency potential, we developed a workflow combining a soil composite from PRISMA (VSWIR) satellite time-series with simulated SBG-TIR (LWIR) data. The SBG-TIR simulation chain included as input a surface emissivity map derived from the airborne HyTES survey. To cover the LWIR wide spectral range (up to 12 µm), the emissivity spectrum was extended using an autoencoder neural network procedure trained on the ECOSTRESS Soil Spectral Library. Top-Of-Atmosphere (TOA) radiance was then simulated using the Radiative Transfer for the TIROS Operational Vertical Sounder (RTTOV-14) model, incorporating the optical depth and cloud/aerosol optical properties coefficients specific to SBG-TIR. Furthermore, these simulated data were atmospherically corrected to produce the target satellite emissivity products according to the TES algorithm.

Soil properties prediction models were developed using supervised machine learning algorithms. We benchmarked two scenarios: 1) the proposed combined approach using PRISMA and the simulated SBG-TIR L2 emissivity product; and 2) a VSWIR-only approach using PRISMA. A quantitative assessment by 10-fold cross-validation using common literature metrics (R², RMSE, RPD) highlighted the benefits of the multi-sensor approach. For SOC retrieval, the standalone VSWIR (PRISMA) model yielded an R² of 0.55 (RPD = 1.5), while the synergistic integration of PRISMA with simulated SBG-TIR data improved the retrieval accuracy, reaching an R² of 0.65 and increasing the RPD to 1.69. This work indicates that, on the agricultural test site of Jolanda di Savoia, the combined use of SVWIR and LWIR spectral range slightly improves the SOC retrieval. Further validation across diverse agricultural scenarios will be essential to test the real advantage of combining next-generation imaging spectroscopy missions.

How to cite: Rossi, F., Casa, R., Marrone, L., Mirzaei, S., Pascucci, S., and Pignatti, S.: Evaluating the combined potential of VSWIR and Thermal Infrared data for soil characterisation., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13611, https://doi.org/10.5194/egusphere-egu26-13611, 2026.

Accurate high resolution wind field prediction is essential for wind resource as-
sessment, renewable energy planning, and regional weather analysis. Although
Numerical Weather Prediction (NWP) models such as the Weather Research
and Forecasting (WRF) model provide physically consistent wind forecasts, their
outputs often suffer from systematic biases arising from uncertainties in surface
characteristics, simplified physical parameterizations, and resolution limitations.
Furthermore, increasing model resolution to the kilometer scale significantly
raises computational cost. To address these challenges, this study presents a
machine learning–based framework for bias correction of WRF-simulated wind
fields over the Southern Tamil Nadu region, with particular focus on the Mup-
pandal wind farm area.
An extensive validation of WRF configurations was first performed using mul-
tiple physics scheme combinations and domain setups, evaluated against ERA5
reanalysis data. The optimal configuration was identified and used to gener-
ate three years (2023–2025) of wind simulations at 3 km × 3 km resolution.
Significant biases were observed in the raw WRF outputs, motivating the appli-
cation of an Artificial Neural Network (ANN) based bias correction approach.
A Random Forest algorithm was employed for feature selection, followed by
Principal Component Analysis (PCA) to reduce dimensionality while retaining
95% of the variance. A feedforward neural network with multiple hidden layers
was trained to correct the U10 and V10 wind components, with the hyperbolic
tangent activation function yielding the best performance. The bias-corrected
wind fields exhibited substantial improvement in mean and extremes, achieving low error metrics and
strong correlation with ERA5 data.
The results demonstrate that combining physically based NWP simulations with
machine learning driven bias correction provides an accurate and computation-
ally efficient approach for generating high-resolution wind fields. This hybrid
framework offers significant potential for wind energy assessment and localized
meteorological applications in data-sparse regions.

How to cite: Pm, V. and Chakravarthy, B.: Bias Correction of Numerical Weather PredictionWind Fields in Southern Tamil Nadu RegionUsing Machine Learning Techniques, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16232, https://doi.org/10.5194/egusphere-egu26-16232, 2026.

Early laboratory experiments of shear flow by Thorpe (Thorpe, 2002) provided evidence of Kelvin-Helmholtz Instability (KHI) billow interactions either due to misaligned adjacent billow cores or varying phases along the adjacent billow axes. Similar evidence has been found in the observations of tropospheric clouds, airglow, and Polar Mesospheric Clouds (PMC) imagery data in the mesosphere. Initial High-Resolution Direct Numerical Simulations (DNS) studies performed at Reynolds Number of 5000 (Fritts et al., 2021a, Fritts et al., 2021b) have demonstrated the that misaligned KH billow cores exhibit strong and complex vortex interactions inducing ‘Tubes and Knots’ (T&K) structures (Thorpe, 2002). These T&K structures were observed to accelerate transition to small-scale turbulence in contrast to previously known notable transitional mechanisms such as secondary KHI and convective instabilities emerging in individual KH billows. Also, the KHI T&K dynamics evidently yield intense turbulence dissipation rates contrasting that of secondary KHI and convective instabilities in billow cores.

More recent high-resolution imaging of OH airglow (Hecht et al., 2021) provide concrete evidence of KHI billows with wavelength ranging between 7-10 km modulated by atmospheric Gravity Waves (GWs) of dominant horizontal wavelengths ∼ 30km and oriented orthogonal to KH billow axes and propagate along the billow cores which result in apparent T&K dynamics rapidly driving KH billow breakdown. Similar evidence has been found in recent PMC imaging. This is the central theme of the idealized DNS discussed in this talk.

We conducted DNS studies to demonstrate the turbulence energetics of KHI billow interactions when subject to modulations due to monochromatic atmospheric gravity waves of small perturbation amplitudes and intrinsic frequency of N/5 (where N is the background Brunt-Vaisala Frequency). Preliminary analyses of our DNS results indicate that GW modes with modest amplitudes promote KHI billow misalignments resulting in complex multi-scale T&K dynamics fixed at specific GW phases. An increase in the GW amplitude resulted in noticeable reduction of KHI billow wavelengths further promoting KH billow misalignments. The resulting turbulence is expected to consist of broader scale ranges of intense turbulence dissipation rate and diffusivity.

References
[Fritts et al., 2021a] Fritts, D. C., Wang, L., Lund, T. S., and Thorpe, S. A. (2021a). Multi-Scale Dynamics of Kelvin-Helmholtz Instabilities . Part 1 : Secondary Instabilities and the Dynamics of Tubes and Knots. pages 1–27.

[Fritts et al., 2021b] Fritts, D. C., Wang, L., Thorpe, S. A., and Lund, T. S. (2021b). Multi-Scale Dynamics of Kelvin-Helmholtz Instabilities . Part 2 : Energy Dissipation Rates , Evolutions , and Statistics. pages 1–39.

[Hecht et al., 2021] Hecht, J. H., Fritts, D. C., Gelinas, L. J., Rudy, R. J., Walterscheid, R. L., and Liu, A. Z. (2021). Kelvin-Helmholtz Billow Interactions and Instabilities in the Mesosphere Over the Andes Lidar Observatory: 1. Observations. Journal of Geophysical Research: Atmospheres, 126(1):e2020JD033414. Publisher: John Wiley & Sons, Ltd.

[Thorpe, 2002] Thorpe, S. A. (2002). The axial coherence of Kelvin–Helmholtz billows. Quarterly Journal of the Royal Meteorological Society, 128(583):1529–1542. Publisher: John Wiley & Sons, Ltd.

How to cite: Doddi, A., Fritts, D., and Lund, T.: Rapid Turbulence Evolution Resulting from Stable Shear layer and Atmospheric Gravity Wave Interactions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4129, https://doi.org/10.5194/egusphere-egu26-4129, 2026.

Socio-economic drought represents the stage at which water stress translates into tangible disruptions to livelihoods, infrastructure, and economic systems, often preceding severe physical water shortages. In India, pronounced climatic variability combined with extreme physiographic heterogeneity leads to strong spatial contrasts in socio-economic vulnerability to drought. Despite this, most drought assessments in the country remain dominated by hydro-meteorological indicators, with limited integration of socio-economic exposure, sensitivity, and adaptive capacity.
This study develops a spatially explicit socio-economic drought risk assessment framework for India by integrating multi-dimensional climatic, environmental, and socio-economic indicators within a Geographic Information System (GIS). Thirteen indicators capturing water availability, agricultural productivity, infrastructure, population pressure, economic activity, and social deprivation are compiled from multi-source datasets and harmonized to a common spatial resolution. The indicators include available soil water, agricultural yield, livestock density, road density, population density, biomass, electricity consumption, Gross Domestic Product (GDP), global surface water availability, digital elevation model, groundwater availability, land use/land cover, and relative deprivation. Indicator weights are objectively derived using the Analytic Hierarchy Process (AHP), with consistency of expert judgments ensured through the consistency ratio criterion (CR < 0.1). A GIS-based weighted overlay approach is then employed to generate a composite socio-economic drought risk index, which is classified into four risk categories to identify spatial patterns and hotspots.
The resulting risk map reveals pronounced regional disparities, highlighting drought-prone agrarian and socio-economically marginalized regions as areas of elevated risk. The proposed framework offers a transferable and scalable decision-support tool for integrating socio-economic dimensions into drought monitoring and preparedness. By explicitly linking water stress to livelihood and infrastructure vulnerability, the study provides actionable insights for risk-informed planning, targeted mitigation, and long-term drought resilience in India.

How to cite: Beerkur, A. K. and Palagiri, H.: A Multi-Criteria GIS Framework for Socio-Economic Drought Risk Assessment across India, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4184, https://doi.org/10.5194/egusphere-egu26-4184, 2026.

Scour, defined as the erosion or removal of sediment from around bridge piers due to flowing water, remains one of the primary causes of hydraulic structure failures worldwide. Local scour around bridge piers poses a serious threat to bridge stability, particularly during high-flow events, as the development of downflow, horseshoe vortices, and wake vortices at the pier base leads to intense sediment removal and foundation instability. To address this challenge, the present study investigates the hydrodynamic behaviour and scour reduction performance of tapered submerged vanes installed upstream of a cylindrical bridge pier as an effective countermeasure against local scour. A combined numerical and experimental approach was adopted to evaluate the influence of tapered submerged vanes on flow structure and scour characteristics. Numerical simulations were carried out using FLOW-3D Hydro to analyse the three-dimensional flow field around the pier–vane system under steady clear-water conditions. The simulations focused on assessing velocity distribution, near-bed shear stress, vortex dynamics, and secondary flow patterns generated by the tapered vanes. Particular attention was given to the formation of leading-edge vortices (LEVs) and their role in modifying erosive flow structures near the pier foundation. Based on the numerical insights, a series of physical model experiments were conducted in a laboratory flume to quantify the scour reduction achieved by the tapered vanes. The experiments aimed to optimize the longitudinal and transverse placement of the vanes relative to the pier. The vanes were installed at a fixed longitudinal distance upstream of the pier, while transverse spacing was systematically varied to examine its effect on sediment transport and scour depth. Bed elevation profiles and maximum scour depths were measured after equilibrium scour conditions were attained. The results demonstrate that tapered submerged vanes significantly alter the near-bed flow field by generating localized leading-edge vortices that effectively deflect high-energy flow away from the pier base. This flow redirection weakens the horseshoe vortex and reduces near-bed shear stress in the vicinity of the pier. Among the tested configurations, the vane arrangement with a longitudinal spacing of 1.5D and transverse spacing of 2D exhibited the best performance, resulting in a 56% reduction in maximum scour depth compared to the no-vane case. Additionally, localized sediment deposition was observed upstream and downstream of the pier, indicating favourable redistribution of sediment induced by the vane-generated secondary currents. By integrating numerical modelling with experimental validation, this study provides valuable insights into the flow mechanisms and optimal placement strategies of tapered submerged vanes. The findings highlight their potential as a practical, efficient, and sustainable solution for mitigating local scour around bridge piers in alluvial channels.

Keywords: Scour, Submerged Vane, Horseshoe Vortices, Wake Vortices, Leading-Edge Vortex (LEV)

How to cite: Karmishtha, K., Behera, R. K., and Singhal, G. D.: Performance of Tapered Submerged Vanes in Mitigating Local Scour Around Bridge Piers, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4200, https://doi.org/10.5194/egusphere-egu26-4200, 2026.

The Piano Key Weir (PKW) has earned recognition for its adaptability for large discharges across weir types of varying heights and with small footprints. Therefore, it has the potential to be a substitute for linear weirs (space being a factor), Ouamane and Lempérière, (2006). Even with the above-mentioned advantages of PKWs, other geometries leave much to be desired. The rectangular PKW and the trapezoidal PKW illustrate a most common inefficiency example. Standard literature describes construction and operational shortfalls such as flowing separation at the inlet key, varying discharge and uneven velocities along the crest, vortex shedding and formation at the key intersections, dead zones in the inlet-outlet, zones of intensified energy dissipation, and lowering weir versatility at high flows. These challenges are combined to mean loss of efficiency in weir discharge capability. In response to these challenges, the present study introduces the Modified Piano Key Weir (MPKW) to assess its performance using 3D computational hydraulic modeling. The Volume of Fluid (VOF) methodology for free surface tracking and the Reynolds-Averaged Navier Stokes (RANS) for turbulence closure modeling characterize pressure gradients, flow accelerations in the several dimensions, and eddies. A systematic numerical investigation was conducted to compare the discharge efficiency of RPKW, TPKW, and MPKW across a range of steady inflow discharges: 0.030, 0.060, 0.090, 0.120, and 0.160 m³·s⁻¹. The MPKW demonstrated consistently superior discharge efficiency over both RPKW and TPKW for all tested cases, without requiring an increase in structural footprint or crest length. The highest relative improvement was observed at 0.060 m³·s⁻¹, which was therefore selected as a representative discharge for in-depth flow diagnostics. Discharge at 0.060 m³·s⁻¹ was applied to determine vorticity structures, turbulent kinetic energy (TKE), and energy dissipation to better understand the flow mechanisms that explain the efficiency of the weir. The MPKW design, with refined geometry and improved inlet–outlet design, rounded key transitions, and adjustable wall skew, was successful in mitigating flow separation at the key inlets and reducing the large-scale vortex formation at the key junctions. The modified sidewall skewed the internal recirculation, and as a consequence, TKE in the stagnation zones was less, and recirculation was more along the crests of the weir, thereby nullifying turbulent structures. While the breakdown of turbulence resulted in localized energy dissipation, the stabilization of the approach flow was improved because the process converted rotational energy of large eddies with a low energy loss to rapidly decaying eddies which do not sustain and produce a recycling of energy. Thus, less energy was concentrated in the vortex cells at the key junctions, the loss due to flow contraction was less, and the nappe cohesion over the crests was improved. MPKW, relative to other configurations, was characterized by a lower level of turbulence and vorticity at the junctions, a greater effective utilization of the crest, and improved pressure recovery. The results confirm MPKW as a hydraulically efficient and economically feasible solution for both new installations and retrofit applications under head or footprint constraints.

How to cite: Kumar, A., Padhi, E., and Mishra, S. K.: CFD-Based Comparative Analysis of Conventional and Modified Piano Key Weirs for Improved Discharge Efficiency, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4951, https://doi.org/10.5194/egusphere-egu26-4951, 2026.

Groundwater is the second largest reserve of fresh water and is an important resource that supports agriculture, industrial and domestic water supplies. Groundwater is facing unsustainable impacts by human activities over the years in different forms. The situation is aggravated by climate change which aggravates groundwater stress through variable precipitation leading to reduced recharge. Thus, highlighting the importance of assessing aquifer potential for sustainable groundwater management. The analysis was carried out in the Manjra and Maner sub-basins, of Godavari river basin, India where data-driven assessments remain limited. In this regard, the present research employs a Multi-Criteria Decision Analysis (MCDA) framework that integrates Geographic Information Systems (GIS) and the Analytical Hierarchy Process (AHP) to define groundwater potential zones (GWPZ) in the Manjra and Maner sub-basins. In a GIS environment, eight thematic layers—geology, land use/land cover, lineament density, drainage density, rainfall, soil, and slope—were examined. These factors were weighted using AHP, and combined using weighted overlay analysis. Area under the Curve (AUC), Receiver Operating Characteristic (ROC) analysis, and groundwater inventory data were used to validate the final GWPZ map. Five classifications of groundwater potential were identified for the research area: very low, low, moderate, high, and very high. The research region's predominance of moderate (45%) to high potential (28%) zones suggests that groundwater availability is generally fair to good. Priority locations for sustainable groundwater development and management are indicated by the high and very high potential zones.

How to cite: Belluti Nanjundagowda, P. and Vema, V. K.: A Geospatial and AHP-Based Approach for Delineating Groundwater Potential Zones in Vulnerable Groundwater Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5098, https://doi.org/10.5194/egusphere-egu26-5098, 2026.

Deep rock masses have complex internal structures, which results in significant discreteness and blocky structures. With the increase in the depth of engineering construction and energy extraction, the unique pendulum-type waves emerge in deep blocky rock masses under the action of dynamic loads from mining and blasting, and they are characterized by low frequency, low velocity, large displacement amplitude and high kinetic energy, distinguishing them fundamentally from conventional seismic waves. Pendulum-type waves can induce alternating stress states of relative compression and separation within blocky rock masses, and may lead to rockburst disasters and even engineering-induced seismicity, thus posing great challenges to the safety of underground engineering such as tunnel construction and mining. In this paper, experimental research is conducted on the mechanical behaviors and typical characteristics of pendulum-type waves of multi-fractured blocky rock masses under static and dynamic loads. Firstly, the strength, deformation and failure mode were analysized based on uniaxial compression tests. The weak structural layers will significantly reduce the uniaxial compressive strength and enhance the ultimate deformation capacity of rock masses. Fractured rock masses have significant nonlinear deformation and may develop macroscopic fractures (vertical splitting failure, with the failure mode transitioning from brittle failure to ductile failure) at the stress level significantly lower than their uniaxial compressive strengths. Subsequently, based on the dynamic impact tests, the dynamic response, overall displacement, wave velocity and the mechanism of anomalously low friction were investigated, and the typical characteristics of pendulum-type waves, including the low frequency (177 Hz and 153 Hz in this experiment, which are much lower than the natural frequency of the rock itself), low velocity (about 900 m/s in this experiment, which is significantly lower than those of P-waves and S-waves), large displacement amplitude (it is more than two orders of magnitude larger than the deformation of an intact rock under an identical load) and high kinetic energy (The total kinetic energy accounts for 40% and 28% of the total energy in this experiment, which has its particularity and cannot be ignored) were quantitatively described. This study holds significant research importance for understanding the nonlinear waves in deep fractured rock masses and their dynamic behaviors, as well as for preventing and controlling engineering disasters in deep rock masses.

How to cite: Jiang, K., Qi, C., and Hu, X.: Research on the mechanical behaviors of multi-fractured blocky rock masses, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5765, https://doi.org/10.5194/egusphere-egu26-5765, 2026.

Predictability of river bank erosion in sinuous alluvial channels requires a combined study of the planform processes, hydraulics processes, sediment transportation, and the geotechnical properties of riverbanks. The research paper provides a detailed analysis of the evolution of channels within the Nabadwip-Kalna stretch of the Bhagirathi-Hooghly River (1990-2020). This analysis combines the synthesis of remote sensing, on-field surveys, lab experiments, and numerical model analysis into a multidimensional analysis. GIS was used through the Digital Shoreline Analysis System (DSAS) to measure changes on the bank-lines using historical satellite images of the same period of time. A two-dimensional migration coefficient (MC) model was used to model spatial-temporal changes in channel centrelines, and an RVR Meander was used to develop a model that takes into consideration depth-averaged flow velocity and reach-averaged hydraulic parameters. The characterisation of cross-sectional bathymetry and near-bank hydraulics was based on ADCP. The results of the geotechnical analysis showed that stratified streambanks showed critical shear stresses of 7.1-7.7 kPa, internal frictional angles of soils less than 30°–34°, and were predominantly affected by either cantilever collapse or piping as a result of varying maximum heights of streams between 5.7 and 6.8 metres. Bank stability through both BSTEM and BEHI was assessed, whereas sediment forecasting combined with SWAT to predict overbank flow and a Genetic Algorithm (GA) to estimate the total load. DSAS analysis on bank-line displacement revealed different erosion patterns within 170 transects, showing different RMSE of 0.090 to 0.162 in predicting zone boundaries. The MC method was able to model the 24-year centreline migration patterns, recording changes in the centreline-geometry parameters. Analysis of five cross-sections instrumented found instability and a factor-of-safety ratio of 0.81-0.95, resulting in 4.07-5.85m/yr and 4.35-7.15 km²/yr, respectively, lateral retreat and the eroded areas. Mean collapse rates were 0.125 to 0.198 m/yr, and the failure angle was 81°–87°. The maximum bank-failure mass was 41.24 kg (seasonal maximum), and the calibrated toe-scour mass was 0.28 kg. The GA model was tried using ten parameterisations and demonstrated the best prediction ability with the coefficient set at ten, where R² = 0.96 and mean relative error (MRE) = 42% gave significantly better performance than the traditional regression analysis (R² = 0.87 and MRE = 40%). There were also considerable changes in the area behind sandbar dynamics, that is, Nandai-Hatsimla increased by 11.87 ha in 1990 to 19.05 ha in 2020; Media by 39.7 ha to 57.68 ha; Char Krishnabati by 82.52 ha to 81.07 ha. Land-use/land-cover (LULC) predictions for 2040 indicated settlement expansion from 13.61% (2020) to 20.19%, with validation accuracy (RMSE = 0.253) confirming model reliability. This combined model shows that the combination of remotely sensed, field, laboratory, and model data provides quantitatively sound estimations of fluvial risks and forms the basis of evidence-based management of high-suspended riverine areas. The modular design can be applied to monsoon-dominated alluvial basins throughout the globe, which will promote adaptive land-use planning and long-term infrastructure development in the vulnerable riparian societies.

How to cite: Ghosh, A.: Unveiling integrated geo-hydraulic assessment of river meandering, bank erosion and sandbar dynamics in Alluvial channels, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7917, https://doi.org/10.5194/egusphere-egu26-7917, 2026.

Acoustic emission (AE) monitoring is widely applied to track damage development in brittle rock, although relating recorded signals to specific fracture mechanisms can remain uncertain, particularly when comparing thermal and mechanical loadings. This contribution presents a preliminary assessment of AE waveform characteristics measured during two heating-only experiments and two uniaxial compressive strength (UCS) experiments performed on 100 mm diameter basalt specimens containing a central axial circular hole. This geometry provides a consistent configuration that promotes stress redistribution and damage localisation around an opening, allowing fracture processes to be compared within a common specimen form.

Full AE waveforms were acquired throughout each test using broadband piezoelectric sensors coupled to the specimen surface, with pre-amplification and digital acquisition. Event features were extracted in the time and frequency domains, including rise angle, duration, hit counts, average frequency, peak frequency, peak amplitude, and amplitude distributions. Feature-space comparisons were then used to evaluate whether thermally and mechanically induced microfracturing exhibit separable signal characteristics.

The thermal experiments were associated with a single dominant fracture initiating along the shortest ligament between the aperture boundary and the nearest specimen edge. In contrast, UCS loading produced a more complex fracture network consistent with mixed tensile and shear microfracturing. Rise angle versus hits per duration plots indicated that thermal events occupied a more restricted region, whereas UCS events displayed a broader spread, which may reflect greater variability in source processes during complex damage evolution. Frequency-based comparisons further highlighted the differences: thermally induced events clustered mainly within a lower-frequency band (approximately 100-300 kHz), while the UCS tests exhibited an additional higher-frequency population (approximately 400-600 kHz), alongside the lower-frequency component. Amplitude distributions were also differed, with thermal events tending toward a narrower amplitude range relative to the wider distribution observed under UCS loading. Collectively, these observations suggest that the combined time-domain, frequency-domain, and amplitude-based AE features support mechanism-informed discrimination between thermally driven tensile fracture and mechanically driven complex fracture networks providing a basis for subsequent statistical or learning-based classification in coupled thermomechanical experiments

How to cite: De Alwis, A., Serati, M., Dyskin, A., Pasternak, E., Martin, D., and Williams, D.: Waveform signatures of acoustic emission from thermally and mechanically induced microfracture in centrally apertured basalt, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8596, https://doi.org/10.5194/egusphere-egu26-8596, 2026.

Partial blockage in open channels and urban drainage systems is a common issue arising from debris accumulation, sediment deposition, and inadequate maintenance, often resulting in reduced flow capacity and increased flood risk. Despite its practical relevance, the hydraulic effects of partial blockage on flow behaviour are not well quantified through controlled experimental studies. This work aims to investigate the influence of partial blockage on flow characteristics in open channels and explore its implications for urban stormwater drainage systems.Laboratory experiments are carried out in a rectangular open-channel flume under steady flow conditions. Velocity measurements are obtained at multiple depths for unblocked conditions and for different partial blockage configurations. Blockages of varying size and location are introduced manually to represent realistic obstructions commonly observed in urban drains. The changes in velocity distribution, water depth, and flow-carrying capacity due to partial blockage are analysed to understand the hydraulic response of the system.

Based on these observations, relationships between blockage extent and hydraulic performance are developed to identify critical blockage conditions.The study framework is applied to urban stormwater drainage networks using SWMM modelling to extend the experimental findings to real-world applications. Blockage scenarios are simulated in selected channels to assess their impact on system performance and flooding behaviour.

The outcomes of this study provide experimental insight into blockage-induced hydraulic effects and highlight the importance of considering partial blockage in urban drainage analysis. The combined experimental and modelling approach offers a practical basis for improving flood risk assessment and maintenance planning in urban stormwater systems.

How to cite: Mishra, A. K., Kumar, H., and Vinnarasi, R.: Assessment of Partial Blockage in Urban Drains for Flood Risk Reduction , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8813, https://doi.org/10.5194/egusphere-egu26-8813, 2026.

Rising flooding, which is exacerbated by both climate change and human behavior, demands proper identification of vulnerable zones. Conventional hydrological analysis can neglect geographical variability. In this study, a combined geospatial and decision-making process is used to determine the levels of vulnerability and risk of flooding in the Koshi River Basin in the state of Bihar.  The research work has developed a susceptible, vulnerable and risk map by integrating GIS, Remote Sensing and AHP. Weightings of eleven physical and hydrological factors and five socio-economic indicators were carried out in a systematic manner using a multi-criteria decision-making framework that allowed appropriate consideration of their relative contributions to flooding. Flood susceptibility, vulnerability and risk maps were created using the GIS environment's Weighted Overlay technique. According to the analysis, population density (41.6%) and literacy rate (24%) are controlling factors for flood vulnerability in the basin, whereas rainfall (23.9%), elevation (14.7%) and drainage density are the main elements that influence flood susceptibility. The Koshi basin is largely covered by the low and moderate classes of flood susceptibility, whereas a very minor amount (0.03%) comes under the high susceptibility classes, according to results from flood susceptibility maps. A significant section (42.87%) of the basin has moderate flood susceptibility due to a combination of exposure and socioeconomic characteristics, according to the results of the flood vulnerability analysis. According to the flood risk results, a significant amount of the basin (84.18%) has moderate flood risk, while a tiny portion has high flood risk in the low-lying, heavily inhabited areas close to the basin's riverbanks.  ROC-AUC for model validation yielded an accuracy of 66.3% and proved that the proposed GIS-AHP model was a reliable. Conclusion from this study underscore an integrating role in both physical and socio-economic considerations with prospects of enhancement through climate scenarios in flood mitigation and planning/early warning maps.

How to cite: Chaudhary, P. and Padhi, E.: Flood Hazard Analysis and Risk Assessment of Koshi River, Bihar (India) using Remote Sensing, GIS and AHP Techniques , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8975, https://doi.org/10.5194/egusphere-egu26-8975, 2026.

Block elbowing, the process in which rotating blocks push neighbouring blocks apart, influences both geological deformation and the stability of mining excavations in blocky rock masses. A clearer understanding of elbowing is essential for improving rock mass modelling and maintaining the safety of engineering structures. To this end, we analyse a chain of stiff blocks connected by springs, with one or two end active (driving) blocks – the blocks whose rotation is externally induced. All other - passive blocks - have translational and rotational degrees of freedom. The results show that block rotation is sequential (starting from driving blocks) producing a rotational wave with strongly configuration-dependent rotational patterns.

Opposite to a single driving block system, a double-driving block system exhibits more complex behaviour, as the active blocks may rotate in the same direction (Case I) or in opposite directions (Case II). In Case I passive blocks can exhibit anticlockwise rotation that is opposite to the clockwise rotating driving blocks, while in Case II all passive blocks do not rotate at all.

Further deformation patterns arise from block geometry, introduced by varying block corner rounding to represent spheroidal weathering. The results reveal a transition from reversible to irreversible passive block kinematics. Reversible responses include either clockwise rotation followed by full recovery or no rotation. The boundary between these types of block behaviour is defined by a linear relationship between the active-passive and passive-passive contact friction coefficients, with the intercept related to block corner rounding. In contrast, irreversible kinematics characterised by residual rotation emerge only for highly rounded blocks. This irreversible behaviour is restricted to short block chains and disappears in chains of five blocks suggesting a critical size of the Cosserat like zone with independent rotational degrees of freedom. This study provides new insights for modelling the stability and long-term evolution of blocky rock masses.

How to cite: Zhang, M., Dyskin, A., and Pasternak, E.: Non-linear rotational waves and complex rotation patterns in a chain of blocks with elbowing, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8985, https://doi.org/10.5194/egusphere-egu26-8985, 2026.

The subject of this study is the process of hydraulic stimulation of a tectonic fault, leading to induced seismicity. We consider a scenario in which fluid injected near ​​an existing fault, causing a localized change in pore pressure and a reduction in effective stresses. This, in turn, initiates slippage of the fault segments and the formation of a slip zone, the size and slip velocity of which determine the magnitude of the resulting seismic events. The goal of this study was to develop a relatively simple model for estimating the potential magnitude of induced seismic events based on a limited set of governing parameters. The primary objectives of the study were to identify the key factors that have the greatest impact on the characteristics of the slip zone and to determine how fluid injection parameters (rate and injected fluid volume) affect earthquake magnitude by changing slip dynamics. The model obtained is based on the results of a series of numerical experiments analyzing the hydromechanical behavior of the fault under various injection conditions. The modeling was performed using a two-parameter rate-and-state friction law, which, unlike a single-parameter model, allows for a wider range of slip regimes to be simulated and accurately describes the transition from stable slip to dynamic failure.

The functional relationships were established between the initial system parameters and the key obtained slip characteristics. It was shown that the final slip zone length is almost linearly related to the length of the initial unstable zone, and the maximum slip velocity increases exponentially with increasing pore pressure rate. At the same time, in the area of high loading rates, the saturation of the sliding velocity is observed at a characteristic level, which leads to a limitation of the possible magnitudes of earthquakes induced by fluid injection.

How to cite: Turuntaev, S., Baryshnikov, N., and Riga, V.: Estimation of potential magnitudes of induced seismic events based on direct numerical simulation of fluid injection near an active tectonic fault., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11339, https://doi.org/10.5194/egusphere-egu26-11339, 2026.

Efficient water use in agriculture is crucial for sustainable water resource management, especially in areas experiencing increasing water scarcity. A critical yet often oversimplified component of irrigation planning is the estimation of water storage soil profile depth, commonly assumed to be 1-1.5 m as the root-zone depth based on practitioner experience rather than validated soil-water dynamics. Such assumptions introduce uncertainty and limit the reliability of irrigation scheduling decisions. This study presents a novel framework for estimating soil profile depth to store maximum water by integrating Richards’ equation, geotechnical soil column concepts, and the Soil Conservation Service Curve Number (SCS-CN) technique to derive an optimal soil profile depth that maximizes storage capacity based on measurable hydraulic and retention soil properties. By linking the water storage soil column depth with the SCS-CN parameter, for practical field applications such as irrigation scheduling and planning. The proposed framework improves model reliability and interpretability by replacing fixed-depth assumptions with soil-specific storage behaviour, thereby reducing uncertainty in irrigation water estimation. It enables consistent evaluation of field capacity, average soil moisture content, and maximum storage potential across soil types, leading to improved irrigation efficiency. By emphasizing physically constrained model selection, data-informed parameterization, and transparent decision-making metrics, this work enhances the reliability of hydrologic modeling and supports robust irrigation management under water-scarce conditions.
Keywords:  Water storage soil profile depth, Richards’ equation, Irrigation water management, Data-informed parameterization, SCS-Curve Number method.

How to cite: Sharma, D., Mishra, S. K., and Pandey, R. P.: From Empirical Assumptions to Data-Informed Decisions: A Reliable Water Storage Soil Depth Estimation Method, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13831, https://doi.org/10.5194/egusphere-egu26-13831, 2026.

The energy cascade rate (ε) depicts the energy transfer in a turbulent system. In incompressible magneto-hydrodynamic (MHD)  turbulence, ε is linked to the third-order structure function (Yaglom vector) via the Yaglom/Politano–Pouquet law in the inertial range. In this study, we compare three estimators of ε in anisotropic MHD turbulence: (1) the lag polyhedral derivative ensemble (LPDE) technique that reconstructs the divergence of the Yaglom vector via tetrahedral linear gradients; (2) a directional-averaged third-order estimator that evaluates the Yaglom vector along a finite number of lag directions and averages over solid angle; and (3) the Yaglom vector on 60 degree with respect to the mean magnetic field direction.  To ensure a fair comparison in more realistic MHD turbulence, we emulate a multipoint virtual mission within anisotropic three-dimensional MHD simulations with a guide field B₀ along the z-axis. This work illuminates the reliable regime for LPDE and directional-averaging methods, and also tests whether 60 degree Yaglom vector is an accurate estimate of ε, providing practical guidance in both simulation and observational turbulence analysis.

How to cite: Gao, Z., Yang, Y., Jiang, B., and Pecora, F.: Anisotropic energy transfer rate quantified by LPDE and directional averaging methods in MHD turbulence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15102, https://doi.org/10.5194/egusphere-egu26-15102, 2026.

Digital elevation models (DEMs) play a fundamental role in hydrological modeling by controlling watershed delineation, stream networks and runoff generation processes. This study assess the impact of global DEM product provided by Shuttle Radar Topography Mission SRTM and the Indian national CartoDEM developed by ISRO-Bhuvan (Indian Space Research Organisation-Bhuvan) on streamflow simulation using the Soil and Water Assessment Tool (SWAT) in the Ong River watershed (4650 sq. km), India. The study area is characterized by forest and cropland. Both DEMs, resampled to 30m resolution, were used as inputs to SWAT, along with meteorological data (IMD), land use/land cover data (Sentinel-2), and soil data (FAO). Streamflow data was sourced from Global Flood Awareness System discharge data (GloFAS). Model calibration (2011-2017) and validation (2018-2020) were performed using SWAT-CUP with the SUFI2 algorithm. Model performance was evaluated using Willmott's index of agreement, Nash-Sutcliffe Efficiency (NSE), R², PBIAS, and RSR. Results showed that both DEMs performed satisfactorily, with CartoDEM exhibiting slightly better performance (higher NSE and R², lower PBIAS and RSR) during both calibration and validation periods. Sensitivity analysis revealed that the runoff curve number was the most sensitive parameter, highlighting the impact of DEM selection on surface runoff simulation. The study concluded that CartoDEM is a preferable choice for hydrological modeling in similar catchments, though further research on stream accuracy and catchment delineation in diverse topographies can be explored.

How to cite: Prashant, P., Kumar Mishra, S., and Kumar Lohani, A.: Assessing the Impact of Digital Elevation Model Selection on Hydrological Predictions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16403, https://doi.org/10.5194/egusphere-egu26-16403, 2026.

Accurate prediction of sediment mobility in open channel flows is essential for effective river engineering and sediment management. This study examines the combined influence of flow depth and sediment grain size on near bed hydraulics and sediment mobility using high-resolution Acoustic Doppler Velocimeter (ADV) measurements in a controlled laboratory flume. Experiments were conducted over uniform sand beds with median grain sizes of d₅₀ = 0.321 mm and d₅₀ = 0.81 mm under four different flow depths (12cm, 15cm,18cm,21cm) and a range of flow velocities. Three dimensional velocity components were measured at multiple vertical locations throughout the flow depth, while water surface elevations were continuously monitored. Depth resolved ADV data were used to compute mean streamwise velocity, Reynolds shear stress, friction velocity, and turbulent kinetic energy for each sediment size and flow depth. Sediment mobility was assessed using the Shields parameter, estimated from ADV-derived bed shear stress, and compared with the critical Shields parameter at multiple velocity points for each depth. The results indicate that coarser sediment beds exhibit increased near-bed turbulence intensity and higher friction velocity across all flow depths, while yielding lower Shields parameter values relative to finer sediment beds. Comparisons across the four flow depths reveal that sediment mobility transitions from stable to mobile conditions depending on the combined effects of flow depth, sediment size, and local velocity magnitude. At lower velocities, Shields parameter values remain below the critical threshold, indicating stable bed conditions, whereas higher velocities at the same depth result in Shields values exceeding the critical limit, signifying active sediment motion. Depth wise velocity and turbulence profiles demonstrate that both flow depth and sediment roughness significantly modify near-bed hydraulic structure and bed shear stress distribution. The findings highlight the importance of accounting for depth-dependent flow structure and sediment characteristics when evaluating sediment mobility. This study provides a robust experimental framework for identifying stable and mobile sediment regimes and estimating sediment transport potential using high-resolution ADV measurements without direct sediment transport observations.

How to cite: Banothu, J. and Devi, K.: Effects of Flow Depth and Sediment Size on Near Bed Hydraulics and Sediment Mobility in Open Channel Flow, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18007, https://doi.org/10.5194/egusphere-egu26-18007, 2026.

Urban flooding poses a growing challenge for rapidly urbanizing cities, where climate change–driven increases in extreme rainfall, expanding impervious surfaces, and limited drainage capacity collectively exacerbate the frequency and severity of surface water inundation. In this context, understanding urban surface flood resilience, defined as the capacity of stormwater drainage systems to withstand, convey, and recover from intense rainfall events, remains essential for effective flood risk management and climate adaptation planning. The present study investigates urban surface flood resilience in Janakpur Sub-Metropolitan City, Nepal, a fast-growing urban center increasingly exposed to pluvial flooding. The study develops an integrated modelling framework using a 3-way coupled MIKE+ hydrodynamic model, integrated with intense spatial analysis using GIS, to evaluate the performance of the existing stormwater drainage system under extreme rainfall conditions. The model represents the urban drainage network and surface flow processes using drainage infrastructure data obtained from field surveys, terrain information derived from a high-resolution digital elevation model, and delineated urban catchments. To characterize rainfall extremes, the analysis employs long-term observed hourly rainfall records spanning 25 years to generate design storm events corresponding to multiple return periods. The modelling framework simulates system response for a representative extreme rainfall event and quantifies inundation dynamics across the urban landscape. The results shows that the coupled approach effectively captures critical flood hazard characteristics, including inundation depth, flow velocity, and the depth–velocity product, allowing for the spatial identification of highly vulnerable catchments and drainage bottlenecks. The findings provide actionable insights into the limitations of existing stormwater infrastructure and support the development of targeted adaptation strategies aimed at enhancing urban surface flood and drainage resilience. Overall, the study underscores the value of integrated hydrodynamic modelling for resolving location-specific flood behaviour and strengthening urban flood resilience assessments under evolving climatic and urbanization pressures.

How to cite: Singh, P., Deopa, R., and Mohanty, M. P.: Assessing urban surface flood resilience using hydrodynamic modelling under extreme rainfall conditions in urban catchment of Nepal, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18481, https://doi.org/10.5194/egusphere-egu26-18481, 2026.

We present a one-layer global energy balance climate model with highly parameterized radiation, convection, and large-scale atmosphere/ocean macroturbulence. Planetary heat content is parameterized by a 2D in latitude-longitude layer characterized by a temperature field and a uniform constant heat capacity. Radiation is parameterized by mean-annual zonal average top-of-atmosphere solar irradiance. Radiative heating and cooling are parameterized by a uniform constant albedo and Stefan-Boltzmann emission with uniform constant emissivity. Convection is parameterized by a temperature threshold for convection which restricts the layer from warming beyond the threshold, effectively cooling the layer. Macroturbulence is parameterized by 2D barotropic turbulence forced at small scales and damped by Rayleigh friction. Energy conservation is maintained by balancing the convective cooling of the layer with the turbulent kinetic energy forcing, resulting in tropical forcing, while the frictional loss of kinetic energy is balanced by frictional heating of the layer. The parameterized energy transforming processes are characterized by timescales, which, for Earth-like planets, are ordered as t_radiation > t_{macroturbulence} > t_convection.

We investigate the model’s equilibrium climate state in terms of the meridional heat transport (MHT), the resulting zonally averaged temperature profile, and their fluctuations by simulating the system over many radiation times. For Earth-like parameters, despite the model’s extremely simplified dynamics, our simulations reveal a MHT profile comparable to the observed, annually averaged MHT on Earth, featuring a maximum in the mid-latitudes of approximately 5PW, a form of Bjerknes compensation. 

How to cite: van Kan, A., Weiss, J., and Knobloch, E.: Global climate dynamics in a highly parameterized radiative-convective-macroturbulent energy balance model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18820, https://doi.org/10.5194/egusphere-egu26-18820, 2026.

High-resolution climate projections are essential for climate impact, vulnerability, and adaptation studies, particularly over regions with strong spatial heterogeneity such as Brazil. Although CMIP6 global climate models (GCMs) provide valuable information on future climate change, their coarse spatial resolutions, typically ranging from 100 to 200 km, limit their direct application at regional and local scales. Statistical downscaling techniques offer computationally efficient alternatives to dynamical downscaling, but their relative performance and added value remain insufficiently assessed over Brazil.

In this study, we compare two statistical downscaling approaches applied to a subset of CMIP6 models previously evaluated by Bazanella et al. (2024) – 10.1007/s00382-023-06979-1 – and identified as skillful in representing Brazilian climate: (i) a bilinear interpolation method followed by percentile-to-percentile bias correction, and  (ii) machine learning–based downscaling approaches. The original GCM outputs are interpolated to a common high-resolution grid of 10 km × 10 km using bilinear weights, providing a physically consistent reference framework. In parallel, ML-based models are trained using historical GCM predictors and high-resolution reference climate datasets to learn nonlinear relationships and generate high-resolution climate fields.

The performance of both approaches is evaluated for the historical period in terms of mean climatology, spatial patterns, and variability. Future projections under the SSP2-4.5 and SSP5-8.5 scenarios are then analyzed to assess regional climate change signals and associated uncertainties. Results assess the extent to which ML-based downscaling provides added value relative to bilinear interpolation, particularly for variables with strong spatial heterogeneity, such as precipitation and temperature extremes, while also evaluating the ability of the approach to preserve the large-scale climate signals projected by the driving CMIP6 models. This comparative analysis provides insights into the applicability, robustness, and limitations of statistical and ML-based downscaling methods for regional climate assessments over Brazil.

How to cite: Jatobá Santos, D., Goracci, G., Alves Martins, M., and Schneider, R.: From bilinear interpolation to machine learning: a comparative assessment of statistical downscaling methods for CMIP6 projections over Brazil, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19471, https://doi.org/10.5194/egusphere-egu26-19471, 2026.

Accurate estimation of evapotranspiration (ET) is critical for various applications in hydrology and agricultural water management. However, direct observations of ET, specially its spatial variation, in time-consuming and cumbersome, thus necessitating the need to use of indirect methods for its estimation. In this study, stomatal conductance data is used in conjunction with bio-physical parameters of wheat crops for deriving the spatially varied estimates of ET (ET_SC) for different irrigation treatments using the Penman-Monteith equation. For this, five treatments, including drip (DI) and flood (FI) irrigated treatments were used in the study, namely fully irrigated (DI)), 50% MAD (maximum allowable deficit) (DI), 50% MAD (FI), farmer fields replication (FI) and rain-fed treatment.

The ET_SC estimates are also compared to the ET estimates derived using a method based on field water balance (ET_WB). When compared with the ET_WB values, the ET_SC estimates compared well particularly for the irrigated treatments. The average root mean square error (RMSE) of ET_SC estimates in comparison to ET_WB values are 0.11, 0.2, 0.23 and 0.26 mm/day for fully irrigated, 50% MAD (FI), 50% MAD (DI) and farmers field replication treatments, respectively. The corresponding RMSE value (0.47 mm/day) for the rain-fed treatment are found significantly higher than the irrigated treatments indicating the limitation of the approach in high water stress conditions. The differences between ET_SC andET_WB values also increase significantly during the end-season stage when the wheat crop is close to maturity. Overall, the results demonstrate the robustness of the proposed approach in estimating the spatial variation of ET using the Penman-Monteith method in conjunction with the on-field field stomatal conductance observations.

How to cite: Upreti, H. and Yadav, M.: Evaluation of Penman-Monteith estimates of evapotranspiration derived using field-collected stomatal conductance observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20031, https://doi.org/10.5194/egusphere-egu26-20031, 2026.

Understanding the spatial distribution and intensity of climate-related hazards is essential for effective risk assessment and adaptation planning.  This study presents a comprehensive analysis of climate hazard indices applied across all IPCC reference regions, using all the available CMIP5-driven regional climate model (RCM) simulations at 25 km resolution over the CORDEX domains, together with Euro-CORDEX simulations at 12 km resolution. The objective is to identify climate hazard hot spots through the formulation of a composite hazard index. 

A subset of hazard indicators representing key climate extremes is selected. Temperature- and heat-stress–related hazards are characterized using TX90p (extreme maximum temperature), TN90p (extreme minimum temperature), and the NOAA Extended Heat Index (HI). Heavy precipitation and drought-related hazards are represented by RX1DAY (maximum 1-day precipitation), P99 (99th percentile of precipitation), and CDD (consecutive dry days).

The composite index integrates both the frequency and intensity of extremes and is computed at both regional and grid-point levels. A normalization approach is used to ensure comparability across regions with diverse climatic characteristics. Results reveal pronounced spatial heterogeneity in hazard intensity, highlighting regions where multiple hazards converge and amplify overall risk. This framework enables systematic identification of global and regional climate hot spots, offering insights into areas that may face heightened climate stress under current and projected conditions. By providing a consistent, region-wide assessment of hazard exposure, this study aims to support comparative climate risk analyses and inform policy-relevant decision-making for climate adaptation and resilience strategies at multiple scales.

How to cite: Zazulie, N., Raffaele, F., and Coppola, E.: Global Hot Spots of Climate Extremes from Composite Hazard Indices, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21173, https://doi.org/10.5194/egusphere-egu26-21173, 2026.

Reliable river discharge simulation generally relies on observed streamflow data for model calibration; however, such observations are often uncertain or unavailable in data-scarce regions, limiting the applicability of conventional hydrological models. This study presents a hybrid modeling framework that uses soil moisture as an alternative calibration variable to improve discharge simulations in the absence of reliable streamflow observations. The framework couples a two-layer version of the daily lumped MISDc (Modello Idrologico Semi-Distribuito in continuo) hydrological model with a Feedforward Neural Network (FFNN), which is employed to enhance parameter calibration by exploiting soil moisture dynamics. The proposed approach is evaluated across three contrasting basins: Tahanaout in semi-arid Morocco, and Colorso (Italy) and Bibeschbach (Luxembourg) in temperate climates. Both in situ and ERA5-Land soil moisture datasets are used as calibration inputs. Model performance is assessed using multiple hydrological metrics, including Mean Absolute Error (MAE), Kling-Gupta Efficiency (KGE), and the correlation coefficient (R). Results show that the hybrid MISDc-FFNN framework substantially improves river discharge simulations compared to the traditional model. Across all basins, MAE is reduced by up to 61%, KGE increases by more than 200%, and R improves by up to 87%, with consistent performance gains observed for both observed and reanalysis-based soil moisture. These findings demonstrate the potential of soil moisture driven calibration strategies to enhance hydrological modeling in data-scarce environments, offering a viable pathway for improved water resources assessment and flood risk management where discharge observations are limited or unreliable.

Keywords: Soil moisture; river discharge simulation; hydrological modeling; machine learning; ERA5-Land; data-scarce regions; feedforward neural network

How to cite: Ait Naceur, K., El Khalki, E. M., Brocca, L., Hadri, A., Jaffar, O., Rachdane, M., Simonneaux, V., Saidi, M. E. M., and Chehbouni, A.: Soil Moisture Based Calibration of a Hybrid Hydrological-Neural Network Model in Data Scarce Basins, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21830, https://doi.org/10.5194/egusphere-egu26-21830, 2026.

Present-day practices of bridge piers design often employ group arrangements of piers in various configurations to modify flow dynamics and mitigate subsequent scour formation around the piers. These group arrangement configurations may vary in aspects of spacing ratio, number of piers, and orientations to alter the flow-structure interaction, and hence the scour development. Investigating the turbulent flow behaviour around various common group arrangements has been a topic of interest for researchers for a few years now. This study presents an experimental investigation aimed at comparing the equilibrium scour depth caused by various four-pier group arrangements. To assess the impact of spacing, the face-to-face distance between piers (G) was taken to values of D, 2D, and 3D, where D refers to the diameter of the circular pier. The scour patterns reveal that the maximum scour depth occurred when spacing G was equal to D. The equilibrium scour depth decreased with an increase in the pier spacing to 2D and 3D, corresponding to an approximate flow intensity of 0.9. The scour contours exhibit the impact of neighbouring piers and how it differs with an increase in pier spacing. Instantaneous velocity data were collected to derive the flow characteristics in the flow field. Velocity vectors depict the influence of different configurations on the flow pattern. The study provides an insight into the spacing effects on equilibrium scour, which can be useful in the design of pier group arrangements.

How to cite: Sahu, C.: Spacing Effect on the Equilibrium Scour and Flow Pattern around Four-Pier group in Different Configurations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22153, https://doi.org/10.5194/egusphere-egu26-22153, 2026.

How to cite: Thurler Rondon da Fonseca, J. F., Schertzer, D., da Silva Rocha Paz, I., and Tchiguirinskaia, I.: Analysis of Vector-Field Multifractal Cascades, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-23043, https://doi.org/10.5194/egusphere-egu26-23043, 2026.

Data-driven global weather models, such as GraphCast, have revolutionized medium-range forecasting but often exhibit systematic limitations in quantitative precipitation forecasting (QPF). Specifically, these models tend to produce over-smoothed blurry rainfall fields and underestimate localized extremes , primarily due to the inherent uncertainties in their reanalysis training data (e.g., ERA5) and the use of mean-squared-error-based loss functions.

To bridge the gap between coarse-resolution global AI forecasts and the need for precise, high-impact weather prediction, we introduce SynQPF-Net, a deep learning framework designed to synergize GraphCast’s dynamical background fields with high-resolution observational analyses. The model employs a dual-stream spatiotemporal encoder to process heterogeneous inputs: the 0.25^o dynamical forecasts from GraphCast and the 0.0625^o precipitation analyses from the China Meteorological Administration Land Data Assimilation System (CLDAS) . A specialized hybrid loss function, combining classification (Dice) and regression (Weighted MSE) objectives, is utilized to jointly optimize the spatial structure and intensity of precipitation.

Evaluated on warm-season events in Southern China, our approach demonstrates significant skill improvements. SynQPF-Net effectively sharpens the forecast, doubling the Critical Success Index (CSI) for heavy rainfall (>=10 mm) at the 6-hour lead time compared to the raw GraphCast output. Crucially, interpretability analysis reveals that the model learns physically consistent meteorological principles: it predominantly relies on extrapolating recent observational patterns for short lead times (<=12 h) and dynamically shifts its focus to large-scale circulation and moisture variables (e.g., 700 hPa specific humidity) as the forecast horizon extends. This work provides a validated pathway for correcting and downscaling global AI weather models, offering a robust solution for short-range extreme precipitation forecasting.

How to cite: Chen, D.: Bridging Global AI Models and Local Extremes: A Dual-Stream Framework for Correcting and Downscaling GraphCast Rainfall Predictions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1413, https://doi.org/10.5194/egusphere-egu26-1413, 2026.

AI-driven climate models are often criticized as “black boxes,” raising concerns about their credibility for scientific and policy-relevant decision making. Explainable artificial intelligence (XAI) is frequently proposed as a solution, focusing on identifying systematic relationships between model input and output data to characterize model behavior. This paper builds on prior work arguing that trust in both dynamical and AI models depends not on such input-output characterizations alone but on scientists’ component-level understanding of their models (O’Loughlin et al. 2025). Component-level understanding refers to scientists’ ability to point to specific model components or parts in the model architecture as the culprit for erratic model behaviors or as the crucial reason why the model functions well.

We argue that component-level understanding plays a distinctive role in establishing credibility because it expands scientists’ ability to answer a wider range of what-if-things-had-been-different questions. For example, when a model exhibits unexpected sensitivity or instability, component-level understanding enables scientists to ask (and design targeted tests to determine) whether the behavior would persist if a specific parameterization, architectural module, or physically informed constraint were altered. We see examples of this in CMIP, e.g., diagnosing the effect of a cloud microphysics scheme on a model’s climate sensitivity (Gettelman et al. 2019; Zelinka et al. 2020) and in AI-driven climate science as well, e.g., attributing model instability to particular architectural choices such as unconstrained neural network layers or inappropriate spectral representations (e.g., Beucler et al., 2019; Bonev et al., 2023). By linking model behavior to specific components or architectural features, scientists are better positioned to diagnose misbehavior, explore counterfactual scenarios, and explain why a model behaves as it does under varying conditions. This explanatory capacity enables scientists to establish credibility with decision-makers by demonstrating when, why, and under what conditions AI-driven climate models can be trusted.

Such explanations will inevitably be incomplete and context-dependent, particularly in complex models whose components interact in nonlinear ways and are often intended to represent emergent climate phenomena. Nevertheless, we argue that credibility is built through explanatory practices involving model successes and failures alike. We conclude by outlining several pathways for strengthening component-level understanding in AI-driven climate science: scientists may develop such understanding themselves; work in close collaboration with AI model builders and domain experts; design model intercomparison projects that explicitly support component-level diagnosis; or adopt evaluation and benchmarking practices that prioritize explanatory and counterfactual insight alongside predictive performance. On this view, establishing credibility requires organizing scientific work so that explanation remains a central and achievable activity.

References

Bonev, B., et al.: Spherical Fourier Neural Operators…, arXiv [preprint], https://doi.org/10.48550/arXiv.2306.03838. 2023.

Beucler, T., et al. Enforcing analytic constraints in neural networks…. Physical review letters, 126(9), p.098302. 2021.

Gettelman, A. et al. High Climate Sensitivity in the Community Earth System Model Version 2 (CESM2), Geophys. Res. Lett., 46, 8329–8337, https://doi.org/10.1029/2019GL083978, 2019

O'Loughlin RJ. Moving beyond post hoc explainable artificial intelligence… https://doi.org/10.5194/gmd-18-787-2025 2025

Zelinka, M. D., et al. Causes of Higher Climate Sensitivity in CMIP6 Models, Geophys. Res. Lett., 47, e2019GL085782, https://doi.org/10.1029/2019GL085782, 2020

How to cite: O'Loughlin, R.: Earning Credibility in AI-Driven Climate Science: The Role of Component-Level Understanding, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3879, https://doi.org/10.5194/egusphere-egu26-3879, 2026.

Accurate weather forecasting is essential for a broad range of socioeconomic activities. While emerging data-driven models match numerical weather prediction accuracy with reduced computational cost, their deterministic nature overlooks uncertainties in initial state estimates, model systematic biases, and stochasticity arising from unresolved subgrid physical processes. This obliviousness results in over-confident deterministic predictions that render uncertainty quantification inaccessible, thereby limiting their utility for risk-based decision-making.

To address these challenges, we present the Generative Ensemble Prediction System (GenEPS), a framework that systematically explores uncertainties in initial states, model formulations, and model stochasticity. GenEPS functions as a foundation model that has explicitly learned the probability distribution of high-dimensional atmospheric states. It provides a plug-and-play solution for ensemble forecasting with arbitrary deterministic models. Specifically, GenEPS utilizes deterministic forecasts as conditions to perform generative sampling, producing an ensemble of states projected back into the realistic atmospheric phase space defined by ERA5. This stochastic sampling process quantifies uncertainties in initial conditions and forecast dynamics while ensuring physical consistency. Crucially, by treating each step as a re-initialization within the valid state space, the framework decouples state evolution from specific model formulations, enabling seamless cross-model integration to mitigate systematic biases.

By explicitly representing all three sources of uncertainty, GenEPS outperforms state-of-the-art numerical ensemble predictions and data-driven predictions when evaluated against ERA5 reanalysis data using both deterministic and probabilistic metrics. GenEPS also enhances extreme event predictions, offering physically consistent forecast fields. These advances establish a new paradigm in ensemble forecasting through multi-model generative integration, combining a surging number of data-driven weather forecasting models and potentially numerical models, to achieve more reliable predictions.

How to cite: Nai, C., Chen, X., Yang, S., Xiao, Z., and Pan, B.: GenEPS: A Generative Foundation Model for Probabilistic Weather Forecasting , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4637, https://doi.org/10.5194/egusphere-egu26-4637, 2026.

Neural networks are powerful and widely used tools in weather and climate sciences, but their reliability under climate change remains uncertain as future conditions may be different from their training distribution. One way to build trust in these models is to assess whether they learn physically meaningful relationships rather than spurious correlations. Here, we present a case study investigating whether a simple convolutional neural network (CNN) predicts the occurrence of heavy rainfall in Western Norway for physically interpretable reasons. Since such rainfall is primarily associated with North Atlantic cyclones, we use explainable AI to assess whether the CNN identifies and uses the “correct” cyclones for its predictions.

Using ERA5 reanalysis data, we train a CNN to predict the occurrence of daily heavy rainfall events up to six days ahead from gridded wind and pressure fields. We apply layer-wise relevance propagation (LRP) to identify which regions of the atmospheric input fields contribute most to the model’s predictions. We find that model relevance is spatially aggregated into a small number of coherent patches, with one to three positive relevance patches dominating the prediction in more than 90% of the cases. Physical consistency is assessed by comparing the relevance patterns to objectively tracked cyclones. Interpreting cyclones as being “used” by the network when they spatially overlap with a patch, we show that cyclones contribute positively to the network’s predictions in about 95% of heavy rainfall events. In addition, we show that cyclones highlighted by the network are physically plausible; their trajectories follow the North Atlantic storm track, shifting from the western and central North Atlantic towards the eastern Atlantic and the Norwegian coast as the prediction lead time decreases. These results demonstrate that the CNN learns physically interpretable large-scale dynamics associated with North Atlantic cyclones, providing evidence that explainable AI methods can be used to assess and build trust in machine learning models for weather and climate applications.

How to cite: Guillaume-Castel, R., Sobolowski, S., and Li, C.: A convolutional network learns about the North Atlantic storm track to predict heavy rainfall in Western Norway, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4957, https://doi.org/10.5194/egusphere-egu26-4957, 2026.

Within the WeatherGenerator project, the Norwegian Meteorological Institute (MET Norway) is applying the project’s foundation model to reconstruct several decades -ideally the most recent 40 years- of atmospheric fields over Scandinavia. The primary objective of this work is to assess the potential of the WeatherGenerator framework for climate monitoring applications.

WeatherGenerator is a pan-European initiative that combines state-of-the-art machine-learning architectures with high-performance computing to develop an open, kilometer-scale foundation model of the coupled Earth system. The project is organized into four thematic areas; the application presented here is one of twenty-two applications developed by project partners within Theme 3.

The reconstructed datasets include near-surface atmospheric variables as well as variables at multiple pressure levels. The approach integrates heterogeneous data sources -ranging from in situ observations and reanalysis products to numerical model output- leveraging the foundation model to generate consistent, high-resolution fields suitable for climate and weather monitoring. The target spatial resolution is 1 km, achieved through data fusion techniques and a task-specific tail network trained to produce gridded analyses at this scale. Multiple temporal resolutions are explored, including hourly data and daily to monthly aggregations.

This contribution represents MET Norway’s first presentation of WeatherGenerator-related results at a scientific conference. The focus is therefore on preliminary results, outlining the overall methodological framework and demonstrating the potential of these novel approaches for high-resolution climate monitoring.

Note: The WeatherGenerator project (grant agreement No101187947) is funded by the European Union. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the Commission. Neither the European Union nor the granting authority can be held responsible for them.

How to cite: Lussana, C., Heilemann Myhre, R., Neuville, A., Nordhagen, E. M., Seierstad, I. A., and Nipen, T. N.: Retrospective reconstruction of 40 years of atmospheric fields in Northern Europe using the WeatherGenerator foundation model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5252, https://doi.org/10.5194/egusphere-egu26-5252, 2026.

This study proposes a novel diagnostic framework for systematically and intuitively evaluating the performance of medium-range weather forecasting models from a multivariate perspective. Traditional evaluation methods have primarily relied on point-wise error metrics (e.g., RMSE) for single variables at specific altitudes, which limits the analysis of inter-variable correlations and the dynamic evolution of forecast structures over lead times. To address these limitations, we present a methodology that integrates multivariate data into a shared, topology-preserving feature space, enabling the comparison and diagnosis of model-specific prediction trajectories.

The framework first represents multivariate atmospheric variables as images to extract semantic feature representations. To ensure robustness against spatial shifts and noise, we employ contrastive learning with data augmentation, effectively capturing the core physical characteristics of the atmospheric state. Subsequently, we apply a parametric manifold embedding specifically designed to preserve both the local neighborhood relationships of the high-dimensional feature space and its temporal continuity. This approach allows for a coherent and aligned comparison of prediction trajectories from diverse forecasting models within a unified coordinate system.

For the experimental setup, the feature space was defined using ERA5 reanalysis data from 2020 to 2024, with the 2025 ECMWF analysis serving as the reference ground truth. We analyzed a total of nine forecast configurations, combining three AI-based models (FourCastNet, GraphCast, and Pangu-Weather) with three operational numerical weather prediction initializations (IFS, KIM, and UM). By tracking trajectories at 6-hour intervals for up to 48 lead times, we visually analyzed model-specific dispersion and bias characteristics. Furthermore, the diagnostic validity of the framework was verified by comparing trajectory evolutions across different pressure levels and analyzing structural changes induced by varying variable compositions.

The proposed framework supplements conventional univariate and direction-agnostic metrics by enabling structure-aware, directional diagnostics in a multivariate feature space. It provides deep analytical insights into model-specific behaviors, serving as a critical diagnostic tool for future research on atmospheric pattern analysis and inter-variable correlation structures.

How to cite: Kim, H. and Cho, J. H.: Topology-preserving Feature-Space Analysis for Diagnostic Comparison of Weather Forecasting Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6319, https://doi.org/10.5194/egusphere-egu26-6319, 2026.

Evaluating regional climate model performance often relies on simple error metrics that assume the largest mean errors matter most, sometimes combined with visual inspection of the large-scale atmosphere circulation fields, that introduces a priori assumptions. However, the influence of spatial error patterns in the large-scale circulation on other surface variables, such as precipitation, is complex, and their link to future projections remains uncertain.

We present a novel computer vision–based framework for regional climate model evaluation that emphasises interpretability and explainability. Rather than treating machine learning as a black box, our approach learns structural characteristics of seasonal mean sea-level pressure fields directly from CMIP6 model data. Using a convolutional variational autoencoder (CNN‑VAE), we construct an interpretable latent space in which large-scale atmospheric patterns cluster according to shared spatial structure.

These clusters enable the identification of systematic differences in how regional climate models represent key large-scale drivers of European climate. Deviations from reanalysis are quantified using simple distance metrics in latent space, allowing the magnitude of structural model errors to be directly compared in a physically meaningful way and without reliance on subjective visual assessment or pointwise error measures.

We further compare end‑of‑century precipitation and temperature projections from models occupying different regions of latent space to assess how distinct large‑scale circulation errors impact future surface climate projections. We present this methodology as a complementary approach to traditional error metrics that can be applied flexibly to any variable of interest in model evaluation.

By linking learned representations to physically interpretable circulation structures, this framework supports more trustworthy use of machine learning in climate model evaluation and provides new insights into how spatial error patterns may influence downstream variables and projections.

How to cite: Palmer, T., Sexton, D., Ellis, A.-L., McNeall, D., and Mercer, G.: A data-driven approach for classifying GCM errors and understanding their impact on climate projections., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6734, https://doi.org/10.5194/egusphere-egu26-6734, 2026.

Weather forecasting for aviation allows tens of thousands of flights to operate daily with prior warning of global hazards including in-flight icing, turbulence, convection, and fog. Machine learning (ML) methods have begun to be utilised across the aviation weather forecasting sector and can provide greater skill, lower false alarm rates, and cheaper running costs than conventional equivalent products. Often these products are built on existing numerical weather prediction techniques, but can also be standalone products that make predictions only based on observations. Aviation is a highly regulated and safety-critical industry, so weather forecasting products must meet stringent quality-control standards, and machine learning processes must be trusted by customers.

The Aviation Applications Team at the Met Office has developed a set of 'Trustworthy AI' principles that ML products must strive to adhere to. These principles have guided the recent development of a range of ML driven weather forecasting solutions for aviation including convective cloud detection at UK airfields, auto-TAF (Terminal Aerodrome Forecast) verification, and global convective forecasting capability. In each of these use cases the aviation sector end-users have been considered to ensure the products are trustworthy and explainable.

This study showcases a range of aviation weather forecasting case studies and how they have utilised trustworthy AI techniques including,  explainable AI (XAI), representative AI, and considered how existing 'research to operations' pipelines can be exploited to add trust to machine learning models. The research group has worked with the UK's aviation regulator, the Civil Aviation Authority, to consider what the industry requires to be able to use machine learning safely in UK aviation operations and what can be learned from the long-standing collaboration between the two organisations in developing trusted weather forecasting products.

Future challenges in operationalising ML driven weather forecasting products in the aviation sector include; sparsity of observations for some hazards, shifting baselines for long-term deployment of products, and regulatory hurdles for the approval of AI products.

How to cite: Chitson, T., Fallon, J., Buchanan, P., and Shapland, J.: Integrating 'Trustworthy AI' Principles into Machine Learning for Aviation Weather Forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7283, https://doi.org/10.5194/egusphere-egu26-7283, 2026.

While AI-based weather foundation models have revolutionized predictive capabilities, their opaque nature and susceptibility to training data biases pose significant challenges for operational trust. A prominent example is Aurora, a state-of-the-art foundation model that demonstrates exceptional hurricane tracking accuracy but consistently underestimates cyclone wind speeds. Because this bias is inherited from the underlying reanalysis data, standard retraining often fails to alleviate the systematic error.

In this work, we propose a novel paradigm for bias correction by adapting AI Steering, a technique recently established for monitoring and adjusting Large Language Model (LLM) behavior, to the domain of climate science. Rather than relying on traditional post-processing or computationally expensive retraining, steering allows us to interrogate and shift the internal neural representations of Aurora without modifying the underlying weights. By identifying the latent features associated with wind speed intensity, we can shift the model’s internal state to align more closely with high-resolution observations.

To evaluate this approach, we run forecasts initialized with IFS-HRES conditions and validate our results against IBTrACS observations. Our results demonstrate that this interpretability-driven approach helps to improve systematic biases by significantly reducing wind speed errors while preserving model integrity and maintaining Aurora’s high-fidelity track accuracy. Furthermore, we show that steering enables a form of "Human-in-the-Loop" oversight, providing a transparent mechanism for meteorologists to adjust model outputs based on physical constraints and domain expertise. By bridging the gap between LLM interpretability and AI-based weather forecasting, we highlight the potential of steering to improve operational forecasts and offer a scalable, transparency-first framework for diagnosing and mitigating failure modes in complex AI-based climate and weather models.

How to cite: Bommer, P. L., Kretschmer, M., Hedstroem, A., Lehmann, F., and Hoehne, M. M.-C.: Guiding the Forecast: Interpretability and AI Steering in Climate Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7781, https://doi.org/10.5194/egusphere-egu26-7781, 2026.

How to cite: Nordhagen, E., Pinon Rodriguez, J., Thøgersen, K., Zeiser, F., Tjøtta, E., and Lappegard, G.: Data-Driven Weather Scenario Generation for Long-Term Energy System Planning in the Nordic Region, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10318, https://doi.org/10.5194/egusphere-egu26-10318, 2026.

Inadequate air quality is still a major cause of illness and premature death, and accurate air pollution analyses and predictions are needed to enhance human and animal well-being, protect natural and agricultural vegetation, and reduce climate change impacts. In a collaborative effort led by ECMWF and funded by the European Union, the WeatherGenerator, a foundation model for Earth system prediction, has been under development for nearly one year. In this study, we train and evaluate an early prototype of the WeatherGenerator model as first steps towards assessing how foundation models can represent chemical transport and regime behavior over extended forecast horizons. While some machine learning models, like Aurora [1], have demonstrated skillful 3-5 day predictions, air quality services require extended predictability windows (10-30 days) for strategic early warning systems and emission mitigation planning, especially for longer-lived species such as CO and background/regional O₃. WeatherGenerator's multi-resolution infrastructure simultaneously ingests datasets at different spatial and temporal resolutions without regridding, enabling integration of CAMS reanalysis chemistry data (0.75°, 3-hourly) with ERA5 meteorological information at synoptic scales (1°, 6-hourly). The model was trained from scratch on 2003-2021 observational data to forecast four reactive chemical species (O₃, CO, NO, NO₂) and three particulate matter size fractions (PM1, PM2.5, PM10). The training was conducted in two stages: pre-training the model on 2-step autoregressive rollouts, followed by fine-tuning on 8-step rollouts. Predictions span the entire tropospheric column including surface-level concentrations and 13 vertical levels from the boundary layer (1000 hPa) to the upper troposphere and tropopause region (50 hPa). We evaluate this proof-of-concept using 30-day autoregressive forecasts initialized from June 1, 2022. The trained model demonstrated stable 30-day continuous prediction of all species across all vertical levels, with comparable 5-day forecast skill to CAMS (at 0.75° resolution). Extended evaluation over June-November 2022 is currently underway to enable direct benchmark comparison. Notably, WeatherGenerator's training and fine-tuning required only 127 hours on 8 NVIDIA A100 GPUs. Ongoing work includes: (1) expanding training data to incorporate CAMS operational analyses at higher spatial resolution, (2) hyperparameter optimization, and (3) quantitative comparison to existing air quality forecast models to contextualize skill relative to operational CAMS and other baseline systems.

[1] Bodnar C, Bruinsma WP, Lucic A, Stanley M, Allen A, Brandstetter J, Garvan P, Riechert M, Weyn JA, Dong H, Gupta JK. A foundation model for the Earth system. Nature. 2025 May 21:1-8.

How to cite: Semcheddine, B. A., Melidonis, S., Schultz, M. G., and Lessig, C.: Tropospheric Transport Simulation with the WeatherGenerator Prototype Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11888, https://doi.org/10.5194/egusphere-egu26-11888, 2026.

How to cite: Hickman, S., Xhonneux, S., Luise, I., Kuehnert, J., Karlbauer, M., Tezcan, K., Perugachi Diaz, Y., Hunter, T., and Lessig, C.: Learning representations from different pre-training strategies in the WeatherGenerator , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13486, https://doi.org/10.5194/egusphere-egu26-13486, 2026.

Foundation models have shown strong potential for data-driven weather and climate forecasting by supporting multiple tasks with limited task-specific engineering. The use of architectures that extract maximum value from large, heterogeneous datasets is important in this approach. WeatherGenerator follows this paradigm by learning from diverse observational and reanalysis sources to encode a latent representation of atmospheric dynamics. In this work, we examined the impact of integrating Mixture of Experts (MoE) layers, developed for large language models, into WeatherGenerator, and assessed how MoE can best be incorporated within its decoder architecture.

The motivation behind MoE is straightforward: during training, a router learns to assign tokens to specialized experts, allowing different parts of the decoder to focus on distinct spatial regions or physical regimes. We build on this idea by introducing spatially aware routing, in which geographic context is provided to the router, and by evaluating loss-aware routing strategies that favor experts by minimizing local prediction errors.

We evaluate four MoE decoder configurations, based on the use of spatial context and loss-aware routing, and compare them against the baseline model. Experiments are conducted using ERA5 reanalysis data, with performance measured using global RMSE and MAE for wind components (u, v), temperature (2t, t850), geopotential height (z500), and specific humidity over three 6-hour autoregressive forecast steps.

Across experiments, MoE architectures consistently improve performance for thermodynamic and large-scale variables. In particular, z500 RMSE is reduced by 26–31% at the first forecast step, with spatially aware routing performing best. Near-surface temperature shows a 7% improvement in RMSE and an 11% improvement in MAE when combining spatial and loss-aware routing. These improvements appear early in training within the first few epochs, indicating efficient use of the available data. On the other hand, MoE variants show limited or slightly negative effects at the second and last forecasting step when evaluating wind component variables, while the baseline performance shows similar or better results, especially on the last forecasting step.

These preliminary results indicate that MoE provides variable-dependent benefits, with notable improvements for slowly varying, large-scale thermodynamic fields, but less impact on highly dynamic momentum variables. Ongoing work will further assess performance across longer forecast horizons, different climatic regions, and training with multiple datasets from different sources.

How to cite: Almikaeel, W. and the WeatherGenerator Team: Mixture of Experts with Spatial Routing in a Weather Foundation Model: Early Results from WeatherGenerator, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14545, https://doi.org/10.5194/egusphere-egu26-14545, 2026.

Foundation models for the Earth system have gained popularity, as they are starting to surpass numerical solvers in the accuracy of predicting Earth’s condition while requiring fewer computational resources. The Earth System Foundation Model (ESFM) contributes to this research direction by further extending the foundation models' flexibility.

The forecasting capabilities of ESFM are achieved in an autoregressive manner, using data from the t₀ - Δt and t₀ timesteps to produce a prediction for t₀ + Δt. This approach is effective on weather timescales. Moreover, we find that it also delivers encouraging results for long-term forecasts, showing reasonable zero-shot subseasonal-to-seasonal (S2S) predictions (15–40 days). 

However, S2S predictions can be further improved while preserving weather skills. This work investigates strategies for this purpose. On such timescales, it is crucial to produce probabilistic predictions to better represent inherent uncertainty. Probabilistic predictions are realised with the introduction of multiple decoder heads (tails) for each variable. Each tail is intended to simulate a different possible trajectory, which, when combined, provides an estimate of the most probable outcome together with the spread of feasible values. To better estimate the distribution of possible values on the S2S timescale, additional trajectories are generated by running multiple predictive rollouts with different initial conditions.

Another strategy to improve S2S rollouts is to fine-tune the model to produce outputs for more distant steps. To this end, we leverage LoRA adapters (Hu et al., 2022), which are trained for each subsequent rollout step. This approach effectively improves predictive performance on long horizons, without significantly affecting training complexity or inference cost.

We also observe that some predictive variables of the model, such as climate forcings, are slowly evolving and can benefit from incorporating inputs from a more distant past than the t₀ - Δt and t₀ timesteps commonly used. To investigate this, we introduce an Attention Temporal Aggregator in the encoder, which leverages learned patch embeddings from an arbitrary number of previous timesteps and attends to those that are most informative for a given variable. In this way, for rapidly changing variables such as wind speed, the model focuses on the most recent data, whereas for slowly evolving variables such as sea surface temperature, it can utilise a broader range of inputs.

Overall, our experiments provide new insights into the development of foundation models for the Earth system, enabling improved predictions on S2S timescales, while conserving performance for weather forecasts.

References:
E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

How to cite: Wilczyński, P., Lehmann, F., Ozdemir, F., Mohebi, S., Cheng, Y., Fuhrer, O., Mishra, S., Salzmann, M., Soja, B., Schemm, S., and Hoefler, T.: Subseasonal-to-Seasonal strategies for the Earth System Foundation Model (ESFM), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14846, https://doi.org/10.5194/egusphere-egu26-14846, 2026.

Foundation models trained on text and images are known to develop abstract internal features that align with human concepts, and that can be directly manipulated via activation steering in order to alter model behaviour. Whether scientific foundation models learn similarly abstract and domain-general representations has remained an open question. Inspired by recent work identifying single directions in activation space which control complex behaviours in LLMs, we show that a Walrus, a large physics foundation model, learns linearly steerable representations of physical phenomena. By computing the delta between activations representing contrasting physical regimes, we identify single directions in activation space that correspond to vorticity, diffusion, and even temporal progression. We find that injecting these concept directions back into the model during inference enables fine-grained causal control: vortices can be induced or removed, diffusion enhanced or suppressed, and simulations sped up or slowed down. Moreover, the concept directions we identified also appear to transfer successfully between unrelated physical systems, indicating that they are domain-general. These results suggest that scientific foundation models indeed learn general representations of physical principles and provides further evidence for the Linear Representation Hypothesis.

How to cite: Fear, R.: Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15558, https://doi.org/10.5194/egusphere-egu26-15558, 2026.

Machine Learning (ML) models, particularly deep neural networks, are often seen as black boxes, offering limited insight into how their predictions are made. This lack of transparency becomes especially important when ML is applied to critical domains such as numerical weather prediction, where traditional models are based on physical laws and differential equations.

Explainable AI (XAI) methods aim to address the black-box behavior by providing tools to interpret and understand model decisions. One such method is Layer-Wise Relevance Propagation (LRP), which traces the output of a neural network backward to assign relevance scores to input features based on their contribution to the prediction.

LRP has since been extended to Graph Neural Networks (GNNs) through the introduction of relevant walks, enabling interpretability in graph-structured data (GNN-LRP). These extensions have shown promise in areas such as image classification, sentiment analysis, and quantum chemistry. At the German National Weather Service (DWD), the AICON forecasting model employs a GNN architecture with message passing, similar in design to the GraphCast model.

In this work, we present an initial exploration of applying GNN-LRP to a simplified, toy version of a GNN model used as a representative of the AICON model. We investigate both saliency map-like visualizations and relevance walks, aiming to identify the most influential input features and their geographical location. While the current results are preliminary and limited in scope, this study tries to lay the groundwork for potential further research into explainability in graph-based weather prediction models.

How to cite: Pruschke, J. and Potthast, R.: Exploring Explainability for Graph-Based Weather Forecasting Models Using Layer-Wise Relevance Propagation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16587, https://doi.org/10.5194/egusphere-egu26-16587, 2026.

The rapid growth of wind and solar power is transforming energy markets, but their inherent variability makes accurate, real-time forecasting more essential than ever. Errors in day-ahead forecasting directly drive up imbalance costs, while the fast-paced nature of intra-day trading requires model inference that is much faster than traditional weather simulations. Foundation weather models such as the WeatherGenerator (WG) offer strong generalization and the potential for low-latency deployment, but their value for the energy sector depends on effective adaptation, as they are not originally designed for plant-specific tasks.

We will present results from applying WG to site-level wind and solar production forecasting in Turkey. The downstream task targets individual plants and is trained and evaluated on historical production observations across a multi-site portfolio. Our focus is on adapting WG for this operational setting by evaluating a spectrum of adaptation strategies, ranging from training task-specific 'tail' networks to fine-tuning the entire model. We report how these choices affect forecast performance and consistency across different sites and conditions, and we describe the resulting workflow in a form that can be carried over to portfolio-scale deployment.

Performance is benchmarked against our current operational baseline, which combines NWP results with machine-learning post-processing. We report MAE as the primary metric and discuss application-oriented indicators that relate forecast improvements to operational value in day-ahead and intra-day settings. The goal is to provide practical guidance on how to translate a foundation weather model into measurable benefits for renewable energy forecasting workflows.

How to cite: Afşar, A. M. and Bölükbaşı, G.: From foundation weather models to renewable operations: Adapting the WeatherGenerator for wind and solar production forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16880, https://doi.org/10.5194/egusphere-egu26-16880, 2026.

With increased availability of high quality diverse weather data, including reanalysis, satellite, surface stations, climate model data, the amount of data-driven foundation models (FM) in the environmental field has increased significantly over the past years with forecasting performances matching and sometimes exceeding physics-based numerical model predictions.  However, most FMs are trained with one dataset or a few datasets with similar sampling and/or resolution properties. While the proposed models achieve remarkable results with the datasets and variables they are trained on; it would be hard to anticipate similar performance under partially missing observations across different dimensions at test time. Similarly, typical design considerations risk limiting usage of these FMs to other heterogeneous datasets concerning the broader Earth sciences community.

We propose Earth System Foundation Model (ESFM), an FM capable of handling heterogeneous observations (i) across different resolutions, (ii) with spatially gridded and non-gridded nature, and (iii) with little to extreme sparsity. We achieve this through simple architectural design considerations and a masked training protocol. Namely, we bin similar ranges of grid resolutions together, while optimizing a different set of tokenizers for significantly different resolution bins to accommodate a single FM for observations across different resolutions. Similarly, we tokenize non-gridded data (i.e., station) separately with a single pixel patch size. Finally, we use variable specific tokenizers, coupled with learnable missing observation tokens, that allow ESFM to naturally accommodate for various subsets of available variables across different spatiotemporal positions. 

In this exploratory study, we show that ESFM is a flexible FM that can achieve impressive forecasting performance under different adverse setups with missing test data across any dimension on ERA5; spatio-temporal and inter-variable. We further test forecasting performance of ESFM in very sparse satellite imagery (3% pixel occupancy) data as well as station data. 

The proposed framework; also compatible for different backbone architectures than the one we experimented with; provides a general approach for integrating diverse Earth system data sources with varying resolutions, sampling patterns, and availability. This makes ESFM particularly relevant for the broader environmental sciences and Earth and space sciences, where challenges related to data heterogeneity and missing observations are central to the development of next-generation data-driven environmental modeling systems.

How to cite: Ozdemir, F., Cheng, Y., Mohebi, S., Lehmann, F., Adamov, S., Trentini, L., Huang, L., Lingsch, L., Zhang, Z., Fuhrer, O., Soja, B., Mishra, S., Hoefler, T., Schemm, S., and Salzmann, M.: ESFM - A foundation model framework for heterogeneous data integration, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18011, https://doi.org/10.5194/egusphere-egu26-18011, 2026.

Dynamics within ecosystems vary in their transience and persistence. While some environmental drivers exert near-instantaneous control on productivity, others exhibit hysteretic responses in which the influence of past conditions persists beyond the moment of exposure. For instance, In gross primary productivity (GPP), incoming shortwave radiation predominantly regulates photosynthesis on a diurnal timescale, whereas temperature, vapour pressure deficit, and soil moisture may exert delayed or cumulative effects associated with acclimation, stress accumulation, and recovery. Accurately representing these temporal dependencies is essential for credible ecosystem modelling, yet remains challenging for both statistical and machine-learning approaches. 

Sequential models, such as Long Short-Term Memory (LSTM) networks, offer a flexible means of learning temporal dependencies directly from observations, without requiring predefined assumptions about lag structure. However, their opaque internal representations raise concerns regarding scientific interpretability and trust, limiting their use beyond prediction. Explainable AI (XAI) methods provide a method of interrogating these learned representations, enabling an assessment of how, when, and for how long different drivers influence model outputs.

Here, we investigate the use of Integrated Gradients (IG) in characterising the temporal structure of driver influence learned by LSTM models trained on combined meteorological and vegetation state datasets. Attribution is examined across input sequences to distinguish short-lived from persistent controls on GPP, and to assess how these patterns vary across environmental conditions and seasonal contexts. This analysis is contrasted with hypothesis-driven "exposure-lag" representations derived from Distributed Lag Non-linear Models (DLNM), highlighting differences in how temporal influence is represented, constrained, and interpreted.

Rather than treating explainability as a post-hoc validation step, this work demonstrates how XAI can function as a scientific diagnostic tool, enabling interrogation of black-box models and supporting exploratory discovery of emergent temporal behaviour. The results illustrate how explainable ML can enhance trust in data-driven ecosystem modelling while offering complementary insights to traditional confirmatory approaches, particularly in complex, high-dimensional settings where temporal dependencies are unknown or context-dependent.

How to cite: Hughes, T.: From Prediction to Understanding: Using Explainable AI to Reveal Temporal Drivers of Ecosystem Productivity, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19181, https://doi.org/10.5194/egusphere-egu26-19181, 2026.

Weather foundation models are increasingly expected to operate under heterogeneous and imperfect observation settings while remaining computationally scalable. Building on the Earth System Foundation Model (ESFM) setting for heterogeneous data integration, we explore how Mixture-of-Experts (MoE) can support robust and efficient learning in multi-modal weather foundation models.

We introduce ESFM-MoE, an exploratory direction that combines conditional computation with climate-semantic routing, a routing principle that encourages expert specialization aligned with meaningful geophysical structure, rather than treating expert selection as a purely generic scaling mechanism. The motivation is that Earth-system data exhibit strong spatial organization, regime-like variability, and modality-dependent uncertainties; MoE offers a natural way to allocate capacity adaptively and promote structured specialization under such heterogeneity.

In this work, we discuss the design space and practical considerations of integrating MoE into Earth-system foundation models, focusing on how routing objectives and inductive biases can shape expert behavior and improve utilization. We highlight potential benefits for robustness to missing observations, scalable training and inference, and outline promising directions for climate-aware expert specialization in next-generation weather foundation models.

How to cite: Cheng, Y., Özdemir, F., Mohebi, S., Lehmann, F., Adamov, S., Trentini, L., Huang, L., Lingsch, L., Zhang, Z., Fuhrer, O., Soja, B., Mishra, S., Hoelfer, T., Schemm, S., and Salzmann, M.: ESFM-MoE: Climate-semantic routing for Earth System Foundation Model (ESFM) , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19620, https://doi.org/10.5194/egusphere-egu26-19620, 2026.

Attention-based deep learning models are becoming a ubiquitous approach for modeling the complex temporal dependencies in many vital Earth observation applications, such as agricultural monitoring. They typically consist of multiple attention heads, with each head containing attention weights that determine how temporal information is combined for the model's prediction. While analyzing the attention weights can provide insights into the model's workings, the existence of multiple heads makes it difficult to comprehend the extracted temporal information by the model. To overcome this issue, we propose an inherently interpretable approach that automatically weights the head importance during the model's forward pass. Our evaluation on the task of crop-type classification shows that the model maintains high accuracy while simplifying interpretation by highlighting only the most significant attention heads.

How to cite: Obadic, I., Rigon, L., and Zhu, X.: An Inherently-Interpretable Approach to Uncover the Head Importance of Attention Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19816, https://doi.org/10.5194/egusphere-egu26-19816, 2026.

Over the course of a decade, a single Earth observation satellite mission, such as Sentinel-2, can generate more than 10 petabytes of data. While this wealth of information offers a unique opportunity for pre-training geospatial foundation models, storing and processing such massive datasets is challenging, even for powerful HPC systems. One potential solution is the use of lossy compression techniques, which remove irrelevant information (e.g., noise and redundancy) while preserving as much relevant content as possible. This approach enables significantly larger training datasets for self-supervised learning, potentially offsetting the loss in data quality and yielding performance gains in downstream tasks. However, at the time of writing, the application of lossy compression in this context remains largely underexplored.

Here, we use Vector Quantized Variational Autoencoders (VQ-VAE) for compression of Sentinel-2 and Sentinel-1 data. We show that VQ-VAE is able to achieve compression rates of up to 65x with minimal reconstruction errors. Compared to classic, general-purpose compression techniques such as JPEG-2000, the VQ-VAE model attains 2-3x higher compression rates for the same reconstruction error. We also present ablation studies on the use of compressed versus uncompressed data for pre-training masked autoencoders (MAE) under a fixed physical storage budget, reflecting the constraints of resource-limited HPC systems. Finally, inspired by previous work in computer vision, we explore using the learned quantization scheme to construct a probabilistic masked autoencoder. Instead of predicting a deterministic reflectance or backscatter value, our probabilistic model predicts a categorical distribution over the learned codewords and is trained using cross-entropy loss. This formulation naturally incorporates uncertainty or bimodality into the masked autoencoding task, for example under cloudy conditions.

Our results demonstrate the potential and feasibility of neural compression techniques for the pre-training of large geospatial foundation models. Looking ahead, we aim to incorporate our findings into the training of WeatherGenerator-Land, an upcoming multi-modal foundation model for Earth's land surface. WeatherGenerator-Land will be used for vegetation forecasting, predicting land-atmosphere interactions, and high-resolution land surface temperature forecasts, with a particular focus on heat waves and urban heat islands.

How to cite: Hoffmann, S., Zehner, M., Benson, V., Wesselkamp, M., Martius, G., and Reichstein, M.: Neural Compression of Remote Sensing Data for the Pre-Training of Geospatial Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20782, https://doi.org/10.5194/egusphere-egu26-20782, 2026.

How to cite: Schilperoort, B., Richardson, R., Kalverla, P., and van den Oord, G.: Precipitation nowcasting over Western Africa using transfer learning with WeatherGenerator, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21352, https://doi.org/10.5194/egusphere-egu26-21352, 2026.

As artificial intelligence (AI) systems transition from research prototypes to operational tools in Earth system science and forecasting, establishing confidence and trust in their predictions becomes increasingly critical. Although the inputs and outputs of AI models are observable, their internal decision-making processes are often highly complex and difficult for humans to interpret, leading to their frequent characterisation as “black boxes” which are potentially untrustworthy.

In this work, we examine a range of explainable artificial intelligence (XAI) techniques designed to provide insight into AI model predictions. Many of these methods have been developed primarily with classification tasks in mind, raising important questions about their suitability for the regression-based problems that dominate geoscientific applications. We investigate the application of XAI methods to a machine learning derived emulator of the Lorenz ’63 system (an archetypal chaotic dynamical model) and review existing case studies that apply XAI in regression settings relevant to Earth sciences.

We highlight key challenges and limitations of current, general-purpose XAI approaches when applied to chaotic, continuous, high-dimensional, and physically constrained systems. Finally, we identify gaps in existing methodologies and discuss future directions for developing XAI techniques better aligned with the context-specific needs of regression problems in geoscientific modelling and forecasting.

How to cite: Higgs, I., Hunt, K., Jones, T., and Ellis, A.-L.: Evaluating explainable AI Methods for geoscientific regression: insights from applications and a chaotic toy model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-23208, https://doi.org/10.5194/egusphere-egu26-23208, 2026.

Senckenberg’s natural history collections encompass over 45 million physical specimens distributed across 11 facilities, with 1.6 million digitized records accessible in 124 collections. Additional digital objects are stored in various infrastructures, such as Edaphobase, a digital repository for harmonized soil information (physical, chemical, and biological), and the WildlIVE Portal, a platform for FAIR (Findable, Accessible, Interoperable, and Reusable) data sharing of biodiversity monitoring with edge sensors such as camera traps. Managing this heterogeneous landscape, ranging from legacy specimen data from the pre-digital era to newly digitized objects, presents significant challenges regarding legal compliance, data sovereignty, and the implementation of FAIR principles.

To address this volume, the FAIRenrich workflow automates the semantic annotation and maintenance of existing digital collection data. Handling the complexity of such enrichment requires AI models suited for partially non-deterministic tasks, incorporating an optional human-in-the-loop mechanism. By executing these workflows on a distributed network of stationary and mobile edge computing devices ('last-mile AI'), the architecture ensures strict adherence to data sovereignty and privacy requirements.

Beyond data curation, FAIRenrich's distributed architecture enables systemic efficiency through resource pooling informed by industry practice. Institutional edge infrastructure is rarely fully utilized; by networking heterogeneous devices, the system dynamically reallocates idle capacity to enrichment tasks. This mirrors industry practice: major technology operators, such as Google, systematically redeploy hardware from their refresh cycles into secondary-use programs.

FAIRenrich extends this model to legacy hardware: rather than treating end-of-life equipment as waste, the system enables cost-effective redeployment for delay-tolerant semantic enrichment tasks, such as inference workloads without strict latency requirements. By aligning workload scheduling to renewable peaks (e.g., photovoltaic installations), the approach implements carbon-aware scheduling principles used by major technology operators, achieving both infrastructure cost reduction and extended hardware lifecycles. This creates a circular-economy model for research institutions, transforming refresh-cycle surplus into productive scientific infrastructure.

This contribution demonstrates how FAIRenrich enables sustainable semantic annotation through a distributed edge architecture that simultaneously ensures data sovereignty, optimizes infrastructure utilization, and can realize cost-effective redeployment of legacy hardware. The approach exemplifies a scalable blueprint for research institutions seeking to decouple semantic enrichment from project-resource limitations through parallelization, temporal flexibility, and circular infrastructure practices.

How to cite: Wolodkin, A., Weiland, C., Grieb, J., and Brylka, R.: FAIRenrich: Distributed semantic annotation at the repository edge, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7020, https://doi.org/10.5194/egusphere-egu26-7020, 2026.

Recent advances in Large Language Models (LLMs) have created opportunities to support reasoning, discovery, and synthesis in Earth Observation (EO) and Earth Sciences, provided domain specificity and reliability can be ensured. In this work, we introduce Earth Virtual Expert (EVE), a comprehensive open-source initiative to develop, evaluate, and deploy a domain-specialized LLM for EO. EVE serves as a testbed for studying domain-adaptive training, grounded generation, and evaluation strategies tailored to scientific use, rather than general-purpose conversational performance.

As part of this initiative, we present EVE-instruct, a text-only, instruction-tuned and aligned LLM specialized for EO. Built on Mistral Small 3.2 (24B parameters) with a 128k context window, it focuses on domain-specific reasoning, question answering, and retrieval– and hallucination-aware generation, without significant tradeoff of general capabilities. We release all data used to train and evaluate EVE-instruct: a large-scale curated EO corpus of 3B tokens, synthetically generated fine-tuning datasets derived from this corpus (4B tokens), and manually-created EO-specific evaluation test sets comprising 7500 samples across multiple-choice and open-ended question answering, and factuality test sets.

To support trustworthy usage and deployment, we further develop a Retrieval-Augmented Generation (RAG) database from the curated corpus and a hallucination-detection module focused on factual consistency and scientific grounding. These components are integrated with EVE-instruct and deployed with a graphical user interface and accessible via API, currently supporting more than 300 users from the EO research and industry field.

All models, datasets, and code are publicly released at: https://huggingface.co/eve-esa and https://github.com/eve-esa.

How to cite: R. Atrio, À., Lopez, A., Rohit, J., Elhouadi, Y., Politi, M., Iyer, V., Bratières, S., Jamil, U., and Longépé, N.: EVE: An Open Source Earth Science LLM for Researchers, Policymakers, and the Public, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12877, https://doi.org/10.5194/egusphere-egu26-12877, 2026.

Large language models (LLMs) promise to make systematic reviews more scalable and less costly, but the validity of LLM-assisted evidence synthesis depends not only on accuracy, but also on which parts of the literature are effectively visible to a deployed model and how reliably they are interpreted. We report a large-scale, domain-specific evaluation of an end-to-end LLM-assisted workflow for a systematic review of national-scale land use/land cover (LULC) prediction research (11,817 records; 11,688 after de-duplication), using a single hosted LLM deployment (Qwen Max) as a concrete case study. At title–abstract screening, the model behaved as a recall-oriented filter, excluding 9,891/11,688 records (84.7%) and routing 1,797 records for human follow-up; compared with the human baseline, it excluded fewer studies (84.7% vs 91.8%) and shifted more records into OK and POSSIBLE (OK: 4.2% vs 1.5%; POSSIBLE: 10.6% vs 5.5%). For full-text extraction, structured fields showed high agreement with expert coding across 342 benchmark papers (mean scores: 0.84 categorical, 0.85 temporal, 0.87 set-based), whereas free-text summaries were more variable (mean 0.79 overall; cosine similarity 0.51–0.87 across narrative fields despite high BERT-F1). In our case study, the workflow was completed in approximately one day on a single workstation for ~US$106 in API costs. Critically, full-text processing also produced explicit refusals: 7/2,084 candidate papers in deep screening and 2/345 papers targeted for insight extraction were blocked as “sensitive” geopolitical content. Although rare, these refusals were non-random and concentrated in contested regions, illustrating how LLM-specific constraints can introduce structured missingness that systematically removes or misinterprets evidence in precisely those settings where land-use conflict and governance are most salient. LLM-assisted reviews can therefore make previously prohibitive syntheses tractable. However, they must be embedded in transparent, human-led workflows that monitor and log model failures including refusals, omissions, and misreadings, and apply targeted auditing to detect and correct systematic blind spots.

How to cite: Derdouri, A. and Masago, Y.: LLM Workflow for Land-Use Prediction Evidence Synthesis: Efficient Screening, Selective Refusals, Reportable Gaps, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15955, https://doi.org/10.5194/egusphere-egu26-15955, 2026.

Large Language Models (LLMs) offer transformative capabilities for scientific workflows, enabling scalable analysis, evidence synthesis, and insight generation. We present an agentic LLM workflow, applied to the EU-funded SWITCH project, which investigates the environmental impact of dietary choices, including CO₂ emissions associated with food consumption.

Validated questionnaire responses from SWITCH participants are securely anonymized and processed using machine learning methods, including clustering and classification, interpreted using ExaplainableAI (XAI) to ensure transparency of feature contributions, to generate food behavioral profiles, including nutritional and environmental habits, while preserving individual privacy. These outputs guide the construction of structured agent directives, enriching contextual information and constraining the LLM to provide scientifically grounded answers and data-driven insights.

Responses are generated within a Retrieval-Augmented Generation (RAG) framework over a curated Data Lake of revisioned documents, including project deliverables, scientific reports, and nutrition-environment datasets covering sustainable diets, CO₂ emissions, and European food policy. The combination of ML-generated profiles and the RAG context acts as a set of constraints, ensuring that LLM outputs remain traceable, grounded, and aligned with verified evidence.

Human-in-the-loop review ensures the quality and correctness of the ML-generated profiles, the construction of agent directives, the LLM outputs, and the revisioned documents used in the RAG framework, while metadata and traceability mechanisms ensure auditability, reproducibility, and risk mitigation.

Our results demonstrate that combining classical machine learning, structured agent directives guided by clustering and classification, RAG grounding, metadata and traceability, and human oversight enables trustworthy, effective, and transformative scientific analysis, highlighting the potential of agentic LLMs for scalable, insight-driven applications in research while ensuring responsible AI deployment.

How to cite: De Carlo, M., Mirto, M., Epicoco, I., Nassisi, P., and Chiriacò, M. V.: Explaining dietary CO₂ impact with trustworthy agentic LLMs, ML, and XAI, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19791, https://doi.org/10.5194/egusphere-egu26-19791, 2026.

In this talk we will outline our learnings from building two production grade agentic systems for data discovery, retrieval and analysis. For both applications trustworthiness, reliability and reproducibility were key criteria that we took into account from the start.

Producing insights and surfacing data in a repeatable and transparent way is not simply a nice-to-have feature, it is indispensable for adoption. In practice, even if the answer from a chatbot is scientifically correct, users will only rely on the outputs for decision making if they trust the system. Users by now have enough experience with chatbots to know that LLMs tend to exaggerate and hallucinate. This has created a healthy skepticism with regard to output from agentic systems. In scientific domains, it is therefore not sufficient to guarantee a correct result, it also has to be presented in a way that is transparent and reproducible. Despite all advances in the LLM domain, this continues to be a challenge in building agentic systems.

We tackled this challenge by adopting a series of techniques that we will illustrate with concrete examples from building production ready agentic applications. One of the main principles is that we rely on the LLM mainly for orchestration of well known tools instead of relying on the generative capabilities of the models. For analysis we built the systems in a way that the transformations on the original data are reproducible. One technique is to use LLMs to write code for analyis that can be stored and used to reproduce the results. This allows end-to-end tracing of where the data is coming from and how it was transformed to produce insights through statistics and charts. We will also mention our approach to evaluation of the agent and share insights from early user research already performed for these systems.

We will illustrate these principles with concrete examples from the two agentic systems outlined below. 

The Destination Earth Digital Assistant, built in collaboration with ECMWF, provides general information about Destination Earth and helps users to discover and retrieve data from the DestinE Digital Twins.

Global Nature Watch, built in collaboration with WRI and the Land and Carbon Lab, provides governments, companies, and communities with trusted, open data and intelligence-driven insights on land conditions and land-use change to enable efficient and evidence-based decisions for nature protection and restoration.

GNW is publicly accessible today and the DestinE Assistant is also planned to be launched publicly before EGU26.

How to cite: Wiesmann, D., Solvsteen, J., Pain, A., Barrett, A., Sweet, C., Mestre, R., da Silva, D., Riany, F., Pérez, F., Goodman, L., Farra, M., Ranjan, S., Mishra, T., Bhangar, S., and Anwar, S.: Making Scientific Data Accessible with LLMs while Preserving Authority and Reliability: Lessons Learned from Building Production Grade Agentic Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19885, https://doi.org/10.5194/egusphere-egu26-19885, 2026.

The exponential expansion of academic literature across complex environmental domains has created a gap where the volume of research outpaces human capacity for effective integration. While Large Language Models (LLMs) offer a transformative solution to bridge this gap, their deployment in rigorous scientific inquiry is frequently compromised by model stochasticity, potential for hallucination, and the opacity of automated reasoning. Addressing the critical imperative for dependable and reproducible AI, this study presents a robust workflow designed to ensure methodological rigor and evidential integrity in the rapid and reliable synthesis of large-scale scientific literature.

We operationalised this framework within the domain of urban water management, specifically to analyse complex Knowledge Transfer (KT) strategies from a corpus of over 1,500 unstructured articles. To mitigate the risks inherent in generative AI, we developed a multi-layered validation protocol. First, we deployed an AI-assisted screening mechanism to filter the initial corpus down to 115 highly relevant articles, ensuring data relevance. Second, we implemented a Human-in-the-Loop design to iteratively synthesise a comprehensive analytical framework. By refining LLM-generated insights against domain expertise, we consolidated 24 operational attributes that specifically characterise the operational mechanisms of learning strategies from the corpus, preventing ungrounded inference while capturing emerging learning dynamics. Third, we addressed model variability through iterative Multi-LLM Triangulation (utilising Gemini, ChatGPT, and Deepseek). By repeatedly coding the 115 articles with the framework, we quantified qualitative insights to analyse how distinct learning strategies manifest their operational mechanisms. Finally, we employed Multiple Correspondence Analysis (MCA) and Hierarchical Clustering (HAC) to analyse the quantified results, categorising the eight identified learning strategies into three distinct clusters based on their functions and usage contexts, thereby effectively harnessing the LLM-generated insights.

Beyond this specific application, this research contributes a methodological blueprint for responsible AI integration in scientific inquiry. It demonstrates that combining theory-driven constraints with statistical verification is essential to elevate LLM-generated insights to the standard of reproducible scientific evidence.

How to cite: Wang, C., Corzo, G., Van Oel, C., and Zevenbergen, C.: Toward trustworthy AI in systematic reviews: a statistically validated AI-augmented framework for analysing knowledge transfer strategies in urban water management, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20948, https://doi.org/10.5194/egusphere-egu26-20948, 2026.

Problem:

The elicitation of user requirements represents a critical first step in successful planning software projects, particularly when a diverse set of use cases must be considered. In practice, this process is often carried out through oral interviews and handwritten notes, which is time-consuming, error-prone, and makes structured processing and subsequent analysis difficult.

Approach:

We present an integrated, automated pipeline framework that supports the systematic collection and analysis of user requirements, from capturing user perspectives, to model-supported analysis. The goal is to gather requirements consistently across different stakeholder roles, including project lead, technical staff, and data scientists. In the first step, requirements are collected using a web-based, structured questionnaire. In the second step, the questionnaire serves as a guideline in follow-up interviews for detailed case description to further refine the requirements.

Figure 1. the pipeline for requirements elicitation, LLM-based analysis, and human-in-the-loop review.

The interviews are recorded and automatically transcribed using a domain-adapted Language Recognition Component (LRC) based on open-source Automatic Speech Recognition (ASR) models. The resulting transcripts are combined with questionnaire responses and initial analysis artifacts, such as charts and diagrams, and processed within a Large Language Model (LLM) pipeline. After requirements have been collected, the pipeline supports the systematic inspection of individual requirements and their consideration in project planning.

Using a dedicated prompt schema, the LLM-based analysis supports the identification of functional and non-functional requirements, highlights open needs, clusters related issues, and organizes the results according to relevant work contexts. A human-in-the-loop review module enables targeted corrections, quality assurance, and iterative improvement of the analysis results.

Implementation test:

To validate our end‑to‑end requirements‑engineering pipeline, we applied it to the IACS‑AI Data‑Management‑Remodeling project. A web‑based survey (Nov–Dec 2024) yielded 53 responses, giving an initial, structured view of user requirements. Subsequently, we held seven interviews with 18 participants (project managers, engineers, data‑scientists), producing > 460 min of video afterward transcribed with the freeware tool Scraibe. 

Prompt‑engineering routine fed these inputs to a Large Language Model (Llama 3.3), which detected semantic clusters, classified requirements, and identified problem statements. For each requirement, we kept the highest‑probability class for further review.  The resulting insight shaped the next milestone: the design and implementation of a data‑pipeline architecture that fulfills the extracted functional and non‑functional requirements. 

Conclusion:

The reproducible design of the pipeline ensures traceability by documenting when, by whom, and in which context requirements were expressed, as well as how project decisions are derived from them. This results in a lightweight yet structured approach to requirements elicitation that improves transparency as well as reproducibility and reduces manual effort and errors.

Because the pipeline is generic, it is ideal for contexts with many stakeholders, heterogeneous use cases, and strong documentation‑traceability needs therefore besides our scientific implementation test it can also be utilized in the field of enterprise software, AI‑driven data projects, e‑government systems, and regulated domains such as healthcare or finance. 

How to cite: Fardhosseini, Z., Ackermann, A., and Oerder, B.: Automating Requirements Elicitation using Large Language Models and Speech Processing, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20967, https://doi.org/10.5194/egusphere-egu26-20967, 2026.

We present Collaborative Agent Reasoning Engineering (CARE), a disciplined methodology for engineering Large Language Model (LLM) agents in scientific domains. Unlike ad-hoc trial-and-error approaches, CARE specifies behavior, grounding, tool orchestration, and verification through reusable artifacts and systematic, stage-gated phases. The methodology employs a three-party workflow involving Subject-Matter Experts (SMEs), developers, and LLM-based helper agents. These helper agents function as facilitation infrastructure, transforming informal domain intent into structured, reviewable specifications for human approval at defined gates. CARE addresses the "jagged technological frontier", characterized by uneven LLM performance, by bridging the gap between novice and expert analysts regarding domain constraints and verification practices. By generating concrete artifacts, including interaction requirements, reasoning policies, and evaluation criteria, CARE ensures agent behavior is specifiable, testable, and maintainable. Evaluation results from a scientific use case demonstrate that this stage-gated, artifact-driven methodology yields measurable improvements in development efficiency and complex-query performance.

How to cite: Ramachandran, R., Jha, N., and Ramasubramanian, M.: Collaborative Agent Reasoning Engineering (CARE): A Structured Methodology for Systematically Engineering AI Agents for Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22068, https://doi.org/10.5194/egusphere-egu26-22068, 2026.

Semantic segmentation is the foundation of a wide range of practical applications, such as urban planning, climate modeling, and environmental protection, all of which have direct socio-economic implications. However, the accelerating densification of metropolitan regions in developing countries complicates accurate mapping of fine-scale urban land uses, as three-band optical imagery often fails to capture spectral variability and the restricted capacity of the CNN-based model to establish spatial and inter-band relationships. Therefore, to address these limitations, we propose a multi-modal architecture built on a SegFormer-B2 backbone. The pipeline integrates auxiliary datasets of DEM for surface information, SWIR for capturing water absorption characteristics, and an ancillary dataset of built-up layers for enhanced urban boundary delineation, along with multi-temporal false-color composites from LISS-4 and Sentinel-2 over the Bengaluru region. The proposed framework integrates convolutional feature extraction with transformer attention to jointly learn local spectral–spatial patterns and global cross-band dependencies.  Attention-guided up-sampling, a hybrid loss function, and cross-attention modules are incorporated to strengthen feature fusion across heterogeneous modalities by establishing a link between the multi-band synergy of the Auxiliary and Ancillary datasets. Empirical evaluation reveals consistent qualitative improvement and higher overall accuracy, with substantial gains for Barren land when incorporating SWIR and vegetation, and when integrating DEM. These results validate the effectiveness of the proposed framework in overcoming spectral insufficiency and spatial ambiguity, as it outperforms baseline models. Overall, the proposed approach offers a scalable and transferable solution for private developers and government agencies seeking robust, fine-resolution mapping to support a sustainable and structured urban environment.

Keywords: Urban mapping, Deep learning architecture, Spectral Feature extraction, Performance Optimization

How to cite: Kumar, V. and Haridas Aithal, B.: Contextual Aware Hybrid Deep learning framework: Assessment with Auxiliary and Ancillary Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1347, https://doi.org/10.5194/egusphere-egu26-1347, 2026.

Growing demand for critical raw materials over the coming decades underscores the need for robust, continental-scale frameworks to identify new mineral resources. Volcanogenic massive sulphide (VMS) deposits supply Cu, Zn, Pb, Au, and Ag, and they play a crucial role in meeting Europe’s growing demand for strategic raw materials. Despite Europe’s long mining history and extensive geological datasets, mineral prospectivity assessments remain largely restricted to national boundaries, limiting the ability to evaluate mineral systems that operate across regional tectonic domains. This study develops the first integrated European-scale prospectivity model for VMS by merging harmonised public geoscience datasets within a mineral-system and machine-learning (ML) framework. This work is carried out as part of the EU-funded TERRAVISION project, which aims to enhance the entire critical raw materials value chain towards implementing sustainable mining practices.

The modelling approach builds on and extends existing regional-scale frameworks by addressing several persistent challenges in regional-scale exploration. A positive–unlabelled training strategy was used to mitigate the lack of reliable negative labels, and ML models capable of estimating uncertainty, along with multiple explainability techniques, were applied. To ensure that predictors capture meaningful geological processes, both data-driven and knowledge-based feature selection were implemented. Model explainability was evaluated through three complementary approaches: (i) built-in feature importance from the ML classifier, (ii) permutation feature importance to assess the robustness of predictor influence, and (iii) SHapley Additive exPlanations (SHAP) values to quantify local and global predictor contributions. Together, these methods provide transparent, interpretable insights into the geological and geophysical variables that indicate prospectivity patterns. The model successfully identified over 97% of known VMS deposits and occurrences with spatial patterns showing strong correlation between high-probability areas and established mineralisation. Importantly, they also highlight prospective trends in regions with limited documented exploration.

The analysis highlights several metallogenic zones that exhibit geological and geophysical signatures consistent with favourable mineral-system conditions, but where known deposits are sparse. These areas represent potential greenfield opportunities at a continental scale. The study also illustrates the value of applying the mineral system concept to regional datasets. Harmonised lithological data and spaceborne geophysical data contribute significantly to mapping crustal-scale structures and tectonic domains with a history of submarine seafloor volcanic activity, a key requirement for VMS formation. More broadly, the proposed framework is transferable to other deposit types and illustrates the strategic potential of continental-scale, process-informed and explainable ML approaches to strengthening Europe’s strategic raw-material knowledge base through consistent, process-informed regional assessments.

How to cite: Dekavalla, M., Tenorio Matanzo, S., López Del Río, M., Papathanasiou, C., and Amditis, A.: Continental-Scale Prospectivity Modelling of Volcanogenic Massive Sulphide Deposits in Europe Using a Mineral-System and Explainable Machine-Learning Framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1610, https://doi.org/10.5194/egusphere-egu26-1610, 2026.

The aim of this work is to propose a new technique for automatic fault tracking from 3D seismic data using the 2D Continuous Wavelet Transform (CWT) method combined with artificial intelligence. Time slices of the variance attribute, derived from the 3D seismic data and chosen by the user, are analysed using the 2D CWT with the 2D Mexican Hat as an analysing wavelet, and the maxima of the modulus of the 2D CWT are mapped for the full range of scales. The ensemble of mapped maxima for the set of time slices is filtered using a Convolutional Neural Network machine. Machine training is performed with a supervised mode using the manually tracked faults as a desired output. Application to real data shows the efficiency and robustness of the proposed method, which can greatly help seismic interpreters in avoiding manual fault tracking, a difficult and time-consuming task.

How to cite: Aliouane, L. and Ouadfeul, S.-A.: Automatic fault tracking from 3D seismic data using the 2D Continuous Wavelet Transform combined with a Convolutional Neural Network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1908, https://doi.org/10.5194/egusphere-egu26-1908, 2026.

Cyanobacterial harmful algal blooms (cHABs) pose serious concerns to drinking-water safety, aquatic ecosystem health, and recreational water use across the globe. cHABs in situ data collection relies on sparse and irregular measurements and hinders the reliable learning of complex ecological processes. Although recently data-driven models have improved bloom prediction skills, the explicit reliance on the black-box nature of these models undermines scientific trust and restricts the actionable value of model outputs. In this study, we developed a mechanism-aligned deep learning framework that embeds ecological process structure directly into the learning architecture using diverse data sources. We incorporated detailed remote-sensing-based atmospheric and environmental variables, including aerosol derived deposition, nutrient wet deposition, and meteorological data. We temporally aggregated this data to reflect both short-term forcing and cumulative conditions over space and time. We evaluated our framework for 2,200 lakes across the continental United States from 2018 - 2023, with a one-week-ahead bloom prediction task. Our model is trained on 2018–2021, validated in 2022, and tested on 2023 dataset. Our preliminary results show stable generalization under diverse spatiotemporal domain shifts (R² = 0.54, RMSE 0.59) with reduced seasonal bias relative to conventional deep learning baselines. In addition to predictive accuracy, our architecture demonstrates high explanation faithfulness (OTA = 0.83) and positive alignment with independent physical proxies (auxiliary physical proxies, R² = 0.36). This further demonstrates that architecture learned representations remain physically consistent despite the absence of direct mechanism labels. Our work advances a new paradigm for trustworthy environmental predictions and provides a novel foundation for actionable bloom management and policy decision support in data-limited inland water systems.

How to cite: Alamdari, N., Imtiaz, S. U., and Nasr Azadani, M.: Deep Learning for Trustworthy Prediction of Cyanobacterial Blooms across CONUS Inland Waters, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3574, https://doi.org/10.5194/egusphere-egu26-3574, 2026.

Stratigraphic forward modeling (SFM) is a geological modeling framework that simulates depositional processes in sedimentary systems (e.g., deep water, fluvial, delta), enabling the generation of stratigraphic architectures and reservoir property distributions. This approach is particularly effective in reproducing the realistic non-stationarity and geological heterogeneity of deep-water reservoirs, which are difficult to capture using conventional geostatistical methods such as two-point and multipoint statistics. However, SFM results are highly sensitive to small variations in initial geological input parameters, making the integration of observational data such as well logs and seismic data challenging and thereby limiting its application at an industrial scale.

In this study, we propose a novel geological model characterization framework that combines SFM with a generative artificial intelligence approach capable of achieving both high generation efficiency and robust geological realism (Fig. 1). First, an SFM-based geological model is constructed and then preprocessed to make it suitable for neural network training. A single-image diffusion model, SinFusion, is then applied to learn the geometric and property distributions of the geological model and to enable multiple equivalent generations. Furthermore, a well data integration strategy is developed using the aforementioned trained SinFusion. By infusing well data during the reverse diffusion process, the proposed method allows seamless conditioning on well data regardless of the number or spatial locations of wells. This enables immediate model updates when new well data become available in the field with no further need for costly model retraining, ensuring high flexibility.

The validity of the proposed method is evaluated through quantitative comparisons of spatial continuity, property distributions, and geometric pattern similarity with the original SFM model. The results demonstrate that the proposed method can efficiently generate multiple geological realizations and is well suited for ensemble-based uncertainty assessment. Moreover, the proposed method has the potential to expand the applicability of SFM toward industrial-scale geological modeling workflows.

Fig. 1. Overview of the SinFusion framework for geological model augmentation and well data integration based on stratigraphic forward modeling

This work was supported by Korea Gas Corporation (RD2025-0071).

How to cite: Park, E., Lee, H., Yoon, J., and Jo, H.: SinFusion-based Geological Model Augmentation and Well Data Integration, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4683, https://doi.org/10.5194/egusphere-egu26-4683, 2026.

Development of effective urban climate adaptation and mitigation strategies requires comprehensive spatial information of rooftops and buildings. The aforementioned information is important for assessing the ecosystem services provided by green and blue infrastructure in urban areas, especially for urban heat island (UHI) mitigation and energy conservation. While green roofs are widely acknowledged as a promising solution for enhancing thermal comfort in urban climate, most existing research tends to focus either on mapping current green rooftops or the potential rooftops to implement green rooftops.

This study presents a modified deep convolutional neural network-based rooftop classification framework, based on the Roofpedia framework originally created by the Urban Analytics Lab at the National University of Singapore (NUS). The model leverages high-resolution aerial imagery and incorporates slope of the rooftop to assess green roof suitability. The proposed model uses publicly available geospatial datasets from Swisstopo such as aerial images from SwissImage dataset, elevation data from the swissALTI3D digital terrain model, and building footprints from the swissTLM3D vector dataset.

When the study applies the implemented model to Bern, Switzerland, the model provides the output with labelling the rooftops into four categories: (1) existing green roofs, (2) rooftops suitable for green roof installation, (3) rooftops with solar panels, and (4) flat rooftops which are unsuitable for roof greening. To improve the accuracy and practicality of the classification, roof slope thresholds derived from terrain model were integrated alongside spectral analysis to reflect real-world installation conditions.

The model demonstrated high predictive performance with training loss of 0.0134, mean Intersection over Union (mIoU) of 0.908, and Matthews Correlation Coefficient (MCC) of 0.901. Validation metric demonstrated the robustness with validation loss of 0.0292, mIoU of 0.843, and MCC of 0.822. Comparison with the original Roofpedia framework, the modified model shows significant improvements in multi-class rooftop classification, particularly in identifying realistic opportunities for green roof expansion.

The inclusion of potential green rooftop class, combined with slope-based constraint, allows for a practical and realistic assessment of rooftop suitability for green roof installation. The modified Roofpedia model assists urban planners and decision makers with evidence-based information to support future green infrastructure deployment in Bern and other Swiss cities. Furthermore, the proposed framework is transferable and can be readily replicated in cities worldwide.

How to cite: Ko Ko, H. Y. and Rast, M.: From Rooftops to Ecosystem Services: Deep Learning–Driven Green Roof Potential Assessment in Bern, Switzerland, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4820, https://doi.org/10.5194/egusphere-egu26-4820, 2026.

The urban-rural fringe (URF) has become the most dynamic area of land use transition and urban-rural factor flows during urbanization, forming a critical focus of sustainable land management. However, existing identification methods have not adequately captured the fine-scale textures and ambiguous transitional boundaries characterizing the URF. Taking Kunming City as the study region, this study developed a lightweight convolutional neural network (UF-Net) to extract spatial textures and boundary features, integrating it with eXtreme Gradient Boosting to construct a hybrid recognition framework. Multisource remote sensing and geospatial datasets were employed to delineate the URF from 2013 to 2023, and stage-specific driving mechanisms were examined using propensity score matching and binary logit models. The results showed that our framework achieved an overall accuracy of approximately 94% for both periods. Over the decade, built-up areas expanded markedly, and the spatial structure evolved from a single-core pattern characterized by fragmented peripheral development to a polycentric configuration with increasingly continuous URF zones. Chenggong and southern Guandu District emerged as major growth frontiers, while URF morphology shifted from linear to ring-shaped and cluster-type forms. Furthermore, the drivers of urban expansion transitioned from dominance by natural terrain and ecological suitability to a regime shaped primarily by human activities and transport accessibility. The proposed hybrid recognition framework, integrating deep feature extraction with ensemble-based classification, establishes a generalizable methodological path for interpreting URF evolution, providing analytical support for optimizing urban spatial structure and sustainable development strategies.

How to cite: Zhu, Y., Hu, T., Li, X., Zhang, D., and Peng, J.: A Deep Learning and Ensemble Decision Method to Identify the Urban–Rural Fringe: A Case Study in Kunming City, China, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5616, https://doi.org/10.5194/egusphere-egu26-5616, 2026.

Self-supervised learning has become a prominent technique for representation learning in Earth observation, largely due to the vast volumes of unlabelled data available in observation archives. However, apart from masked auto-encoding (MAE) techniques and contrastive learning, the diversity of geospatial self-supervised learning schemes in the existing literature remains limited.

This work explores the task of structure and texture disentanglement as an alternative route to self-supervised learning in the domain of Earth observation. Inspired by the Swapping Autoencoder architecture, this pipeline involves an encoder tailored to extract disentangled textural and structural information from an image and reconstruct it back to the image domain. Crucially, it includes an augmentation step that swaps texture and structure embeddings from different samples. This synthetic generation is driven by adversarial training, employing two discriminators: one responsible for assessing the likelihood of the image as a whole being real, and the other for assessing whether individual patches in the image are consistent with the source texture vector.

The texture embedding extracted from the image acts as a global vector describing the aggregated statistics of local features, while the structure embedding represents how these features are distributed in space. This preliminary work explores the potential of this approach on a domain where image labels are particularly scarce: cloud formation types in high-resolution optical imagery. The pipeline is tested on a large collection of cloudy Sentinel-2 images with the goal of identifying observational clusters of cloud formations that share similar properties, as part of the Clouds Decoded project. This work introduces a foundational architecture for this framework along with several methods of analysis that leverage the resulting deep neural network.

How to cite: Czerkawski, M., Francis, A., Borne--Pons, P., Bertozzi, B., and Campbell, J.: Disentanglement of Structure and Texture Representations as a Method of Self-Supervision for Earth Observation Data: A Case Study on Cloud Type, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7605, https://doi.org/10.5194/egusphere-egu26-7605, 2026.

Remote sensing time series monitoring plays a vital role in capturing the dynamic evolution of the Earth’s surface. Recent deep learning based temporal change detection (TCD) methods have achieved remarkable progress under cloud-free optical image sequences. However, optical imagery is frequently affected by clouds and cloud shadows, resulting in pervasive and irregular data gaps that disrupt temporal continuity and sampling regularity. Consequently, current TCD approaches struggle to cope with highly dynamic surfaces and long-term or irregularly missing observations, often leading to inaccurate change detection results. To address these challenges, we propose UniRT, a unified framework that jointly performs time-series reconstruction and change detection, enabling robust monitoring from image sequences with missing observations. Specifically, a temporal-adaptive module is seamlessly embedded into a spatiotemporal learning framework while maintaining a lightweight architectural design. In addition, a time-aware decoder is introduced to better capture temporal dependencies and enhance robustness and generalization capability under irregular sampling conditions. Extensive experiments conducted on DynamicEarthNet and SpaceNet7 demonstrate that UniRT consistently outperforms state-of-the-art methods in temporal change detection, particularly in challenging scenarios characterized by severe data gaps and highly dynamic surface changes.

How to cite: Huang, H., Shao, Z., Zhong, C., Zhu, D., and Zhang, W.: UniRT: A Unified Framework for Time-Series Remote Sensing Image Reconstruction and Change Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7703, https://doi.org/10.5194/egusphere-egu26-7703, 2026.

The research presented in this study addresses the subject of large-scale soil type classification. It is based on multispectral data from the Sentinel-2 satellite along with recent advances in deep learning for tabular data analysis. Initially we created a soil dataset aligned with the World Reference Base (WRB) classification system. This dataset was created by integrating Sentinel-2 spectral bands with different indices regarding vegetation, exposed soil conditions, mineralogical composition, and moisture dynamics. The study assesses the performance of different classification models, and some hybrid approaches using ensemble learning techniques. We applied and assessed several techniques for data balancing and augmentation to address the uneven class distribution that often exists in soil datasets. The results show that combining multispectral satellite features with specific spectral indices and various learning methods offers an effective and scalable way to generate WRB-consistent soil maps from Sentinel-2 data.

Acknowledgment:

This work is supported by the project "Romanian Hub for Artificial Intelligence-HRIA", Smart Growth, Digitization and Financial Instruments Program, MySMIS no. 351416.

How to cite: Bacu, V.: Deep Learning–Based Soil Classification from Sentinel-2 Multispectral Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8216, https://doi.org/10.5194/egusphere-egu26-8216, 2026.

Although numerous satellites have been developed and launched recently, low Earth orbit satellites offer high spatial resolution but long revisit cycles, resulting in low temporal resolution, whereas geostationary satellites offer high temporal resolution but low spatial resolution. As a result, there are still limitations in reliably acquiring satellite images with high spatiotemporal resolution. To overcome these limitations, research on super-resolution fusion using various satellite images is underway. However, challenges such as data loss due to clouds, differences in revisit cycles between satellites, and sensor characteristic mismatches make it difficult to produce fused super-resolution images. Therefore, this study proposes a multi-satellite-based fusion framework that addresses these issues and reliably generates spatiotemporal super-resolution fusion images.

To this end, this study utilized various satellite images, including GK2A (high temporal frequency, 2km resolution), MODIS (high temporal frequency, 500m resolution), GOCI-II (high temporal frequency, 250m resolution), Landsat-8 (30m resolution), Sentinel-2 (10m resolution), PlanetScope (2.8m resolution), and KOMPSAT-3(3m resolution). Each satellite image underwent preprocessing steps, including geometric correction, radiometric correction, BRDF (Bidirectional Reflectance Distribution Function) correction, and normalization, to ensure spatial alignment and radiometric consistency.

Subsequently, a deep learning model based on DeepLabV3+ ResNet101 was used to generate cloud mask label data, creating mask labels for clouds and missing areas in the video. These labels were then used to apply a gap-filling technique to fill in the cloud and missing regions. Finally, a step-by-step resolution enhancement image fusion method based on spatial resolution was employed to produce a spatiotemporal super-resolution fused image.

The final super-resolution fused image will be validated using spectral data collected from a ground observation tower located in Naju, Jeollanam-do, South Korea. The multi-satellite fusion framework proposed in this study can efficiently overcome the limitations of spatiotemporal resolution by utilizing deep learning and physics-based models during various processing stages. The fusion results are expected to be applicable in various remote sensing fields, such as detecting climate change and environmental variations.

Acknowledgements: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00515357).

How to cite: Han, D., Hahn, S., Ryu, Y., Jeong, S., Ha, J., and Yeom, J.: A Deep Learning and Physics-Based Multi-Satellite Fusion Framework for Spatiotemporal Super-Resolution Image Generation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9172, https://doi.org/10.5194/egusphere-egu26-9172, 2026.

Mountainous regions of the Kumaon Himalayas are particularly prone to landslides due to steep terrain, weak geological conditions, intense monsoon rainfall, and increasing human activity. In such environments, continuous ground-based monitoring is often difficult because of poor accessibility, dense vegetation cover, and frequent cloud conditions. Microwave remote sensing, especially satellite-based Synthetic Aperture Radar (SAR), offers a reliable, weather-independent means of monitoring surface deformation over large areas and long time periods.

This study applies an integrated multi-temporal InSAR (MT-InSAR) and deep learning framework to investigate surface deformation and landslide activity in the Nainital region, Kumaon Himalayas, India. Sentinel-1A SAR data (2020–2025) were processed on the ASF Vertex HyP3 cloud platform using the GAMMA Small Baseline Subset (SBAS) processing chain. The cloud-based workflow automates key interferometric steps, enabling efficient processing of multi-year SAR archives without the need for local high-performance computing facilities.

Time-series inversion and analysis were carried out using MintPy in a GPU-enabled OpenSARLab environment. Weighted least-squares inversion was applied to generate line-of-sight (LOS) deformation time-series and mean LOS velocity maps. In addition, ENU decomposition was performed, and the vertical (Up) component was used for subsequent analysis. The resulting five-year deformation record highlights marked spatial variability across the study area, reflecting deformation associated with slow-moving landslides, slope creep, and other forms of localized instability.

To focus on actively deforming areas, pixels were objectively selected using Otsu thresholding applied to long-term displacement metrics derived from the MT-InSAR time-series. This data-driven approach allowed stable and deforming areas to be separated without relying on subjective thresholds, capturing both known unstable slopes and newly emerging deformation zones. The selected high-deformation pixels were then used for short-term deformation forecasting using two models: a Long Short-Term Memory (LSTM) network and a Temporal Convolutional Network (TCN).

Model performance was assessed using five-fold time-series cross-validation. The TCN model showed consistently better performance, achieving an R² of ~0.95 and an F1-score of ~0.97, compared to the LSTM model (R² ~0.93, F1 ~0.92), indicating improved representation of long-range temporal dependencies and non-linear deformation behaviour.

To examine the spatial evolution of instability, K-means clustering was applied to both five-year historical and twelve-month forecasted displacement time-series, producing deformation cluster maps for each period. Areas showing transitions from lower to higher deformation classes were identified as emerging landslide hotspots. The observed deformation patterns and hotspot distribution show strong agreement with previous MT-InSAR-based landslide studies and regional landslide inventories from the Himalayan region, providing independent validation of the proposed framework for landslide hazard assessment and risk management.

How to cite: Basu, A., Dikshit, O., and Tiwari, A.: Predictive Modelling of Landslide Hotspots in Nainital using MT-InSAR and Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9267, https://doi.org/10.5194/egusphere-egu26-9267, 2026.

High-resolution X-ray microtomography (XMT) imaging of rock deformation experiments at micrometer scale provide valuable insights into the coupled evolution of pores, cracks, and fluid pathways (Noiriel & Renard, 2022). A critical step in XMT data processing is the removal of ring artefacts, which is attributed to malfunctioning detector components within the acquisition system (Vo et al., 2018). These artefacts appear as stripes in the raw acquired sinogram domain and concentric circles in reconstructed images. Ring artefacts can adversely affect downstream analyses such as pore and fracture segmentation, and digital volume correlation (DVC) (Mahdaviara et al., 2025). Advances in GPU computing and ML-based image processing have led the synchrotron community to explore deep learning architectures including ResUNET (Fu et al., 2023), and attention-based variants (Zhang et al., 2022) to suppress ring artefacts. Most ML-based denoisers rely on single-domain, pixel-based loss functions, such as L₁, L₂, or structural similarity index measure (SSIM), applied either in the sinogram or reconstructed image domain.

This study investigates a dual-domain loss function that combines loss terms in the sinogram domain with those in the corresponding Fast Fourier Transform magnitude (FFT amplitude) domain, aiming to improve generalization of trained U-Net variants. Existing artefact-free XMT images of basalt were used to simulate stripe artefacts in its raw sinogram domain. Stripe artefact generation was controlled using three parameters: pixel thickness, amplitude, and number of stripes per sinogram. A total of 5,000 paired noisy and clean sinograms were generated and split into training, validation, and test datasets. Three UNET-based architectures were evaluated: a baseline U-Net (baseUNET), a residual U-Net (ResUNET), and a residual U-Net with attention gates (AG-ResUNET). Models were trained for 100 epochs using the Adam optimiser and their performances were assessed using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and qualitative inspection of sinograms, reconstructed images, and FFT amplitude spectra. The study compares single- and dual-domain loss functions in terms of ring artefact suppression and generalization beyond the training data distribution, discusses limitations related to the dynamic range of the training data and its implications for denoising experimental XMT datasets.

References

• Fu, T., Wang, Y., Zhang, K., Zhang, J., Wang, S., Huang, W., Wang, Y., Yao, C., Zhou, C., & Qin, Y. (2023). Deep-learning-based ring artifact correction for tomographic reconstruction. Journal of Synchrotron Radiation, 30(3).
• Mahdaviara, M., Mousavi, M., Rafiei, Y., Raoof, A., & Sharifi, M. (2025). Improving numerical fluid flow simulation by ring artifact removal in micro-CT images of porous media using attention autoencoder–decoders. Transport in Porous Media, 152, 57.
• Noiriel, C., & Renard, F. (2022). Four-dimensional X-ray micro-tomography imaging of dynamic processes in geosciences. Comptes Rendus Géoscience, 354(G2), 255–280. https://doi.org/10.5802/crgeos.137
• Vo, N. T., Atwood, R. C., & Drakopoulos, M. (2018). Superior techniques for eliminating ring artifacts in X-ray micro-tomography. Optics Express, 26(22), 28396. https://doi.org/10.1364/oe.26.028396
• Zhang, J., Niu, Y., Shangguan, Z., Gong, W., & Cheng, Y. (2022). A novel denoising method for CT images based on U-net and multi-attention. Computers in Biology and Medicine, 152, 106387. https://doi.org/10.1016/j.compbiomed.2022.106387

How to cite: Dhanapal, S., Cordonnier, B., and Renard, F.: Improving Generalization of Deep Learning–Based Ring Artefact Removal in X-ray Microtomography Imaging of Geomaterials, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10211, https://doi.org/10.5194/egusphere-egu26-10211, 2026.

SAR-to-optical image translation (S2OIT) aims to transform the complex backscattering characteristics of Synthetic Aperture Radar (SAR) into more interpretable optical appearances. However, existing methods often suffer from over-smoothed structural details, generation of pseudo-textures caused by inconsistencies between generated textures and real optical images, and insufficient global consistency in complex scenes. To address these challenges, we propose a Physics-Aware and Frequency-Regularized Generative Adversarial Network (PAFM-GAN) for SAR-to-optical translation. Specifically, we extract local statistical and edge structural cues from SAR images and inject them into the generator as additional guidance, which enhances structural authenticity and mitigates the impact of speckle noise. To mitigate spectral misalignment and suppress high-frequency artifacts, we further transform both the generated and real optical images into the Fourier frequency domain and perform spectral distribution alignment between them. We also introduce a frequency-domain discriminator to suppress unrealistic high-frequency components, thereby effectively reducing spurious details in the synthesized results. In addition, to capture long-range dependencies under high-resolution scenarios with low computational overhead, we integrate a Mamba-based state space module (SSM) into the generator for efficient global context modeling, improving scene-level style coherence and overall consistency. Extensive experiments on the SAR2Opt, SEN1-2, and QXS-SAROPT demonstrate that PAFM-GAN consistently outperforms representative SAR-to-optical baselines across five metrics, including PSNR, SSIM, FID, LPIPS, and FSIMc.  In addition, the results of multiple ablation experiments validate the effectiveness of the proposed method.

How to cite: Wang, L., Liu, Y., Zhao, J., and Lu, Z.: PAFM-GAN: Physics-Aware and Frequency-Regularized GAN for SAR-to-Optical Image Translation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10786, https://doi.org/10.5194/egusphere-egu26-10786, 2026.

How to cite: Piedboeuf, F., Girard, M., Jervis, D., McKeever, J., and Sampson, J.: Automatic Detection and Segmentation of Methane Plumes in GHGSat Imagery , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11947, https://doi.org/10.5194/egusphere-egu26-11947, 2026.

Earth Observation data (EO) can support food-security decision making in Sub-Saharan Africa, yet operational crop-type mapping in dryland smallholder systems remains challenging under rainfed-season cloud cover, heterogeneous cropping calendars, and small, irregular fields. Model performance is further degraded by scarce and noisy labels, mixed-cropping and intercropping practices, and strong domain shift across agro-ecologies. A particularly consequential failure mode is confusion between cropped and fallow parcels, where vegetated fallows can mimic crop spectral–temporal signatures and bias cropland statistics and downstream indicators used for early warning, input targeting, and program planning. A hierarchical, stage-wise sequential transfer-learning framework built on Convolutional Recurrent Neural Networks (ConvRNNs/ConvLSTMs) to improve robustness in data-scarce smallholder landscapes is proposed. The approach learns reusable spatiotemporal representations in a coarse-to-fine curriculum and transfers them across tasks of increasing label granularity. Stage 1 produces a cropland mask by classifying cropland versus other land uses (explicitly including fallow), targeting the crop–fallow confusion that dominates errors in dryland settings. Stage 2 refines cropland into agronomic family groups (e.g., cereals, legumes, vegetables), preserving interpretable subclass structure that is often sufficient for operational monitoring when fine labels are sparse. Stage 3 resolves fine-grained crop types and mixed-dominant intercropping states. The ConvLSTM backbone is trained stage-wise: parameters learned at a coarser stage initialize the next stage, while stage-specific classification heads are optimized for the current hierarchy level. The framework is demonstrated in Senegal using Planet NICFI monthly composites (~5 m; RGB+NIR) and in situ polygon labels collected during the 2020 and 2023 rainfed seasons. Training samples are built as ~0.5 ha image patches (14×14 pixels) extracted from interior points within polygons, with sampling density scaled by polygon area to better represent large fields while maintaining coverage of small parcels. The dataset, 6,978 labeled polygons in 2020 and 5,827 in 2023 generate 18,380 and 18,378 patches for September and October 2020 (no August imagery), and 13,733/13,623/13,524 patches for August/September/October 2023. To address severe long-tail imbalance typical of regional crop inventories, offline quota-based corpus curation, online weighted sampling, and consolidate ultra-rare fine-grained labels into an “OTHER” class at Stage 3 to stabilize training, are combined. The staged framework is benchmarked against machine and deep learning baselines (Random Forest, XGBoost, CNN, and single-stage recurrent models) using macro-averaged metrics and precision–recall behavior, selecting operating points that favor higher precision for operational mapping. Results show robust cropland maps with stable accuracy under limited labels and small, irregular fields, while preserving subclass structure; cereals and legumes remain identifiable at Stage 2 (validation accuracy ≈ 0.59). At Stage 3, precision is highest for major crops, Groundnut 0.83, Millet 0.72, and Maize 0.69, and moderate for Cowpea 0.51 and Rice 0.42. Remaining errors are primarily driven by data imbalance, mixed-cropping systems, and spectral confusion, highlighting priority areas for improving long-tail supervision and intercropping representation.

How to cite: Gelaye, K. K., Sarr, M. A., Gumma, M. K., Traore, P. C. S., Bassene, C. B. E., Mbemgue, F., and Mutuku, J. M.: Stage-wise ConvLSTM Sequential Transfer Learning for Hierarchical Crop Type Mapping in Senegal’s Smallholder Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12575, https://doi.org/10.5194/egusphere-egu26-12575, 2026.

GRACE and GRACE-FO satellite missions offer an observation-based perspective on terrestrial water storage anomalies (TWSA), which is valuable for assessing climate variability and human influence on large-scale water resources. In practice, however, the short duration of the GRACE record and its coarse spatial resolution make it difficult to build long, spatially consistent storage information that are needed to study basin-scale responses to hydrologic extremes such as droughts and floods. To address this limitation, we develop a multi-model hindcasting framework that reconstructs monthly GRACE TWSA from hydro-climatic predictors and evaluates both predictive performance and hydrologic plausibility using independent evidence related to extremes.

We compare four models representing three methodological families: (i) tree-based machine learning (Extreme Gradient Boosting and Extra Trees Regressor), (ii) spatio-temporal deep learning (Convolutional LSTM), and (iii) an efficient Transformer architecture for long-sequence forecasting (Informer). All models are trained and tested over the observational GRACE period (2002–2025) using strict time-block splits, and then applied to hindcast historical monthly TWSA. Predictor co-variables include precipitation, evapotranspiration, temperature, soil moisture, and land data assimilation–based storage components from GLDAS products (Noah and CLSM), enabling the models to learn storage persistence and hydro-climatic controls beyond what can be inferred from GRACE alone. The performance of the models is assessed using standard error and agreement metrics (RMSE and correlation) as well as hydrologically oriented measures (Nash–Sutcliffe Efficiency and Kling–Gupta Efficiency), with additional diagnostics targeting the representation of seasonal and inter-annual variability.

Transformer model shows the best performance on testing periods with R² of 0.81, CC of 0.89, NSE of 0.86 and KGE of 0.88, and RMSE of 61.3 mm, while ETR showed the least R² of 0.713, CC of 0.76, NSE of 0.68 and KGE of 0.73. To move beyond statistical agreement with GRACE, we evaluate physical reliability through multi-source validation across all major sub-basins of Türkiye using (1) groundwater table observations and (2) independent flood information, including a flood potential indicator and mapped flood extents where available. All model families capture key groundwater storage variability, while Informer generally provides the highest predictive skill and better preserves persistence and the seasonal cycle than ConvLSTM and the tabular learners. Periods of elevated reconstructed storage are consistently associated with historical record of higher flood potential, while the lower extremes of TWSA record identifying historical droughts, supporting the hydrologic realism of the hindcast products. At the same time, the tree-based models—particularly XGBoost—remain attractive due to their low computational cost and, their stronger ability to reproduce observed flood-extent spatial patterns in some basins while maintaining extreme behavior comparable to transformer architecture.

Overall, the inter-comparison highlights practical trade-offs among accuracy, robustness to extremes, and computational efficiency, and provides guidance for scalable GRACE TWSA hindcasting on cloud platforms. The validation approach is transferable and supports the use of reconstructed storage fields for drought–flood assessment and basin-scale water resources analysis.

How to cite: Karam, W., Yüksel, K., Gümüş, A., and Gündüz, O.: From conventional machine learning to Transformers: multi-model hindcasting of GRACE terrestrial water storage anomalies (TWSA) with multi-source validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14810, https://doi.org/10.5194/egusphere-egu26-14810, 2026.

High-sulfidation (HS) orebodies are typically characterised by advanced argillic alteration and strong structural control. However, their 3D delineation remains challenging due to the inherent complexity of alteration facies and the limitations of discrete drillhole observations. We present a comprehensive 3D machine-learning workflow to delineate HS orebodies at the Čukaru Peki deposit (Eastern Serbia) by integrating drill-core SWIR spectroscopy with geological and geochemical constraints.Alteration mineralogy was characterised from SWIR spectra using The Spectral Geologist (TSG), extracting diagnostic sulfate–clay signatures (e.g., alunite-group minerals) and spectral scalars (e.g., ~2.20 μm absorption depth and white-mica crystallinity). To bridge the gap between discrete samples and a continuous volume, we constructed voxel-scale attribute fields using CatBoost regression. Unlike conventional distance-based interpolation, CatBoost learns nonlinear spatial dependencies conditioned on coordinates and geological context (lithology, alteration facies, and fault proximity), enabling data-driven 3D inference across the entire modelling volume.Subsequently, a Transformer encoder was employed for voxel-wise evidence fusion on the stacked 3D attribute layers. The model captures the nonlinear mapping of "multi-evidence interaction → HS mineralisation probability" to output a probabilistic targeting volume. The model was trained on labelled exploration drilling data (604 samples) and rigorously validated against an independent in-mine dataset (2,850 samples). Performance evaluation using confusion matrices and ROC curves consistently suggests that sulfur enrichment, alteration intensity, and structural proximity jointly govern HS distribution. This approach provides a robust, interpretable basis for 3D orebody modelling and drill targeting in complex porphyry–epithermal systems.

How to cite: Wang, G., Zhang, G., Wang, Z., Yang, S., and Wang, Y.: 3D HS orebody delineation integrating CatBoost modelling and Transformer-based evidence fusion: A case study from Čukaru Peki, Serbia, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15741, https://doi.org/10.5194/egusphere-egu26-15741, 2026.

Abstract：Open-pit mines, typical land-surface features shaped by intensive human activities, require rapid identification of their spatial distribution for effective mineral resource supervision, ecological disturbance assessment, and land inspection. Optical remote sensing imagery, with its wide coverage, convenient acquisition, and rich spatial details and textures, provides intuitive morphological and contextual cues for open-pit mine identification and is therefore widely employed in routine monitoring and rapid assessment. Nevertheless, open-pit mines often bear strong visual similarities to quarries, bare land, construction-disturbed zones, and waste dumps. Meanwhile, slender structures (e.g., pit boundaries, bench slopes, and haul roads) tend to be smoothed out in multi-scale representations, which makes it challenging to balance global shape characterization with precise local boundary localization. To address these issues, we propose GLSNet (Global-Local State-space Network) , a feature-enhancement framework for open-pit mine detection, consisting of three synergistic modules. First, an Adaptive Scale-aware Spatial Pyramid Pooling Fast (A-SPPF) module is introduced to adaptively select effective contextual ranges, suppress confusing background interference, and improve scale robustness. Second, a Low-resolution State-Space Modeling (LS-SSM) module is designed to efficiently model long-range dependencies and scene structural relationships, enhancing discrimination between open-pit mines and visually similar land-surface units. Third, a Scale-adaptive Global–Local Fusion (SGF) module is proposed to jointly strengthen global structural constraints and local boundary details, thereby balancing holistic morphology representation and key boundary localization, and improving detection stability and cross-region generalization.We evaluate our method on the public Open Pit Mine Object Detection Dataset and compare it with Faster R-CNN, YOLOv5, YOLOv8, YOLOv10, RTMDet, RT-DETR, DEIM, and Mamba-YOLO. Results demonstrate that GLSNet achieves superior overall detection performance, with particularly notable advantages in resisting background-induced confusion under complex conditions and in recognizing small-scale targets, while maintaining high inference efficiency, thereby validating the effectiveness and synergy of the proposed modules.

Keywords：open-pit mine detection; state-space models (SSM); multi-scale features; global-local fusion.

How to cite: Chen, Z., Zheng, Y., Yao, D., and Zhao, J.: GLSNet: State-Space-Enhanced Open-Pit Mine Detection With Global-Local Information Fusion, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16478, https://doi.org/10.5194/egusphere-egu26-16478, 2026.

Machine learning has seen widespread adoption across the geosciences. In particular, deep learning methods have proven effective for producing gridded datasets of climate parameters. Convolutional neural networks are commonly used, but their performance can be limited by the availability and structure of data, especially for sparse or irregularly sampled climate observations. Graph neural networks can handle irregular spatio-temporal data, but their reliance on local interactions restricts their ability to capture large-scale climate processes.

An alternative approach is DeepKriging, originally proposed by Chen et al., which embeds the spatial domain using basis functions centered at knot points across the region of interest. By using these basis functions as input features for a neural network, DeepKriging provides an efficient and flexible representation of both spatial and temporal domains, making it suitable for irregular data and capable of capturing both large-scale and local effects. However, DeepKriging requires basis functions to be manually defined before model training, which can require extensive work from the practitioner to fine-tune the model. This also limits the model’s ability to adapt to varying spatio-temporal patterns.

Here, we propose several extensions to DeepKriging, primarily by allowing basis functions to be updated throughout model training. The resulting model dynamically adapts to diverse spatio-temporal patterns while converging on a basis function representation that is optimal for the current data. We further improve the flexibility of the spatial embedding through a mesh generated via constrained Delaunay triangulation. This approach is applied to multiple climate variables, including precipitation and wind data for Ireland, demonstrating an improved performance compared with the original DeepKriging as well as several state-of-the-art deep learning and geostatistical gridding methods.

Finally, we also show how basis function representations are particularly well suited for datasets with limited availability, such as sparsely sampled climate parameters like relative humidity or soil moisture. This flexibility can be leveraged across a range of machine learning frameworks, including transfer learning with DeepKriging models or more lightweight algorithms such as Random Forests and XGBoost.

How to cite: O'Sullivan, B. and Coonan, B.: Basis Functions Representation for Deep Learning Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16765, https://doi.org/10.5194/egusphere-egu26-16765, 2026.

Detecting wildebeest from very-high-resolution (VHR) satellite imagery enables large-area population monitoring in a single acquisition, avoiding aircraft-induced disturbance and reducing sampling bias caused by transect-based surveys. However, wildebeest appear as extremely small objects in satellite images, and direct application of classical object detectors (e.g., YOLO-style detectors) often yields poor performance. In particular, high-density aggregation areas suffer from severe missed detections due to scale mismatch and limitations in post-processing for densely packed small objects.

To address these challenges, we develop a targeted detection solution that integrates (1) an adaptive sliding-window strategy to better capture local context under varying density conditions, (2) resolution–detector adaptation to mitigate scale mismatch between object size and detector design, and (3) improved post-processing modules, including an enhanced non-maximum suppression (NMS) tailored for dense small-object scenarios. We evaluate the proposed framework using WorldView-2 and WorldView-3 imagery over the Serengeti acquired in 2022 and 2023. The overall F1-score improves from 0.727 to 0.770 in 2022 and from 0.682 to 0.756 in 2023. Notably, in high-density areas in 2022, the F1-score increases from 0.330 to 0.821, demonstrating that our approach effectively reduces missed detections in dense small-object scenarios that commonly lead to substantial omissions in traditional pipelines.

Beyond wildebeest monitoring, our results highlight a generalizable pathway for adapting classical detectors to dense small-object detection in VHR satellite imagery, where objects are tiny and crowded. 

How to cite: Xu, Z., Wu, Z., Duporge, I., Lee, S., and Wang, T.: An adaptive window and resolution-aware detection framework for dense small-object mapping from very-high-resolution satellite imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17458, https://doi.org/10.5194/egusphere-egu26-17458, 2026.

Change detection from high-resolution remote sensing images lies as a cornerstone of Earth observation applications, yet its efficacy is often compromised by two critical challenges. First, false alarms are prevalent as models misinterpret radiometric variations from temporal shifts (e.g., illumination, season) as genuine changes. Second, a non-negligible semantic gap between deep abstract features and shallow detail-rich features tends to obstruct their effective fusion, culminating in poorly delineated boundaries. To step further in addressing these issues, we propose the Frequency-Spatial Synergistic Gated Network (FSG-Net), a novel paradigm that aims to systematically disentangle semantic changes from nuisance variations. Specifically, FSG-Net first operates in the frequency domain, where a Discrepancy-Aware Wavelet Interaction Module (DAWIM) adaptively mitigates pseudo-changes by discerningly processing different frequency components. Subsequently, the refined features are enhanced in the spatial domain by a Synergistic Temporal-Spatial Attention Module (STSAM), which amplifies the saliency of genuine change regions. To finally bridge the semantic gap, a Lightweight Gated Fusion Unit (LGFU) leverages high-level semantics to selectively gate and integrate crucial details from shallow layers. Comprehensive experiments on the CDD, GZ-CD, and LEVIR-CD benchmarks validate the superiority of FSG-Net, establishing a new state-of-the-art with F1-scores of 94.16%, 89.51%, and 91.27%, respectively. The code will be made available at https://github.com/zxXie-Air/FSG-Net.

How to cite: Xie, Z., Miao, S., Jiang, Y., Zhang, Z., Yao, J., Li, X., Huang, J., and Ghamisi, P.: FSG-Net: Frequency-Spatial Synergistic Gated Network for High-Resolution Remote Sensing Change Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17651, https://doi.org/10.5194/egusphere-egu26-17651, 2026.

Reservoir modelling in heterogeneous carbonate systems is often constrained by sparse well control and labor-intensive interpretation, which increases uncertainty when extrapolating between wells. We present an enhanced Pix2Geomodel.v2 workflow that reframes facies and petrophysical modelling as paired image-to-image translation. Facies and petrophysical properties are exported from a reference reservoir model, converted into paired 2D training images, and used to train a Pix2Pix-style conditional generative adversarial network (cGAN). The architecture couples a U-Net generator with a PatchGAN discriminator, enabling the model to learn spatial relationships directly from examples. To reduce data requirements while retaining geological heterogeneity, the workflow operates on a streamlined grid of 54 vertical layers and targets complex facies distributions. Preliminary results show stable training and predictions that reproduce the main geological patterns of the reference data. In facies-to-property translation, the network learns meaningful mappings to porosity, permeability, and volume of shale.

How to cite: Al-Fakih, A., Hanafy, S., Saraih, N., Koeshidayatullah, A., and Kaka, S.: Data-efficient enhanced Pix2Geomodel.v2 for complex facies settings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2220, https://doi.org/10.5194/egusphere-egu26-2220, 2026.

Capturing cross-property correlations while preserving spatial continuity is essential for reliable reservoir characterization, especially in heterogeneous reservoirs where facies architecture controls petrophysical variability. In this study, we evaluate Pix2Geomodel.v2 as a bidirectional image-to-image translation framework that learns mappings between facies and petrophysical properties using paired 2D slices exported from a reference reservoir model. To reduce data demands while maintaining geological complexity, the workflow operates on a streamlined grid of 54 vertical layers, enabling efficient training and rapid experimentation without removing key stratigraphic and facies patterns. The approach is based on a conditional generative adversarial learning strategy. A U-Net generator is trained to synthesize target facies or property maps from input images, while a PatchGAN discriminator encourages locally realistic textures and geologically plausible transitions. The paired-slice formulation allows the model to learn both large-scale structural organization and fine-scale heterogeneity directly from examples. We investigate two complementary directions: (i) facies-to-property translation, where facies maps are used to predict continuous property fields such as porosity and permeability, and (ii) property-to-facies translation, where petrophysical images are used to reconstruct discrete facies distributions. Beyond conventional forward mapping, the reverse translation experiments are particularly informative because they test whether the model captures meaningful cross-property dependencies rather than superficial patterns. The reconstructed facies maps recover coherent large-scale facies trends and geologically consistent connectivity, indicating that the learned representation encodes relationships between depositional architecture and petrophysical response. Spatial realism is further examined using experimental variograms, providing a continuity-based check that generated outputs qualitatively align with the reference model in terms of spatial correlation structure. Overall, the results suggest a data-efficient route to robust forward and reverse translations that can support faster reservoir model prototyping, property population guided by facies, and consistency checking between facies and petrophysical interpretations.

How to cite: Kaka, S., Al-Fakih, A., Saraih, N., Koeshidayatullah, A., and Hanafy, S.: Bidirectional translation + spatial continuity validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2222, https://doi.org/10.5194/egusphere-egu26-2222, 2026.

Data generated in the field of geoscience has unique properties that can be characterized by high complexity, sparsity, and site-specific variability. Owing to the unique characteristics, applying general artificial intelligence frameworks and achieving model generalization in geoscience still remains a challenging problem. In this work, we introduce the KIGAM GeoAI Platform, an integrated AI environment designed to bridge the gap between advancing AI technology and practical geoscience research. The platform supports the entire research workflow through a user-friendly, web-based interface, systematically covering essential stages: data uploading, preprocessing, model development, testing, validation, and the final deployment of analytical applications. By providing a centralized online environment for collaborative research, the platform aims to reduce technical entry barriers for geoscientists who may not be AI experts, while establishing a robust foundation for data-driven cooperation. We plan to continue improving and scaling this platform to ensure it remains a stable, accessible, and high-performance tool for both domestic and international geoscience communities. Through these efforts, the KIGAM GeoAI Platform is expected to accelerate digital transformation and foster a more integrated global research ecosystem in the field of geoscience.

How to cite: Kwon, J.:  Innovate with Ease: Introducing the KIGAM GeoAI Platform, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2290, https://doi.org/10.5194/egusphere-egu26-2290, 2026.

Artificial intelligence (AI) tools increasingly enhance the efficiency and consistency of seismic interpretation, particularly in structurally complex areas or areas where data quality is reduced by acquisition limitations. As a result, interpretations can become difficult and time-consuming, especially in the context of structural interpretation and fault tracking. To evaluate the performance of AI-based fault detection, we applied Geoplat AI software to a 3D seismic volume from the Drmno Basin, located at the southeastern margin of the Pannonian SuperBasin in Serbia.
A conventional structural interpretation was first performed by mapping the major fault systems, then minor fault systems, generating fault sticks and polygons for all visible faults and developing a structural model to illustrate the basin's opening and evolution. Subsequently, AI-based workflows were applied in order to enhance the quality of the seismic data. This involved removing noise, restoring reflections, highlighting fault zones, and applying smoothing filters. The final step was the utilization of a fault tracking tool that segments the seismic data, recognizes fault zones, traces them, identifies structural patterns, and calculates a probability field. The AI-derived fault interpretation was then compared with the manual interpretation.
The results indicate that the Drmno basin was developed under an extensional tectonic regime during the Early Miocene, which formed a large Morava detachment fault and opened accommodation of the basin. The basin itself has complex architecture in the syn-rift phase, with many synthetic and few antithetic faults, oriented from the east to the west. During the stage of the rift climax, the dominant fault systems remained consistent, with most syn-rift structures continuing to accommodate the subsidence formed by the Morava detachment. The shift in the tectonic conditions in the post-rift stage leads to the formation of systems of parallel faults in the younger sediments, adjusting strike-slip movements in a compressional tectonic field. The younger structures are dominantly oriented in the north-south direction, or reactivated older fault structures.
The AI tool effectively interpreted fault systems in the younger geological units, benefiting from higher data quality, and clearly indicated younger fault systems with a high level of certainty. However, in the lower part of the seismic cube, the basement structures remain unclear or unrecognized. Reactivated fault surfaces and a significant fault zone are evident in the interpretation. In areas with low-quality seismic data, the AI tool struggled to trace faults accurately, resulting in geologically inconsistent fault patterns.
Overall, the AI-based 3D fault tracking tool proved effective in resolving the main structural framework of the basin. The dominant fault directions are clearly identifiable, and the main geological structures have been mapped with reasonable precision. The AI-supported interpretation successfully captures the main structural trends and provides a solid basis for evaluating the tectonic evolution. This case study demonstrates the potential of AI to support structural interpretation and tectonic analysis of complex sedimentary basins.

How to cite: Ninić, A., Radivojević, D., and Đurić, D.: The application of artificial intelligence in fault tracking on 3D seismic data – A case study from Drmno Basin (SE Serbia), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2380, https://doi.org/10.5194/egusphere-egu26-2380, 2026.

Climate Service Recipes (HACID-CSR) is an agentic system designed to assist providers of climate services in developing their advice for a wide range of clients. HACID-CSR guides providers by navigating the large and ever-increasing corpus of knowledge as well as an area without established standards and with limited access to scientific experts. It automatically generates detailed workflows (or “recipes”) by leveraging both a large language model’s internal reasoning and contextual knowledge from a domain knowledge graph for climate services (CS-DKG). The CS-DKG is an expert-curated ontology of climate service concepts with mapped relationships between climate variables, emission scenarios, indices, hazards, sectors, and key datasets (CORDEX, CMIP5, UKCP18), built as part of the Horizon Europe-funded HACID project (Hybrid Human Artificial Collective Intelligence in Open-Ended Decision Making).

The HACID-CSR architecture consists of a memory-enabled supervisor agent orchestrating multiple specialised agents. A planning agent first proposes an initial workflow outline, and a preliminary recipe agent uses only the LLM’s knowledge to draft answers to key workflow steps. The system then engages a knowledge graph retrieval sequence: a class selection agent identifies relevant classes in the CS-DKG, an instance selection agent finds specific instances (entries) highly relevant to the query within those classes following a two-stage selection process, i.e. semantic similarity based pre-selection and LLM-enabled refined selection, and a subgraph extraction agent retrieves the corresponding subgraph of related knowledge entities. Next, a recipe generation agent creates each step of the workflow by combining the LLM’s reasoning with the retrieved graph context using graph retrieval-augmented generation (GraphRAG). Finally, a recipe refinement agent compares the preliminary LLM-only solution with the knowledge-enhanced solution and refines the output, yielding a diverse and context-aware workflow.

By using this multi-agent approach, HACID-CSR increases the diversity of solutions and fills the knowledge gap between climate information and domain specific applications, helping experts to identify suitable methodologies and datasets. The resulting workflows are more traceable and transparent, improving user trust compared to answers from a general-purpose chatbot. We have also developed a bespoke automatic evaluation method to complement human expert validation of the generated recipes. We highlight the potential of the HACID-CSR approach for multi-hazard climate service design, and discuss remaining challenges and opportunities for further refinement of this agentic LLM-based system.

How to cite: Abele, A., Xie, H., Biswas, A., Dong, H., Fung, F., and Williams, H.: Climate Service Recipes: automatic multi-hazard climate information workflow generation using agentic Large Language Models (LLMs) and knowledge graphs, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2819, https://doi.org/10.5194/egusphere-egu26-2819, 2026.

Earth system science relies on the integration of knowledge from many branches of geoscience, including climate dynamics, hydrology, ecology, land use and biogeochemical cycles. However, the scientific literature informing these domains has become vast and increasingly difficult to navigate due to its rapid development and disciplinary spread. This complexity makes it difficult to maintain an integrated overview of relevant findings and to identify scientific connections in a systematic manner. Recent advances in generative artificial intelligence (AI) and large language models (LLMs) provide opportunities to support these tasks, particularly when combined with retrieval methods and transparent source attribution.

Here we propose a retrieval-augmented AI platform designed to assist scientific knowledge integration in Earth system science. The platform is conceived as a living system, built on a continuously expanding and updateable knowledge base that aggregates scholarly literature from major scientific databases. User queries initiate targeted retrieval of relevant documents followed by the generation of concise, source-linked summaries using locally hosted open-weighted LLMs. By explicitly grounding outputs in retrieved literature, the platform alleviates the need for manual screening and limits hallucination risks that currently constrain the use of general-purpose LLMs in geoscientific research.

Evaluation of the initial prototype demonstrates that domain-specific retrieval-augmented generation systems can provide reliable, traceable synthesis of Earth system knowledge and help address the growing gap between accelerating publication rates and the need for timely, verifiable scientific assessment.

How to cite: Kart Tokmak, Ö., Caesar, L., and Sakschewski, B.: A Living AI Platform for the Earth System Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3316, https://doi.org/10.5194/egusphere-egu26-3316, 2026.

This study investigates the domain adaptation of the Vision-Language Model (VLM) for road damage assessment, focusing on a fine-tuning strategy optimized for resource-constrained engineering environments. Unlike conventional object detection models that operate within fixed label spaces, VLMs provide superior semantic understanding and generalization in complex scenarios. To facilitate practical deployment, this research systematically analyzes key variables of Parameter-Efficient Fine-Tuning (PEFT) to mitigate the high computational demands inherent in large-scale VLMs.

In the experimental phase, hyperparameter tuning was conducted using the Low-Rank Adaptation (LoRA) technique. The primary variables included LoRA ranks (16, 32, 64, and 96), training data scale, and image resolutions (1,024ⅹ28ⅹ28 vs. 1,536ⅹ28ⅹ28). A comprehensive dataset of 26,796 images comprising six damage categories and negative samples was established, utilizing a 7n sampling strategy (n=500, 750, 1,000) to address class imbalance. The impact of data volume was evaluated by augmenting the 7,000-sample set (corresponding to n=1,000) to match the full dataset size of 26,796, with zero-shot inference serving as the performance baseline.

Experimental results demonstrated substantial improvements over zero-shot inference, indicating that performance positively correlates with increased data scale with augmentation and higher image resolution, while lower LoRA ranks (16, 32) proved most effective for this domain. Furthermore, the introduction of specialized ad-hoc metrics, M_mAP and M_F1, verified a stable trade-off between False Positives and False Negatives. Notably, to minimize safety-critical False Negatives, a prompt engineering-based 'Double Check' mechanism and multi-turn interactions were utilized. This approach successfully leveraged the model’s inherent reasoning capabilities to refine damage identification through iterative feedback.

Acknowledgements This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2025-25437298)

How to cite: Kim, D. and Youn, H.: Optimizing Vision-Language Model for Robust Road Damage Assessment via Parameter-Efficient Fine-Tuning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3674, https://doi.org/10.5194/egusphere-egu26-3674, 2026.

Nitrate contamination of groundwater is often addressed as a diffuse agricultural management problem, yet monitoring in Denmark depicts that exceedance risk at abstraction and observation wells is spatially structured and closely linked to surrounding land-use composition and configuration. This suggests a land-use policy opportunity: if landscape fractions and fragmentation patterns help drive nitrate vulnerability, interventions could be spatially targeted and tailored rather than uniformly applied. In this study, we present a scenario-based planning framework for policy appraisal, enabling regulators, municipalities, and water utilities to test alternative policy packages and targeting rules and quantify their expected effects on groundwater nitrate hotspot risk. The system operates a predictive pipeline that relates nitrate outcomes to land-use fractions and landscape configuration metrics computed within configurable protection zones. Model outputs are formulated as a binary hotspot classification (hotspot vs. non-hotspot) based on exceedance of a drinking-water nitrate threshold, producing vulnerability maps to prioritize locations for intervention and prevention. The core functionality is a “what-if” engine built on an AI-based ensemble that generates a baseline nitrate-risk probability map and re-predicts risk under user-defined scenarios. Scenario levers are organized into two policy bundles: (i) land-use policy and management, implemented as controlled reallocations among land-cover fractions (e.g., reducing large contiguous cropland blocks, increasing wetland/riparian woodland cover, restricting impervious expansion) while enforcing feasibility constraints; and (ii) agricultural management, implemented as proportional reductions or caps on nitrogen surplus and fertilizer inputs. For each scenario, the system outputs an updated probability map and a different map relative to baseline, supporting spatial prioritization, instrument design, and transparent justification of differential targeting. By combining ex-ante scenario testing with ex-post monitoring of hotspot transitions after implementation, the framework supports adaptive groundwater governance and moves from risk mapping toward operational, spatially explicit nitrate-reduction policy design.

How to cite: Naghibi, A., Ahmadi, K., and Berndtsson, R.: Policy-oriented Land-Use and Agricultural Management Scenarios for Groundwater Nitrate Hotspot Mitigation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4407, https://doi.org/10.5194/egusphere-egu26-4407, 2026.

AI-driven weather and climate prediction has become highly visible in recent years, and most Earth scientists are now familiar with learned forecast models. Far fewer are aware that the same advances in artificial intelligence are producing general-purpose systems that can autonomously review literature, write and debug code, design experiments, and carry out extended research tasks with minimal supervision. These capabilities may ultimately have a greater impact on everyday scientific practice than any single prediction model.

AI-based forecasting represents only a narrow entry point into a broader transformation driven by hybrid intelligence, in which domain-specific Earth system models are combined with general AI systems such as large language and multimodal models and autonomous agents. In practice, this hybrid intelligence already spans simulation, data assimilation, downscaling, and analysis, while general AI systems increasingly handle coding, synthesis, and workflow orchestration. Together, these systems function less as isolated tools and more as adaptive research partners. Drawing on examples from NVIDIA’s Earth-2 research program and related international efforts, this talk examines how this shift reconfigures the human role toward problem formulation, validation, interpretation, and ethical governance, and highlights practical AI-assisted workflows already reshaping research productivity. Framing AI for environmental prediction within this wider context invites a broader discussion of how hybrid intelligence should be integrated thoughtfully into future Earth system science.

How to cite: Hall, D.: Beyond the Forecast: Hybrid Intelligence as a Force Multiplier for Earth Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5926, https://doi.org/10.5194/egusphere-egu26-5926, 2026.

The Yellow River Basin (YRB) serves as a critical repository of literature for understanding human-earth systems, yet existing automated metadata-level review methods suffer from deep semantic loss and deficiencies in spatial representation: They neither capture fine-grained logic chains from full texts nor possess the capability to extract the spatial and hierarchical attributes of geographic entities. However, rapid developments in Large Language Models (LLMs) provide a technological opportunity for the automated extraction of full-text knowledge. To this end, this study proposes the Geo-Knowledge Infused Reasoning Framework (GK-IRF), coupling full-text semantics with multi-level spatial indexing. Methodologically, we first construct an ontology-based full-text parsing mechanism based on 8,493 YRB-related papers (2015-2024), utilizing LLMs to accurately extract structured semantic triplets. Simultaneously, we introduce an adaptive multi-level GeoHash indexing model to map textual toponyms into hierarchically nested grid sets, reconstructing the spatial coverage and multi-scale associations of geographic entities. Validations against a manually annotated dataset indicate that GK-IRF achieves an F1-score comparable to human performance in full-granularity semantic extraction; furthermore, the Spatial Coverage Accuracy of the multi-level grids for the YRB substantially outperforms traditional geocoding methods, effectively resolving the challenge of multi-scale coverage representation.

How to cite: Wu, S. and Wang, H.: Coupling Full-Text Semantics with Multi-Level Spatial Indexing: A Knowledge Representation Framework for Yellow River Basin Literature, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6201, https://doi.org/10.5194/egusphere-egu26-6201, 2026.

Seismic interpretation of mass transport deposits (MTDs) relies heavily on expert knowledge and conceptual reasoning yet remains difficult to formalize and scale. While recent artificial intelligence (AI) methods have shown strong capabilities in seismic pattern recognition, most approaches operate as black boxes and remain poorly aligned with the interpretative frameworks used by geoscientists, limiting transparency and trust.

This study proposes a geoscience-aware hybrid intelligence framework that integrates expert knowledge graphs (KGs) with large language models (LLMs) to support interpretable seismic interpretation of MTDs. The approach builds upon the conceptual methodology of Le Bouteiller et al. (2019), which organizes MTD interpretation through causal relationships linking environmental controls, mass transport properties, and observable seismic descriptors across trigger, transport, and post-deposition phases.

The KG provides a structured reference for interpretation that constrains vocabulary, causal direction, and temporal logic. Our workflow reads scientific papers, identifies relevant descriptors and processes, checks them with LLMs, and evaluates how well they support interpretation. In this setup, seismic descriptors give different levels of support (weak to strong) for geological processes, like how experts reason under uncertainty.

Preliminary results show that ~68% of expert defined concepts are recovered in the inferred graph, with a semantic validation score of 0.73, indicating good conceptual alignment. However, descriptor matching based on textual similarity remains difficult, with average scores around 0.41. This gap highlights the difference between semantic agreement (conceptually correct) and textual agreement (exact wording), mainly due to synonymy and variable phrasing in the literature. We plan to address this by using domain-specific LLMs and ontology-based synonym expansion to improve semantic matching in future iterations

How to cite: Talbi, F. B., Armitage, J., Charléty, J., Rabaute, A., Bouziat, A., Vittaut, J.-N., and Leroy, S.: Geoscience-Aware AI for Interpretable Seismic Interpretation of Mass Transport Deposits Using Knowledge Graphs and Large Language Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6459, https://doi.org/10.5194/egusphere-egu26-6459, 2026.

The assessment of environmental and resource performance of energy transition technologies relies on quantitative information scattered across heterogeneous sources, including scientific articles, patents, and industrial reports, such as ESG (Environmental, Social, and Governance) disclosures. These documents contain key data for Life Cycle Inventory (LCI) and Material Flow Analysis (MFA), such as material and energy intensities, water consumption, mining production volumes, emissions, and technological descriptors. However, this information is predominantly embedded in unstructured PDF documents optimized for human reading, making large-scale, traceable data aggregation difficult and costly when performed manually.

This work presents an automated and modular methodology designed to extract and contextualize quantitative LCI and MFA data from three major categories of technical documentation. The approach combines large-scale document collection, relevance screening, and multimodal artificial intelligence within a reproducible and auditable workflow.

• Scientific Articles

Peer-reviewed articles are collected through automated scraping workflows based on structured search outputs. Documents are screened for LCI/MFA relevance using domain-specific keywords, methodological markers, and quantitative signal density. Relevant articles are then processed using a multimodal AI-based extraction core in which each page is analyzed through a combined text and image input. This enables robust extraction of numerical values from tables, text and figures while preserving contextual information such as units, methodological assumptions, and source location.

• Patents

Patent documents contain information about future trends on technologies and metal uses. Patents are collected via dedicated scraping pipelines and processed separately from scientific articles. The workflow focuses on extracting and structuring patent metadata, including publication year, country, and technology class, in order to characterize technological activity related to energy transition technologies. While quantitative LCI/MFA extraction from patents is not yet systematically performed, the pipeline enables descriptive statistical analyses of patent dynamics, including temporal trends and geographical patterns of technological development.

• Mining technical and ESG Reports

Official mining companies reports, with a specific focus on ESG ones, are processed through a screening module acting as a gatekeeper. The screening relies on sequential text parsing and, when necessary, geometric reconstruction of tables to identify reports containing sufficiently granular and structured quantitative information. Following human validation of the screening results, selected reports are analyzed using a IA-multimodal vision–language model combining page images and extracted text, enabling structured extraction of industrial metrics with associated context and traceability.

This automated methodology addresses one of the core challenges of data collection and significantly improves the granularity, consistency, and verifiability of LCI datasets and MFA inputs. The application of methodology is illustrated through examples related to battery and hydrogen technologies based on scientific articles and patents, and through case studies on copper and nickel production with a focus on mining based on industrial report. Although applied for LCA and MFA, the approach can also support the extraction of other types of data and indicators relevant to environmental and resource analyses. The tool provides automated and reliable support for researchers aiming to extract comprehensive foundational data from heterogeneous sources.

How to cite: Bejjit, C. E., Monfort, D., Muller, S., Lai, F., Beylot, A., and Hennioui, D.: Mining and raw materials sector: Automated Data Extraction and Contextualization for Life Cycle Inventory (LCI) and Material Flow Analysis (MFA) Across Scientific Articles, Patents and Mining companies Reports, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6501, https://doi.org/10.5194/egusphere-egu26-6501, 2026.

Compound drought and heat extremes (CDHEs) exert impacts that exceed the sum of their individual components. With global warming amplifying the associated risks, CDHEs have become a critical threat to agricultural production. Thus, identifying and monitoring CDHEs in cropland systems is key for food security. As CDHEs formation and evolution are shaped by climatic factors, hydrological cycles, and ecosystem feedbacks, their fine- and large-scale identification in agricultural areas poses substantial challenges.

Our study reviews existing methods for identifying CDHEs, including combined threshold approaches, comprehensive index methods, traditional machine learning techniques, and improved mechanistic modeling. We summarize the current limitations of these methods as follows: (1) Combined threshold and comprehensive index methods often focus on a single aspect of CDHEs, failing to systematically describe the complex processes of compound events. (2) While traditional machine learning methods attempt to integrate characteristics of the hazard-bearing body, disaster-causing factors, and hazard-inducing environment to establish complex nonlinear relationships between multiple elements and compound event indices, their "black-box" nature lacks mechanistic interpretability. Furthermore, these methods rely heavily on large volumes of high-quality samples to achieve satisfactory accuracy. (3) Improved mechanistic models, typically based on classical agricultural process models such as APSIM and AquaCrop, introduce CDHE impact modules to address the oversimplification of these effects in original models. Nevertheless, these mechanistic models require extensive input parameters, and their calibration processes depend on substantial amounts of measured data. Additionally, the computational resources needed for simulations are considerable, making the cost of analyzing CDHEs over large farmland areas under various future climate scenarios prohibitive for individual researchers.

To address these challenges, this study highlights the potential of physics-informed neural network models for identifying compound events and proposes future research directions regarding mechanistic constraints, neural network architecture design, and experimental plans: (1) Farmland CDHEs are essentially phenomena of water and heat imbalance within the soil-crop-atmosphere (SCA) system. Utilizing the Richards equation and the Penman-Monteith formula can characterize this process by constraining the water and heat environmental factors at the two key interfaces: root-soil and leaf-atmosphere. (2) Solar-induced chlorophyll fluorescence (SIF), a byproduct of vegetation photosynthesis closely related to GPP, responds rapidly to physiological damage caused by stress. Utilizing multi-band SIF data can provide a detailed depiction of crop physiological responses to stress from the perspective of the hazard-affected body. (3) Automated design of model architectures incorporating mechanistic information for farmland compound events can be achieved through distillation learning. (4) Future work should integrate ground-based water and heat control experiments with site-specific hyperspectral SIF observation data. Through continuous combinatorial experimental design, this approach can lead to the development of accurate and efficient physics-informed neural networks. Coupled with large-scale satellite and reanalysis data products, this framework aims to enable the large-area identification of farmland CDHEs under future climate scenarios.

How to cite: Xia, H., Wu, J., Zhou, L., and Du, R.: Towards a Mechanism-Informed Intelligent Framework for Identification of Compound Drought-Heat Extremes in Croplands, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8579, https://doi.org/10.5194/egusphere-egu26-8579, 2026.

Electromagnetic methods are among the most widely used techniques in the geophysical exploration industry due to their efficiency and non-invasive nature. However, their data processing workflows are highly time-consuming and strongly dependent on expert intervention. With the rapid and broad success of deep learning, applying deep learning techniques to electromagnetic methods to overcome the limitations of traditional approaches has become an active area of research. The effectiveness of deep learning methods, however, largely depends on the quality of the dataset, which directly influences model performance and generalization capability. Existing applications typically rely on self-constructed datasets composed of randomly generated one-dimensional models or structurally simple three-dimensional models, which fail to capture the complexity of realistic geological environments. Moreover, the absence of a unified and publicly available three-dimensional geoelectrical model repository has further constrained the development of deep learning for three-dimensional electromagnetic exploration. To address these challenges, we introduce OpenEM, a large-scale, multi-structural three-dimensional geoelectrical model repository that incorporates a wide range of geologically plausible subsurface structures.

OpenEM comprises nine categories of geoelectrical models, encompassing a wide spectrum of subsurface structures ranging from simple to complex. These include models of homogeneous half-spaces with embedded anomalous bodies, as well as configurations featuring flat stratigraphy, curved stratigraphy, planar faults, curved faults, and their variants containing anomalous bodies. The resistivity values span from 1 to 2000 Ω·m, with the number of layers ranging from three to seven. In models containing anomalous bodies, the number of anomalies varies from one to five, and both regular and irregular geometries are considered to enhance dataset diversity and realistic representativeness. In addition, OpenEM is accompanied by a three-dimensional model generator that enables fully controllable model construction, allowing users to customize structural configurations, including resistivity magnitudes, fault geometries and locations, as well as the size, shape, and placement of anomalous bodies.

OpenEM provides a unified, comprehensive, and large-scale dataset for common electromagnetic exploration systems, thereby promoting the application of deep learning methods in electromagnetic prospecting.

How to cite: Wang, S., Wang, X., Deng, F., and Jiang, P.: OpenEM: Large-scale multi-structure 3D dataset for electromagnetic methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8976, https://doi.org/10.5194/egusphere-egu26-8976, 2026.

Typically, to start working on a remote sensing–based application, various analyses and insights are needed from domain experts. A significant amount of time and effort goes into preprocessing, structuring, and analyzing the data, which can be a repetitive task, especially when a multi-sensor approach is involved. This often takes away time that could otherwise be invested in innovation or research. To address this, training an LLM to understand and process the context of remote sensing tasks can improve efficiency and reduce human-induced errors.

In this work, we develop an AI agent that can reason and think like a remote sensing expert. This agent uses a RAG-based foundational model (FM) and is equipped with various image processing tools to complete a task. We use gpt-4.1-mini as the FM and the Agno framework to deploy the agent. The knowledge base provided to this agent is specially curated with relevant research articles, books, and remote sensing methodologies. This knowledge base helps the model break down a problem into logical steps that can be performed using the tools available within the agent.

These tools can download data, process it, and provide relevant statistics and visualizations. The user can prompt the agent to download multi-sensor (optical and SAR) data, perform time-series analysis for forest monitoring, and identify deforestation hotspots. The agent can fetch data from Google Earth Engine (GEE), plan processing workflows, dynamically generate Python code, and complete the prompted tasks. This approach highlights the feasibility of integrating LLMs with domain-specific knowledge bases and geospatial processing tools to create autonomous, context-aware systems. Figure 1 depicts the overall workflow of the proposed agentic system, illustrating the interaction between the user, the knowledge base, the foundational model, and the integrated processing tools. The framework is directly usable for operational forest monitoring applications and can be further fine-tuned and extended to support a broader range of environmental monitoring and geospatial analytics use cases.

                                        Figure 1: Workflow of the Agentic AI system

How to cite: Jain, A. and Sabir, A.: Development of a Context-Aware AI Agent for Forest Applications Using Multi-Sensor Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10390, https://doi.org/10.5194/egusphere-egu26-10390, 2026.

Multi-agent GIS systems are increasingly emerging as a general paradigm for complex geospatial tasks. However, many existing approaches rely on text-only large language models (LLMs) as the primary reasoning substrate. In the absence of explicit geometric constraints and verifiable evidence, spatial relations are often indirectly represented through linguistic statistical correlations. This makes LLMs prone to inconsistency when interpreting and inferring topological, directional, and distance relations in geospatial data, and leads to error accumulation across multi-step tool invocations and long-horizon decision-making, ultimately degrading the accuracy and efficiency of task reasoning and execution. In this work, we propose VisCritic-GIS, a multi-agent framework for geospatial task reasoning and execution driven by visualized evidence review. VisCritic-GIS introduces a Visualization Generation Agent and a Visualization Critic Agent into conventional multi-agent GIS pipelines. The generation agent renders key spatial data and intermediate results into 2D maps, explicitly externalizing spatial relations in a visual form. The critic agent leverages multimodal LLMs to read and critically review these map-based evidence, producing textual feedback on spatial relations, anomalous results, and reasoning deviations, which constrains and drives iterative refinement of other agents’ reasoning trajectories and toolchain configurations. We build evaluation protocols over representative remote sensing and geospatial tasks, and systematically demonstrate that VisCritic-GIS improves task accuracy, execution efficiency, and interpretability. Overall, our framework provides a mechanism for shifting geospatial reasoning from “text-only probabilistic completion” toward “visually grounded, verifiable inference,” thereby strengthening the robustness of spatial relation understanding in multi-agent GIS systems.

How to cite: Lan, Q., Hu, L., Wu, S., and Du, Z.: VisCritic-GIS: A Visualization-Critic–Empowered Framework for Multi-Agent Geospatial Task Reasoning and Execution, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10473, https://doi.org/10.5194/egusphere-egu26-10473, 2026.

We present SeaScope, an explainable AI agent that accelerates interaction with complex Earth Observation (EO) workflows. Users express analytical questions in natural language, which are transformed into transparent, executable EO analyses. By combining generative AI, vision–language models, and Retrieval-Augmented Generation (RAG), SeaScope links scientific literature, satellite data descriptions, and validated analysis methods to automatically generate, execute, and explain EO workflows. For example, a query such as “Detect vessel activity and possible oil spills in May 2025” triggers dataset selection, code generation, cloud execution, and map outputs with traceable reasoning.

SeaScope is designed as a geoscience-specific AI agent that supports both rapid decision-making and accelerated research. Non-technical users can obtain EO-based insights in time-critical situations without continuous involvement of expert programmers, while researchers benefit from faster hypothesis testing, automated pipeline generation, and reproducible workflows. Human expertise remains central: users inspect retrieved sources, review generated code, and validate analytical steps, ensuring scientific control and accountability. This setup combines domain knowledge with AI-driven scalability, addressing challenges such as sensor-specific scripts and fragmented tools.

As a pilot use case, SeaScope is applied to maritime EO in the Mediterranean region, supporting environmental monitoring and marine activity analysis using satellite data. Beyond the application, the project delivers research insights on generative and vision-based AI for EO, including lessons learned from benchmarking LLMs for code generation, evaluating vision-language models for image understanding, and comparing different RAG and knowledge ingestion strategies. The findings highlight practical trade-offs in accuracy, robustness, explainability, and user validation in real-world workflows.

How to cite: Sekas, C., Philippopoulos, K., Agathangelidis, I., Cartalis, C., Neophytides, S., and Mavrovouniotis, M.: Agentic AI for Earth-Observation-Driven Maritime Monitoring - the SeaScope Project, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13454, https://doi.org/10.5194/egusphere-egu26-13454, 2026.

Large Language Models (LLMs) have emerged as powerful tools for text and data processing, with potential extending far beyond conversational interfaces. We demonstrate that integrating LLMs into agentic workflows enables automated climate and oceanographic data analysis while minimizing hallucinations through strict reliance on real data sources.

ClimSight combines LLMs with climate model data to deliver localized climate insights for decision-making. Specialized agents consult external databases, extract variables from climate models, generate Python scripts for post-processing, and validate outputs through visual analysis. The workflow iteratively corrects errors until reliable results are achieved.

PANGAEA GPT enhances accessibility to the PANGAEA data repository through a supervisor agent that interprets queries, delegates tasks to domain-specific subagents, and coordinates data extraction, statistical analysis, and visualization of oceanographic and atmospheric datasets.

Both systems leverage automatic Python execution and image analysis for quality control. By constraining outputs to verifiable data sources and implementing multi-agent verification, we demonstrate that LLMs can play a significant role in geoscientific data pipelines and automated research workflows.

How to cite: Kuznetsov, I., Pantiukhin, D., Grassi, J., Shapkin, B., Jung, T., and Koldunov, N.: Integrating Large Language Models into Climate and Geoscientific Data Workflows, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13612, https://doi.org/10.5194/egusphere-egu26-13612, 2026.

Generative AI is now woven into the daily study practices of geoscience students, often more deeply than educators acknowledge. This study examines how bachelor students in Earth Sciences (GEOL1008, NTNU) and master students in Engineering Geology (TGB4200, NTNU) use AI tools to understand literature, analyse data, synthesise research findings, and prepare written and oral assignments. The analysis draws on two structured surveys designed to map the extent and character of AI use in both cohorts.

Preliminary results indicate that AI has become the default support tool. Students turn to it to decode complex concepts, troubleshoot coding tasks, analyze data, structure reports, and polish presentations. Many see little distinction between traditional digital tools and generative AI, and the boundary between personal work and AI-augmented work is increasingly blurred. At the same time, students express uncertainty and worry about ethical expectations, disclosure practices, and the legitimacy of relying heavily on AI in academic work.

These trends have immediate consequences for assessment. Home exams, reports, and pre-prepared presentations no longer reliably reveal individual understanding, since nearly all students now use AI during preparation. Emerging evidence from portfolio-based courses suggests grade inflation and reduced differentiation between students, not because learning outcomes have improved, but because AI elevates the baseline quality of submitted work. In practice, written in-person exams and oral examinations remain among the few ways to assess unassisted reasoning.

The findings underscore a need to rethink teaching and assessment in geoscience education. AI is not a future challenge but a present reality, and universities must adapt if they aim to evaluate what students actually know rather than what their tools can produce.

How to cite: Fredin, O.: Geoscience Education in the Age of Generative AI: What Do Students Actually Learn?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16861, https://doi.org/10.5194/egusphere-egu26-16861, 2026.

Large Language Models (LLMs), a class of contemporary artificial intelligence systems, are increasingly used in scientific practice to support research workflows, accelerate discovery, and automate routine administrative tasks. This contribution identifies and analyzes three underexplored aspects of LLM adoption in scientific research. The first aspect concerns the uneven adoption of LLMs among scientists and the inconsistent application of established best practices. The second examines how LLMs can be employed to improve the robustness and reproducibility of scientific practices. The third addresses institutional strategies by which large scientific organizations—such as universities and research networks—can reduce dependence on commercial technology providers while increasing trust in LLM-based systems.

The findings found in our contribution are the partly summarization of István Bozsó’s experiences serving in the role of an “AI ambassador” at the Institute of Earth Phyisics and Space Science (EPSS) of the Hungarian Research Network (HUN-REN).

In our experience many scientists are still skeptical of using LLMs in any capacity or lack the time to invest in learning these technologies. These barriers are primarily sociotechnical rather than purely technical in nature, and they require, on one hand materials that teach best-practices and show motivating examples for using LLMs, on the other hand services provided by research organisations.

Recent advances in open-weight LLMs enable self-hosting within institutional computing infrastructures, which means research institutes can run these models on their own hardware and thereby ensuring that sensitive data and research materials remain within the organization’s controlled digital environment. This also ensures that the LLM usage stays independent of Large Technology corporations and builds trust with colleagues.

Regarding motivating examples, we wish to focus on two areas which can be addressed with the help of LLMs. One area is scientific communication. LLMs can easily generate materials (primarily text, sound and video) which can be used to inform the wider public on new scientific discoveries and push back against misinformation and disinformation campaigns. The involment of scientists is paramount in the review and finalization of such materials to ensure they represent accurate scientific information.

The other area is scientific programming. Many scientists are not trained as professional software engineers and often lack the time and background to apply software development best practices. In many cases this results in software artifacts that are fragile, difficult to reproduce, and challenging to maintain and usually only work on the machine of the researcher who developed the package. LLMs can help out in these situations by suggesting and even implementing best practices and giving programming advice to the researcher during the development of the scientific code.

The common theme in these examples is that the LLM is not meant to replace the scientist but enhance their capabilities with the goal of increasing the robustness, transparency, and sustainability of the scientific research process.

How to cite: Bozsó, I., Horváth, A., and Kuslits, L.: Bridging the gap between scientists and large language models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17176, https://doi.org/10.5194/egusphere-egu26-17176, 2026.

Urban public spaces are highly dynamic systems where traffic patterns, pedestrian flows, and human activities vary strongly across temporal scales. Capturing these dynamics at high temporal resolution remains challenging, particularly using low-cost and reproducible observation methods. In this study, we present an automated workflow for continuous urban activity monitoring based on publicly available webcam imagery and deep learning–based object detection.

A public webcam overlooking Augustusplatz, a central urban square in Leipzig (Germany), is continuously accessed, and still frames are extracted from the video stream at one-minute intervals. Each frame is processed using the YOLO11 object detection model to identify and count relevant object classes, including passenger vehicles and pedestrians. The detection results are converted into structured JSON records and enriched with metadata such as timestamp and geographic location. All data are stored in an InfluxDB time-series database and visualized and statistically analyzed using Grafana.

This setup enables near-real-time and long-term analysis of urban activity patterns across multiple temporal scales. Distinct signatures of recurring and episodic events can be identified, including daily commuting cycles, evening rush hours, road closures, public celebrations, and large seasonal events such as Christmas markets. The minute-scale resolution allows for detailed investigation of short-term dynamics, while continuous operation over longer periods enables comparative and trend analyses.

The presented approach demonstrates how publicly available visual data and open-source tools can be combined into a scalable and transferable framework for urban monitoring. Potential applications include event detection, urban mobility analysis, validation of traffic models, assessment of public space usage, and integration with other environmental or socio-economic datasets. The method provides a cost-efficient complement to traditional urban sensing infrastructures and offers new opportunities for data-driven urban and environmental research.

How to cite: Oesen, B., Wagner, R., and Goblirsch, T.: High-Temporal-Resolution Urban Activity Monitoring Using Public Webcams and Deep Learning: A Case Study from Leipzig, Germany, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17514, https://doi.org/10.5194/egusphere-egu26-17514, 2026.

Climate change, land-use change, and increasing socio-economic pressures are reshaping water and environmental systems, while the volume and heterogeneity of available data—from in situ observations to reanalysis products, remote sensing, and citizen-generated sources—continue to grow. Machine learning (ML) has become an important component of hydro-environmental modelling for forecasting, classification, and pattern discovery. However, in practice, many ML applications remain highly case-specific and dependent on implicit expert decisions related to problem formulation, predictor selection, validation design, and interpretation, which are rarely made explicit or transferable across regions and users.

This contribution presents a human-in-the-loop hybrid intelligence framework that integrates ML workflows with Large Language Models (LLMs) to support structured reasoning during environmental model development and evaluation. Rather than using LLMs for automated optimisation or model selection, the framework positions them as a guidance and scaffolding layer that helps make modelling assumptions, choices, and limitations explicit and traceable, while retaining expert control over all final decisions.

Methodologically, the framework combines (i) hands-on ML pipelines, ranging from baseline statistical models to more advanced learning algorithms for forecasting and classification, and (ii) an LLM-based guidance layer that structures expert reasoning through prompts, checklists, and decision logs. This guidance supports key stages of the modelling process, including the definition of modelling objectives, assessment of data quality, selection of environmentally meaningful predictors, and the design of validation strategies. Particular emphasis is placed on encouraging validation schemes that account for temporal dependence and spatial heterogeneity, such as blocked or spatial cross-validation, rather than default random data splits.

The framework is currently being developed and iteratively evaluated through expert-led case studies using real hydro-environmental datasets, rather than through formal classroom deployment. Initial applications focus on groundwater level analysis and hydro-environmental forecasting problems in Greece, including collaborative work in Crete, where the framework has been used to structure modelling choices and interpret model behaviour under non-stationary conditions. Additional exploratory applications using existing datasets have been used to stress-test the transferability of the workflow across contrasting environmental settings. Ongoing extensions include the application of the framework within coastal erosion modelling activities currently being developed in Colombia.

The LLM layer supports explicit reasoning about why a model performs well or poorly under specific conditions, how assumptions propagate into uncertainty, and where data-driven learning diverges from physical expectations. This reflective use of hybrid intelligence helps expose failure modes and modelling sensitivities that are often hidden in automated pipelines.

Results from the expert-led evaluations indicate that the proposed framework improves the transparency and reproducibility of modelling decisions, facilitates comparison across case studies, and supports more consistent interpretation of ML results across regions and scales. At the same time, the approach lowers the entry barrier for non-specialists without removing expert oversight or domain judgement.

The framework is being developed within the context of the Erasmus+ AI-LEARN project (Project reference: 2025-1-NL01-KA220-HED-000355215), where it serves as a methodological backbone for future training and capacity-building activities in water and environmental intelligence.

How to cite: Corzo P, G. A., Varouchakis, E., Kamińska-Chuchmała, A., Agioutanti, R., and Dominguez, V.: Integrating Machine Learning and Large Language Models for Next-Generation Water & Environmental Intelligence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18439, https://doi.org/10.5194/egusphere-egu26-18439, 2026.

The increasing availability of satellite remote sensing data has made automatic land cover change detection a persistent research focus. However, real-world applications show that single AI models struggle to cope with the combined challenges of spatial-temporal complexity, feature diversity, and evolving engineering requirements. Consequently, the accuracy of automatically extracted land cover changes is often compromised, making the results insufficient for direct engineering application. Guided by practical application need, this paper focuses on how to utilize satellite remote sensing data, various knowledge and AI technologies to improve the accuracy and efficiency of automatic land cover extraction. This research focuses on the key technologies involved in the complete land cover monitoring process. Central to this study is the proposal of a progressive intelligent change detection technology for satellite remote sensing, characterized by a “identify all, discriminate precisely, refine extraction” workflow. Specifically, the “identify all” step extracts all potential change patches using models such as generic binary change detection. Building on these results, the “discriminate precisely” step filters out patches that are not of current interest. Finally, the “refine extraction” step employs models like semantic segmentation to further screen the results and enhance overall accuracy. An application demonstration in Shanxi Province, China, for new PV facilities, buildings, and roads demonstrated a recall rate of 89.3% for automatic extraction. The high-quality outputs confirm the practical applicability of the results. Consequently, this research affirms the technology as both a valuable and transferable solution for land cover monitoring.

How to cite: You, S., Du, L., He, Y., and Ye, F.: Research on Key Technologies for High-Precision Land Cover Change Monitoring Using Satellite Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19954, https://doi.org/10.5194/egusphere-egu26-19954, 2026.

This study explores the integration of EVE (Earth Virtual Expert), a Large Language Model specialized in Earth Observation (EO) and Earth Sciences, developed under ESA’s Φ-lab #AI4EO initiative in collaboration with Pi School. The primary objective is to enable EVE to connect ESA’s EO platforms and data clusters, creating an integrated ecosystem for the community. This approach leverages agentic capabilities, allowing EVE to dynamically interact with EO tools, databases, and APIs to reason and act autonomously.
To demonstrate this concept, we present a use case where EVE operates within an agentic framework to interact with the EO Dashboard, a joint initiative by ESA, NASA, and JAXA that provides global indicators and narratives derived from multi-mission EO data. Using the MCP protocol, this work enables dynamic connectivity between EVE and the Dashboard, allowing the model to interpret and summarize narratives, extend insights with additional context, and facilitate advanced information retrieval across datasets and stories. In addition, the study considers potential directions for agentic behaviors, assessing early-stage possibilities and limitations for features such as autonomous task chaining. These capabilities enable EVE to perform multi-step reasoning, for example, by interpreting quantitative trends in dashboard indicators such as air quality changes, greenhouse gas concentrations, or land cover dynamics. This links EVE to underlying datasets and enables the generation of scientifically grounded responses. This proof-of-concept demonstrates EVE’s potential to foster interoperability and accelerate Earth system science by improving knowledge accessibility and enabling more effective use of EO data resources.

How to cite: Gmelich Meijling, E., D'Ercole, R., Anghelea, A., Cocchiara, C. M., and Longepe, N.: Building Connected Earth Observation Ecosystems with Agentic AI using EVE, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20206, https://doi.org/10.5194/egusphere-egu26-20206, 2026.

The scale and heterogeneity of modern geospatial datasets, coupled with expanding suites of statistical and dynamical models, produce analysis outputs that are increasingly difficult to navigate and synthesise. We present a practical case study on using a large language model (LLM)-assisted coding tool (GitHub Copilot with GPT-5 mini within Visual Studio Code) to accelerate the development of a lightweight, HTML-based platform that visualises results from pre-calculated climate indicators.

Our starting point was a dataset comprising more than 130 climate indicators derived from gridded observations spanning over 60 years. These indicators originate from multiple meteorological variable groups (e.g., temperature, precipitation) and are aggregated at several temporal resolutions (e.g., annual, seasonal). Downstream analyses include spatiotemporal  statistics, extreme value analyses and statistical significance testing, yielding hundreds of figures that are difficult to navigate and analyse. To make these outputs tractable, we prompted Copilot to generate a simple web application for visualisation and analysis purposes. The pre-generated plots from the climate indicator workflow were displayed there in an organised way, allowing for quick filtering through all indicators and different temporal resolutions, comparing different plots next to each other and using a subpage to concisely display aggregated group plots.

The platform is embedded and deployed via a GitLab CI pipeline, ensuring reproducible updates and immediate web accessibility for collaborators and users, thereby enabling rapid and easy access to vasts amount of output results. Our process of prompting a LLM to generate a visualisation platform offers a convenient and transferable workflow to aid geospatial data analysis.

How to cite: Lehner, S. and Schlögl, M.: Using copilot for the rapid generation of a visualisation platform to aid geospatial analyses, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20992, https://doi.org/10.5194/egusphere-egu26-20992, 2026.

Land subsidence has long been a critical environmental hazard along the southwestern coast of Taiwan, with Yunlin County being one of the most severely affected areas. In this study, Long Short-Term Memory (LSTM) neural networks are employed to develop predictive models for land subsidence. Cumulative land subsidence, groundwater-level variations, and lithological layering are considered as input features to investigate the predictive performance of the models from both temporal and spatial perspectives.

As long-term groundwater monitoring data often suffer from missing values, this study further introduces a Cue Wasserstein GAN with Gradient Penalty (CWGAIN-GP) to impute missing groundwater-level data, thereby improving the stability and completeness of subsequent prediction models. Artificial masking experiments, including continuous missing periods ranging from one month to one year and random removal of 10%–50% of the data. The results show that the average Nash–Sutcliffe efficiency (NSE) achieved by the imputation model reaches 0.897.

For temporal prediction, the land subsidence model is trained using different training lengths (one year and seven years) and variable combinations to forecast cumulative land subsidence over the following one to two years. The most recent six months of observations are used as input to predict the monthly land subsidence increment. The results indicate that longer training periods and more comprehensive input variables lead to improved model performance. The coefficient of determination (R²) for the first prediction year reaches 0.945, while for the second year—under conditions of three consecutive months of missing data—the R² remains as high as 0.923.

For spatial prediction, a multi-station training and single-station validation strategy is adopted. When predicting a target station, the three nearest neighboring stations are selected, and their observations from the most recent three months are used as inputs to predict the monthly land subsidence increment at the target station. This increment is then combined with the known cumulative subsidence from the previous month to estimate the current cumulative subsidence. The results show that the average R² for single-month predictions reaches 0.966. Even when cumulative subsidence is estimated iteratively by adding predicted monthly increments over six consecutive months, the average R² remains around 0.90, demonstrating strong spatial generalization capability of the proposed model.

Fig.1 Monthly vertical profiles of cumulative land subsidence at different depths for the Huwei (MW_HWES) station in 2021.

Overall, this study demonstrates that cumulative land subsidence can be effectively predicted by integrating temporally and spatially informed LSTM models with vertically stratified hydrogeological information. Although cumulative subsidence is used as the primary prediction target, the inclusion of groundwater-level variations and lithological layering enables the model to capture the vertical characteristics of aquifer systems and their influence on subsidence processes. The results highlight the importance of incorporating stratified subsurface information when modeling land subsidence and provide a robust framework for spatiotemporal subsidence prediction under realistic data availability constraints.

How to cite: Hsu, Y.-Y., Lo, W., Lee, J.-W., and Huang, C.-T.: Predicting Cumulative Land Subsidence and Its Spatiotemporal Relationship Using Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2950, https://doi.org/10.5194/egusphere-egu26-2950, 2026.

Soil moisture downscaling is a challenging geospatial regression task that requires accurately capturing complex spatiotemporal relationships across scales. In this study, we conduct a preliminary applicability assessment of denoising diffusion probabilistic models (DDPMs) for continuous-value geospatial regression, exploring the potential of generative modeling frameworks for soil moisture downscaling. The model learns the relationships between coarse-resolution soil moisture observations and multi-source auxiliary features, enabling the generation of high-resolution soil moisture estimates.

During training, the model uses 36 km resolution satellite soil moisture data and conditions on auxiliary variables, including normalized difference vegetation index (NDVI), land surface temperature, surface albedo, precipitation, and digital elevation model (DEM). A conditional embedding strategy is introduced to incorporate temporal information, spatial location information, and in-situ statistics into the diffusion network via feature-wise linear modulation (FiLM), enhancing the model’s ability to capture complex spatiotemporal structures while maintaining stability. During inference, a two-stage “generation–correction” pipeline is employed: high-resolution (1 km) auxiliary features are first used to generate initial predictions through the diffusion model, which are subsequently bias-corrected using in-situ station data.

The applicability assessment combines quantitative and qualitative evaluation. Quantitative metrics include unbiased mean squared error (UMSE), root mean square error (RMSE), mean absolute error (MAE), and R², while qualitative evaluation focuses on spatial pattern consistency and temporal trend representation. Experimental results indicate that the diffusion-based generative model produces reasonable, spatially coherent, high-resolution soil moisture results and successfully captures major temporal variations. These findings demonstrate the applicability of generative frameworks for geospatial regression and their potential as a geospatial regression modeling paradigm, providing a foundation for further refinement and evaluation.

How to cite: Yu, X., Hu, L., Su, C., Yan, Y., Wu, S., and Du, Z.: Long-term Soil Moisture Downscaling Based on Diffusion Models: Applicability Assessment of Generative Models for Geospatial Regression Tasks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5255, https://doi.org/10.5194/egusphere-egu26-5255, 2026.

This study presents a comprehensive sensitivity analysis framework to disentangle the drivers of predictive uncertainty in spatial interpolation and how they ultimately affect spatial predictions. Developed within a Global Sensitivity Analysis context, the proposed approach is model-independent and generic, allowing for broad application across diverse spatial interpolation workflows.

The framework is demonstrated using groundwater Sulfate concentration in the Paris Basin, a dataset characterised by sparse and highly clustered sampling across six distinct aquifers according to the French "BD LISA" hydrogeological system (https://bdlisa.eaufrance.fr/). We represent the underlying spatial process as a  Gaussian Random Field, leveraging Integrated Nested Laplace Approximations for computationally efficient Bayesian inference. This allows for a  probabilistic treatment of uncertainty even within complex spatial structures.

We systematically evaluate the impact of several key uncertainty factors related to both data and model configuration: (1) the number of monitoring stations and their spatial distribution; (2) the selection of the environmental covariates  and the functional form of their effects (linear vs. non-linear); (3) the treatment of censored data (values below detection limits); and (4) structural assumptions regarding the spatial covariance function, specifically the estimation of variogram hyperparameters such as range, sill, and nugget effects and their prior specification. By propagating these uncertainty sources through our framework, we derive domain-wide aggregated sensitivity measures. These metrics quantify how specific data topologies—including sampling density, clustering effects, and censoring rates—govern the stability and accuracy of the resulting spatial interpolations.

Finally, the results facilitate an in-depth discussion on the limitations of purely probabilistic methods in data-poor scenarios. We provide an outlook on the potential of extra-probabilistic approaches, such as imprecise or interval-based kriging, to more robustly address the wide range of epistemic uncertainties inherent in environmental monitoring.

We acknowledge financial support of the French National Research Agency within the HOUSES project (grant N°ANR-22-CE56-0006).

How to cite: Perchtold, C., Rohmer, J., Thomas, A., Lions, J., and Wieskotten, M.: Global Sensitivity Analysis of Spatial Interpolation for Sparse, Clustered, and Censored Data: A Case Study of Groundwater Sulfate in the Paris Basin, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8056, https://doi.org/10.5194/egusphere-egu26-8056, 2026.

Predicting building attributes—such as functional classification, socioeconomic status, and energy efficiency—is a fundamental task in urban science. The current paradigm involves leveraging domain knowledge to extract attribute-specific morphological or topological features for supervised modeling. However, this heavy reliance on manual feature engineering often leads to task-specific models where features must be redefined for each attribute. Consequently, the field lacks a unified, generalizable framework capable of multi-attribute building prediction.

Inspired by recent advances in Regression Language Models (RLMs), which cast continuous prediction as a text-to-text task, we propose Buildings as Text (BaT). BaT serializes structured building representations (e.g., GeoJSON) into raw text and enables end-to-end text-to-text regression. To mitigate the spatial sensitivity of building data, we introduce a Topology-Preserved Coordinate (TPC) strategy that removes each building text’s absolute positional information. Specifically, TPC applies a global coordinate shift to the serialized geometry, suppressing absolute-location bias while preserving local shape and topology. By operating directly on raw text, BaT eliminates manual feature engineering and allows the model to learn a “spatial syntax” from the underlying geometric descriptions.

We validated the BaT framework through a case study on informal settlement (slum) classification. The results demonstrate that our model achieves superior performance and higher adaptability compared to traditional morphology-based methods. While validated on slum detection, this research offers a universal and scalable paradigm for urban building analysis, suggesting that Large Language Models can effectively "read" urban forms for diverse prediction tasks beyond specific domains.

How to cite: Wang, C.-C. and Luo, P.: Buildings as Text: A Universal Regression Paradigm for Building Attribute Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8402, https://doi.org/10.5194/egusphere-egu26-8402, 2026.

The archaeological landscapes of northern Oman host thousands of funerary monuments of different periods and morphologies, forming one of the densest and least explored burial regions of the Arabian Peninsula. Within the framework of LAA&AAS (Laboratorio di Archeologia Analitica e Sistemi Artificiali Adattivi) and MASPAG (Missione Archeologica della Sapienza nella Penisola Arabica e nel Golfo), a multidisciplinary project supported by Sapienza University of Rome and the Italian Ministry of Foreign Affairs, we developed a reproducible geo-AI workflow to classify and analyse funerary structures based on remote-sensing and spatial-context information.

The first dataset, encompassing 185 tombs mapped in the Southwestern Cemetery near the village of Muslimat, in the region of Wadi al-Maʿawil (ca. 70 Km southwest of Muscat) was used to test a machine-learning pipeline designed to discriminate between morphological classes (“tombs” vs “non-tombs”, and within-type subclasses) from high-resolution satellite imagery and derived spatial metrics. Two Random Forest models were compared: a geometry-only baseline using shape descriptors (area, compactness, circularity, elongation), and an extended model incorporating spatial-context features such as kernel density, nearest-neighbour distances, Moran’s I local autocorrelation and cluster membership. The integration of these contextual descriptors increased overall accuracy from 59 % to 76 %, improving model reliability and reducing false positives in morphologically ambiguous contexts. The workflow includes systematic feature importance analysis and confusion-matrix evaluation to assess interpretability and class-imbalance effects.

Beyond the single-site test case, this approach aims to address a broader spatiotemporal challenge: learning and transferring morphological–contextual patterns across different archaeological regions. During 2025 field campaign (20 October – 20 December 2025), more than 500 new tombs were surveyed and georeferenced in the area of the Western Cemetery, expanding the available dataset and enabling large-scale testing of model scalability and transferability. This new phase will assess whether models trained in Wadi al-Maʿawil can generalize to nearby valleys with comparable geomorphological and cultural settings, supporting semi-automated mapping and predictive modelling of funerary features.

The presented pipeline, implemented in an open-source environment (Python, QGIS, and scikit-learn), is designed for reproducibility and transparent parameter tracking. All processing steps—from data preparation and feature extraction to model training and evaluation—are logged and versioned, facilitating cross-project reuse. The workflow thus bridges archaeological and geospatial domains, demonstrating how spatially aware machine learning can improve the detection, classification, and interpretation of complex cultural landscapes.

This contribution highlights the potential of AI and ML in managing spatiotemporal archaeological data and in advancing reproducible analytical frameworks. The methodological approach developed for the Omani funerary landscapes can be generalized to other MASPAG regions, supporting comparative analysis of desert landscapes and long-term dynamics of human–environment interaction across the Arabian Peninsula.

How to cite: Meneses Pineda, A. S., Solinas, M., Ramazzotti, M., Musacchio, M., and Buongiorno, M. F.: A preliminary study of the morphology and spatial distribution of funerary elements in Oman, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11193, https://doi.org/10.5194/egusphere-egu26-11193, 2026.

Understanding the spatial dependence of residuals is important for interpreting and diagnosing spatial machine learning models. Spatial autocorrelation in the residuals suggests that the model has not fully captured the data's spatial structure. This may imply that the model is missing crucial spatial context or interactions, and that, in effect, it is spatially biased, leading to underestimation in some areas and overestimation in others.

Moran's I is a commonly used statistic for the diagnosis of spatial autocorrelation in spatial predictions, providing a single-value quantitative measure with a straightforward interpretation. This measure quantifies the degree of spatial autocorrelation, indicating whether similar values are clustered together or dispersed across space. The information provided by Moran's I has been used in various ways in studies applying machine learning: to evaluate model performance, interpret results, understand model limitations, and compare different modeling approaches.

Unlike standard model performance metrics, such as R2 or RMSE, Moran's I depends not only on the values of residuals but also on the spatial context—especially the study area's extent, the sampling strategy used, and the specification of spatial weights. However, there is a lack of a comprehensive understanding of how these factors influence the results of Moran's I calculation in the context of spatial machine learning, and of how to best use this measure for model evaluation and comparison.

Using simulated data with controlled spatial properties, we investigated how testing set size, sampling strategy, and the specification of spatial weights influence Moran's I computed on model residuals. Our results show that Moran's I, calculated based on k-nearest neighbors approach,  primarily reflects the spatial structure of values in the testing set rather than the residual autocorrelation across the full prediction domain, often underestimating fine-scale spatial patterns. These findings have various implications: weight-matrix definitions must be clearly reported, calculations on sparsely distributed or clustered samples should be avoided, Moran's I is generally not directly comparable across studies due to differences in spatial extents and sampling, and its values are inherently scale-dependent.

With this contribution, we aim to present the behavior of Moran's I calculated from residuals of spatial machine learning models under different conditions, outline best practices for selecting and reporting spatial weights, and discuss how to interpret Moran’s I. 

How to cite: Nowosad, J., Meyer, H., and Schmidinger, J.: Using Moran's I for assessing residual spatial autocorrelation in machine learning models , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11342, https://doi.org/10.5194/egusphere-egu26-11342, 2026.

How to cite: Görgen, D. A., Heilig, S., Meyn-Grünhagen, L., Fischer, A., Lederer, J., and Meyer, H.: Area of Applicability for Deep Learning: Exploring Latent Space Geometry of Earth Observation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12275, https://doi.org/10.5194/egusphere-egu26-12275, 2026.

Near-real-time forest monitoring is a critical component for managing climate risk and resource conservation in the Brazilian Legal Amazon (ALB). The DETER program, managed by the National Institute for Space Research (INPE), has played a pivotal role for over two decades by producing warnings of deforestation and forest degradation to support environmental enforcement by agencies such as IBAMA. However, the effectiveness of these warnings is highly dependent on temporal efficiency — the speed at which a disturbance is detected and published after the activity begins.

Recent objective evaluations of DETER’s performance regarding selective logging — a major driver of forest degradation — revealed a significant median delay of approximately 312 days between the start of logging activities and the corresponding warning publication during 2022-2023. This temporal gap highlights the challenge of applying traditional monitoring to complex spatiotemporal datasets, where factors like cloud cover and sensor resolution can hinder early detection.

To address these challenges, this research proposes a novel approach within the framework of AI and Machine Learning in Spatiotemporal Contexts. We leverage Foundation Models and Deep Learning architectures designed to process the complex temporal dynamics of tropical forests using Harmonized Landsat-Sentinel (HLS) time series. A key contribution of using foundation models in this pipeline is their ability to learn robust representations from large-scale data, significantly reducing the requirement for vast volumes of manually annotated samples — a known bottleneck for AI-based remote sensing monitoring systems. By applying these models to HLS data, we aim to improve spatiotemporal predictions and the reliability of the modeling pipeline, facilitating the production of more agile and efficient early warnings.

This work contributes to the development of the next generation of forest monitoring systems, focusing on interpretability and transferability across the Amazonian landscape. By reducing the detection lag of selective logging, this approach seeks to enhance technological sovereignty in environmental monitoring and provide more effective decision-making support for forest preservation.

How to cite: Taquary, E. and Aragão, L.: Enhancing Near-Real-Time Forest Monitoring: Foundation Models and Harmonized Landsat-Sentinel (HLS) Time Series for Selective Logging Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16001, https://doi.org/10.5194/egusphere-egu26-16001, 2026.

The integration of Earth Observation (EO) data with machine learning (ML) has transformed the mapping of Deprived Urban Areas (DUA). Despite these technical advances, persistent disconnect remains between research outputs and their operational uptake by local stakeholders. In parallel, advances in ML and deep learning (DL), together with new satellite missions have improved the extraction of building footprints and urban morphology. Nevertheless, DUA mapping studies, which largely depend on these physical indicators, often prioritize benchmark performance over the robustness, transparency, or usability required in real-world decision-making contexts. One of the main reasons for this gap is spatial data quality (SDQ), which fundamentally limits model performance and generalization. When data quality is poor, due to inaccuracies, incompleteness, or inadequate provenance, models become unreliable, regardless of architectural complexity. Furthermore, many studies rely on validation strategies that ignore spatial autocorrelation, thereby yielding overoptimistic accuracy estimates that mask poor generalization to new local contexts.

To address these challenges, this paper argues for a shift toward a systematic assessment of spatial data quality. We first conduct a scoping review of 50 state-of-the-art DUA mapping studies published between 2017 and 2025. Our analysis reveals a high dependence on very-high-resolution imagery (72%), a widespread lack of publicly accessible data and code (92%), and a critical deficiency in operationalizing semantic definitions of DUAs with 90% of studies failing to provide mapping rules (for visual interpretation) or ground rules (for in-situ collection). Most studies also fail to assess user needs (90%) or do not consider the ethical implications of using DUA data (88%), which is highly sensitive due to risks such as forced evictions. Building on these findings and established international standards from ISO and the OGC, we propose a comprehensive Spatial Data Quality (SDQ) framework tailored to transparently document supervised image classification in DUA mapping. This framework integrates established practices such as adherence to the Findable, Accessible, Interoperable, Reusable (FAIR) principles and assessment of acquisition, measurement and spatial-temporal quality with novel dimensions addressing semantic consistency, sampling representativeness, human factors in annotation, learning shortcut risk, user needs validity, ethical considerations, and transparent reporting of the dataset’s potential failure modes or uncertainties. By operationalizing SDQ as a living, extensible framework, this work aims to better align advances in ML and DL with sustained societal impact, ensuring that DUA mapping products, or any relevant application domain, are fit for use by local communities and decision-makers.

How to cite: Campomanes V, F., Kuffer, M., Stein, A., Dijkstra, A. M., Trento Oliveira, L., and Belgiu, M.: A framework for assessing the quality of spatial data applied in supervised image classification of deprived urban areas, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19025, https://doi.org/10.5194/egusphere-egu26-19025, 2026.

Greenhouse agriculture has become a crucial element of agricultural practices in Morocco, yet its spatial and temporal evolution remain insufficiently quantified. This study aims to map greenhouse structures at the Souss-Massa region scale in order to assess the progress of covered agriculture and examine its relationship with socio-economic development in Morocco. Using hand-annotated greenhouse data from the Chtouka region as ground truth, we develop a deep learning–based detection framework relying exclusively on open-source tools. Multispectral Sentinel-2 satellite imagery at 10 m spatial resolution is used as input to a U-Net convolutional neural network, which is trained, validated, and tested for greenhouse segmentation. The proposed model achieves an overall accuracy of up to 94%, demonstrating strong generalization capability. The resulting plug-and-play methodology enables scalable, cost-effective, and open-source greenhouse mapping, and provides valuable insights into the dynamics of covered agriculture and its role in Morocco’s agricultural and socio-economic development.

How to cite: El hachemy, S., Aglagal, C., Ait-Ichou, H., Elhaid, I., Zlaiga, J., Hssaisoune, M., Bouchaou, L., and Belaqziz, S.: Satellite imagery for greenhouse mapping in Morocco using U-net model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19057, https://doi.org/10.5194/egusphere-egu26-19057, 2026.

Accurate forecasting of vector-borne diseases such as dengue is often challenged by limited and noisy spatiotemporal data. This study evaluates the effectiveness of data augmentation techniques in enhancing the robustness and predictive accuracy of machine learning models. We assess multiple augmentation strategies applied to weekly dengue case data across countries in South and Central America (2014–2022). Results show that augmentation substantially improves short-term forecasting performance, particularly in regions with sparse or irregular observations, yielding higher R² values and lower relative errors compared to non-augmented baselines. These findings demonstrate that well‑designed augmentation can mitigate data scarcity and strengthen the generalization of graph‑based deep learning frameworks for epidemiological forecasting. Overall, the study highlights augmentation as a practical and scalable approach for improving spatiotemporal ML applications in disease surveillance.

How to cite: Siabi, N., Son, R., Thomas, M., Irrgang, C., and Saynisch Wagner, J.: Improving Dengue Forecasting with Spatiotemporal Data Augmentation and Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19452, https://doi.org/10.5194/egusphere-egu26-19452, 2026.

At continental scale, crop classification needs models that capture phenology through temporal analysis without degrading field boundaries. We introduce a decoupled architecture that uses static foundation‑model features across multi‑sensor time series and fuses them with high‑resolution spatial features. The temporal stream ingests paired multispectral and SAR sequences plus a static DEM and metadata, extracts foundation model token features per timestep, and compresses them with a Perceiver‑style bottleneck that cross attends from a fixed latent bank to the full foundation model token volume. Such heavy compression collapses sequence length by orders of magnitude, which makes longer temporal windows and larger batches ingestible on consumer‑grade GPU memory constraints while preserving the temporal signatures needed to separate crops with similar single‑date appearance.
The spatial stream stays purely static --- it selects a single high‑quality multispectral reference frame and passes it through a high‑resolution backbone to retain fine geometry and crisp boundaries. The two streams are joined in a query‑based decoder, where dynamic queries generated from the compressed temporal latents attend to multi‑scale spatial features, aligning phenological signatures with precise field edges. This fusion mechanism prevents coarse temporal features from blurring geometry and makes delineation robust to shifts in timing or crop management practice. In fact, temporal queries encode crop‑specific growth signatures, while the spatial stream supplies the pixel‑level evidence for boundary localization, whereas the decoder enforces instance‑aware segmentation through iterative cross‑attention and masked refinement.
We evaluate on EuroCrops crop‑class labels, achieving a Micro Recall of 84.1% and a Segmentation Quality of 84.2%. Transferability is tested with a spatial holdout protocol using geographically disjoint train/test regions, reliability is summarized by aggregate metrics on these strict splits, and uncertainty is communicated through per‑class performance variability and label‑noise sensitivity analyses that bound achievable scores.

How to cite: Vinholi, J. G., Sleimi, R., Werner, F., and Abelló, A.: A Two‑Stream Spatiotemporal Architecture with Foundation‑Model Features Applied to Crop Classification, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19669, https://doi.org/10.5194/egusphere-egu26-19669, 2026.

Spatial Data Infrastructures (SDIs) contain a lot of spatial data from various organizations and data producers. Metadata is intended to enable the discovery of the data, yet finding the relevant data can be challenging. The challenges include rigid keyword-search, complex search interfaces in geoportals, map-based search that require some geographic knowledge, and language differences between user queries and the metadata.

The development of Large Language Models (LLMs) offers new opportunities to improve spatial data discovery. LLMs demonstrate strong language understanding and generation capabilities and have been used in information retrieval tasks. They can overcome semantic differences and language barriers between user queries and the needed information. However, their internal knowledge is limited and they are prone to hallucinations. Unless the datasets in SDIs, or the web pages describing them are indexed by search engines, LLMs with internet search tools cannot find them. 

Retrieval-Augmented Generation (RAG) offers a solution for the knowledge limitations, by connecting an LLM with an external and up-to-date knowledge base. However, RAG mainly works in the textual domain and excels at retrieving external information that is semantically relevant to a user query. Queries for geographic data have a spatial aspect yet the spatial reasoning capabilities of LLMs are limited. For a query like “forest data for Vienna”, RAG can identify the relevant forest data from a pool of metadata, regardless of the language or words used to describe the data. However, identifying datasets that meet the spatial intent is a problem. DCAT metadata, the most popular metadata standard, defines the spatial extent of spatial datasets using bounding box coordinates or as links to gazetteers. Naive RAG is based on semantic similarity approaches.  An LLM can identify “Vienna” as a location, but would struggle to identify datasets relevant to the location, as there is little semantic similarity between the location name and coordinate digits or gazetteer links.  There is thus a need to incorporate spatial indexing techniques for improved spatial reasoning.

With our contribution we present an approach that combines LLMs, RAG, and spatial indexing techniques to overcome existing challenges in discovering spatial data in SDIs, and improve spatial data discovery through natural language queries.

How to cite: Ondieki, J. O., Rieke, M., and Jirka, S.: Using Large Language Models to Enhance Spatial Data Discovery in Spatial Data Infrastructures, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20221, https://doi.org/10.5194/egusphere-egu26-20221, 2026.

The recent rise of foundation models in Earth Observation (EO) has reshaped how remote sensing tasks are approached, particularly by allowing strong downstream performance with comparatively limited labeled data. These models have reported impressive results in applications such as land cover classification and semantic segmentation. However, performance gains alone do not resolve a central concern: whether the resulting predictions can be trusted. In practical EO scenarios—including disaster response and environmental monitoring—miscalibrated confidence estimates may lead to incorrect decisions even when overall accuracy appears high.

Motivated by this gap between accuracy and reliability, this study focuses on the uncertainty calibration behaviour of fine-tuned EO foundation models. Using TorchGeo for consistent data handling and the Lightning-UQ-Box framework for uncertainty quantification, we construct an evaluation pipeline that contrasts Vision Transformer–based pretrained models with conventional convolutional neural networks trained from scratch. Experiments are conducted across both image classification tasks (e.g., EuroSAT) and dense prediction settings such as semantic segmentation.

Rather than assuming superior representations automatically yield better-calibrated predictions, we explicitly examine how calibration properties change after fine-tuning large pretrained models. In addition, we evaluate a spectrum of uncertainty quantification approaches, from lightweight post-hoc methods like temperature scaling to more computationally demanding techniques, including Monte Carlo Dropout, deep ensembles, and Laplace approximation. Calibration quality is assessed using expected calibration error and reliability diagrams, alongside predictive accuracy.

By analysing the trade-offs between computational cost, accuracy, and calibration, this work provides practical insight into which UQ strategies are most effective for EO foundation models. Our findings aim to support the deployment of remote sensing systems in operational settings where reliable uncertainty estimates are as critical as raw predictive performance.

How to cite: Wei, Y.: Uncertainty Quantification for Earth Observation Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20813, https://doi.org/10.5194/egusphere-egu26-20813, 2026.

Localized rapid urban expansion and global climate change have contributed to land use and land cover (LULC) dynamic modifications, which further links to changed land surface temperatures (LST). This study proposes an integrated approach of machine learning (ML) models in assessing decadal LULC changes and future prediction in a city in the Mekong region. To achieve an accurate LULC map object-based classification strategies were implemented using various ML techniques across observed years with four main land cover categories: built-up areas, water bodies, paddy fields/shrubs, and orchards, together with LST extraction. The findings reveal that Random Forest classifier works superior to other classifiers, achieving the best overall accuracy of 81%. There have been substantial land usage changes, with the percentage of developed areas rising from 8% in 2014 to approximately 12% in 2024. Urbanization is correlated with rising temperatures , while, vegetation, on the other hand, helps alleviate this heat by providing shade and cooling. With an overall accuracy of 85% in the patch-generating land use simulation (PLUS) model, by 2030, under the impacts of both natural and socio-economic drivers, an apparent increase in the proportion of built-up areas to 15% and a slight variation in other categories could be seen in line with planning objectives. The urban expansion could be clearly seen in the highly dense districts with an increase to 42% by 2030 from an initial stage of merely 27% in 2014. The primary forecast conversions in LULC observed were vegetated lands transforming into construction areas for urbanization, yet maintaining agricultural practices for food security. The integrated approach has proven its suitability in intricate land usage patterns evaluation and optimization.

How to cite: Nguyen, L. and Daou, D.: Harnessing an Integrated Machine Learning based Approach in Monitoring and Predicting Dynamic Spatiotemporal Land Use and Land Cover Changes. A case study in a Mekong city , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21002, https://doi.org/10.5194/egusphere-egu26-21002, 2026.

Ecosystems are a key component of biodiversity, providing vital services to humans and the economy. Anthropogenic pressures driving environmental change result in widespread ecosystem degradation and loss. The area and spatial distribution of ecosystem types, referred to as ecosystem extent, provide a critical entry point for assessing ecosystem condition, functioning, and associated services, and therefore require detailed and spatially explicit monitoring.

Despite advances in geospatial analysis, consistent mapping and delineation of ecosystem extent remain challenging. Map products on ecosystem extent should, therefore, be supported by uncertainty assessments, ideally in a spatially explicit manner. According to best practices in related fields, the minimum requirement for uncertainty quantification for thematic maps is the aggregated estimation of per-class accuracy and per-class area uncertainty, following a validation procedure based on independent reference data. However, the standard practice remains spatially implicit. To date, there is no established practice for spatially explicit uncertainty quantification procedures.

This study presents a spatial solution for estimating the uncertainty of maps produced using machine-learning algorithms. The approach builds on the standard map-validation procedure and extends it to pixel-wise assessments using conformal prediction. While conformal prediction can be applied to any machine learning algorithm, ecosystem extent mapping poses domain-specific challenges, including a high-dimensional multi-class setting and hierarchical class structures. This study, therefore, focuses on developing solutions to ensure robust class-specific coverage, exploring different conformal prediction implementation variants, and adapting them from flat to hierarchical mapping scenarios.

To assess the feasibility and applicability of our approach, we tested it on the Oslo-Viken municipality in Norway. In this case study, we developed an ecosystem extent map for 2024 and quantified and mapped its uncertainty at pixel-level. This analysis helped to evaluate the practical application and performance of the approach on real-world cases.

How to cite: Tregubova, P., Clappe, S., Marielle Mienna, I., Smets, B., Buchhorn, M., Remelgado, R., and Meyer, C.: Spatially-explicit uncertainty assessment of ecosystem extent mapping, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21265, https://doi.org/10.5194/egusphere-egu26-21265, 2026.

Developing operational Digital Twins for large-scale agro-hydrological systems presents significant challenges regarding data heterogeneity, computational efficiency, and the integration of Earth Observation (EO) with process-based modeling. This study presents the development and initial testing of DT-Agro, a spatially explicit Digital Twin of the Greek agro-hydro-system, designed to support sustainable water management and agricultural planning at the national scale.

DT-Agro integrates a high-resolution spatial database, a hybrid meteorological forcing scheme, and a distributed agro-hydrological model (AgroHydroLogos) recoded in C++ and Python for enhanced performance. A key innovation in the process simulation is the development of a novel, impervious-aware SCS-CN formulation. Unlike traditional lumped approaches, this method explicitly decomposes each grid cell into pervious and impervious fractions using high-resolution Copernicus Imperviousness Density data. This allows for a physically consistent representation of runoff generation in mixed landscapes, capturing the hydraulic response of small impervious patches that are often lost in standard gridded models.

Furthermore, to address the chronic fragmentation of ground-based monitoring networks, the system introduces a "virtual station" meteorological framework. Recognizing that raw global reanalysis products (e.g., AgERA5) often exhibit significant biases in Greece’s complex terrain, we developed a hybrid correction workflow. AgERA5 time series are sampled at the locations of historical stations and bias-corrected using station-specific regressions. This creates a network of "virtual stations" that provide continuous, homogenized daily records, filling temporal gaps while preserving local climatological characteristics. These records drive a dynamic spatial interpolation scheme that accounts for temperature and precipitation gradients, ensuring physically consistent meteorological forcing across the national domain.

We present results from the initial national-scale application of the system. The testing phase focused on quantifying irrigation water abstractions and their spatial-temporal drivers. Initial simulations estimate the long-term average national irrigation abstraction at approximately 6,600 hm³/year, with significant inter-annual variability (6,000–7,800 hm³) driven by climatic conditions. Validation against theoretical net irrigation requirements for major crops (maize, cotton, alfalfa) yielded consistent depths (380–420 mm), confirming the biophysical realism of the model core. These results demonstrate DT-Agro’s capability to provide a robust, evolving representation of the Greek agro-hydro-system for climate adaptation planning.

How to cite: Palli Gravani, S., Soulis, K., Soulis, X., and Kalivas, D.: DT-Agro, a digital twin of the Greek AgroHydroSystem, development and testing, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1994, https://doi.org/10.5194/egusphere-egu26-1994, 2026.

Destination Earth has given rise to a new generation of global climate models capable of resolving kilometre-scale processes. These storm-resolving models are deployed operationally to produce multidecadal climate simulations within Climate Adaptation Digital Twin. OBSALL conceptualise and implemente an Earth observation-based system for monitoring the quality of the climate simulations. Technically, it is a one-pass algorithm in which observation models operate online on the state vector of the simulation model and generate a full-resolution trace in the observation space. This enables real-time monitoring of the simulation and posterior evaluation right after its completion, thus enhancing the overall resilience of the Climate DT workflow.

OBSALL has the projection and monitoring pipeline implemented for three observation modalities: surface variables with SYNOP weather stations, vertical profiles with TEMP radiosoundings, and satellite products with AMSU-A. OBSALL consumes climate model states in the form of stream of Generic State Vectors, projects the model states in to the observation space spanned by activated observation modalities, stores the projected states in Observation DataBase files and intermittently produces a suite of monitoring plots comparing the climate simulation to climatology of observations.  All modalities share the same general workflow structure facilitating implementation of new modalities or features. Finally, OBSALL has been designed to allow deployment on a range of environments from personal computer to HPC infrastructures.

This functionality will be part of the EU’s Climate Adaptation Information Service and it is intended to benefit Users of the Service, although it also holds strong research appeal. From the User perspective, presenting the simulation in observation space is advantageous. Earth observations, especially the in-situ components, are intuitive and easy to interpret. Therefore, observation-space projections provide a concrete handle on adaptation information and enhance the User relevance of the DestinE Climate DT. These projections tend to lower the threshold for users to engage with the Information Service. Therefore, we see future development of user-oriented tools using the projection data as input as a promising strategy to attract new users. Extending the projection to other user-relevant observation types, such as long-term wind mast measurements, would be beneficial.

Here, we showcase the current capabilities of the climate model projection and monitoring software OBSALL from the view-points of runtime simulation monitoring through different observation types, and observation-based posterior validation, which offers a versatile way to validate process-level features in storm-resolving climate models.

How to cite: Bouvier, C., Tuppi, L., Räisänen, J., and Järvinen, H.: OBSALL: DestinE climate simulations in observation space, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5404, https://doi.org/10.5194/egusphere-egu26-5404, 2026.

This contribution presents the system architecture, data pipelines, and modelling logic at a conceptual level for a Digital Twin prototype named JÄÄTwin. The aim of JÄÄTwin is to integrate near-real-time observations from multiple heterogeneous data sources, with numerical models, and computational infrastructure into a live “river ice condition” digital twin. While large-scale initiatives focus on continental and global domains, digital twin implementations at the river scale remain limited, specifically for cryo-hydrological systems. JÄÄTwin integrates in-situ monitoring data, remote sensing products, and meteorological forecast inputs through a modular backend that supports data ingestion, preprocessing, model orchestration, and visualization. The emphasis of this contribution is on system architecture, data integration logic, and operational workflow design.

The Kiiminkijoki River in northern Finland is used as a pilot to demonstrate how a river-scale digital twin can be implemented using existing monitoring infrastructure. The presentation discusses design choices, integration challenges, and transferability considerations relevant to future Earth system digital twin developments and to emerging European digital twin initiatives.

How to cite: Jalali Shahrood, A.: River-Scale Digital Twin for Cryo-Hydrological Systems: Architecture and Integration Principles from the JÄÄTwin Framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6018, https://doi.org/10.5194/egusphere-egu26-6018, 2026.

As Destination Earth (DestinE) matures, the capability to simulate not just natural phenomena but also the "related human activities" becomes critical for delivering actionable insights on sustainable development. This work presents TRANSITION, an operational Digital Twin application designed to model the complex socio-environmental dynamics of land-use change, renewable energy integration, and agricultural sustainability within the DestinE ecosystem.

While traditional Earth system digital twins excel at forecasting physical variables (e.g., crop yields or solar irradiance), they often lack the behavioral fidelity to predict how human actors will respond to these changes. TRANSITION bridges this gap by integrating Earth Observation (EO) data with a Multi-Level Agent-Based Modelling (ML-ABM) system driven by Reinforcement Learning (RL). In this framework, autonomous agents—representing farmers, landowners, and policymakers—make spatially explicit decisions based on environmental suitability, economic incentives, and social factors (PECS framework).

We demonstrate the application of this digital twin through three core stakeholder-co-designed use cases:

• Climate Change Adaptation Strategies: Simulating long-term land-use shifts under various CMIP climate scenarios to identify regions at risk of agricultural abandonment or suitable for crop diversification.
• Green Credit & Policy Simulation: allowing policymakers to "stress-test" interventions—such as subsidies for photovoltaics (PV) or green credits—in a risk-free virtual environment to assess adoption rates and potential conflicts between food and energy production.
• Renewable Energy Optimization: Utilizing Sentinel-derived analytics and high-resolution Digital Elevation Models (DEMs) to identify optimal deployment zones for renewable infrastructure while accounting for socio-economic acceptance.

To ensure these insights are scalable and actionable, the application is architected to eventually run on Destine’s Platform, with the potential to utilize High-Performance Computing (HPC) for heavy agent training and the DestinE Data Lake for seamless access to Sentinel and ERA5 datasets. By coupling high-precision physical modeling with realistic human behavior, TRANSITION offers a robust decision-support tool for evidence-based policymaking, directly contributing to the European Green Deal’s vision of a resilient and adaptive society.

How to cite: Paraskevas, C., Gousios, G., Mamouka, T., Vourlioti, P., Kasampalis, D., Kotsopoulos, S., and Vitolo, C.: Simulating the human dimension in Destination Earth. An EO-Informed digital twin application for climate-adaptive policy planning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6871, https://doi.org/10.5194/egusphere-egu26-6871, 2026.

Mountain glaciers are critical elements of the Earth’s hydrological and climate systems. The rapid changes in glaciers due to climate change hinder our ability to monitor and address associated risks effectively. To address these challenges, we present the Digital Twin Component for Glaciers (DTC Glaciers), part of ESA’s Digital Twin Earth (DTE) programme. In this presentation, we will demonstrate the early prototype of the DTC Glaciers system, developed through close co-design with our stakeholders in the hydropower and water sectors. The demonstration will highlight current capabilities, including regional glacier mass-balance assessment, runoff estimation, and user-informed scenarios. We will also share key lessons learned from Phase 1, focussing on data assimilation of heterogeneous datasets, the practicalities of building adaptive architectures, and the challenges of meeting diverse user needs within a unified framework. Despite these challenges, DTC Glaciers offers a well defined test case to assess the transformative potential of digital twins in climate risk assessments.

How to cite: Maussion, F., Bizon, J., Gampierakis, N., Gourmelen, N., Jakob, L., Michael, C., Nagler, T., Nussbaumer, S., Schmitt, P., Schwaizer, G., and Zemp, M.: Lessons learned from the pilot phase of the Digital Twin Component for Glaciers project (DTC Glaciers), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8395, https://doi.org/10.5194/egusphere-egu26-8395, 2026.

The AI4Clouds project develops a deep-learning-based enhanced short-term (up to 12 h) cloud fields for the Destination Earth (DestinE) Weather-Induced Extremes Digital Twin (Extremes DT). By fusing high-resolution Extremes DT simulations with EUMETSAT satellite observations (SEVIRI / FCI), the system learns to correct systematic model biases and provide enhanced cloud-related fields, such as cloud cover, optical depth, and top height fields, which are key variables for renewable-energy and weather applications.

AI4Clouds follows a multi-stage training strategy: it first pre-trains on ERA5 reanalysis data collocated with satellite datasets to capture large-scale dynamics, then fine-tunes on Extremes DT forecasts, also collocated with satellite datasets. It employs stretched-grid Graph Neural Network–Transformer architectures implemented within ECMWF’s Anemoi framework. Probabilistic forecasts are produced via an ensemble approach that quantifies aleatory and epistemic uncertainty. All data retrieval, preprocessing, training, and serving workflows are deployed on the DestinE Data Lake using its HDA, ISLET, and Stack services, ensuring reproducibility and operational integration through MLOps pipelines.

Validation relies on the open-source AQUA framework, extended with cloud-forecast and deep-learning diagnostics (e.g., RMSE, bias, CRPS, spectral metrics). An industrial partner from the solar-energy sector provides user-driven evaluation across several Iberian sites.

By integrating Earth-observation data, high-resolution numerical forecasts, and deep learning within DestinE’s infrastructure—a cloud-native environment—AI4Clouds demonstrates a scalable path toward building actionable applications for weather-sensitive sectors.

How to cite: Iglesias-Suarez, F., García, M., Heredia, I., Perez, A., Portilla, S., Sáinz-Pardo, J., and San-Martin Segura, D.: AI4Clouds: Enhancing Short-Term Cloud Fields in DestinE for the Solar-Energy Sector, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10081, https://doi.org/10.5194/egusphere-egu26-10081, 2026.

Climate resilience is a defining challenge of the 21st century, yet public health authorities continue to face difficulties in operationalising state-of-the-art geospatial and environmental science. In Portugal, as in Europe more broadly, extreme temperatures have already increased in frequency and severity, contributing to substantial excess mortality and morbidity. These impacts are often amplified by the simultaneous degradation of air quality. However, evidence has largely been event-specific, fragmented across case studies of individual heatwaves, cold waves, or air-quality episodes, limiting our ability to implement early-warning systems. The ESA-funded AIR4health project, developed under the Early Digital Twin Components initiative, addresses these gaps by designing innovative algorithms to predict human mortality and morbidity during compound extreme events. The project develops two Machine Learning (ML)–based AIR4health Risk Algorithms focusing on (1) Heat & Ozone and (2) Cold & Nitrogen Dioxide, using a two-decades-long, high-resolution healthcare database for mainland Portugal. These indicators will integrate EO data, in-situ air-quality records from the EEA, and CAMS/C3S model outputs. Satellite and model data are dynamically downscaled using approaches previously demonstrated for air-temperature modelling in Lisbon, enabling daily, spatially detailed (municipal-level) time series of compound extreme events. AIR4health advances beyond current country-level systems by implementing fully spatiotemporal exposure–response modelling. Its dynamic and continuous framework will deliver a prototype DTC capable of providing fine-scale early warning for combined climate and air-quality extremes. By benchmarking results against European-level datasets, AIR4health will support scalable pathways towards relevant practices in planetary health and climate-preparedness, while contributing to the broader European Digital Twin ecosystem. 

How to cite: Villa de Brito, A., Oliveira, A., Marques, B., Fonteles, C., Pereira, É., Silva, F., Girão, I., Figueiredo, L., Lima, M., Cunha, R., Oliveira, M., Nogueira, P., Santos, B., Trancoso, R., and Teresa, V.: Climate and Environmental Digital Twins for Human Health: Leveraging Earth Observation for Compound Climate and Air Quality Extremes Early Warning , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13491, https://doi.org/10.5194/egusphere-egu26-13491, 2026.

Spring-time frost events pose a significant risk for the agricultural sector. Vineyards, apricots and other crops might be damaged by freezing temperatures after vegetation season onset. Frost events occur commonly during cold outbreaks in April or May after prolonged periods of relatively high air temperatures in late-winter or early-spring. The damage can be significantly reduced by suitable measures, provided a reliable forecast is available. As frost intensity is often highly spatially variable, its forecast can benefit from hectometric-scale numerical atmospheric modelling. We present results related to frost events detection and impact modelling within the Destination Earth On-Demand Extremes (DEODE) project. The detection scheme uses the ECMWF ensemble forecasts of 2-m air temperature and cloud cover to identify regions of potential high-impact frost events. The Global Digital Twin 2-m air temperature sums since the start of the year are used to delimitate areas where vegetation is not active yet. These areas are masked from the risk assessment. Finally, the frost damage risk is expressed in 0–5 scale and may serve as guidance for hectometric-scale simulation domains. After a hectometric simulation is completed, the output is ingested into a downstream frost impact model. The impact model estimates current development phase of several crops and corresponding temperature thresholds for frost damage. These thresholds are compared with hectometric-scale forecasted temperatures to obtain the magnitude of excess of critical temperature for each crop. The impact model has been evaluated during several pilot frost events in the Czech Republic and Spain.

How to cite: Matějka, M., Hájková, L., Možný, M., Musilová, A., and Vlach, V.: Frost event detection and impact modelling within the Destination Earth On-Demand Extremes framework    , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13552, https://doi.org/10.5194/egusphere-egu26-13552, 2026.

Operating in the polar regions is growing increasingly important due to expanding research needs, emerging economic opportunities and efforts to protect national sovereignty.

Ships operating in the Arctic and Antarctic face heightened risks and more severe consequences in the event of an accident due to sea ice, icebergs, harsh sea states, low visibility and extreme temperatures. These hazards are compounded by sparse infrastructures and remoteness, with immediate assistances not readily available during emergencies.

While climate change is quickly reducing the amount of sea ice, this does not necessarily translate to a reduction in risk in the short-term. On the contrary, uncertainly due to changing sea ice conditions, and an increase in icebergs due to melting glaciers, can increase risk.

The operational sea ice community provides information to lessen risk to life, property and the environment and to improve operational efficiency. That information is used by ships to avoid or navigate through sea ice, by ship operators in planning polar voyages, and by policy makers to assess the impact of climate change on future decisions regarding polar operations. Beyond safety, these services also increase operational efficiency by enabling optimized routing, reduced fuel consumption, and shorter transit times.

The information provided by the operational sea ice community comes from in-situ measurements, satellite earth observation data, and sophisticated models of the atmosphere, oceans, sea ice, and icebergs. These data streams have historically been assembled and interpreted by highly trained human analysts. Given the rapid increase in the amount of data that needs to be analyzed and the constraints on workforces due to government fiscal reductions, their work is increasingly being assisted by artificial intelligence. Furthermore, because sea ice is highly dynamic and can change within a matter of hours, automation and AI are indispensable for providing real-time and forecast information.

The operational sea ice workflows have many of the attributes of an Earth System Digital Twin:

• Utilization of detailed digital models of relevant earth systems (atmosphere, ocean, sea ice),
• Continuous incorporation of near-real-time data from in-situ sensors and earth observation satellites,
• Use of artificial intelligence and machine learning,
• Generation of predictive models and forecasts, and
• Provision of advance analytics and decision support tools to allow end-users to optimize their choices.

In fact, much of the information both used and generated by the operational sea ice community is available in Destination Earth, the initiative of the European Commission to develop a digital twin of the Earth.

This presentation will examine the provision of operational sea ice information as an example of the application of digital twins to provide actionable insights on climate adaptation and disaster risk reduction. It will present current capabilities, AI-enabled workflows, and lessons learned from operational implementation, with a focus on supporting safe and sustainable polar shipping.

How to cite: Arthurs, D., Rabenstein, L., Dolezalova, T., Puestow, T., Rasmussen, T., and Korosov, A.: Operational Sea Ice Information as a Digital Twin, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14321, https://doi.org/10.5194/egusphere-egu26-14321, 2026.

How to cite: Lo Iacono, M., Tona, C., Cortese, M., and Scarda, B.: DestinE Platform: Transforming Earth System Science into Actionable Services, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17401, https://doi.org/10.5194/egusphere-egu26-17401, 2026.

The  GET-it project (Geohazards Early Digital Twin Component) is devoted to build a Digital Twin for geohazards in the framework of the ESA Early Digital Twin Component (DTC) programme. The Geohazard Digital Twin is transitioning from a demonstration prototype to a shareable, pre-operational Earth Observation (EO)-based service platform. This transition has been initiated through the integration of the GET-it scenario modules within the Geohazards TEP (https://geohazards-tep.eu), a long-standing operational platform developed under ESA, and will continue through their future integration and federation with the Destination Earth (DestinE) framework. The project explores how EO data can effectively drive DTCs for volcanic and seismic geohazards within the Destination Earth framework.

The Geohazard DTC developed in GET-it is designed as a modular and customizable environment, capable of integrating multi-sensor EO data, primarily from the Copernicus programme, with established physical models and EO-driven analysis tools. The adoption of standardized input and output formats ensures interoperability and facilitates the uptake of DTC products by diverse user communities with different operational needs. These characteristics enable repeatable, timely EO-driven simulations and facilitate the integration of the Geohazard DTC into downstream pre-operational workflows.

The current prototype includes several scenario modules operating at increasing levels of EO data exploitation: GEOMOD, for modelling EO-derived geodetic signals; FALL3D, for volcanic ash and SO₂ dispersion constrained by EO observations; GPUFLOW, for lava flow modelling based on EO-derived effusion rates and topography; and DAMSAT, for EO-based change and damage detection. Together, these modules act as building blocks for pre-operational services, empowering stakeholders to explore realistic emergency scenarios and assess potential mitigation and adaptation strategies. 

A central aspect of GET-it is the systematic integration of stakeholder requirements into the DTC design. A structured engagement process, based on questionnaires and direct interactions, involved a broad range of public and private stakeholders, including civil protection authorities, aviation and transport stakeholders, infrastructure managers, insurance companies, energy providers, and decision-makers. All the scenario models were considered relevant, with FALL3D emerging as the most requested service. The collected feedback directly informed the definition of the scenario modules and the organization of a dedicated Demonstration Day, focused on the validation of EO-driven what-if scenarios.

The Geohazard DTC has been demonstrated through representative multi-hazard use cases, including the 2018 Mount Etna eruption, the 2021 La Palma eruption, and the 2016 Central Italy earthquake sequence. Based on these activities, a medium- to long-term roadmap is being defined, focusing on enhanced EO-driven simulations, advanced decision-support tools, interoperable visualization interfaces, scalable services, and extension to additional geohazards, in alignment with other DTC initiatives and community-building actions. This roadmap aims to consolidate the Geohazard DTC as a sustainable pre-operational platform, ready for future operational uptake.

How to cite: Ganci, G. and Stramondo, S. and the GET-IT team: Advancing the EO-based Geohazard Digital Twin from Prototype to Pre-operational Platform, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17698, https://doi.org/10.5194/egusphere-egu26-17698, 2026.

Meteorological imagery occupies a special role in Earth observation: its high temporal frequency and broad spectral coverage are indispensable for weather forecasting and climate modelling, while its spatial resolution remains limited. Technological advances, however, are driving a trend toward higher spatial detail and open new application domains beyond traditional meteorology.

The Flexible Combined Imager (FCI) aboard Meteosat-12 represents the latest generation of geostationary weather sensors and images the entire Earth disk at 10-minute intervals. Its 16 spectral channels span the visible to longwave infrared and offer native spatial resolutions ranging from 2000 m down to 500 m — a configuration well suited to super-resolution techniques.

The AIDE project develops a methodology to increase the spatial resolution of all FCI channels to up to 500 m while preserving the radiometric integrity of the data. The approach employs a purpose-built deep-learning model that is augmented with an estimate of its own predictive uncertainty, thereby enabling safe downstream use of the enhanced products in demanding applications and quantitative analyses. The method is implemented as a demonstrator within the Destination Earth (DestinE) Data Lake, demonstrating that appropriately designed machine-learning approaches can be deployed reliably in critical operational contexts.

This contribution presents the concept and development of the method, summarizes promising project results, and discusses the limitations and potential of super-resolution approaches in Earth observation.

How to cite: Fessel, A. and Krummrich, D.: AIDE: Trusted Deep-Learning Super-Resolution for MTG FCI within Destination Earth, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18330, https://doi.org/10.5194/egusphere-egu26-18330, 2026.

Extreme heat is increasingly recognized as one of the most severe climate-related risks affecting urban populations, with disproportionate impacts on public health, energy systems, and vulnerable communities. As heatwaves intensify under climate change, cities require near-real time, high-resolution and actionable information to support early warning systems, preparedness, and long-term adaptation. Addressing this challenge at urban-local scale demands not only methodological innovation, but also robust digital infrastructures capable of delivering consistent and interoperable climate intelligence across regions.

Destination Earth (DestinE), a strategic initiative of the European Union, represents a transformative step in this direction by providing global, high-resolution climate and weather simulations, through Digital Twins of the Earth System. By coupling advanced numerical models, Earth Observation (EO) data, and high-performance computing, DestinE establishes a common backbone for next-generation climate services. However, translating these powerful datasets into locally relevant, operational products for cities remains a critical challenge.

DE_395-Urban Heat Health Forecasting (UHHF) project addresses this gap by demonstrating how DestinE Extremes Digital-Twin outputs can be transformed into urban-scale user-oriented heat-health indicators through the operational use of Machine Learning (ML). The project applies ML-based downscaling techniques to near-surface air temperature (T2m) and relative humidity (RH) forecasts, enhancing spatial resolution from kilometre-scale to approximately 200 m. These downscaled fields are subsequently used to derive human-biometeorological indicators such as the Universal Thermal Climate Index (UTCI) and Thermal Stress Duration (TSD), supporting health-oriented risk assessment.

The UHHF framework integrates DestinE atmospheric drivers with EO-derived and geospatial predictors describing urban form, land cover, vegetation, and topography, including Local Climate Zones. Quality-controlled crowdsourced observations from citizen weather stations are combined with WMO reference data to constrain and validate the ML models, ensuring robustness under both average and extreme conditions. The approach is being implemented across four climatically and socio-environmentally diverse Functional Urban Areas, e.g. Naples, Chicago, Santiago, and Cape Town, enabling a systematic evaluation of models across continents.

By building directly on DestinE and complementary European programmes led by ECMWF, ESA, and Copernicus, drawing on both their data assets and operational services, UHHF aims to illustrate how these can be leveraged to develop affordable, scalable, and reproducible urban-scale climate information and services. The project highlights the strategic importance of climate data platforms in bridging the gap between global simulations and local decision-making, contributing to the development of interoperable urban climate and health services aligned with European and international resilience frameworks.

How to cite: Girão, I., Paixão, J., Castro, M., Miranda, V., Souza Silva, F., Siemen, S., Di Napoli, C., and Oliveira, A. P.: Urban Heat Health Forecasting with Destination Earth: Leveraging Digital Twins and Machine Learning for Scalable Urban Climate Services, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19009, https://doi.org/10.5194/egusphere-egu26-19009, 2026.

The Global Fish Tracking System (GFTS) and Pangeo-Fish integrate biologging data with high-resolution environmental data in a digital-twin framework to address key challenges in marine conservation and fisheries management. Linking fish movement models with climate projections from Europe’s Destination Earth (DestinE) Climate Change Adaptation digital twin yields an evidence-based tool for decision support in habitat conservation and fisheries management under climate change. The implementation is built on the open-source Pangeo ecosystem and deployed on the DestinE platform.

Pangeo-Fish is an open-source package that ingests multiple biologging data types—including archival tags, pop-up satellite archival tags (PSATs), and acoustic telemetry detections. It processes time series observed by fish (e.g., depth, temperature, and light) together with geolocation constraints derived from external sources such as acoustic receiver networks and tag-based positioning. These heterogeneous observations are harmonised in a cloud-native workflow to support scalable track reconstruction and downstream habitat-relevant products.

Tracks are reconstructed using a Hidden Markov Model (HMM) geolocation approach that combines tag-recorded time series (e.g., depth, temperature, and light) with external geolocation constraints (e.g., acoustic detections), together with priors such as bathymetry and release/recapture information. Processing leverages cloud-native tools (Jupyter, Dask, Xarray) and chunked cloud-optimised storage (Zarr) for scalable analysis. A key design choice is the use of HEALPix as the base spatial grid and indexing scheme from ingestion to visualisation, enabling efficient path-likelihood evaluation on an equal-area, iso-latitude grid while avoiding distortive resampling. Environmental reference fields are primarily sourced from the Copernicus Marine Service, but the workflow can also ingest user-defined datasets. In addition, in situ observations (e.g., Argo float temperature) can be incorporated to represent uncertainty in the ocean physics fields used by the geolocation model and better account for model–data discrepancies.

The Pangeo-Fish workflow yields most-probable tracks and daily presence-probability maps, while GFTS supports aggregation of distributions for management-relevant analyses. GFTS intersects these outputs with climate projections from the DestinE Climate Change Adaptation Digital Twin to assess potential habitat exposure under climate change. GFTS has so far been demonstrated for Atlantic applications, while Pangeo-Fish has been extended to the Pacific Ocean, enabled by portable cloud-native processing and the availability of cloud-accessible datasets. As an early DestinE platform use case, this work illustrates how Earth system digital twins can be operationalised for reproducible, scalable biologging analytics to inform marine conservation and sustainable fisheries management.

How to cite: Odaka, T. E., Cap, E., Mazouni, Q., Hue, C., Delouis, J.-M., Woillez, M., Fouilloux, A., Ragan-Kelley, B., and Wiesmann, D.: Pangeo-Fish and the Global Fish Tracking System: Scaling Biologging Analytics with Earth System Digital Twins for evidence-based policy support, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19040, https://doi.org/10.5194/egusphere-egu26-19040, 2026.

Urban mobility and air quality are tightly coupled in cities: travel demand, network performance, and the spatial distribution of activities shape transport-related emissions and accessibility, and therefore underpin policy instruments such as low-emission zones, speed regulation, network reconfiguration, and land-use adjustments. Urban transport systems, however, operate within a broader set of constraints that extend beyond traffic and emissions management. Flooding and other water-related disruptions are a recurrent and increasingly relevant challenge for cities, with the potential to severely affect the transport network functionality. These challenges are often addressed through separate analytical tools, reflecting isolated approaches to mobility, environmental quality, and climate resilience topics. This fragmentation limits the ability of decision-support systems to effectively evaluate interventions across interrelated domains, motivating the need for modular and extensible services capable of integrating multiple processes within a single analytical framework.

CityNexus Pro, an Advanced Application Service (AAS) onboarded on Destination Earth, addresses this need through an operational and modular architecture designed to support integrated and cross-cutting scenarios analysis. The service builds on CityNexus, initially developed to support scenario-based assessment of mobility patterns, transport-related emissions, and air quality, and extends this baseline by incorporating an interoperable module for modelling mobility disruptions caused by flooding.

Within CityNexus Pro, urban mobility dynamics are represented by coupling data-driven origin–destination estimation with a traffic simulation engine, generating high-resolution spatio-temporal traffic flows. These flows are translated into emission estimates and linked to air quality models to quantitatively assess pollutant concentrations under alternative urban scenarios. Model assumptions, data sources, and parameterisations are explicitly documented to ensure transparency and reproducibility. Flood modelling outputs derived from hydrodynamic simulations based on the SFINCS model are mapped onto road network elements and incorporated into the mobility simulation chain, enabling dynamic modification of traffic conditions through configurable speed-reduction and road-closure thresholds.

The service, which already supports a range of policy-relevant scenarios—including low-emission zones implementation, speed limit changes, partial or full road inaccessibility, land-use reallocations, and shifts in mobility demand—is further complemented by compound scenario analysis in which flood hazards and mobility-related policy interventions are evaluated jointly. Users can configure flood scenarios by adjusting parameters such as precipitation timing and duration, river and sea discharge, and the presence of flood defences. In addition, CityNexus Pro is designed to integrate forecast products from the DestinE Digital Twin for Weather-Induced Extremes, enabling the impact assessment of extreme rainfall and flooding on mobility patterns and accessibility. 

CityNexus Pro is operational in Copenhagen, Seville, Bologna, and Aarhus, and is currently being deployed in Bucharest and Vitoria-Gasteiz. The service demonstrates how modular urban analytics on Destination Earth can be incrementally enhanced to address compound climate risks and support non-siloed, policy-relevant scenario analysis for urban resilience.

How to cite: Feliciotti, A., Lizama, A., Lemma, L., Ruess, A., Marconcini, M., Altenkirch, A., Stark, J., and Fratini, S.: Integrating Flood-Induced Mobility Disruption Modelling into CityNexus as a New Advanced Application Service on Destination Earth, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19782, https://doi.org/10.5194/egusphere-egu26-19782, 2026.

The satellite-based monitoring solution operated by GEO4A B.V. and powered by GeoVille GmbH is commercially called HARVIC. This abbreviation stands for “Harvest in control”. HARVIC offers comprehensive in-season crop growth monitoring and evaluation at the field level, tailored specifically for potato crops in central Europe, with the potential for global expansion. This service provides a detailed, continuous assessment of potato growth dynamics throughout the growing season, using high-resolution satellite (Sentinel-1 & Sentinel-2) and meteorological data as input to the crop growth module.

The main benefit of the HARVIC service is an EO data-supported transition of stakeholders in the potato industry towards digitalisation. Critical decisions can be made based on objective and timely data. HARVIC therefore identifies and evaluates critical growth stages, including emergence, vegetative growth, maturity, and senescence, ensuring that users receive timely updates and insights into potatoes' unique development cycle. The service also provides yield forecasting and quality information for each growing stage to improve logistics and reduce costs, as well as unnecessary greenhouse gas emissions.

The HARVIC service has been selected as a high-priority agricultural service for onboarding and implementation within the European Commission’s Destination Earth (DestinE) initiative. The basic service is published, and registered users of the DestinE community are eligible to take advantage of this unique solution for the potato industry. Furthermore, the consortium of GeoVille GmbH and GEO4A B.V. is about to include monitoring services to track and trace regenerative agricultural practices, such as cover cropping and tillage. In addition, the latest climate data (ECMWF) will be used to enhance HARVIC, potentially delineating future growing regions for potato cultivation in the coming decades.

How to cite: Pichler, J., Kolitzus, D., and Santbergen, P.: Satellite-based monitoring solution for the potato industry, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22126, https://doi.org/10.5194/egusphere-egu26-22126, 2026.

Desert locusts are known as the world’s most destructive migratory pest. A single swarm can count 80 million locusts, traveling up to 150 km daily and consuming the same amount of food as 35.000 people per day. They are cause of major mid-to-long-term impacts on the economy, quality of life and the environment. Climate change is amplifying the occurrence of such pests: the increase of extreme events such as cyclones creates ideal conditions for locust breeding.

The Desert Locust Monitoring Service (DLMS) is the new system for preventing upsurges of desert locusts across Eastern Africa, Southwest Asia and northwest of India. It leverages the power of satellite observations, model outputs, in-situ measurements and advanced AI techniques to monitor the locust threat, help mitigate crop damage, and safeguard essential food supplies.

Onboarded on the Destination Earth (DestinE) platform, the service consists of two layers: the Early-Stage Locust Appearance and the Locust Swarm Migration. The first layer relies on a customized Maxent model, a statistical approach widely used in species distribution modeling (SDM): it allows identifying areas with favorable environmental conditions for desert locust breeding. There are two available modes: the Safe Mode makes use of 50 days of climate data from ERA5-Land (soil water content, precipitation, and temperature) and normalized difference vegetation index (NDVI) from Sentinel-3 to generate daily probability forecasts; the Experimental Mode extends forecasts to two days thanks to the advanced projections of DestinE’s Weather Extremes Digital Twin. The variables used are: skin temperature, total precipitation and runoff.
The second layer forecasts adult locust swarm migration under biological and climatic conditions, accounting for their rapid and unpredictable long-distance movements. It takes as primary input the Early-Stage layer output, which provides initial location predictions under specific environmental conditions. Environmental variables, including the Leaf Area Index (LAI) from Copernicus, as well as wind velocity components and temperature from DestinE’s Climate Change Adaptation Digital Twin, inform the model of conditions that may trigger migration events. Swarm behavior is then represented using a stochastic model, which simulates an environment-biased random movement on a 2D lattice, generating batches of diverse potential scenarios. In this framework, locusts move based on environmental cues, including climate conditions and the availability of resources, such as vegetation. Finally, the model performs a statistical analysis across all generated scenarios to produce output maps estimating future locations of adult locusts and the size of their swarms.
Both Maxent and stochastic models were trained using a presence-only dataset provided by FAO’s Locust Watch. Prediction results have been validated with FAO data on desert locust activity and independent data provided by the International Center of Insect Physiology and Ecology (ICIPE).

Enabled by Destination Earth’s cutting-edge modeling and data infrastructure, the DLMS offers high-quality insights to anticipate risks, mitigate impacts and support the protection of crops, communities, and ecosystems in the most affected regions. Forecasting maps are available to all DestinE registered users, with some features reserved to users with upgraded access.

How to cite: Outmani, S., Grassi, A., Azami, W., Bellotto, K., and Houël, M.: The Desert Locust Monitoring Service: a new Destination Earth service for Environmental Pest prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22287, https://doi.org/10.5194/egusphere-egu26-22287, 2026.

Policy makers at local, national, and international levels are increasingly being required to make decisions that mitigate the effects of climate change on society and the economy. Earth Observations (EO) are already an important source of data to support such decisions, but this data represents only one aspect of the broader socio-technical systems that decision makers seek to influence. Policy effectiveness depends not only on environmental conditions, but also on household behaviour, technology adoption decisions, economic constraints, and feedbacks across scales. Capturing these dynamics requires modelling approaches that explicitly represent human decision-making alongside EO-derived inputs. This paper will present the results from the development of a scenario-planning digital twin (SPDT) designed to support decision-making processes related to local energy policies. The new SPDT will demonstrate how EO datasets can be integrated with multilevel agent-based models (MABMs) to enable specific scenarios to be used to support policy decisions.

The use case in this work enables policy makers to (i) model residential heating demand, (ii) test policy levers that might best encourage the uptake of low-carbon heating, and (iii) assess the implications for energy use and fuel poverty. Specifically, the MABM simulates hourly residential energy demand for space heating at the household level, accounting for building characteristics, policy levers, occupancy patterns, retail energy prices, and external ambient temperature. The MABM supports baseline demand estimation at fine spatial granularity (individual households), the assessment of new technologies (such as retrofit measures or heating controls/meters) and energy price variations, and counterfactual analysis (e.g., setpoint shifts, tariff changes, warm/cold snaps). Earth observation data is used to inform medium to long-term climate trends for the use case region. The project results presented in this paper have been developed to serve the city of Newcastle upon Tyne in the UK. We discuss results for Newcastle developed so far, and possible new future scenarios that could be developed using this type of methodology.

How to cite: Wagg, D. J., Malleson, N., Beltran, A., Tipuric, M., and Arribas-Bel, D.: Using Earth Observation Informed Agent-Based Models to build a Scenario Planning Digital Twin for Local Energy Policies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22325, https://doi.org/10.5194/egusphere-egu26-22325, 2026.

The UrbanSquare service within Destination Earth aims to deliver an operational digital twin of cities, enabling urban planners to explore and assess environmental processes and intervention scenarios. One key component of this digital twin is the representation of the urban heat island (UHI) effect. UrbanSquare relies on land surface temperature (LST) observations derived from thermal satellite imagery, primarily Landsat data resampled and distributed at 30 m spatial resolution.

However, actionable urban digital twins require both finer spatial detail and the ability to simulate “what-if” scenarios driven by land-cover change. In particular, UHI mitigation planning calls for high-resolution thermal information (well below 30 m) and dynamic coupling between land-cover configurations and surface temperature responses.

We present a three-stage framework for generating scenario-ready LST maps at 5-meter resolution.

In Stage 1, a Random Forest model upscales Landsat LST from 30 m to 5 m using Sentinel-2 spectral bands and indices (B2, B3, B4, B8, B11, B12, NDVI, albedo), solar geometry variables, and ERA5 meteorological predictors. Sentinel-2 data are harmonized to 30 m for model training, then applied at super-resolution (5 m) for inference. Cross-validation assesses predictive performance in the absence of in situ measurements.

Stage 2 applies pixel-wise linear regression between the super-resolved LST time series and synchronous ERA5 air temperature across a summer period. This normalization removes temporal and meteorological variability and enables LST generation for user-defined air-temperature scenarios, ensuring consistent thermal comparisons.

Stage 3 constructs a lookup table of thermal signatures for each land cover class. When users modify land cover, pixels are reassigned the corresponding thermal signature. A diffusion process accounts for lateral heat dispersion, producing delta-temperature maps, uncertainty layers, and decomposed contributions from different factors (vegetation, albedo increase, etc.).

Soon to be integrated into the UrbanSquare digital twin, this framework enables exploration, comparison, and quantification of UHI mitigation strategies, supporting evidence-based urban planning and climate adaptation decisions.

How to cite: Poupard, H., Fernandez, F., Gonzalez Fradejas, G., and Castel, F.: High-Resolution Thermal Mapping and Simulation Scenarios for Land Cover Intervention Planning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22710, https://doi.org/10.5194/egusphere-egu26-22710, 2026.

CALIFE (Certification of LIfestyle and Environment) is an operational service designed to make Earth observation–derived environmental information directly accessible to citizens. The service delivers hyper-local “quality of life” reports for any address, synthesising complex Earth observation (EO), in-situ and model data into a clear, visual and actionable format. CALIFE indicators cover key dimensions of everyday living conditions, including air quality, climate and weather comfort, water and vegetation resources, biodiversity and access to green spaces, mobility and accessibility, and exposure to selected climate hazards. Results are presented through intuitive scores, maps and short recommendations, allowing users to understand how their local environment influences well-being.

A core ambition of CALIFE is to address the general public rather than professional or expert users. The service targets non-specialists with no background in EO, climate science or geospatial data. CALIFE explores how highly technical datasets—such as Copernicus services and Destination Earth digital twins—can be transformed into information that is meaningful at the scale of daily life. By lowering technical barriers and focusing on usability, CALIFE represents an attempt to put Earth observation “into the hands of citizens”, fostering environmental awareness, transparency and informed decision-making at neighbourhood level.

Beyond technical and scientific challenges, CALIFE also serves as a laboratory for experimenting with sustainable business models for citizen-oriented EO services. The long-term viability of such services remains a key open question. CALIFE initially explored a micro-payment model for individual users, but has since evolved towards a hybrid approach: free access for the general public with controlled usage (limited number of reports per user), combined with paid bulk report generation for public authorities and private actors requiring large-scale analyses. This contribution will discuss both the service design choices and the lessons learned regarding sustainability, scalability and citizen engagement when deploying EO-based services for non-expert audiences.

How to cite: Castel, F., Lavigne, C., Semlal, R., Cauneau, M., and Pialot, L.: CALIFE, a service to discover the quality of life in your neighbourhood, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22715, https://doi.org/10.5194/egusphere-egu26-22715, 2026.

Earth science data are typically highly heterogeneous which leads to mixed determined inverse problems and poses challenges to extract process-level information. For example, ocean sediment cores from the International Ocean Discovery Program (IODP) contain hundreds of millions of measurements across multiple geophysical properties, but usable datasets are only 5-10% complete due to missing data. We present a semi-supervised variational autoencoder with masked encoding that simultaneously imputes missing measurements and predicts lithology, enabling more complete utilization of legacy IODP archives. We train a masked variational autoencoder on the LILY database (89 km of core, 34 million observations, 42 IODP missions) to learn joint distributions across bulk density, magnetic susceptibility, RGB reflectance, and natural gamma ray attenuation. The model uses selective masking during training to learn imputation strategies for missing modalities. Crucially, the learned latent representations are constrained to recover lithological labels from unseen cores without retraining. We demonstrate that the model both captures the nonlinearities contained in the training data and is able to reconstruct the test data (R2_avg=0.86) and that data lithology (AUC_avg=0.9), while also providing descriptive embedding vectors (ARI=0.2). Additionally, the underlying data contains strong non-linear relationships that are not captured by simpler models on reconstruction (e.g., a typical LASSO-based regression (R2=0.24)). Our work represents a step towards scalable cross-modal assimilation and representation of existing earth datasets.

How to cite: Aiken, J. M., Liu, D., Gilpin, W., and Becker, T.: A multi-modal semi-supervised model for ocean sediment lithology, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1492, https://doi.org/10.5194/egusphere-egu26-1492, 2026.

This presentation details the "Enhancement Digital Twin Application" initiative led by the United Nations Office for Outer Space Affairs (UNOOSA) and UN-SPIDER. Designed to support the United Nations' "Early Warnings for All" (EW4All) agenda, this project leverages cutting-edge space technologies to bolster disaster resilience in Small Island Developing States (SIDS). A primary challenge in the Pacific region is the significant "data gap" - specifically the lack of building footprints that include height information, which is critical for accurate disaster modeling. While global datasets exist, they often lack vertical data, and regional initiatives like PCRAFI have limited coverage. To bridge this gap, this project utilizes 30cm high-resolution satellite imagery combined with Deep Learning AI models to construct cost-effective 3D Digital Twins. The methodology employs advanced techniques, including NeRF and Gaussian Splatting, to generate models ranging from LOD1 (for GIS analysis) to LOD3 (for high-fidelity visualization). The core of the presentation focuses on the "Tonga Disaster Preparedness Platform," a pilot project implemented in 2024. This platform integrates these 3D geospatial models with real-time environmental data from IoT rain gauges and water-level sensors installed on the ground. This fusion enables precise, real-time simulations of sea-level rise and flood scenarios. A key innovation is the system's ability to optimize evacuation routes dynamically; by analyzing real-time flood depth data, the digital twin can identify safe passage corridors and update evacuation directions instantly, a capability that static hazard maps cannot provide. Finally, the presentation outlines the roadmap for expanding these capabilities to the Cook Islands and the Republic of Palau. It demonstrates how satellite-derived digital twins can revolutionize the entire Disaster Risk Management (DRM) cycle - spanning prevention, mitigation, response, and recovery - providing a scalable, data-driven framework for climate adaptation in vulnerable "big ocean" states.

How to cite: Takami, J.: Enhancement Digital Twin Application for Disaster Management by UNOOSA/UN-SPIDER, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1580, https://doi.org/10.5194/egusphere-egu26-1580, 2026.

Deep geological repositories rely on robust, transparent, and scientifically based safety concepts to ensure the long-term safety of radioactive waste. As safety cases become increasingly data-rich and computationally integrated, Digital Twins are emerging as a powerful tool to represent, test, and communicate the behavior of complex geosystems over geological timescales. A core requirement for such Digital Twins is the explicit quantification of parameter uncertainties and sensitivities, ensuring that the model is both reliable and efficient in reproducing key safety functions.

In this contribution, we introduce a workflow designed to assess uncertainties and sensitivities associated with radionuclide retention in geological host formations. Our approach combines geostatistical as well as geochemical simulation and global sensitivity analysis. Mineralogical heterogeneity is represented using geostatistical realizations generated through custom Python implementations of Markov-chain methods and truncated Gaussian random field simulations, producing spatially realistic mineral distributions. These mineralogical scenarios are then propagated through a geochemical modelling step using Geochemist Workbench, in which the distribution coefficient (K_d) is computed for each realization to quantify the effect of mineralogical and geochemical variability on uranium retention.

To identify the key indicators of variability, the workflow incorporates variance-based sensitivity analysis (SA) based on a custom Python toolbox. The SA reveals both first- and second-order effects, highlighting the influence of individual parameters on the resulting K_d values as well as pairwise parameter interactions. In almost all cases, the identified sensitivities and interactions can be explained by underlying chemical and physical processes. Additionally, this approach enables targeted dimensionality reduction, a critical step for constructing Digital Twins that maintain scientific robustness while remaining computationally tractable.

The workflow is presented for crystalline host rocks, where we focus on uranium retention within granitic systems governed by solid–liquid interactions: sorption, aqueous speciation, precipitation, and dissolution. A key advantage of our workflow is its modular structure. Each component, geostatistical simulation, geochemical modelling, and sensitivity analysis, can be independently adapted, extended, or replaced. This makes the framework readily transferable to other host rocks such as salt or clay, which exhibit fundamentally different retention mechanisms, as well as to other radionuclides with distinct sorption, solubility, or redox characteristics.

Our results highlight (i) the magnitude of uncertainty introduced by mineralogical heterogeneity, (ii) the non-linear sensitivity of uranium retention to coupled mineral–solution systems, and (iii) the potential to substantially reduce model complexity by focusing on a small subset of high-impact parameters. Overall, the workflow provides a structured and scalable method for quantifying uncertainties and identifying the parameters most relevant to long-term safety. In this way, it provides the essential, uncertainty-aware input data required for the generation of reliable and computationally efficient Digital Twins in geological disposal scenarios.

How to cite: Duckstein, A., Pospiech, S., Brendler, V., Bok, F., Tolosana-Delgado, R., Plischke, E., and Abdelhafiz, M.: Towards Digital Twins: Uncertainty and Sensitivity Analysis for Safety-Case Modelling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1624, https://doi.org/10.5194/egusphere-egu26-1624, 2026.

High-resolution, kilometer-scale information on regional climate impacts is critical for effective adaptation and mitigation strategies. The European Commission’s Destination Earth (DestinE) Climate Adaptation Digital Twin (Climate DT) aims to address this need; however, actionable impact assessments remain limited by incomplete representation of key Earth system components and their interactions. The Horizon Europe funded TerraDT project tackles these limitations by developing a state of the art Digital Twin focused on the cryosphere, land surface, aerosols, and their coupled processes, fully interoperable within the DestinE ecosystem.

TerraDT pursues three objectives: (1) build and deploy new Digital Twin Components (DTCs) to strengthen process realism and enable impact assessments; (2) deliver a modular, scalable, interoperable platform integrating advanced software, high-performance computing, and data workflows that can host physical models and Artificial Intelligence (AI)/Machine Learning (ML) emulators; and (3) foster user uptake through early engagement and a User centric Interface (UI).

In its first year, TerraDT achieved several milestones:

• Cryosphere: A prototype Land-Ice DTC was established by coupling Elmer/Ice with ICON climate model via YAC coupler, supported by curated glacier dynamics datasets. Development of the Sea-Ice DTC (FESIM) began in mid-2025, including YAC-mediated coupling and an AI sea-ice emulator capable of ~100-day to multi-year rollouts, producing smoother fields than physical models. 
• Land Surface: A prototype time-varying land use dataset was generated for ECland and ICON land surface models. 
• Aerosols: A simplified Aerosol DTC was tested, with integration into (open) Integrated Forecasting System (IFS). ML components were prototyped in HAM-LITE to capture advanced aerosol physics (e.g., hygroscopicity) at reduced computational cost.

Impact modelling advanced across multiple domains:

• Sea-ice: Assessments of ice season duration, severe condition probabilities.
• Forest: Integration of 3PG and Prebasso models, calibration across European ecosystems, ML emulation of Prebasso, and characterization of old-growth forests.
• Urban: A carbon-sequestration emulator validated in Helsinki, with planned extensions to Lisbon, Barcelona, Munich, Paris, and Zurich. Key data sets required are prepared in combination with ML methods, and will be applied to build advanced Urban impact models for assessing climate extremes.

Infrastructure and interoperability were strengthened through YAC based coupling (ICON-Energy Balance Firn Model-Elmer/Ice on LUMI and Levante Supercomputers), and Sea Ice DTC I/O plans were aligned with DestinE workflows. A map-based UI architecture was designed to expose high resolution impact assessments for decision support.

By advancing new DTCs, AI/ML emulators, and generic coupling interface, TerraDT is being developed for full integration into the DestinE framework, ensuring compatibility and enhancing the overall ecosystem’s capability to inform climate adaptation and mitigation strategies. This presentation will summarize first year progress, outline objectives, and present the roadmap toward fully coupled simulations, validation, and dissemination of impact indicators through TerraDT UI for policy and stakeholder communities.

How to cite: Devaraju, N., Kontkanen, J., Poutanen, J., Tonttila, J., Bockelmann, H., Schmidt, H., Koldunov, N., Klocke, D., Tourigny, E., Giuffrida, M., Kokkola, H., Zwinger, T., Acosta, M., Laakso, A., and Garavelli, S.: Developing a new Digital Twin for Destination Earth: Technical Progress of TerraDT in its First Year, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1780, https://doi.org/10.5194/egusphere-egu26-1780, 2026.

The long-term safety of Deep Geological Repositories (DGRs) requires rigorous assessments capable of predicting radionuclide transport over million year timescales. While the Digital Twin (DT) concept offers a robust framework for such assessments, the traditional requirement for bidirectional, real time communication is currently unfeasible due to the absence of active physical repositories. We propose a modular DT prototype application framework designed to evolve from a high fidelity simulation environment into a fully synchronized system as field data emerge.

At its core, this framework utilizes standardized data schemas to harmonize heterogeneous, site-specific field data from crystalline host rock including mineral composition, pore water chemistry, and surface properties. These standardized datasets are integrated via a specialized API into a modular orchestration pipeline that connects 1D and 2D fracture simulations with reactive transport codes such as PHAST, OpenGeoSys, and PFlotran. By containerizing these secondary physics models into Docker environments, the framework ensures high computational flexibility and reproducibility. This approach allows for the seamless integration of Machine Learning models and complex physics-based workflows while maintaining isolated execution environments.

Acknowledging the post-closure reality of a DGR where sensors may fail or lose power supply this framework prioritizes the characterization of source term evolution (radionuclide fluxes) through a "build fill close abandon" logic. Current  focus is on building features to establish resilient data formats and interface protocols to create a future proof foundation for geological safety. We demonstrate how containerization and robust interface design can transform divergent research projects into a unified, reproducible DT framework, applicable to any domain where long term predictive modeling is required despite limited real-time data.

How to cite: Ravichandran, S., Pospiech, S., Brendler, V., and Juckeland, G.: An Initial Digital Twin Architecture for Long-Term Radionuclide Transport Modeling in Deep Geological Repositories, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3381, https://doi.org/10.5194/egusphere-egu26-3381, 2026.

The EU CARMINE project (https://carmine-project.eu/) aims to support urban and surrounding metropolitan communities to become more climate resilient. The project focuses on heat, wildfires, flooding, pollution and drought across eight case study areas in Europe. Birmingham, located within the West Midlands Combined Authority (WMCA), serves as the UK case study area. High priority climate hazards for Birmingham are extreme heat, as well as pluvial flooding caused by extreme precipitation events. Increasing urban tree cover to alleviate these hazards could be a promising nature-based solution. However, a large amount of newly planted trees tends to wilt or die due to drought stress.

To assist the Council and community volunteers in maintaining the young trees during drought events, we are developing a digital twin framework to identify when and where young trees across Birmingham need watering. Common indicators include daily plant available water and the level of drought/wetness for the last few weeks. These indicators are based on soil moisture content, usually at different depths. Unfortunately, such measurements are sparse or absent in urban areas. We use the Joint UK Land-Environment Simulator (JULES) model forced by the UK weather forecasting model UKV at a spatial resolution of 1.5km, to estimate soil moisture content. Using machine learning techniques, we emulate JULES outputs to provide soil moisture estimations in a faster, more efficient and more flexible way. Platforms developed through the CARMINE project allow us to communicate the need for watering to interested communities. This approach is an important step to support communities and city authorities to improve the management of urban trees and resilience of cities to climate hazards like heat waves and flooding.

This approach explores how a digital twin, combined with an emulation of JULES soil moisture using ML techniques, could provide drought information for young trees more efficiently. It has the potential to scale beyond the case study area of Birmingham and transfer the digital twin to other urban areas.

How to cite: Urhausen, S., Hemming, D., Brettle, D., Ferranti, E., and Greenham, S.: Developing a Digital Twin to support the resilience of young trees to drought, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3775, https://doi.org/10.5194/egusphere-egu26-3775, 2026.

This study developed a Geospatial Artificial Intelligence (Geo-AI)–based framework to estimate and visualize the three-dimensional (3-D) distribution of ultrafine particles (PM₀.₁) and associated population exposure across Taichung City, Taiwan. An unmanned aerial vehicle (UAV) platform equipped with a P-Trak Ultrafine Particle Counter was deployed to collect high-resolution 3-D PM₀.₁ concentration data across varying altitudes and land-use types. These 3-D PM₀.₁ data were integrated with multi-source geospatial datasets, including 3-D building models, meteorological variables, and emission inventories. The SHapley Additive exPlanations (SHAP) method was then employed to identify key predictors for machine-learning modeling. The optimized model was applied to map the continuous 3-D pollution field and used to estimate and visualize population exposure for each floor level. The resulting Geo-AI model achieved strong predictive performance, with R² values of 0.95 for training and above 0.85 for validation, demonstrating high robustness and predictive capability. Visualizations reveal a nonlinear vertical structure of PM₀.₁ in 3-D space, characterized by near-ground peaks in industrial and traffic zones alongside persistent localized hotspots at mid-to-high elevations. Population exposure assessments highlighted that, despite lower concentrations at higher elevations, the total exposure burden remains significant in mid-to-high-rise residential buildings due to higher population density. This research presents an advanced framework for assessing 3-D air pollution exposure risks in dense urban environments, demonstrating the potential of Digital Twin technologies in supporting air quality management and public health decision-making.

How to cite: Hsu, C.-W., Su, J.-J., Yang, R.-Z., Wijaya, C., Chen, Y.-C., Lung, S.-C. C., Hsiao, T.-C., Lin, C.-H., and Wu, C.-D.: Geo-AI-Based Assessment of 3-D Ultrafine Particle Distribution and Population Exposure: A Digital Twin Approach in Taichung, Taiwan, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5493, https://doi.org/10.5194/egusphere-egu26-5493, 2026.

Sediment transport is a fundamental process shaping landscapes and posing significant hazards in mountainous regions. However, traditional field monitoring and simulation approaches, such as grain size sampling and numerical modeling, are often costly and time consuming. Recent advances in physics-based models and machine learning have substantially improved spatial and temporal resolution. These achievements enable the development of digital twins to explore what-if scenarios and to better understand the dynamic processes involved.

In this work, we combine the probabilistic sediment cascade model (SedCas) with the machine learning–based event detection model (Flow-Alert) to develop a digital twin of a catchment. The former relies solely on climate forcing to simulate sediment dynamics, whereas the latter uses seismic signals to identify extreme sediment transport events, such as debris flows. We address three key questions. First, how to design a digital twin framework that captures the physical components of sediment transport, including erosion on hillslopes, hillslope to channel transfer, and channel transport to the catchment outlet, at hourly and even sub hourly temporal resolution. Second, how to fuse predictions from the physics-based model SedCas and the machine learning based model Flow-Alert to merge and balance the strengths of these two modeling approaches. Third, how to reduce uncertainty when translating insights from the virtual entity back to the physical entity. We demonstrate that the digital twin framework enables potential users, such as governmental agencies and local stakeholders, to explore what if scenarios and better understand how climate change and human interventions influence sediment transport dynamics.

How to cite: Zhou, Q., Tang, H., Hirschberg, J., and Walter, F.: stTwin: A digital twin framework for catchment-scale sediments transport, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6409, https://doi.org/10.5194/egusphere-egu26-6409, 2026.

Digital twins as complex decision-making systems are increasingly used in climate adaptation and sustainability planning. However, most of the applications currently available remain largely sector-oriented, thus limiting their capacity to capture interactions between different domains with multiple interrelated indicator systems. This constraint is particularly applicable at the neighborhood scale, where planning interventions are applied, and trade-offs between competing objectives become most visible.  

This work introduces the prototype of a neighborhood-scale digital twin system, which is designed to support integrated, scenario-based analysis of urban ecology and energy systems. The digital twin implemented in the post-war residential neighbourhood of Twekkelerveld, Enschede, the Netherlands, attempts to solve major issues, such as the ageing building stock, limited green infrastructure, and relatively high energy demand. The framework incorporates open and municipal datasets, including tree inventories, green spaces, urban heat potential, building geometry, energy-use intensity, energy estimation, solar electricity potential, and carbon footprint. The system explicitly represents interactions among ecological and energy interventions at the neighbourhood level, unlike the existing tools.

The digital twin is designed to facilitate interactive "what-if" exploration of typical urban interventions across multiple domains. Ecological scenarios, such as tree planting strategies and green facade deployment, enable users to assess the impacts on greenness, urban heat mitigation, carbon sequestration, and investment costs. Energy scenarios include building insulation improvements, rooftop solar deployment, heat pump transitions, and local energy sharing, measured by the indicators on the level of the neighborhood and buildings. The interrelation module explicitly connects the ecological and energy measures, which allow the comparison of the combined effects on cooling, energy demand, emissions, and overall performance.

Instead of making sustainability planning a one-sector endeavour, the prototype assists in the exploration of options: what changes, what gets better, and what gets worse when various measures are combined. Presenting baseline and scenario outcomes side by side makes trade-offs clearer across ecological, energy, and environmental indicators. The work shows how neighbourhood-scale digital twins can operationalise multi-domain data and scenario logic in a form that is usable by urban planners, municipalities, and local decision-makers. This complements Earth system-scale digital twins, which are centered around the local level where interventions are discussed and implemented.

How to cite: Athid, R., Koeva, Dr. M. N., and Nourian, Dr. P.: Digital Twin For Improvement of The Sustainability of Neighbourhoods Through Scenario Planning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7450, https://doi.org/10.5194/egusphere-egu26-7450, 2026.

How to cite: Exarchou, E., Rodriguez Pinilla, M., Martin Gomez, V., Benitez Benavides, M., Senande Rivera, M., Bueso, D., Baladima, F., Canaleta, G., Borràs, M., Toli, E., and Koltsida, P.: A Digital Twin for Wildfire risk adaptation planning: DT-WILDFIRE, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7482, https://doi.org/10.5194/egusphere-egu26-7482, 2026.

How to cite: Tomami, K., Okamoto, A., and Omori, T.: Depth super-resolution of rock CT images based on latent diffusion models by deep learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9426, https://doi.org/10.5194/egusphere-egu26-9426, 2026.

The development of digital twins for subsurface applications faces several challenges, in this contribution we are focusing on the issue of providing near real-time predictions for numerical multi-physics applications describe by partial differential equations. Even when fronted against state-of-the-art high-performance computing infrastructures, conventional multi-physics simulations are not real-time compatible because of their huge computational demand. At the same time, they are subject to uncertainties from, for instance, the geometry, material properties, and boundary conditions.

To address the computational demand, we introduce the usage of surrogate models. Surrogate models comprise data-driven and physics-based approaches. While data-driven techniques, such as neural-networks, well capture complex system responses, they typically lack interpretability, hindering the degree of reliability of the model outcomes. This, in turn, poses challenges for the integration into digital twins especially in applications where risks need to be assessed. In contrast, physics-based approaches are fully interpretable, but often limited to elliptic and parabolic partial differential equations. Hence, they cannot capture the full complexity of the systems dynamics. To overcome the limitations of both data-driven and physics-based techniques, we introduce a hybrid approach namely the non-intrusive reduced basis method within the class of projection-based model order reduction techniques.

In this contribution, we demonstrate for a geothermal case study how this interpretable physics-based AI method can be used to reliably and efficiently accelerate the high-fidelity numerical multi-physics simulations. Furthermore, we illustrate their integration into a Bayesian uncertainty quantification framework, including hierarchical approaches. At last, we discuss possibilities to extend the aforementioned approaches to allow for a continuous integration of observational data.

How to cite: Degen, D., Gruzdeva, Y., Hayek, N., Faber, M., Siegel, C., and Cacace, M.: Perspective of Interpretable Physics-Based AI method for Digital Twins of Geosystems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9537, https://doi.org/10.5194/egusphere-egu26-9537, 2026.

Wildfires are a major type of disaster and challenge for economic prosperity, public health and safety around the globe. Decision Intelligence, particularly AI based scenario analysis, can make a significant difference [1] in disaster mitigation efforts. Data-driven methods have shown promise in various downstream applications [2]. Still, reference data remains a significant bottleneck across domains such as fire behaviour modeling.

We develop three data-driven decision intelligence tools: a novel machine learning based fire spread model, a fire break placement recommender, and triage decision support.
We make use of data from OroraTech’s global near-real-time fire monitoring network, which provides hotspot data from both public and proprietary satellites, in addition to burned area products.
We have created a novel dataset with thousands of fires from the US, Chile and Europe between 2022-2025. We enriched the thermal hotspot-based fire perimeters with a variety of EO (land cover, soil moisture, elevation, previously burned area, vegetation index) and non-EO (wind, temperature, relative humidity, dew point, and precipitation) data.

With this dataset, we train fire spread prediction models based on leading DL architectures. Graph Neural Networks (GNN) are particularly promising, since they have excelled in related domains such as weather forecasting [3], and showed promising spatial generalization properties for fire spread [4]. To mitigate uneven satellite overpass intervals, we treat the time gap between input-target images as an additional learning signal.
A major hurdle in the operational use of fire intelligence tools is a lack of user trust. Therefore, we incorporate explainability metrics in all three of key contributions.

The use of fire breaks - creating “barriers” of non-burnable materials to prevent fires from spreading - is a significant tactic in wildfire management. Scenario analysis tools are essential to inform the placement of fire breaks. Despite recent progress, significant challenges remain in this domain, such as reliance on basic fire spread simulators, and a complex action space for fire break placement [1]. We aim to close this gap by coupling our improved fire spread model combined with reinforcement learning, a promising approach pioneered in a recent case study [1] for fire break recommendations.

In conclusion, we present a novel fire dataset and operational tools for global, real-time fire spread modeling and  firebreak placement supporting wildfire management worldwide. 

References

[1]Murray,L.,Castillo,T.,Carrasco,J.,Weintraub,A.,Weber,R.,deDiego,I.M.,...&GarcíaGonzalo,J.(2024).Advancing Forest Fire Prevention: Deep Reinforcement Learning for Effective Firebreak Placement.arXiv preprint arXiv:2404.08523.

[2]Bot,K.,&Borges,J.G.(2022).A systematic review of applications of machine learning techniques for wildfire management decision support.Inventions,7(1),15.

[3]Lam,R.,Sanchez-Gonzalez,A.,Willson,M.,Wirnsberger,P.,Fortunato,M.,Alet,F.,...&Battaglia,P.(2023).Learning skillful medium-range global weather forecasting.Science,382(6677),1416-1421.

[4]Rösch,M.,Nolde,M.,Ullmann,T.,&Riedlinger,T.(2024).Data-Driven Wildfire Spread Modeling of European Wildfires Using a Spatiotemporal Graph Neural Network.Fire,7(6),207.

How to cite: Laux, D., Wahbe, J., Rovó, D., Pratik, P., Pörtge, V., Liesenhoff, L., and Gottfriedsen, J.: FireAID - Real-time Wildfire Spread Modeling with Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9595, https://doi.org/10.5194/egusphere-egu26-9595, 2026.

Geothermal energy plays a key role in energy transition by offering a clean baseload alternative to fossil fuels for space heating. Long-term geothermal production is subject to inherent uncertainty due to the heterogeneity of geological formations that host the geothermal resource, and the limited data available to characterize and quantify these heterogeneities. It is insufficient to explore and quantify such uncertainty based on a single concept or interpretational scenario. The TU Delft campus geothermal project has been initiated to provide a dedicated research environment with the vision to scale-up the deployment of geothermal energy as well as providing and storing heat for the TU Delft campus. Inspired by the reservoir that hosts the geothermal resource at TU Delft - a channelised fluvial system - we are presenting a framework of an open-source digital twin for geothermal reservoirs that aims to integrate geological scenario modelling, production simulation, uncertainty analysis, and data assimilation to mitigate operational risks, reduce maintenance costs, extend reservoir longevity, and enhance the overall sustainability for geothermal production.

We propose a scenario-based geological modelling approach using Rapid Reservoir Modelling (RRM), in which channelised fluvial layer templates are stacked and constrained by facies information along well trajectories. Multiple geological scenarios with distinct channel distributions are generated. Heterogeneous petrophysical properties are then assigned to different facies in the reservoir models. Uncertainties in both, reservoir architecture and petrophysical properties, are captured. The flow and thermal simulations are performed with the open-source Delft Advanced Research Terra Simulator (open-DARTS), and production uncertainty is quantified by evaluating the impact of reservoir architectures and petrophysical heterogeneities. The Ensemble Smoother with Multiple Data Assimilation (ESMDA) is then applied across these scenarios to constrain production and reservoir forecasts using well temperature and pressure observations, tracer tests, and related monitoring data. Scenarios that fail to reproduce the observations after data assimilation are falsified, while data-worth analysis is conducted on the remaining plausible scenarios to provide a reliable evaluation of data acquisition strategies and identify the most cost-effective options for reliable assessment of geothermal production.

Our digital-twin framework enables us to explore a broader range of geological uncertainties and constrains production uncertainties, thereby enabling a more reliable assessment of geothermal reservoir performance and production forecasts, both of which are essential for optimizing operational strategies and supporting informed decision-making for geothermal systems.

How to cite: Song, G., Voskov, D., Abels, H. A., Vardon, P. J., and Geiger, S.: Towards a digital twin for modelling geothermal reservoirs in channelised fluvial systems , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10068, https://doi.org/10.5194/egusphere-egu26-10068, 2026.

Digital Twins of the Earth are required to represent future scenario- and trajectory-based hazards that obey physical laws and realistic dynamics in an interpretable and actionable manner, understandable not only by experts but also by non-expert stakeholders and local authorities, to support efficient decision-making, adaptation planning, and emergency management. Machine learning has substantially advanced generating landslide susceptibility maps (LSM). However, LSMs typically provide static, abstract, expert-oriented snapshots that are difficult for non-expert audiences to interpret and are poorly aligned with the interactive, immersive visualization needs of Digital Twin and Augmented Reality (AR)/Virtual Reality (VR) environments, thereby limiting their effectiveness for anticipatory risk communication and decision support.

We present a physics-aware generative framework that transforms predictive landslide modeling into photorealistic satellite imagery of future events, enabling intuitive “what-if” hazard exploration within Digital Twin architectures.

Our approach integrates Landslide Physics-Aware Neural Networks (LPANNs) with conditional Generative Adversarial Networks (GANs) to generate synthetic, post-event satellite images. These generate synthetic images conditioned on multi-attribute probability maps (physics-informed predictions) resulting from embedding geotechnical, hydrological, geomorphological, and geometric constraints, ensuring physical plausibility. Our developed conditional GAN is trained based on pre- and post-event real images, with annotated landslide areas. Different supervised and self-supervised deep learning are used for large-scale landslide detection.

By conditioning generative part of the approach on physics-informed predictions, the proposed Digital Twin component mitigates hallucinations typical of generative AI and synthetic images and trustworthy hazard visualizations. The resulting synthetic imagery is scenario-consistent and bridges the gap between numerical susceptibility outputs and human-centered decision support, enhancing interpretability for policymakers, emergency managers, and non-expert stakeholders.

How to cite: Ghorbanzadeh, O. and Crivellari, A.: From Physics-Aware AI to Digital Twins: Generating Photorealistic Satellite Imagery of Future Landslides for Predictive Hazard Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10764, https://doi.org/10.5194/egusphere-egu26-10764, 2026.

WBGeo (WorkBench for Digital Geosystems) aims at automating the workflow from geological data integration to structural modeling, mesh generation, numerical simulation, and visualization. The framework is designed as a collaborative project, enabling the systematic and reproducible development of geoscientific models while reducing manual intervention across the entire modeling pipeline.

One of the core components of WBGeo is the generation of computational meshes tailored to complex geoscientific workflows. The framework supports three mesh representations: implicit structured meshes, explicit structured meshes, and explicit unstructured meshes. This flexible design allows users to select an appropriate meshing strategy based on model complexity, data availability, and computational requirements.

Implicit structured meshes are generated from volumetric structural models in which lithological information is defined on a regular grid. The meshing procedure operates directly on the implicit representation of the structural geological model and produces a structured hexahedral mesh suitable for numerical simulations based on finite element or finite volume/difference methods.

For explicit structured meshes, vertices are first extracted directly from the geological surfaces as provided by the structural model. Each geological layer is first discretized using a uniform, user-defined number of interpolated points to ensure consistent lateral resolution across all layers. Subsequently, vertical refinement between adjacent layers is performed using a user-defined number of subdivisions, allowing controlled resolution along the depth direction. To preserve mesh quality and avoid numerical instabilities, the minimum vertical distance between corresponding points in adjacent layers is evaluated using a user-defined threshold. If this distance falls below the specified limit, one of the points is adjusted vertically by a predefined amount to enforce the minimum separation. Following this correction step, hexahedral elements are constructed, resulting in a structured mesh suitable for efficient numerical simulations.

For explicit unstructured meshes, vertices are obtained directly from the structural model geometry. The surfaces are then interpolated and discretized, and the resulting geometry is passed to the Gmsh Python API for mesh generation. After determining intersections between surfaces and performing geometric fragmentation, tetrahedral elements are generated. One of the main components of unstructured meshes in the workflow is the inclusion of fault planes, and engineering objects such as wells, mining shafts, point sources, or additional internal planes, which are difficult to represent within a structured mesh framework.

By supporting both structured and unstructured meshing strategies within a unified workflow, WBGeo enables users to balance computational efficiency and geometric complexity while maintaining reproducibility and consistency across geosystem modeling applications. The generated meshes can be exported to different formats such as exodus, abaqus, feflow to be used by different commercial and open-source simulation packages.

How to cite: Cacace, M., Baes, M., von Harten, J., Lüpges, A., Degen, D., Niederau, J., Rolf, T., Scheck-Wenderoth, M., Wellmann, F., Rumpe, B., Koltzer, N., and Virgo, S.: WBGeo: An Automated Framework for Geosystem Modeling with Advanced Mesh Generation Capabilities, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10926, https://doi.org/10.5194/egusphere-egu26-10926, 2026.

Accurate and efficient modelling of geothermal reservoirs is important for sustainable energy production and for the reliable assessment of operational risks. Predicting thermo-hydraulic (TH) system evolution under varying injection and production scenarios remains computationally challenging, particularly when physical knowledge of the subsurface system is incomplete and observational data are sparse. High-fidelity finite-element simulators are typically used to provide physics-based predictions of coupled flow and heat transport governed by complex partial-differential equations (PDEs). Such full-order simulations are, however, often prohibitively expensive for real-time forecasting, which is essential, for instance, in the context of digital twins.

Physics-based machine-learning (PBML) approaches, such as the non-intrusive reduced basis (NIRB) method address this challenge by constructing physics-consistent surrogate models that project full-order simulation outputs onto a low-dimensional subspace learned from representative snapshots. By retaining only the dominantbasis functions, the NIRB surrogate enables orders-of-magnitude speedup in parametric predictions while staying consistent with the physical transport mechanisms and structural assumptions on fracture networks encoded in the full-order model. Despite these advantages, classical NIRB surrogates are intrinsically limited to the physical regimes represented by the governing PDEs, and consequently by the training simulations. If the surrogate does not fully capture the observed system behaviour, it is important to detect and adapt to missing or misrepresented local physics revealed by observational data, such as unmodeled convective heat transport or flow channelling arising from fracture activation.

To address this need, we propose a complementary residual-learning framework that augments a baseline NIRB surrogate with parameter-to-state maps of residual temperature and pressure fields learned by Kolmogorov-Arnold Networks (KANs). The residual, defined as the difference between observed data (or a synthetic reference solution) and the NIRB model prediction, is interpreted as a proxy for missing or misrepresented physics not explicitly captured by the baseline model. KANs represent mappings as sums of learned univariate functions and provide explicit access to the functional structure of parameter dependence. Thereby, KANs could act as interpretable discrepancy models by learning the residual between observations and NIRB predictions. By analysing the dominant functional families emerging in the learned residual, such as linear dependence characteristic of conduction-dominated regimes or exponential dependence associated with convection, KANs can provide diagnostic insight into missing thermo-hydraulic processes and their relevance across parameter regimes.

We validate the proposed approach synthetically by comparing a conduction-only NIRB surrogate against synthetic reference observations generated with an advection–diffusion model. We expect that KAN-based residual learning both improves predictive accuracy and reveals clear functional signatures of missing convective physics, even when only pointwise information is available. As an outlook, we aim to apply this workflow to real geothermal case studies, where sparse temperature and pressure measurements are available at well locations. In such settings, functional-family learning of residuals offers a promising pathway to improve surrogate predictions and to enhance the physical interpretability of geothermal systems, ultimately supporting more reliable assessments of reservoir behaviour.

How to cite: Faber, M., Cacace, M., and Degen, D.: Detecting Unrepresented Physics in Hybrid Machine Learning Surrogates of Geothermal Systems using Kolmogorov-Arnold Networks , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11263, https://doi.org/10.5194/egusphere-egu26-11263, 2026.

Deep aquifers offer significant potential for diverse energy and storage applications, these opportunities also will require synergistic multi-user subsurface management. To maximize these resources, operators require flexible modeling tools capable of rapidly evaluating how independent but concurrent projects might interact hydraulically over time. Traditional grid-based numerical models are robust but can be computationally demanding when rapid scenario testing is required across large, heterogeneous regions. We propose a modular Physics-Informed Neural Network (PINN) framework designed to provide a flexible, faster alternative for evaluating regional pressure interference between co-located subsurface activities.

Our proposed architecture treats the aquifer as a continuous volumetric field. We define injection and extraction points as dynamic operational conditions (e.g., transient rate or pressure constraints) that can be positioned anywhere in the domain. The neural network is trained to satisfy the 3D transient diffusivity equation, learning to map the relationship between these sources and the resulting pressure field without relying on fixed meshes We address this by introducing a  "modular" architecture: by training separate sub-networks for each activity type, we aim to mathematically isolate or "de-mix" the pressure contribution of specific projects from the total regional signal.

This research focuses on a case study in the Campine Basin (Belgium). We are developing the framework to infer effective aquifer properties from sparse historical monitoring data and to simulate interference patterns specifically between gas storage and geothermal operations. The expected outcome is a spatial scenario analysis tool that allows future users to dynamically test new project locations and optimize setback distances within a Subsurface Digital Twin environment. By decoupling the geological parameterization from specific well locations, we aim to provide a scalable engine that supports adaptive planning and de-risks decision-making in multi-activity aquifers.

How to cite: Rodriguez, J. D., Piessens, K., and Welkenhuysen, K.: A Modular Physics-Informed Neural Network Framework for Quantifying Pressure Interference Between Concurrent Deep Subsurface Activities, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11702, https://doi.org/10.5194/egusphere-egu26-11702, 2026.

In recent years, advancement in computational infrastructures has made possible to start exploiting the implementation and use of digital twins in climate science. A growing number of studies and prototypes have already appeared, aiming at modelling single or multiple components of the Earth system. Among them, it is worth mentioning the European Commission's Destination Earth initiative with the ambition of realizing a digital replica of the Earth. While the development of digital twins seems straightforward and is proceeding at fast pace, there are still some key conceptual issues and challenges to overcome and go beyond classic numerical models and digital shadows. Realising a continuous bidirectional data flow between the virtual system and the real one is among them. Together with innovative approaches in data assimilation and the integration of physics-consistent machine learning, there is the need to conceptualize what a continuous data loop means at time scales covering the coming years and decades. Furthermore, the need to address the "human-in-the-loop" requirement remains central to allow for actionable "what-if" scenario testing. In this contribution, we discuss these open issues as well as the minimum requirements twins should have. We conclude by proposing pathways to fulfil the ambition of having a digital twin of the Earth system. 

How to cite: Toreti, A., Hrast Essenfelder, A., and Lucarini, V.: Digital twins in climate science: challenges and opportunities , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13008, https://doi.org/10.5194/egusphere-egu26-13008, 2026.

Disaster risk management (DRM) faces increasing challenges due to urbanisation, environmental degradation, and the growing complexity of interacting hazards. Digital Twins (DTs), defined as digital representations of physical systems connected through continuous data exchange, have gained attention for their potential to support monitoring, simulation, and decision-making. However, their application to disaster contexts remains limited, as many DT implementations depend on uninterrupted automated data streams, predefined control mechanisms, and automated interventions that are often unavailable or impractical during disasters.

In this study, the Digital Risk Twin (DRT) is introduced as a paradigm specifically designed for DRM. The DRT extends DT concepts by integrating automated and manual data collection methods, such as IoT, remote sensing, surveys and field observations, while incorporating human-in-the-loop decision-making for flexible and effective interventions, maintaining real-time virtual simulations, and addressing disaster scenario challenges. To demonstrate its practical relevance, an example of how a DRT can be conceptualised for a multi-hazard response case study is formulated, illustrating how DRT can support effective DRM.

The DRT integrates diverse data sources such as remote sensing, in situ observations, field surveys, and community-based reporting, while supporting both automated analysis and expert-driven interpretation. A defining feature of the framework is the explicit inclusion of human decision-making within the digital representation. Rather than aiming for full automation, the DRT enables iterative interaction between digital models and stakeholders, supporting context-aware decisions under uncertainty. This is particularly important in disaster situations where data gaps, infrastructure damage, and rapidly changing conditions constrain the effectiveness of purely automated systems.

Digital Risk Twins represent a conceptual advancement over original Digital Twins by addressing the socio-technical nature of disaster risk. The proposed framework and multi-hazard conceptualisation provide a foundation for future operational implementations, with the potential to strengthen adaptive capacity and resilience to cascading and compound hazards.

How to cite: Ghaffarian, S.: Digital Risk Twins: The Next Generation of Digital Twins for Complex Disaster Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14599, https://doi.org/10.5194/egusphere-egu26-14599, 2026.

Accurately representing three-dimensional (3D) canopy structure is essential for Earth System Models (ESMs) and radiative transfer schemes that link vegetation to climate–carbon feedback. Leaf area density (LAD) and related structural metrics are widely retrieved from remote sensing using Beer–Lambert (BL) transmittance inversions, yet these approaches commonly assume randomly distributed foliage and woody material. In real canopies, plant material is spatially aggregated (clumped), violating random mixing and introducing systematic LAD bias. Although clumping has been corrected using canopy or crown scale clumping indices (CI), voxel-based LAD retrievals from terrestrial laser scanning (TLS) and other 3D sensing approaches require clumping information that is defined at the same spatial scale as the inversion. The lack of a physically grounded voxel-resolved CI remains a key methodological gap, particularly for dense and heterogeneous canopy regions.

Here, we develop a voxel-scale effective reference clumping index (CI_ref) retrieval method that is structurally consistent with voxel-based BL retrievals. We used digital twin 3D tree meshes from the RAMI-V benchmark forest scenes, spanning six contrasting crown forms and six leaf inclination angle distribution (LIAD) variants (36 canopy geometries). Each tree was partitioned into regular voxel grids at four sizes (0.2, 0.5, 1.0, and 2.0 m). Within each voxel, we performed multi-directional (18 bin viewing angle) ray tracing on every voxel-clipped mesh to directly quantify within-voxel gap probability, leaf projection function G(θ), and path-length statistics required for transmittance-based LAD inference. Directional CI estimates were derived for each viewing angle and then aggregated through a hierarchical pooling strategy that reduces sampling noise and directional variability (all angles → azimuth pooled → zenith-pooled). This procedure yields a single, robust CI_ref per voxel that is independent of viewing angle and suitable as a reference label for operational LAD retrieval algorithm development from LiDAR data.

We then quantified the practical impact of voxel-scale clumping correction on BL LAD retrieval using simulated TLS point clouds. LAD was estimated per voxel under two assumptions: (i) the conventional random-foliage case (CI = 1) and (ii) clumping-corrected inversion using CI_ref. Across all crown forms, LIAD variants, and voxel sizes, the CI = 1 assumption produced predominantly negative LAD errors relative to mesh-derived reference LAD, consistent with systematic underestimation when clumping is ignored. Incorporating CI_ref shifted LAD errors toward zero and improved agreement, evidenced by reduced bias and normalized RMSE. Improvements were most pronounced for planophile canopies, where directional foliage aggregation is strongest and for coarser voxel sizes (1.0–2.0 m), where greater within-voxel heterogeneity amplifies departures from random mixing, demonstrating that clumping-induced bias is strongly scale dependent.

These results provide practical recommendations for 3D canopy modelling: specifically, that voxel-scale clumping correction becomes increasingly essential as voxel size increases, especially when within-voxel heterogeneity grows. The proposed CI_ref framework strengthens scale consistency between local canopy structure and voxel-based radiative transfer, enabling unbiased LAD retrievals and providing physically grounded labels for future deep learning model-based CI prediction from TLS point clouds.

How to cite: Aryal, R. R., Devereux, T., Rivory, J., Eaton, G., Phinn, S., and Woodgate, W.: Digital twin–based voxel-scale clumping index (CI) improves leaf area density (LAD) retrieval from simulated terrestrial laser scanning (TLS), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15220, https://doi.org/10.5194/egusphere-egu26-15220, 2026.

Recently, the active Eifel volcanic region has received increasing interest due to the occurrence of deep low-frequency earthquakes, often interpreted as a sign of rising volatiles in the crust. Additionally, recent tomographic models have resolved vertically inclined low-velocity anomalies beneath the Laacher See volcano, which may indicate enhanced fluid ascent. These observations raise the question of whether volcanic activity in the region is increasing and whether such activity may be beneficial for geothermal exploration.

To address these questions, the DEGREE project is developing a digital laboratory that enhances predictive capabilities by combining geophysical data with geological and numerical models. The laboratory includes workflows that couple data assimilation, geological modeling, and numerical simulations into a single process. A key challenge is the propagation of uncertainties in the input data and parameters through the entire workflow. This allows us to obtain quantitive uncertainties for derived quantities to support decision making.

The foundation of the laboratory is a collection of diverse datasets compiled during the project. An extensive seismic dataset acquired by the Eifel Large-N network, deployed between September 2023 and September 2024, is used to investigate subsurface structure and active geodynamic processes in the Eifel region. We employ seismic tomography methods to resolve crustal thickness variations and velocity anomalies, together with moment tensor inversion to constrain fault geometries and deformation mechanisms.

Surface geological maps, digital elevation models, and geological cross-sections are used to build 3D structural geological models using the open-source software GemPy. Model construction follows a stepwise approach, starting from a simplified stratigraphic framework and gradually adding geological complexity, such as time-equivalent units and major fault structures. Although the steps are applied sequentially, the geological model is constructed from the input data and is therefore reproducible, enabling integration into subsequent workflow steps.

GemPy addresses uncertainty in geological models by generating ensembles of realizations through sampling input parameters from probability distributions. These ensembles serve as inputs for numerical simulations of physical quantities. Computing adjoint sensitivity kernels allows us to assess how each realization affects model outputs and to identify which models best match available observations, integrating structural uncertainty with process-based simulations. The numerical simulations are performed with LaMEM and its bindings for the Julia programming language. As GemPy is written in Python, the GemPy.jl package has been developed to expose its functionality in Julia.

The resulting geological and geophysical models may serve as the basis for a Play Fairway Analysis (PFA) which identifies regions with high potential for geothermal exploration. Crucially, this type of analysis requires uncertainty estimations for the modeled physical quantities, which our workflow provides.

From an implementation perspective, the digital laboratory consists of three main parts: a repository to collect data and models and their metadata, workflows and infrastructure for automatic processing, and an interface for visualization and interaction with the results. To demonstrate the feasibility, we develop the first prototype in JupyterLab which accommodates different computing environments and enables an interactive development process.

How to cite: Volk, M., Frasunkiewicz, J. A., Laumann, P., and Rahimi, A.: The DEGREE Project: A Digital Laboratory for Geothermal Exploration in the Eifel Region, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16511, https://doi.org/10.5194/egusphere-egu26-16511, 2026.

Subsurface Digital Twins rely critically on assimilation of data frameworks to continuously integrate multi-source, multi-type observations. While numerous methods have been developed to improve quantitative subsurface predictions, there is currently no clear consensus or standardised guidance on their appropriate computational deployment within digital twin workflows. Instead, research communities often adopt specific algorithms primarily because they are prevalent within their discipline, rather than because they are demonstrably optimal for the problem at hand. This lack of consensus reflects our limited understanding of how to rigorously characterise the mathematical structure of subsurface assimilation problems involving coupled multi-physics processes, multiple spatial and temporal scales, and heterogeneous data streams. As a result, current efforts frequently focus on empirical experimentation with algorithms rather than on the design of problem-adapted methodologies. This challenge extends to the formulation of the inverse problem itself, including parameterisation, parameter ranges, objective functions, and performance metrics, as well as to the selection of optimisation or inference strategies in multi-source data environments. Furthermore, comprehensive uncertainty quantification through global multi-factor sensitivity analysis is often infeasible due to the prohibitive computational cost of large-scale problems. To address these challenges, we propose Orison, a modular data assimilation environment designed to support systematic benchmarking and comparative analysis of classical model update algorithms for subsurface digital twin workflows. Orison enables controlled experimentation across a range of thematical problems, facilitating insight into algorithm performance and robustness. We demonstrate the capabilities of Orison through representative case studies in geothermal systems and groundwater management, illustrating how such a benchmarking framework can support more transparent methodological choices and contribute to the development of reliable, pragmatical subsurface digital twins.

How to cite: Lohier, T., Armandine les Landes, A., Rohmer, J., and Chassagne, R.: Orison: A modular data assimilation environment for subsurface digital twins, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16891, https://doi.org/10.5194/egusphere-egu26-16891, 2026.

An informed-decision making in managing risks due to climate-driven hazards for emergency response, designing preventive interventions, or policymaking for future requires either short-term and scenario-based assessments or long-term and uncertain assessments. Data requirements, spatial and temporal scales, observations required, and modelling techniques employed change drastically depending on the scope of the risk assessment. Digital twins (DT) in applications for natural hazards provide a great opportunity for significant improvements in disaster management. What makes DTs possible today is various technological advancements such as embedded sensors, cloud computing, edge computing, IoT. However, DTs also require a digital representation of the physical counterpart, mostly in the form of a computational or a data-driven model, to be able to predict future states. The utilisation of complex computational models in DTs is generally hindered by their relatively high computational budget and runtimes. A pathway to involve such models in (near) real-time decisions in DTs for geohazards is surrogate modelling. They are statistically valid representations of the computational model, into which physical laws and constraints can be embedded. Physics-compliant, physics-based or physics-informed surrogate models can facilitate DTs with i) instantaneous predictions, ii) the ability to conduct uncertainty quantification and sensitivity analysis to ensure reliability, iii) online updating of model parameters based on advanced calibration routines, iv) increased trust due to explainability based on physical laws. We present herein surrogate modelling as an enabler to replace computational models predicting the runout behaviour of geophysical flows. We investigate their applicability in uncertainty quantification, global sensitivity analysis, Bayesian parameter estimation, Bayesian model selection, and optimal experimental design. We demonstrate our workflow with two open-source computational models, r.avaflow 4.0 and synxflow, with synthetic and real-world case studies.

How to cite: Yildiz, A. and Kowalski, J.: Surrogate modelling as enabling methodology for predictive Digital Twins in geohazards, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17851, https://doi.org/10.5194/egusphere-egu26-17851, 2026.

A key prerequisite for reliable geoscientific process simulations is the calibration of uncertain model parameters against field observations. In practice, both measurements and simulation outputs are subject to uncertainty, arising from the observational errors, limited knowledge of material properties and inexact physical models. Bayesian inference provides a framework to explicitly acknowledge multiple sources of uncertainty by encoding modelling assumptions in prior distributions and updating them against observational data through the likelihood to obtain posterior estimates. However, applying Bayesian methods remains challenging in coupled multiphysical applications, including thermo-hydro-mechanical problems, as computational costs of repeated forward evaluations grow rapidly with model complexity.  

To address these limitations, we develop a hierarchical simulator for Bayesian calibrations that dynamically combines fast low-fidelity surrogate models with accurate high-fidelity finite-element simulations during the sampling stage. The core of the method stems from a fidelity-selection policy embedded directly in the probabilistic model, which transparently accounts for both surrogate-induced bias and the computational cost associated with high-fidelity simulations. We provide and compare several scenarios, that represent different optimization strategies for balancing posterior accuracy and computational efficiency. The resulting hierarchical Bayesian workflow is highly modular, and it can be coupled with external high-fidelity solvers through a unified forward interface and hence applicable to a wider range of geoscientific problems.

How to cite: Gruzdeva, Y., Degen, D., and Cacace, M.: A Hierarchical multi-fidelity approach for Bayesian inference for numerical process simulations , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17948, https://doi.org/10.5194/egusphere-egu26-17948, 2026.

One of the key challenges addressed by digital twins (DT) is the long-term modelling and monitoring of subsurface system behaviour. Existing DT technologies primarily rely on physics-based models capable of simulating dynamic processes. Long-term forecasting often suffers from uncertainty in data, modelling equations and their parameters, initial conditions and accumulating errors.

DT for natural systems remains an unexplored opportunity at a juvenile stage. Challenges with DT design for natural systems are largely related to their complex and uncertainty multi-physics nature.

We propose algebraic approach for DT design, where system parameters/attributes are represented as constraints rather than as specific values. This approach enables generation of subsurface scenarios and analysis of possible occurrence of critical system states/event.

We model the system as a collection of interacting entities (agents), whose states are defined by sets of attributes. For instance, a geological layer is considered as an agent characterised by its geometry represented by a 3D mesh (X0) , elasticity (E) , porosity (φ), thermal conductivity (T), and other relevant attributes. The initial state S0 of the agent can be presented as a set of constraints.

S0:        E1≤E≤E2 Ʌ  F1≤φ≤F2 Ʌ T1≤T≤T2 Ʌ X0 ,

The geometry X0 can also be represented as a set of constraints that take into account structural/mesh uncertainty. Thus, constraints can be specified for the set of all agents/layers interacting with each other.

We define the semantics of the agent's actions using formalized transitions that changes the constraints on the attributes/agent's state. An example of such a transition is the change in the layer state according to a function that is constructed from a combination of the equilibrium equations F, the constitutive equation Q, which relates the stress σ and the strain ε, and the kinematic equation of the strain D.

S1=G(S0, F(X0,σ), Q(φ,E,T), D(X0, ε)) .

The next state S1 is determined by the change of the agent state with the modelling of this transition and also represents the conjunction of constraints. The resulting new state is checked for compatibility with the critical state Z(σ, σ_max) following the threshold constraint (eg fracture):

σ <= σ_max .

If conjunction  S1 Ʌ Z(σ, σ_max) is satisfied, then there are such layer attribute values for which it is true. Such attributes are represented by the corresponding constraints generated by the solver. Having such constraints, we can obtain scenarios by the method of backward modelling, which will lead to the initial state.

Formalized transitions can be built by considering other parallel processes that affect the change in the state of the agent, in particular thermal, chemical, fluid flow.

This approach increases capability for long-term forecasting because it operates with subsurface states/events constraints/conditions rather than parameter specific simulations.

DT can combine algebraic modelling with neural networks that classify the predictions of a certain event. Algebraic modelling of the agent's behaviour from the classified state will confirm the correctness of the classification and build the corresponding explanatory scenario.

How to cite: Demyanov, V. and Letychevskyi, O.: Digital twins for subsurface systems based on algebraic models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19643, https://doi.org/10.5194/egusphere-egu26-19643, 2026.

Jordan’s dryland watersheds face acute water stress alongside increasing land degradation. Intense, short-duration storms cause flash runoff that accelerates soil erosion and sediment delivery to downstream infrastructure while groundwater, which is Jordan’s primary strategic water source, remains under long-term pressure. Rainwater-harvesting (RWH) interventions, including Vallerani micro catchments on damaged hillslopes, and Marab/flood-spreading and check-dam systems along ephemeral waterways, are increasingly used in restoration efforts. However, basin-scale planning is often limited by uncertainties in hydrological trade-offs and a gap between model outputs and stakeholder-ready, spatially explicit decision support.

This study develops a basin-scale hydrological Digital Twin (DT) for the Mujib Basin located in central Jordan by transforming process-based simulation findings into an interactive, scenario-driven dashboard. The DT combines a hydrological modelling core (SWAT) with harmonized in-situ and Earth Observation (EO) datasets to represent both water and land-surface responses. Physiographic inputs such as topography, soils, and land use, together with meteorological forcing derived from ERA5 reanalysis, and complemented by EO time series including Sentinel-2 vegetation indices, evapotranspiration products, and soil moisture to support the ecohydrological context.

Four intervention scenarios are represented - baseline, Vallerani, Marab, and combined - and evaluated using indicators relevant to water security, including surface runoff, sediment yield, and groundwater recharge, alongside vegetation/ET-related metrics. Outputs are produced at the sub-basin level and visualized through a web-based 3D dashboard, allowing users to visualize and compare different scenarios. The DT also enables "what-if" scenario testing by combining suitability-driven intervention placement with adjustable weather perturbations, allowing users to explore combined management and climate futures.

Beyond single-variable maps, the DT adds a decision layer for intervention targeting through a composite appropriateness framework matched with actual restoration goals: (1) Marab/check-dam suitability, which emphasizes high runoff generation, terrain controls, and proximity to channel networks; (2) infiltration-focused suitability, which highlights zones where slowing and spreading flow can increase recharge.  This study shows how digital twins can support hydrological decision-making in data-scarce dryland settings by bridging modelling outputs and implementation-oriented planning, usin Mujib Basin as a case study.

How to cite: Procheta, N., Koeva, M. N., and Aguilar, R. R.: Developing a Digital Twin Framework for Watershed Restoration Scenario Analysis:A Case Study in Mujib Basin, Jordan, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20401, https://doi.org/10.5194/egusphere-egu26-20401, 2026.

How to cite: Storrøsten, E., Carlton, B., Magni, V., Ragu Ramalingam, N., Gibbons, S. J., and Løvholt, F.: Multi-Source Tsunami Hazard Assessment for Digital Twin Workflows, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20704, https://doi.org/10.5194/egusphere-egu26-20704, 2026.

The Climate Change Adaptation Digital Twin (Climate DT), developed as part of the Destination Earth Initiative, produces global multi-decadal kilometer-scale simulations (5 – 10 km) in a new operational framework. A significant achievement in Climate DT is the capability to automatically process the hourly model output with impact applications which provide insights for users. Such applications include for instance the analysis of flood risks, renewable energy generation and wildfire risks. Therefore, Climate DT data can provide direct insights into potential adaptation requirements. Additionally, the Climate DT runs with multiple climate models (IFS-FESOM, IFS-NEMO and ICON) which led to the implementation of a standardized data portfolio on HealPix meshes, further benefiting data users in analyzing the data.

In this presentation we will introduce the operational Climate DT framework as well as the workflow that enables us to perform the climate simulations with automatic post-processing by multiple applications including scientific evaluation. Other aspects that will be introduced are the standardized data-portfolio and the simulations that have been performed so far as part of Climate DT.

How to cite: Kiszler, T., Kontkanen, J., Sigurdsson, B., de Paula Kinoshita, B., Bretonniere, P.-A., Narayanappa, D., Acosta, M., Polade, S., Sievi-Korte, O., Jung, T., Klocke, D., Doblas-Reyes, F., Koldunov, N., Gaya-Àvila, A., von Hardenberg, J., Davini, P., Frueh, B., Thober, S., Milinski, S., and Roura Adserias, F. and the Climate DT team: From climate simulations directly to actionable insights: The Climate Change Digital Twin, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21582, https://doi.org/10.5194/egusphere-egu26-21582, 2026.

Objectives:

In this study, we aim to develop a data-driven petrophysical inversion technique in the context of CO₂ sequestration. By integrating reservoir flow simulation, petroelastic modeling, and Graph Neural Networks (GNNs), CO₂ saturation can be estimated within models with multi-grid resolutions. The goal is to enhance accuracy and resolution adaptability in predicting CO₂ plume behavior in subsurface geological formations, thus improving carbon capture and storage (CCS) strategies.

Methodology:

We generated 100 2-dimensional synthetic reservoir models using a sequential indicator simulation algorithm for facies simulation, each populated by heterogeneous porosity and permeability. Flow simulations were conducted for 11 years using a central well with a constant injection rate. Petroelastic modeling was then performed to compute changes in P-wave and S-wave velocities and density every six months. The models were resampled to mimic a varying resolution scenario, with higher resolution near the well. A GNN model handled multi-resolution inputs and outputs, representing each grid as a node linked to its nearest eight neighbors, using direction and distances as edge attributes.

Results, Observations, and Conclusions:

The integrated modeling approach successfully predicted CO₂ plume migration within geological formations, demonstrating high predictive accuracy and robustness. Petroelastic modeling revealed significant changes in reservoir properties such as P-wave and S-wave velocities and density due to CO₂ injection. The Graph Neural Network (GNN) model, optimized through hyperparameter tuning, effectively utilized these changes to predict CO₂ saturation with a Mean Squared Error (MSE) of 0.0217 and a Coefficient of Determination (R²) of 0.981, confirming its high reliability in practical scenarios. In comparison, the Multilayer Perceptron model (MLP) achieved an MSE of 0.0260 and an R2 of 0.9695, processing data without considering spatial connections, underscoring the GNN's superior computational efficiency and spatial data integration. Furthermore, visual assessments confirmed the model’s accuracy, closely aligning predicted and actual CO₂ saturation levels, especially in dynamically changing reservoir zones. The study concludes that combining static property modeling, flow simulation, petroelastic modeling, and GNNs provides a valuable tool for enhancing CO₂ sequestration strategies, improving the prediction accuracy of CO₂ behavior in the subsurface, and significantly advancing CCS technologies.

Novel/Additive Information:

Our work leverages Graph Neural Networks (GNNs) to predict changes in CO₂ saturation from elastic properties, integrating flow dynamics with petroelastic modeling and deep learning via adaptive meshing grids. This novel approach addresses the limitations of conventional neural networks in adapting to mesh variations. Our project uniquely targets the complex challenges of CO₂ monitoring, advancing sequestration monitoring technologies by bridging seismic monitoring and dynamic flow simulation, providing a tool to predict CO₂ saturation from elastic properties.

How to cite: Alfayez, H.: Physics-Informed Graph Neural Networks for Multi-Resolution CO₂ Saturation Estimation in Subsurface , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22140, https://doi.org/10.5194/egusphere-egu26-22140, 2026.

Digital Twins (DTs) are increasingly used as integrative frameworks to combine data streams, numerical models and automated workflows for monitoring complex systems and supporting decision-making. In the field of seismic risk management, operational DTs must rely on fast, robust and reproducible modelling approaches, capable of assimilating real-time observations despite strong epistemic uncertainty. This contribution presents an operational earthquake DT implemented on the VIGIRISKS platform, and illustrated through two complementary rapid-response tools: SEISAid, dedicated to territorial-scale impact assessment, and B-Wave, focused on near real-time structural damage monitoring.

Rather than relying on detailed physics-based representations of subsurface processes, the proposed DT is built upon empirical ground-motion models and vulnerability models, which can be considered as meta-models linking observed seismic signals to expected ground motion and damage. Real-time seismic data from regional and national monitoring networks are continuously ingested through Pulsar approach. Seismic intensity fields are generated using the USGS ShakeMap framework, which embeds data weighting and uncertainty propagation to combine ground-motion prediction equations, instrumental recordings, macroseimic observations, and site-effect information. These ShakeMap products are then encapsulated within the VIGIRISKS infrastructure, where they trigger automated impact assessment workflows.

At the territorial scale, SEISAid exploits ShakeMap outputs and empirically calibrated vulnerability models to estimate building damage and potential human losses within 15–30 minutes after earthquake detection. Calculations are performed using reproducible scientific codes hosted on VIGIRISKS, and results are automatically aggregated and disseminated to decision-makers through standardized notification reports. This workflow supports rapid situational awareness and early operational decision-making under uncertainty.

At the structural scale, B-Wave extends the DT by integrating recorded dynamic responses from instrumented buildings. Damage assessment relies on data-driven signal processing methods, such as continuous wavelet transform–based frequency identification, to detect changes in structural dynamic properties. These changes are empirically related to damage states aligned with European EMS-98 classes, enabling near real-time alerts on the condition of critical structures without requiring detailed mechanical models.

A key characteristic of the framework is its event-driven and iterative cycle: each new earthquake updates data, models and outputs, progressively enriching the DT. By embedding empirical modelling, uncertainty handling and updating (via ShakeMap), and automated decision support within a unified infrastructure, this work illustrates how DT concepts can be operationally implemented for natural risk applications, contributing methodological insights relevant to subsurface-related DT workflows focused on data integration and decision support. Although this contribution focuses on the event-driven DT cycle triggered by real earthquakes, the proposed framework also enables “what-if scenario” based impact assessments, illustrating the flexibility of the DT for both operational response and prospective risk analysis. 

How to cite: Negulescu, C., Gehl, P., Auclair, S., Bertil, D., Legendre, Y., Guidez, R., Ramambazafy, H., Chan Thaw, F., Gracianne, C., Hoste Colomer, R., Roulle, A., and Grandjean, G.: An Operational Earthquake Digital Twin Based on Empirical Ground-Motion Models and Period Estimation: Integration of SEISAID and B-Wave within the VIGIRISKS Platform, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-23011, https://doi.org/10.5194/egusphere-egu26-23011, 2026.

Semantic Change Detection (SCD) plays a crucial role in understanding land surface dynamics, from urban expansion and deforestation to disaster impact assessment. However, despite the success of deep learning models in SCD tasks, their real-world deployment faces two critical challenges: significant domain shifts between geographically distinct regions and the prohibitive cost of data annotation for new locations. Models trained on public benchmarks, predominantly from developed countries, experience substantial performance degradation when applied to regions with different characteristics, such as Indian cities, due to inter-domain variance in sensors, atmospheric conditions, and landscapes. Additionally, substantial intra-domain variance within target regions compounds this problem, necessitating robust solutions that operate with limited labels.

To address these challenges, we propose SSLCD-Adapt, a novel hierarchical framework for label-efficient, cross-domain SCD that tackles inter-domain, intra-domain, and label constraints through a three-stage process. First, we employ Change-Enhanced Self-Supervised Pre-Training, where change representations are learned directly from unlabeled bi-temporal image pairs from the source domain using the FSC-180K benchmark dataset. Through the application of a Barlow Twins objective to fuse features from distorted views, the model learns invariant characteristics of change without manual annotation, providing superior initialization compared to ImageNet pre-trained models, which differ significantly from remote sensing imagery.

Second, Domain Alignment bridges the data distribution gap between the source (FSC-180K) and the target (six Indian cities) domains. The source encoder remains frozen while the target encoder undergoes training in only three layers within an adversarial setup. We employ a Domain-Adversarial Neural Network (DANN) that incorporates a Gradient Reversal Layer (GRL) with a Maximum Mean Discrepancy (MMD) loss to align feature distributions without requiring target labels. The domain classifier maximizes the H-divergence between the source and target domains, while GRL reverses the gradients, forcing target encoders to generate features similar to those of the source encoders, thereby achieving alignment in feature space and minimizing inter-domain variance.

Third, the trained target encoder undergoes Progressive Domain-specific fine-tuning using limited target labeled data. The encoder trains for one-third of the epochs on target data, then for two-thirds of the epochs using city-specific batches with domain-specific batch normalization for each city, effectively minimizing intra-domain variance between the six Indian cities. Figure 1 demonstrates the complete SSLCD-Adapt architecture.

Figure 1: Proposed SSLCD-Adapt Architecture

How to cite: Srivastava, N. and Jain, K.: SSLCD-ADAPT: A hierarchical framework for label-efficient cross-domain semantic change detection in complex environments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-548, https://doi.org/10.5194/egusphere-egu26-548, 2026.

Large-scale foundation models trained on multi-sensor satellite imagery has been driving recent advances in Earth Observation (EO) tasks. Although such models achieve impressive transferability across diverse downstream tasks, their computational and memory demands hinder accessibility, reproducibility, and deployment in resource-constrained environments. This work explores a compact and efficient alternative, introducing a metadata-aware Mixture-of-Experts Masked Autoencoder (MoE-MAE) for EO representation learning (Albughdadi, 2025).

The proposed MoE-MAE is a self-supervised transformer-based architecture with only 2.5 million parameters. It combines sparse expert routing and geo-temporal conditioning. The sparse routing allows token specialization while keeping active computation low. The geo-temporal conditioning injects information about latitude, longitude, and cyclic temporal attributes directly into the model. The proposed design enables the algorithm to exploit spatial and temporal regularities inherent in EO data without requiring dense, and computationally costly transformers.

The model is pretrained in the BigEarthNet-Landsat (BEN-LS) (Corley et al., 2025) dataset using a masked reconstruction loss function augmented with auxiliary unmasked and load-balancing losses to encourage stable expert utilization. The learned encoder representations are then evaluated using linear probing on two benchmark datasets: (1) BEN-LS, a multi-label land-cover dataset with explicit metadata, and (2) EuroSAT-Landsat (EuroSAT-LS) (Corley et al., 2025), a single-label classification datasets without metadata. Despite the encoder’s small size (~2.3 M parameters), the proposed MoE-MAE achieves competitive results with models’ orders of magnitude larger. On BEN-LS, the frozen encoder achieves a micro mean average precision of 0.767, comparable to SSL4EO-L ViT-S/16 MoCo v2 (0.775) (Stewart et al., 2023). On EuroSAT-LS, the model maintains strong transferability, achieving 84.2% accuracy, even in the absence of geo-temporal metadata.

Expert specialization across spatial patterns is revealed through adequate ablation and visualization studies, which show that some experts respond primarily to vegetation, others to water or textured regions. This demonstrates interpretable behaviour and complementary feature learning. Additionally, only about half of the model’s expert feed-forward capacity is activated per token, confirming computational sparsity in practice. These findings suggests that such models can retain strong representational power while substantially reducing training and inference costs.

This work presents a first step toward small-scale architectures for EO representation learning by integrating metadata, and leveraging sparse computation to approach the performance of massive transformers. Future work will extend this framework to multi-sensor and multi-temporal datasets to capture dynamic Earth processes efficiently.

Albughdadi, M. (2025). Lightweight Metadata-Aware Mixture-of-Experts Masked Autoencoder for Earth Observation.arXiv:2509.10919.

Stewart, A. J., Lehmann, N., Corley, I. A., Wang, Y., Chang, Y.-C., Braham, N. A. A., Sehgal, S., Robinson, C., & Banerjee, A. (2023). SSL4EO-L: Datasets and Foundation Models for Landsat Imagery. arXiv:2312.05241.

Corley, I., Sharma, L., and Crasto, R. (2025). Landsat-Bench: Datasets and Benchmarks for Landsat Foundation Models.arXiv:2506.08780.

How to cite: Albughdadi, M., Antonacci, M., Baousis, V., Fornari, F., Kaprol, T., and Pisa, C.: Efficient Earth Observation Representation Learning Using Metadata-Aware Mixture-of-Experts Masked Autoencoder, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2530, https://doi.org/10.5194/egusphere-egu26-2530, 2026.

Recent progress in Earth Observation (EO) foundation models has raised expectations that large-scale pretraining will yield general-purpose representations comparable to those in natural language processing and computer vision. In this work, we show that this promise has not yet been realized. We introduce spectral stability as a principled criterion for foundation models, measuring the extent to which the principal singular subspaces of pretrained weights are preserved during fine-tuning. Through this lens, we conduct a comparative analysis of several EO foundation models, including AnySat and Presto, alongside established models from vision and language, namely DINOv2 and BERT. Our analysis reveals a stark contrast between domains. BERT and DINOv2 exhibit strong spectral stability, with fine-tuning primarily inducing rotations within a small low-rank subspace. In contrast, EO models display severe spectral instability, where fine-tuning substantially rewrites their dominant singular directions. We show that this instability explains two key limitations of current EO foundation models. First, pretraining does not consistently accelerate downstream learning. Second, low-rank adaptation methods such as LoRA can fail or collapse, as the pretrained subspaces are only partially useful. Using extensive experiments on the TimeMatch benchmark for cross-regional crop classification, we demonstrate that despite strong performance claims, pretrained EO models yield inconsistent or marginal improvements over random initialization and do not achieve state-of-the-art performance. These findings indicate that current EO models lack the representational universality characteristic of true foundation models. We conclude that spectral stability is a critical property for robust transfer learning in Earth Observation, and we argue that future EO foundation models should prioritize spectral coherence through improved pretraining objectives and architectural designs that better capture the underlying structure of geospatial data.

How to cite: Turkoglu, M. O. and Aasen, H.: Are Earth Observation Foundation Models Really Foundation Models? Investigation Based on Spectral Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2602, https://doi.org/10.5194/egusphere-egu26-2602, 2026.

Effective protected area management requires frequent habitat monitoring to respond to rapid climate change and disturbances, yet traditional manual mapping methods cannot provide the temporal resolution needed for evidence-based policy decisions. We present a practical implementation of AI-driven change detection developed in collaboration with Gesaeuse National Park administration, Austria, to support operational habitat monitoring and management planning.

We address critical challenges in deploying AI technologies for complex environmental contexts: fuzzy class boundaries in natural habitats, highly imbalanced classes, and limited training data typical of protected areas. Using 15 years of high-resolution multimodal data (RGB, NIR, LiDAR, terrain attributes) covering 4,480 documented habitat changes across 15.3 km², we compare emerging geospatial foundation models (Clay v1.0, Prithvi-EO-2.0) against established U-Net architectures to identify the most robust approach for real-world application.

Results demonstrate that foundation models show superior cross-temporal robustness (Clay: 33% accuracy vs U-Net: 23% on unseen temporal data), a critical factor for operational monitoring systems. Integrating LiDAR improves detection accuracy from 30% to 50%. While overall accuracies are lower than in homogeneous agricultural landscapes, they reflect realistic performance for complex alpine environments and provide actionable information for park management.

To further enhance practical applicability for environmental agencies, we integrate object-based post-processing and physical constraints to filter misclassifications, making outputs directly usable for management decisions. This case study demonstrates practical strategies for implementing AI technologies in complex environmental monitoring contexts where traditional approaches face significant challenges. Building upon this work from the Habitalp 2.0 project, the BioDivAI project will extend these habitat mapping approaches to predict biodiversity impacts under various land use and land cover change scenarios, providing decision-makers with tools to assess trade-offs between economic activities and ecosystem protection.

How to cite: Kristen, H., Kulmer, D., and Hirschmugl, M.: Habitat and Land Cover Change Detection in Alpine Protected Areas: A Comparison of AI Architectures, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7472, https://doi.org/10.5194/egusphere-egu26-7472, 2026.

Earth observation datasets, especially those derived from remote sensing, are often characterized by significant data gaps. However, the pretraining of Geospatial Foundation Models requires mostly complete samples, leading to very selective sampling strategies that leave out large parts of the original observations. The problem is exacerbated in spatiotemporal data where these restrictions apply to the entire time series. Systems like Prithvi-EO-2.0 allow very small gap regions that can be addressed with interpolation during preprocessing. But a solution for integrating significant gap areas (>20% of the sample) into the pretraining process is yet to be established. We introduce an architecture that builds upon the random masking strategies in popular MAE-style architectures by additionally force-masking patches that contain gaps. Doing so requires a BERT masking scheme where masked patches are encoded instead of being removed from the sequence. Custom loss functions are introduced to account for the gaps in both the targets and the masked patches. While the resulting encoder-only architecture does not benefit from the reduced computational complexity in MAE-style masking, we mitigate this effect by using factorized space-time attention in the Video Vision Transformer (ViViT) backbone, thus creating a simple and lightweight model that is easily scalable. We demonstrate the potential of the architecture by performing spatiotemporal representation learning in a multivariate setup involving global Land Surface Temperature (LST) observations. The model is embedded in a framework that provides customizable sampling strategies for large-scale Earth observation datasets, including control over parameters like the maximum gap ratio per sample, the sampling strides, and the involved variables in shared-grid datasets like Earth System Data Cubes (ESDC). This flexibility in sampling enables the generation of training datasets with millions of samples, thus exposing the full volume of information stored in Earth observation data to Geospatial Foundation Models.

How to cite: Zimmer, C., Umlauft, J., Kraemer, G., Montero, D., and Mahecha, M. D.: Gap-Aware Transformer-Based Foundation Model Pretraining for Spatiotemporal Earth Observation Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7697, https://doi.org/10.5194/egusphere-egu26-7697, 2026.

Accurate and scalable mapping of mariculture facilities is essential for coastal resource management, environmental monitoring, and sustainable aquaculture development. However, existing remote sensing–based segmentation approaches heavily rely on large amounts of annotated data or manual interaction, limiting their scalability and generalization. Recently, foundation models such as the Segment Anything Model (SAM) have demonstrated strong generalization ability across diverse visual domains. Nevertheless, SAM’s performance in remote sensing applications remains constrained by its reliance on manually selected prompts, which is impractical for large-scale or automated mapping tasks.

In this study, we propose an AutoPrompt-enhanced SAM framework (AutoPrompt-SAM) for the automated segmentation of mariculture facilities, specifically floating rafts and net cages, from high-resolution PlanetScope imagery. The proposed framework eliminates the need for human-provided prompts by introducing an AutoPrompt module that automatically generates high-quality point prompts for SAM, enabling prompt-free semantic segmentation in a fully automated manner.

As a foundation for this work, we construct a large-scale, high-quality mariculture facility segmentation dataset consisting of more than 1,000 manually annotated PlanetScope image patches with a spatial resolution of 3 m. Each sample is cropped to 256 × 256 pixels and includes pixel-level labels for floating rafts, net cages, and background. To the best of our knowledge, this dataset represents one of the first publicly usable high-resolution semantic segmentation benchmarks for mariculture facilities based on PlanetScope imagery.

The proposed AutoPrompt module learns to generate representative prompt points directly from image features, without requiring any human interaction during inference. These automatically generated prompts are then fed into SAM to produce segmentation masks. By leveraging SAM’s powerful pre-trained visual representations, our method effectively combines the generalization capability of foundation models with task-specific structural cues learned by the AutoPrompt module. Experimental results demonstrate that AutoPrompt-SAM achieves competitive performance compared with manually prompted SAM, while completely removing the need for human intervention.

Beyond mariculture mapping, we further investigate the transferability of the proposed framework. Without additional labeled data, AutoPrompt-SAM shows strong generalization performance when applied to other remote sensing segmentation scenarios, indicating that the learned prompt generation strategy captures transferable spatial and structural patterns. This highlights the potential of AutoPrompt-SAM as a label-efficient and domain-adaptive segmentation framework, capable of extending SAM to broader remote sensing applications.

Overall, this work makes three key contributions: (1) the construction of a large-scale, high-resolution PlanetScope mariculture facility segmentation dataset; (2) the proposal of an AutoPrompt-driven SAM framework that enables fully automated, prompt-free semantic segmentation while effectively exploiting SAM’s pre-trained knowledge; and (3) a demonstration of the framework’s strong transferability, offering a new pathway for reducing human intervention and annotation dependency in remote sensing segmentation tasks. The proposed approach provides a practical solution for adapting foundation models to large-scale Earth observation applications and paves the way toward more autonomous and scalable remote sensing analysis.

How to cite: Xu, Y. and Lu, L.: Towards Prompt-Free Segmentation of Mariculture Facilities Using an AutoPrompt-Enhanced Segment Anything Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9080, https://doi.org/10.5194/egusphere-egu26-9080, 2026.

The work presented herein showcases early results of COP-GEN, a general-purpose diffusion model supporting flexible zero-shot translation between a number of popular data modalities related to the Copernicus programme: Sentinel-2 (both L1C and L2A), Sentinel-1 RTC, Copernicus DEM-30, Land Use Land Cover Maps, Cloud Masks, geospatial coordinates, and timestamps.

COP-GEN is designed as a diffusion model with a transformer backbone, which offers two concrete advantages. Firstly, the diffusion formulation respects the stochastic nature of cross-modal translation tasks; so that nearly every conditional generation query can be satisfied by a diverse range of plausible outputs rather than a single deterministic sample. Secondly, the sequence-based architecture facilitates the integration of diverse data modalities by flattening their latent representations, along with modality-specific diffusion timesteps, into a single sequence of tokens. Consequently, COP-GEN is capable of synthesising missing data from any subset of modalities in a zero-shot manner.

The model is pre-trained at global scale on MajorTOM, using over one million paired, geographically distributed samples spanning diverse climate zones, land-cover types, and acquisition conditions. By training jointly on matched data modalities, COP-GEN can, for example, estimate Land Use Land Cover, cloud coverage, atmospheric correction, and the spatiotemporal context of the available observations.

The first set of results indicates strong generative capability and high output diversity across modalities. The work concludes by discussing the available open-source implementation along with potential use cases.

How to cite: Espinosa, M., Gmelich Meijling, E., Marsocci, V., Crowley, E. J., and Czerkawski, M.: COP-GEN: Stochastic Generative Modelling of Copernicus Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10800, https://doi.org/10.5194/egusphere-egu26-10800, 2026.

The growing availability of Earth Observation (EO) data enables monitoring of terrestrial ecosystems at unprecedented spatio-temporal resolutions. In practice, however, effective use of EO data remains constrained by substantial technical barriers. Working with raw, multi-modal EO imagery requires specialised domain expertise, large data transfers, access to high-performance computing infrastructure, and advanced machine learning (ML) skills. These requirements limit the accessibility of EO-based analytics for many downstream applications.

Geospatial foundation models (GFMs) provide a promising alternative by learning general-purpose representations from large volumes of unlabeled EO data. By decoupling representation learning from downstream task modelling, GFMs allow users to exploit expressive features from modern deep learning models with limited EO or deep learning expertise and modest computational resources.

In this work, we investigate embedding-based GFM workflows for forest disturbance monitoring, where timely inference and regional customisation are often more critical than maximising absolute predictive accuracy. Forest disturbances such as logging, windthrows, fires, pests, and diseases can occur abruptly and require rapid detection to support conservation, policy-making, and risk-management efforts.

Machine learning methods for forest disturbance detection from EO data are well established and have shown strong performance on regional benchmarks. However, much of this work remains confined to academic demonstrations and is rarely translated into operational monitoring systems. Existing forest monitoring tools, including those aggregated by Global Forest Watch, typically rely on region- and sensor-specific models with limited feature expressivity. These systems may benefit from the rich multi-modal and spatio-temporal representations learned by GFMs, provided such embeddings can be accessed through scalable and practical deployment pipelines.

We build on a pipeline designed to deliver on-demand, location- and time-specific geospatial embeddings as a service. Embeddings are generated server-side from raw EO data, compressed, and distributed as lightweight representations. End users interact only with these embeddings, which can be analysed using simple models such as linear probes or small decoders. This approach removes the need for the user to manipulate raw EO data, download large multi-modal datasets, or train and deploy large deep learning models, enabling rapid adaptation to local contexts with limited annotations and modest computational resources.

We present preliminary results demonstrating the feasibility of this approach for forest disturbance detection and discuss its strengths and limitations relative to bespoke, fully supervised image-based models. While GFMs may not be optimal for applications with abundant annotations and stringent accuracy requirements, embedding-based services are particularly well suited to time-sensitive and regionally adaptive monitoring scenarios. Overall, this work illustrates how releasing geospatial embeddings as a product or service can lower barriers to EO-based forest monitoring and support faster, more inclusive environmental decision-making.

How to cite: Robert, D. and Wegner, J. D.: Forest Disturbance Monitoring with Geospatial Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11247, https://doi.org/10.5194/egusphere-egu26-11247, 2026.

The emergence of foundation models marks a transformative era in Earth observation, delivering powerful and adaptable tools to tackle the complexities of processing massive satellite imagery. Currently, land cover mapping faces two primary obstacles: 1) the prohibitive cost and high reliance on high-quality labels for data annotation; and 2) the significant spectral and spatial variability of identical ground objects caused by differences in temporal phases, locations, and sensors. Visual Foundation Models (VFMs), with their potent generalization capabilities, offer a means to effectively bridge the domain gap. Inspired by it, a high-resolution remote sensing foundation model leveraging knowledge distillation and feature fusion HiD-FM is proposed. Specifically, HiD-FM undergoes self-supervised pre-training on a dataset of one million high-resolution unlabeled images. By synergizing knowledge distillation with feature fusion, it integrates the generalization power of pre-trained VFMs into a semi-supervised learning framework, thereby boosting performance on unlabeled data and enhancing fine-grained feature representation. Extensive experiments on semantic segmentation tasks demonstrate that HiD-FM consistently outperforms some RSFMs (such as RVSA, SMLFR and CMID), particularly in data-scarce scenarios. On the LoveDA and GID-15 datasets, our method surpasses both specialized models and existing foundation models across various labeling ratios. Notably, using only 30% of the training data, HiD-FM achieved OA of 83.19% on the GID-15 dataset. Furthermore, transfer learning experiments on GF-2 imagery across diverse spatiotemporal contexts yielded superior visualization results. HiD-FM enables rapid and cost-effective adaptation to target domains, thereby significantly advancing the field of remote sensing interpretation.

How to cite: Xu, G., Shen, H., Li, X., Xu, M., Lin, D., and Jiang, T.: HiD-FM: A High-Resolution Remote Sensing Foundation Model with Knowledge Distillation and Feature Fusion for Image Semantic Segmentation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11394, https://doi.org/10.5194/egusphere-egu26-11394, 2026.

Recent advancements in Earth embeddings have opened up new frontiers for geosciences, enabling efficient analysis of vast volumes of geospatial data. However, the practical utilization of these embeddings is often hindered by complex software environments and the requirement for specialized computational expertise. To help democratize access to Earth embeddings, we introduce EarthEmbeddingExplorer, an open-source, web-based application designed to enhance the accessibility, understanding and interactivity of Earth embeddings for the broader geoscience community. 

EarthEmbeddingExplorer integrates multiple state-of-the-art foundation models, including SatCLIP, FarSLIP, and SigLIP, to support cross-modal retrieval of Sentinel-2 imagery via text, image, and geographic location queries. Our implementation leverages the MajorTOM Core-S2L2A dataset as the primary data source; we pre-computed approximately 250,000 embeddings per model based on a uniform spatial sampling of the MajorTOM grid. This approach ensures a representative global coverage of 1.2% of the Earth's land surface. To ensure accessibility, all models and datasets are hosted on open-source frameworks, specifically ModelScope and Hugging Face. The application provides an intuitive interface for visualizing the geographical distribution of the retrieved results, rendering top-match thumbnails, and exporting comprehensive metadata. Such transparent and low-cost access to large-scale embedding analysis is essential for identifying model-specific advantages and limitations. By enabling instant cross-model comparisons within specific spatiotemporal contexts, EarthEmbeddingExplorer allows users to evaluate model performance for their unique monitoring needs and domains of interest.

Ongoing development focuses on expanding EarthEmbeddingExplorer’s capabilities by integrating additional embedding models such as DINOv2, and increasing global spatial coverage. We are further implementing FAISS-based vector similarity search to enable near-instantaneous queries across tens of millions of global embeddings. Future iterations will prioritize modular software architecture, standardized APIs, and detailed documentation to facilitate community-driven contributions of new embedding models and datasets. The web applications are accessible at https://huggingface.co/spaces/ML4Sustain/EarthExplorer and at https://www.modelscope.cn/studios/VoyagerX/EarthExplorer.

How to cite: Zheng, Y., Wu, W., Wu, B., Li, G., Czerkawski, M., and Klemmer, K.: Sentinel-2 Image Retrieval with Global, Cross-modal Embeddings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11672, https://doi.org/10.5194/egusphere-egu26-11672, 2026.

Recent research in geospatial machine learning has demonstrated that models pretrained with self-supervised learning on Earth observation data can perform well on downstream tasks with limited training data. However, most of the existing geospatial benchmark datasets have few data modalities and poor global representation, limiting the ability to evaluate multimodal pretrained models at global scales. To fill this gap, we introduce MMEarth-Bench, a collection of five new multimodal downstream tasks with 12 input modalities, globally distributed data, and both in- and out-of-distribution test splits. We benchmark a diverse set of pretrained models on MMEarth-Bench and find that multimodal models generally perform best. While pretraining tends to improve model robustness in limited data settings, geographic generalization abilities remain poor and using multimodal inputs at test time can sometimes lead to geographic overfitting. In order to facilitate model adaptation to new downstream tasks and geographic domains, we propose a model-agnostic method for test-time training with multimodal reconstruction (TTT-MMR) that uses all the modalities available at test time, regardless of whether the pretrained model accepts them as input. We show that TTT-MMR improves model performance on both random and geographic test splits, and that geographic batching (TTT-MMR-Geo) leads to a good trade-off between regularization and specialization during TTT. Our dataset, code, and visualization tool are linked from the project page at https://lgordon99.github.io/mmearth-bench.

How to cite: Gordon, L., Belongie, S., Igel, C., and Lang, N.: MMEarth-Bench: Global Environmental Tasks for Multimodal Geospatial Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12736, https://doi.org/10.5194/egusphere-egu26-12736, 2026.

Vision-language foundation models (VLFMs), such as CLIP, have demonstrated remarkable generalizability across diverse downstream tasks, including both cross-modal and vision-centric tasks. Leveraging large-scale textual supervision, VLFMs capture a broad spectrum of visual concepts and achieve breakthrough performance in zero-shot image understanding. However, current remote sensing (RS)-specific VLFMs, while performing well on image-level tasks, exhibit limited capability in fine-grained tasks such as open-vocabulary semantic segmentation (OVSS). This limitation stems from their adherence to the CLIP training paradigm, which aligns image and text features only at the global level, thereby degrading performance in tasks requiring high-quality visual representations at local level. Moreover, existing VLFMs that incorporate fine-grained alignment mechanisms still exhibit limited performance on remote sensing tasks, whether through direct transfer to RS scenarios or fine-tuning on RS image-caption datasets. This further underscores the need for developing RS-tailored fine-grained VLFMs.

To address this, we construct the first multi-granularity RS image-text dataset, MGRS-200k (Figure 1). MGRS-200k contains approximately 200k RS images, each annotated with both short and long global captions, as well as multiple object-level bounding boxes with corresponding categories, totaling over one million instances. We further investigate existing fine-grained VLFM training methods and find that their explicit region-text alignment strategies often disrupt semantic coherence, as their underlying assumptions do not hold in RS scenarios,  and thus degrade fine-grained understanding.

Building on these, we propose FarSLIP, a Fine-grained Aligned RS Language-Image Pretraining framework (Figure 2). FarSLIP first employs patch-to-patch self-distillation to align local and global visual cues, enhancing feature discriminability while preserving semantic coherence. It then applies CLS token-based region-category alignment using the MGRS-200k dataset to further improve spatial awareness. FarSLIP achieves state-of-the-art performance in zero-shot RS image understanding, excelling not only on image-level tasks such as scene classification and image-text retrieval, but more importantly on fine-grained tasks like OVSS. Additionally, it serves as a strong foundation for multimodal large language models (MLLMs) in RS image comprehension.

Figure 1.  Examples of our proposed MGRS-200k dataset. 

Figure 2. Overall rchitecture of FarSLIP. The model is trained in a two-stage manner. In Stage I, FarSLIP is optimized with image-caption alignment and patch-to-patch self-distillation. In Stage II, image-caption alignment and region-category alignment are jointly employed on the MGRS-200k dataset.

How to cite: Li, Z., Zhang, X., Xiao, P., and Zhu, X.: FarSLIP: A Vision-Language Foundation Model for Fine-Grained Remote Sensing Understanding, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12871, https://doi.org/10.5194/egusphere-egu26-12871, 2026.

The current landscape of geospatial AI models is expanding rapidly, with new open-source models being released nearly every month. Consequently, with this rising number of potential general-purpose models, many of which claim state-of-the-art performance, their suitability for specific tasks or spatiotemporal contexts is often difficult to judge. A key step towards the democratisation of model benchmarking can be made by releasing large-scale datasets of pre-computed embeddings, such as those shared within the Major TOM project.

Yet, even with easy access to global and dense embeddings of a given model, it is not clear how to evaluate them on a global scale, given the scarcity and spatiotemporal biases of high-quality labels. This work explores a set of evaluation tests that can be conducted on a global scale, moving beyond canonical use cases to understand the inherent biases of individual models.

First, a set of proxy tasks with worldwide coverage is introduced. In this benchmark prototype, several sensitivity variables are tested, including time, location (estimation of spatiotemporal context), and VIIRS nightlights data (estimating a proxy for human activity). Despite not being traditional downstream tasks, these three variables have the advantage of uniform quality across the entire dataset. This allows for standardised, fair evaluation of representations extracted from Sentinel-2 and Sentinel-1 data across a range of pre-trained encoders as part of the Major TOM Embedding suite.

Secondly, a suite of techniques for comparing internal representation geometries of latent space vectors from multiple models is introduced to evaluate the similarities and differences between individual models. This approach does not require any reference labels, enabling a deeper understanding of geospatial semantic relationships encoded by different architectures.

Ultimately, this work advances the large-scale evaluation of deep learning models for Earth observation data, utilizing these model comparisons to develop a set of recommendations for future benchmarking efforts within the Earth Science community.

How to cite: Czerkawski, M., Kluczek, M., and Bojanowski, J. S.: Time, Space, and Nightlights: Global Evaluation of Major TOM Earth Embeddings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13425, https://doi.org/10.5194/egusphere-egu26-13425, 2026.

In a context of rapid environmental change, delivering robust tree species mapping is essential.  It enables better quantification of forest biomass, facilitates climate change adaptation through better forest management and supports biodiversity preservation. However, the scarce existing ground-truth datasets suffer from geographic sparsity, semantic inconsistencies and class imbalance, making current methods overfit to context and unsuitable for accurate large-scale tree species mapping. Therefore, it is imperative to design methods that learn spatially invariant representations for tree species mapping.

The surge of Earth Observation missions has unlocked vast amounts of Satellite Image Time Series (SITS) which capture phenology and spectral dynamics that are an asset for tree species classification. Leveraging this data, an increasing number of Foundation Models (FM) pre-trained using Self-Supervised Learning (SSL) have been introduced.  Yet, due to the prevalence of patch-level annotations in tree species datasets,  FMs are primarily evaluated on classification tasks instead of segmentation, preventing the production of pixel-level maps. Furthermore, spatial generalization remains largely unexplored, partially explained by the geographic sparsity of the labels. As a result, current models often overfit to local context: they perform well on training areas but fail to generalize to new spatial domains.  Therefore this work focuses on rigorous spatial generalization evaluation and the development of methods to produce large-scale pixel-level tree species maps overcoming current spatial domain shifts. 

To quantify this generalization gap, we propose a spatial zero-shot domain adaptation evaluation protocol, where frozen FMs are linearly probed through a segmentation task on a geographical region and tested on geographically distinct, unseen regions. We aligned 3 datasets in Europe (TreeSatAI, PureForest and a regional dataset covering Poland) into 6 classes to benchmark state-of-the-art FMs (AnySat, ALISE, Presto) pre-trained on SITS and introduce a new architecture addressing current limitations.
We propose a SSL framework based on the TimeSFormer backbone. It captures complex spatio-temporal dynamics using divided space and time attention. The model is pre-trained as a Masked Auto-Encoder on a European-scale unlabeled Sentinel-2 dataset to learn robust phenological features. To mitigate the observed spatial generalization gap, we investigate different strategies such as auxiliary conditioning and thermal temporal positional encoding. 

Our evaluation protocol reveals a significant accuracy drop of state-of-the-art models when applied to unseen regions. This decline suggests that current FMs capture geographically-dependent features rather than intrinsic tree species characteristics, resulting in a spatial generalization gap. 
Experiments confirm that the proposed architecture learns semantically rich features, evidenced by its high capacity to reconstruct missing time steps of satellite time series.  

By quantifying the spatial domain shift, proposing a resilient SSL architecture and applying domain adaptation strategies this work addresses the important challenge of generalization in label-scarce regimes. It supports high-resolution forest monitoring, a prerequisite for precise carbon accounting and forest biodiversity conservation.

How to cite: Brood, S., Dumeur, I., Anger, J., de Truchis, A., Sean, E., Fayad, I., d'Aspremont, A., and Ciais, P.: Addressing Geographical Domain Shift in Tree Species Mapping via Foundation Models using Satellite Image Time Series, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13546, https://doi.org/10.5194/egusphere-egu26-13546, 2026.

How to cite: Augustyn, B., Kluczek, M., Bojanowski, J., and Czerkawski, M.: Similarity Search of Earth Observation Data Using Foundation Model Embeddings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14376, https://doi.org/10.5194/egusphere-egu26-14376, 2026.

This study presents a comparative analysis of two diverse Geospatial Foundation Models (GFMs) developed by consortia under the European Space Agency (ESA): THOR and TerraMind. THOR introduces a compute-adaptive architecture designed to handle heterogeneous sensors and variable patch sizes. This enables flexible compute–accuracy trade-offs and high performance in limited training data regimes. It is also the first GFM to extensively include Sentinel-1, -2, and -3 data. TerraMind, in contrast, is a multimodal GFM with both discriminative and generative capabilities, pretrained with a dual‑scale scheme that fuses token‑level context and pixel‑level detail, enabling any‑to‑any cross‑modal generation and Thinking‑in‑Modalities (TiM) to infer missing modalities during fine‑tuning and inference. The cross-comparison, aimed to understand the level of maturity of European technologies in AI4EO, covers a collection of Earth Observation use cases provided by the two consortia, encompassing several tasks (segmentation, change detection, and classification), across diverse and overlooked domains, including climate disaster analysis, methane leak detection, forest biomass monitoring, and sea ice mapping. To ensure consistent preprocessing and evaluation of the two models and use cases, we benchmarked them in two very widespread and acknowledged framework: PANGAEA and TerraTorch. The analysis focuses on task coverage, architectural capabilities, and performance metrics, highlighting differences in adaptability, modality integration, and downstream application effectiveness. Results provide insights into the strengths and limitations of current GFMs for various scenarios, making it possible to grasp insights on different GFMs approaches, not limited to THOR and TerraMind.

How to cite: Gmelich Meijling, E., Marsocci, V., Schindlegger, F., Bounegta, K., and Longepe, N.: TerraMind vs. THOR: A comparative analysis of ESA’s Geospatial Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18737, https://doi.org/10.5194/egusphere-egu26-18737, 2026.

How to cite: Sleimi, R., Vinholi, J., Werner, F., and Abelló, A.: A Geometry-Aware Multi-Task Framework for Parcel Delineation with Geospatial Foundation Models , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19446, https://doi.org/10.5194/egusphere-egu26-19446, 2026.

Machine learning models for Earth system prediction tasks differ substantially in their training strategy and the type of context they encode. Task-specific models are trained from scratch using a limited set of variables assumed to directly influence the prediction target. These models lack broad spatial and cross-variable context. In contrast, Earth system foundation models are pre-trained on large and heterogeneous data sets and are expected to capture richer environmental context that can be transferred to downstream prediction tasks.

In other machine learning domains, such as natural language processing, fine-tuning pre-trained foundation models has become standard practice due to consistent performance gains over models trained from scratch. Whether similar benefits arise for Earth system time-series prediction tasks remains unclear.

To address this gap, we compare task-specific transformer encoder models operating on pixel-level time series with fine-tuned Earth system foundation models across a set of time-series prediction tasks describing vegetation response to environmental change, including Gross Primary Productivity. This comparison isolates the effect of pre-training on predictive performance by keeping the prediction targets fixed. 

Our aim is to determine which modelling approach yields higher predictive accuracy for environmental time-series analyses.

How to cite: von Blohn, A. L., Mahecha, M., and Peters, J.: Foundation versus Task-Specific Models for Environmental Time-Series Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19581, https://doi.org/10.5194/egusphere-egu26-19581, 2026.

Monitoring and understanding Earth system dynamics and their response to climate change and human activity requires innovative approaches to analyse complex and multivariate remote sensing data. However, the current trend is towards large models that require a lot of memory and computational power to be trained. The DeepFeatures project addresses this challenge by developing an embedding approach to create Feature Data Cubes, which capture the underlying spatio-temporal ecosystem dynamics as a low dimensional representation in latent space. These reduced representations enable the use of simpler, resource-efficient downstream models, which are easier to train and require minimal computational resources.

Specifically, the project builds on the rationale that each spectral index (SI), which is calculated from spectral bands and represents certain surface properties such as vegetation greenness,  reflects a specific aspect of ecosystem behaviour. Despite the development of over two hundred spectral indices, current studies often narrow their focus to individual SIs, overlooking the broader context of land surface processes represented by not considered SIs. The DeepFeatures project addresses this challenge by adopting a spatio-temporal multivariate approach. The SIs are derived from Sentinel-2 observations to generate a SI Data Cube. A deep learning embedding algorithm is applied to reduce the SI dimension and extract a latent space to create the Feature Data Cubes.

To demonstrate the potential of the Feature Data Cubes, the project focuses on inference across a range of scientific applications, including modelling gross primary production, analysing  tree mortality and greening trends, biodiversity  monitoring for conservation, comparing phenological features using satellite and crowd-sourced data, and studying the ecological impacts of open-pit lignite mining.

DeepFeatures emphasises the deployment of transparent and reproducible workflows, from generating Sentinel-2 derived Training Data Cubes to creating Feature Data Cubes. It aims to have an accessible, extensible, and modifiable framework for diverse applications, fostering broad community engagement and enabling open exploration of Earth system dynamics.

This presentation will showcase the methodology, scientific cases, and transformative potential of the DeepFeatures framework, highlighting its contributions to Earth observation and climate research.

The project DeepFeatures is funded by ESA’s AI4Science activity. Website: https://rsc4earth.de/project/deepfeatures/ 

How to cite: Mora, K., Peters, J., Ntokas, K., Reinhardt, M., Brandt, G., Kattenborn, T., Kraemer, G., Montero, D., Mosig, C., and Mahecha, M. D.: DeepFeatures: Learning Latent Representations from Spectral Indices for Ecosystem Monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19720, https://doi.org/10.5194/egusphere-egu26-19720, 2026.

Pretraining geospatial foundation models (FMs) is expensive, and architectural choices control inductive bias for multiscale context, cross-resolution behavior, and band/sensor variation. Therefore, benchmarking reduces the risk of scaling the “wrong” base. A model benchmarking framework for geospatial image segmentation is a critical prerequisite for developing robust and scalable geospatial FMs. In the emerging era of Earth observation FMs, success hinges on strong, well-characterized base architectures that can generalize across sensors, modalities, and geographies. The extreme heterogeneity of Earth observation vision data (different spectral bands, resolutions, and regions) makes such generalization especially challenging, underscoring the need for systematic, controlled benchmarking across diverse model families to identify viable architectures for different scenarios.
Our work rigorously evaluates a broad spectrum of segmentation architectures and backbones under consistent conditions. We benchmark classical convolutional architectures (U-Net, DeepLab, UPerNet, FPN, PAN, and LinkNet) alongside modern transformer-based models (Dense Prediction Transformer (DPT) and SegFormer). For this comparison, we use representative backbones from both CNNs (ResNet and MobileNet) and Mix Vision Transformer (MiT). By comparing these heterogeneous models on equal footing, we determine which architectural patterns and hybrid combinations yield representations most conducive to generalization. This diversity in evaluation identifies well-founded architectural bases for geospatial FMs.
To guide architecture selection and pipeline design, we deploy a comprehensive suite of metrics covering both accuracy and efficiency. We evaluate segmentation accuracy via IoU, Dice, and boundary F1-score, and also measure efficiency (convergence speed and inference latency). These holistic benchmarks reveal critical trade-offs. For instance, some lightweight CNN models excel in speed, while transformer models boost boundary F-1 score. By capturing such nuances, our benchmark informs which architectures are best suited as general-purpose base models. It highlights how certain encoder–decoder combinations optimally balance performance and efficiency, and flags architectures with high transfer-readiness for new tasks and domains.
The result is a reproducible, transferable model landscape that serves as a blueprint for FM development. Our benchmark framework effectively preconditions the FM pipeline, enabling researchers to enter the scaling phase with proven architecture candidates that have demonstrated cross-task and cross-sensor robustness. This “model landscape” allows subsequent large-scale pretraining to confidently build on architectures that ensure broad downstream generalization even in agentic (autonomous) deployment scenarios.
Finally, we situate this work within the broader trend toward sensor-agnostic, self-supervised FMs in Earth observation. We argue that intelligent architecture search must precede any massive self-supervised pretraining effort. Early vetting of architectures under diverse conditions ensures that large-scale training resources are invested in the most promising designs. In summary, we frame this hybrid benchmarking framework as a strategic new layer in the geospatial FM ecosystem. The insights extend beyond segmentation, providing a reference point for building fine-tunable, sensor-agnostic foundation models that can be readily adapted to various downstream tasks and even deployed onboard satellites or other edge platforms. By solidifying architecture evaluation as an essential step, this work makes a serious scientific and strategic contribution toward the next generation of Earth observation AI.

How to cite: Alizadeh Pirbasti, M., McArdle, G., and Akbari, V.: Segmentation Model Benchmarking: A Strategic Prerequisite for Robust Geospatial Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21700, https://doi.org/10.5194/egusphere-egu26-21700, 2026.

The current generation of hybrid machine learning and physics-informed machine learning is often limited by the missing availability of comprehensive differentiable models: either strongly simplified models have to be used or machine learning (ML) can’t be integrated natively into process-based models and must be trained separately. Here, we present the ongoing development of SpeedyWeather.jl: A general circulation model that’s differentiable, GPU-capable and ready for ML simulations. SpeedyWeather.jl is a spectral atmospheric GCM with a primitive equation core on flexible grid implementations from Gaussian to HEALPix. It contains simple yet interactive representations of ocean, land and sea ice for coupled climate simulations. With a user interface made for modularity and interactivity, it’s ideally suited as a framework for hybrid atmospheric models. For example, new parameterizations can be defined without any lines of code for GPU or differentiability specifics, yet integrate seamlessly into those. We document the process to achieve differentiability of our model using the general purpose automatic differentiation library Enzyme, problems we encountered and solutions we found. We demonstrate the differentiability with a sensitivity analysis of our model, initial developments of data-driven parameterizations, and give an outlook on the development of differentiable Earth system models. 

How to cite: Gelbrecht, M., Klöwer, M., Groenke, B., and Boers, N.: Differentiable Atmospheric Modelling with SpeedyWeather.jl , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1927, https://doi.org/10.5194/egusphere-egu26-1927, 2026.

Canopy resistance is a key parameter in the Penman–Monteith (P–M) equation for calculating evapotranspiration (ET). In this study, we compared a machine learning algorithm–support vector machine (SVM) and an analytical solution (Todorovic, 1999) for estimating canopy resistances. Then, these estimated canopy resistances were applied to the P–M equation for estimating ET; as a benchmark, a constant (fixed) canopy resistance was also adopted for ET estimations. ET data were measured using the eddy-covariance method above three sites: a grassland (south Ireland), Cypress forest (north Taiwan), and Cryptomeria forest (central Taiwan) were used to test the accuracy of the above two methods. The observed canopy resistance was derived from rearranging the P–M equation. From the measurements, the average canopy resistances for the grassland, Cypress forest, and Cryptomeria forest were 163, 346, and 321 (s/m), respectively. Our results show that both methods tend to reproduce canopy resistances within a certain range of intervals. In general, the SVM model performs better, and the analytical solution systematically underestimates the canopy resistances and leads to an overestimation of evapotranspiration. It is found that the analytical solution is only suitable for low canopy resistance (less than 100 s/m) conditions.

How to cite: Hsieh, C.-I., Huang, I.-H., and Lu, C.-T.: Estimating Canopy Resistance Using Machine Learning and Analytical Approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2240, https://doi.org/10.5194/egusphere-egu26-2240, 2026.

Machine learning models have been used with success to produce accurate river discharge forecasts at multiple lead times. However, almost no research has been done to show if they are physically consistent across lead times. In the deterministic problem setting, where models output a single forecast with multiple leadtimes, these models are known to be mean-seeking, predicting the most likely river flow for each day, regardless of how likely the resulting trajectory is to occur. This is important for forecasters who need to look at the multi-day properties of a forecast, such as the accumulated flow or number of days over threshold. When each leadtime is described as an independent distribution, the model provides no insight into how to connect the uncertainties at each lead time, as an ensemble forecast would. In this paper, we show that temporal consistency in machine learning forecasts cannot be assumed, and develop two methods for enforcing temporal consistency, the Conditional-LSTM and Seeded-LSTM. Through this, we create ensemble forecasts that successfully predict temporal properties of the 10-day hydrographs. We find that by explicitly training the model to treat the prediction of previous lead times as truth, our model better predicts temporal properties of 10-day hydrographs than other standard methods. Our approach allows users to efficiently generate as many ensemble members as desired, and we use our results to highlight the important of developing temporally consistent ensembles.

How to cite: Ruparell, K., Hunt, K., Cloke, H., Prudhomme, C., Pappenberger, F., and Chantry, M.: AI-generated ensemble river flow forecasting: Using rollout and an additional noise input to build ensemble forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2804, https://doi.org/10.5194/egusphere-egu26-2804, 2026.

How to cite: Dacre, H., Charlton-Perez, A., Driscoll, S., Gray, S., Harvey, B., Harvey, N., Hodges, K., Hunt, K., and Volonte, A.: Midlatitude Cyclone Intensity Biases in Machine Learning Weather Prediction Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3361, https://doi.org/10.5194/egusphere-egu26-3361, 2026.

The use of machine learning to represent small-scale processes, such as subgrid-scale (SGS) dynamics, is now well established in weather forecasting and climate modelling. Recent advances have demonstrated that SGS models trained via "online" end-to-end learning - where the dynamical solver operating on the filtered equations participates in the training - can outperform traditional physics-based approaches. However, most studies have focused on idealised periodic domains or spheres, neglecting mechanical boundaries present in systems such as planetary interiors. To address this issue, we introduce a pseudo-spectral differentiable solver for the study of two-dimensional quasi-geostrophic turbulence in a rapidly rotating, axially symmetric bounded domain. A key advantage of the online learning approach is its implicit correction of the commutation errors arising from the irregular Chebyshev grid used in the radial direction, achieved through the estimation of correction terms for the filtered equations. In addition, since Chebyshev polynomials are not boundary-preserving, we project training data extracted from the high-resolution direct numerical simulation (DNS) from the fine grid onto the coarse grid using a Galerkin approach that ensures compatibility with the boundary conditions. 

We examine three configurations, varying the geometry (between an exponential container and a spherical shell) and the rotation rate. The flow is driven by a prescribed analytical forcing that mimics a network of pumps, allowing precise control over the energy injection scale and an exact estimate of the power input. For each case, we evaluate the accuracy of the online-trained SGS model against the reference DNS using integral quantities and spectral diagnostics. In all configurations, we show that an SGS model trained on data spanning only one turnover time remains stable and accurate over integrations at least a hundred times longer than the training period. Moreover, we demonstrate the model's remarkable ability to reproduce slow processes occurring on time scales far exceeding the training duration, such as the inward drift of jets in the spherical shell geometry, which exhibits a quasi-periodic recurrence time of O(10) turnover times. These results suggest a promising path towards developing SGS models for planetary and stellar interior dynamics, including dynamo processes. They indicate that costly DNS may need to be run only for short durations to generate training data, enabling subsequent long-term simulations with the trained model at a negligible computational cost.

How to cite: Fournier, A., Frezat, H., and Gastine, T.: Online learning of subgrid-scale models for quasi-geostrophic turbulence in planetary interiors, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4279, https://doi.org/10.5194/egusphere-egu26-4279, 2026.

Weather forecasting is critical for a range of human activities including transportation, agriculture, industry, as well as the safety of the general public. Over the last two years, machine learning models have shown that they have the potential to transform the complex weather prediction pipeline, but current approaches still rely on numerical weather prediction (NWP) systems, limiting forecast speed and accuracy. In this talk, I will give some of the background on these developments. I will then introduce a machine learning model which can replace the entire operational NWP pipeline. Aardvark Weather, an end-to-end data-driven weather prediction system, ingests raw observations and outputs global gridded forecasts and local station forecasts. Further, it can be optimised end-to-end to maximise performance over quantities of interest. I will show that the system outperforms an operational NWP baseline for multiple variables and lead times for gridded and station forecasts. These forecasts are produced with a remarkably simple neural process model using just 8% of the input data and three orders of magnitude less compute than existing NWP and hybrid AI-NWP methods. We anticipate that Aardvark Weather will be the starting point for a new generation of end-to-end machine learning models for medium-range forecasting.

How to cite: Turner, R.: Aardvark weather: end-to-end data-driven weather forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4536, https://doi.org/10.5194/egusphere-egu26-4536, 2026.

The inherent limitations of individual geophysical methods and the sparsity of observational data often render inversion results unstable and non-unique. Joint inversion of multiphysics data exploits the complementary sensitivities of different physical fields regarding depth, resolution, and boundary features, thereby significantly mitigating the ambiguity of single-method inversion and enhancing interpretation reliability. Traditional joint inversion approaches primarily fall into two categories: spatial structure-based and physical parameter-based constraints. The former relies on the similarity of property distribution patterns, which struggles to decouple non-homologous anomalies, while the latter is often constrained by the unreliability of empirical relationships under complex geological conditions. Recently, deep learning methods based on the U-Net architecture have achieved joint inversion by establishing constraints based solely on spatial structural similarity (Hu et al., 2025) or physical parameter correlations (Guo et al., 2021). Although promising, these methods often fail to accurately characterize non-homologous anomalies in complex geological environments.

This study proposes a dual-stream 3D U-Net architecture incorporating a hybrid attention-gating mechanism. In terms of methodology, we first construct a training dataset based on rock physics data that encompasses both statistical correlations and structural discrepancies. Regarding the network architecture, independent encoders are employed to extract 3D features from gravity and magnetic data, respectively. A cross-attention module is then utilized to capture deep structural correlations, thereby enhancing cooperative inversion in homologous regions. Subsequently, a gated fusion module is introduced as an adaptive feature selector to effectively disentangle inconsistent features in non-homologous regions. Finally, the prediction models are generated through independent decoders.

During the joint inversion implementation phase, the network takes preliminary independent inversion results as input to predict high-fidelity models that integrate physical and geological priors. We incorporate these predicted models as reference models into the regularization term of the joint inversion objective function, constructing a deep-prior-based constraint. During iterative optimization, this constraint guides the inversion trajectory toward the fine geological structures predicted by deep learning by minimizing the discrepancy between the inverted and reference models, while ensuring the fit to observational data. This mechanism achieves an organic integration of data-driven and physics-driven approaches.

References

• Hu, Y. Su, X. Wu, Y. Huang and J. Chen, "Successive Deep Perceptual Constraints for Multiphysics Joint Inversion," in IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1-14, 2025, Art no. 5907114. 
• Guo, H. M. Yao, M. Li, M. K. P. Ng, L. Jiang and A. Abubakar, "Joint Inversion of Audio-Magnetotelluric and Seismic Travel Time Data With Deep Learning Constraint," in IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 9, pp. 7982-7995, Sept. 2021.

How to cite: Xi, B. and wang, Z.: A Hybrid Attention-Gating Deep Learning Framework for 3D Joint Inversion of Gravity and Magnetic Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5241, https://doi.org/10.5194/egusphere-egu26-5241, 2026.

In light of the urgent need to accelerate measures for the adaptation to and mitigation of climate change, accurate Earth system models are more important than ever, for technology assessment and the identification of the most effective climate protection strategies. Global climate models have successfully projected consequences of different future scenarios, but the spread in projections remains large, with subgrid-scale parametrizations being the main origin of these uncertainties. Recently, machine learning-based hybrid models have successfully enhanced parametrizations - their directly data-driven structure can more effectively capture the empirical aspects of the parametrizations. Especially for the more complex parametrizations, such as microphysics or turbulence, which we study here, quantum computing could bring decisive further improvements as a part of hybrid models. Atmospheric turbulence strongly affects weather and climate because it determines the rates of exchange of heat, moisture, and momentum between the earth surface and the atmosphere. However, due to the chaotic nature of turbulence and the wide range of turbulent regimes in the atmospheric boundary layer from deep convection to nearly laminar stable conditions, it is notoriously hard to predict and model.
Here, we develop a prototype of a quantum machine learning-based subgrid-scale parametrization for the vertical temperature flux caused by atmospheric turbulence based on semi-idealized Large-Eddy-Simulations. We run experiments with dry convective boundary layers with the PALM model system. The setups span an 8x8 km² domain with a resolution of 10 m and horizontal periodic boundary conditions and an imposed surface heat flux, combining runs with different surface heat fluxes and geostrophic winds in our training data set. We train quantum and classical neural networks with different architectures, and find that quantum models based on parametrized circuits with just 2 or 3 qubits achieve accuracies similar to classical models with the same number of trainable parameters, highlighting the possibility to use quantum computing for parametrizations in the near future. In contrast, the Smagorinsky closure deviates strongly from the true flux in this setup. Our quantum and classical cell-based models both generalize well to data from PALM runs with unseen parameters close to the seen range. We further analyze the feature importance in quantum and classical models and find that most of our quantum models show better stability of the Shapley values with respect to varying the random initial conditions of the training runs. Since the number of required qubits to capture the idealized setting is low, it is promising to extend our model to more complex settings with realistic topography and varied weather conditions in the future, e.g. by using ICON boundary conditions in PALM, opening the possibility to exploit quantum advantages anticipated by the more stable interpretability of our prototype models.

How to cite: Dogra, L., Klamt, J., Eyring, V., and Schwabe, M.: Quantum machine learning-based parametrization for boundary layer turbulence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5888, https://doi.org/10.5194/egusphere-egu26-5888, 2026.

How to cite: Whan, K., van der Wiel, K., and Mücke, N.: Emulating transient climate simulations with generative AI , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6079, https://doi.org/10.5194/egusphere-egu26-6079, 2026.

Over the past two years, ECMWF has rapidly developed and operationalised two machine-learned forecasting systems: AIFS Single, a deterministic model, and AIFS-ENS, a fully probabilistic forecasting system. Both systems are trained on ERA5 reanalysis data and further fine-tuned using operational IFS analyses. In this talk, we briefly introduce the AIFS framework and present ongoing research aimed at further improving its forecast skill.

Current efforts are driven by several research directions, including increasing spatial resolution and incorporating observational data. The current AIFS models operate at the native ERA5 resolution of approximately 30km. While higher resolutions could significantly improve forecast skill in surface variables, available datasets, such as the operational IFS analysis at 9km, are only available for a limited number of years. To address this, we explore a cross-resolution fine-tuning strategy in which AIFS is first pretrained on ERA5 at coarse resolution and subsequently fine-tuned on six years of recent operational IFS analyses at 9 km. We present promising early results showing that this approach enables stable fine-tuning down to 9 km and leads to significant gains in surface forecast skill.

A second research direction investigates the use of alternative datasets to improve total precipitation forecasts. ERA5 is known to exhibit deficiencies in the representation of precipitation, particularly in the tropics. We therefore fine-tune AIFS using the Integrated Multi-satellitE Retrievals for GPM (IMERG) dataset, which has been shown to better capture precipitation characteristics in this region. Early results indicate that incorporating IMERG data can significantly improve total precipitation forecast skill in AIFS, with the largest benefits observed, as expected, in tropical regions.

How to cite: Moldovan, G., Prieto Nemesio, A., Pinnington, E., Lang, S., Polster, J., O'Brien, C., Santa Cruz, M., Alexe, M., Cook, H., Forbes, R., and Chantry, M.: Improving AIFS Forecast Skill through Fine-Tuning across Spatial Resolutions and Datasets, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6950, https://doi.org/10.5194/egusphere-egu26-6950, 2026.