ITS2 – Responsible and intelligent data modelling, AI and machine learning strategies for climate, environment, planetary and space sciences

EGU22-486 | Presentations | ITS2.1/PS1.2

Enhancing planetary imagery with the holistic attention network algorithm 

Denis Maxheimer, Ioannis Markonis, Masner Jan, Curin Vojtech, Pavlik Jan, and Solomonidou Anezina

The recent developments in computer vision research in the field of Single Image Super Resolution (SISR)

can help improve the satellite imagery data quality and, thus, find application in planetary exploration.

The aim of this study is to enhance planetary surface imagery, in planetary bodies that there are

available data but in a low resolution. Here, we have applied the holistic attention network (HAN)

algorithm to a set of images of Saturn’s moon Titan from the Titan Radar Mapper instrument in its

Synthetic Aperture Radar (SAR) mode, which was on board the Cassini spacecraft. HAN can find

correlations among hierarchical layers, channels of each layer, and all positions of each channel, which

can be interpreted as an application and intersection of previously known models. The algorithm used

in our case-study was trained on 5000 grayscale images from HydroSHED Earth surface imagery dataset

resampled over different resolutions. Our experimental setup was to generate High Resolution (HR)

imagery from eight times lower resolution (x8 scale). We followed the standard workflow for this

purpose, which is to first train the network enhancing x2 scale to HR, then x4 scale to x2 scale, and

finally x8 scale to x4 scale, using subsequently the results of the previous training. The promising results

open a path for further applications of the trained model to improve the imagery data quality, and aid

in the detection and analysis of planetary surface features.

How to cite: Maxheimer, D., Markonis, I., Jan, M., Vojtech, C., Jan, P., and Anezina, S.: Enhancing planetary imagery with the holistic attention network algorithm, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-486,, 2022.

EGU22-692 | Presentations | ITS2.1/PS1.2

Autonomous lineament detection in Galileo images of Europa 

Caroline Haslebacher and Nicolas Thomas

Lineaments are prominent features on the surface of Jupiter's moon Europa. Analysing these linear features thoroughly leads to insights on their formation mechanisms and the interactions between the subsurface ocean and the surface. The orientation and position of lineaments is also important for determining the stress field on Europa. The Europa Clipper mission is planned to launch in 2024 and will fly by Europa more than 40 times. In the light of this, an autonomous lineament detection and segmentation tool would prove useful for processing the vast amount of expected images efficiently and would help to identify processes affecting the ice sheet. 

We have trained a convolutional neural network to detect, classify and segment lineaments in images of Europa returned by the Galileo mission. The Galileo images that make up the training set are segmented manually, following a dedicated guideline. For better performance, we make use of synthetically generated data to pre-train the network. The current status of the work will be described.

How to cite: Haslebacher, C. and Thomas, N.: Autonomous lineament detection in Galileo images of Europa, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-692,, 2022.

EGU22-1014 | Presentations | ITS2.1/PS1.2

Automatic detection of the electron density from the WHISPER instrument onboard CLUSTER II 

Emmanuel De Leon, Nicolas Gilet, Xavier Vallières, Luca Bucciantini, Pierre Henri, and Jean-Louis Rauch

The Waves of HIgh frequency and Sounder for Probing Electron density by Relaxation
(WHISPER) instrument, is part of the Wave Experiment Consortium (WEC) of the CLUSTER II
mission. The instrument consists of a receiver, a transmitter, and a wave spectrum
analyzer. It delivers active (when in sounding mode) and natural electric field spectra. The
characteristic signature of waves indicates the nature of the ambient plasma regime and, combined
with the spacecraft position, reveals the different magnetosphere boundaries and regions. The
thermal electron density can be deduced from the characteristics of natural waves in natural mode
and from the resonances triggered in sounding mode, giving access to a key parameter of scientific
interest and major driver for the calibration of particles instrument.
Until recently, the electron density derivation required a manual time/frequency domain
initialization of the search algorithms, based upon visual inspection of WHISPER active and natural
spectrograms and other datasets from different instruments onboard CLUSTER.
To automate this process, knowledge of the region (plasma regime) is highly desirable. A Multi-
Layer Perceptron model has been implemented for this purpose. For each detected region, a GRU,
recurrent network model combined with an ad-hoc algorithm is then used to determine the electron
density from WHISPER active spectra. These models have been trained using the electron density
previously derived from various semi-automatic algorithms and manually validated, resulting in an
accuracy up to 98% in some plasma regions. A production pipeline based on these models has been
implemented to routinely derive electron density, reducing human intervention up to 10 times. Work
is currently ongoing to create some models to process natural measurements where the data volume
is much higher and the validation process more complex. These models of electron density
automated determination will be useful for future other space missions.

How to cite: De Leon, E., Gilet, N., Vallières, X., Bucciantini, L., Henri, P., and Rauch, J.-L.: Automatic detection of the electron density from the WHISPER instrument onboard CLUSTER II, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1014,, 2022.

EGU22-2765 | Presentations | ITS2.1/PS1.2

Extrapolation of CRISM based spectral feature maps using CaSSIS four-band images with machine learning techniques 

Michael Fernandes, Nicolas Thomas, Benedikt Elser, Angelo Pio Rossi, Alexander Pletl, and Gabriele Cremonese

Spectroscopy provides important information on the surface composition of Mars. Spectral data can support studies such as the evaluation of potential (manned) landing sites as well as supporting determination of past surface processes. The CRISM instrument on NASA’s Mars Reconnaissance Orbiter is a high spectral resolution visible infrared mapping spectrometer currently in orbit around Mars. It records 2D spatially resolved spectra over a wavelength range of 362 nm to 3920 nm. At present data collected covers less than 2% of the planet. Lifetime issues with the cryo-coolers prevents limits further data acquisition in the infrared band. In order to extend areal coverage for spectroscopic analysis in regions of major importance to the history of liquid water on Mars (e.g. Valles Marineris, Noachis Terra), we investigate whether data from other instruments can be fused to extrapolate spectral features in CRISMto these non-spectral imaged areas. The present work will use data from the CaSSIS instrument which is a high spatial resolution colour and stereo imager onboard the European Space Agency’s ExoMars Trace Gas Orbiter (TGO). CaSSIS returns images at 4.5 m/px from the nominal 400 km altitude orbit in four colours. Its filters were selected to provide mineral diagnostics in the visible wavelength range (400 – 1100 nm). It has so far imaged around 2% of the planet with an estimated overlap of ≲0.01% of CRISM data. This study introduces a two-step pixel based reconstruction approach using CaSSIS four band images. In the first step advanced unsupervised techniques are applied on CRISM hyperspectral datacubes to reduce dimensionality and establish clusters of spectral features. Given that these clusters contain reasonable information about the surface composition, in a second step, it is feasible to map CaSSIS four band images to the spectral clusters by training a machine learning classifier (for the cluster labels) using only CaSSIS datasets. In this way the system can extrapolate spectral features to areas unmapped by CRISM. To assess the performance of this proposed methodology we analyzed actual and artificially generated CaSSIS images and benchmarked results against traditional correlation based methods. Qualitative and quantitative analyses indicate that by this novel procedure spectral features of in non-spectral imaged areas can be predicted to an extent that can be evaluated quantitatively, especially in highly feature-rich landscapes.

How to cite: Fernandes, M., Thomas, N., Elser, B., Rossi, A. P., Pletl, A., and Cremonese, G.: Extrapolation of CRISM based spectral feature maps using CaSSIS four-band images with machine learning techniques, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2765,, 2022.

EGU22-2994 | Presentations | ITS2.1/PS1.2

Interpretable Solar Flare Prediction with Deep Learning 

Robert Jarolim, Astrid Veronig, Tatiana Podladchikova, Julia Thalmann, Dominik Narnhofer, Markus Hofinger, and Thomas Pock

Solar flares and coronal mass ejections (CMEs) are the main drivers for severe space weather disturbances on Earth and other planets. While the geo-effects of CMEs give us a lead time of about 1 to 4 days, the effects of flares and flare-accelerated solar energetic particles (SEPs) are very immediate, 8 minutes for the enhanced radiation and as short as about 20 minutes for the highest energy SEPs arriving at Earth. Thus, predictions of solar flare occurrence at least several hours ahead are of high importance for the mitigation of severe space weather effects.

Observations and simulations of solar flares suggest that the structure and evolution of the active region’s magnetic field is a key component for energetic eruptions. The recent advances in deep learning provide tools to directly learn complex relations from multi-dimensional data. Here, we present a novel deep learning method for short-term solar flare prediction. The algorithm is based on the HMI photospheric line-of-sight magnetic field and its temporal evolution together with the coronal evolution as observed by multi-wavelengths EUV filtergrams from the AIA instrument onboard the Solar Dynamics Observatory. We train a neural network to independently identify features in the imaging data based on the dynamic evolution of the coronal structure and the photospheric magnetic field evolution, which may hint at flare occurrence in the near future.

We show that our method  can reliably predict flares six hours ahead, with 73% correct flaring predictions (89% when considering only M- and X-class flares), and 83% correct quiet active region predictions.

In order to overcome the “black box problem” of machine-learning algorithms, and thus to allow for physical interpretation of the network findings, we employ a spatio-temporal attention mechanism. This allows us to extract the emphasized regions, which reveal the neural network interpretation of the flare onset conditions. Our comparison shows that predicted precursors are associated with the position of flare occurrence, respond to dynamic changes, and align with characteristics within the active region.

How to cite: Jarolim, R., Veronig, A., Podladchikova, T., Thalmann, J., Narnhofer, D., Hofinger, M., and Pock, T.: Interpretable Solar Flare Prediction with Deep Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2994,, 2022.

EGU22-5721 | Presentations | ITS2.1/PS1.2

Magnetopause and bow shock models with machine learning 

Ambre Ghisalberti, Nicolas Aunai, and Bayane Michotte de Welle

The magnetopause (MP) and the bow shock (BS) are the two boundaries bounding the magnetosheath, the region between the magnetosphere and the solar wind. Their position and shape depend on the upstream solar wind and interplanetary magnetic field conditions.

Predicting their shape and position is the starting point of many subsequent studies of processes controlling the coupling between the Earth’s magnetosphere and its interplanetary environment. We now have at our disposal an important amount of data from a multitude of spacecraft missions allowing for good spatial coverage, as well as algorithms based on statistical learning to automatically detect the two boundaries. From the data of 9 satellites over 20 years, we identified around 19000 crossings of the BS and 36000 crossings of the MP. They were used, together with their associated upstream conditions, to train a regression model to predict the shape and position of the boundaries. 

Preliminary results indicate that the obtained models outperform analytical models without making simplifying assumptions on the geometry and the dependency over control parameters.

How to cite: Ghisalberti, A., Aunai, N., and Michotte de Welle, B.: Magnetopause and bow shock models with machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5721,, 2022.

EGU22-5739 | Presentations | ITS2.1/PS1.2

Deep learning for surrogate modeling of two-dimensional mantle convection 

Siddhant Agarwal, Nicola Tosi, Pan Kessel, Doris Breuer, and Grégoire Montavon

Mantle convection plays a fundamental role in the long-term thermal evolution of terrestrial planets like Earth, Mars, Mercury and Venus. The buoyancy-driven creeping flow of silicate rocks in the mantle is modeled as a highly viscous fluid over geological time scales and quantified using partial differential equations (PDEs) for conservation of mass, momentum and energy. Yet, key parameters and initial conditions to these PDEs are poorly constrained and often require a large sampling of the parameter space to find constraints from observational data. Since it is not computationally feasible to solve hundreds of thousands of forward models in 2D or 3D, some alternatives have been proposed. 

The traditional alternative to high-fidelity simulations has been to use 1D models based on scaling laws. While computationally efficient, these are limited in the amount of physics they can model (e.g., depth-dependent material properties) and predict only mean quantities such as the mean mantle temperature. Hence, there has been a growing interest in machine learning techniques to come up with more advanced surrogate models. For example, Agarwal et al. (2020) used feedforward neural networks (FNNs) to reliably predict the evolution of entire 1D laterally averaged temperature profile in time from five parameters: reference viscosity, enrichment factor for the crust in heat producing elements, initial mantle temperature, activation energy and activation volume of the diffusion creep. 

We extend that study to predict the full 2D temperature field, which contains more information in the form of convection structures such as hot plumes and cold downwellings. This is achieved by training deep learning algorithms on a data set of 10,525 2D simulations of the thermal evolution of the mantle of a Mars-like planet. First, we use convolutional autoencoders to compress the size of each temperature field by a factor of 142. Second,  we compare the use of two algorithms for predicting the compressed (latent) temperature fields: FNNs and long-short-term memory networks (LSTMs).  On the one hand, the FNN predictions are slightly more accurate with respect to unseen simulations (99.30%  vs. 99.22% for the LSTM). On the other hand, Proper orthogonal decomposition (POD) of the LSTM and FNN predictions shows that despite a lower mean relative accuracy, LSTMs capture the flow dynamics better than FNNs. The POD coefficients from FNN predictions sum up to 96.51% relative to the coefficients of the original simulations, while for LSTMs this metric increases to 97.66%. 

We conclude the talk by stating some strengths and weaknesses of this approach, as well as highlighting some ongoing research in the broader field of fluid dynamics that could help increase the accuracy and efficiency of such parameterized surrogate models.

How to cite: Agarwal, S., Tosi, N., Kessel, P., Breuer, D., and Montavon, G.: Deep learning for surrogate modeling of two-dimensional mantle convection, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5739,, 2022.

EGU22-6371 | Presentations | ITS2.1/PS1.2

STIX solar flare image reconstruction and classification using machine learning 

Hualin Xiao, Säm Krucker, Daniel Ryan, Andrea Battaglia, Erica Lastufka, Etesi László, Ewan Dickson, and Wen Wang

The Spectrometer Telescope for Imaging X-rays (STIX) is an instrument onboard Solar Orbiter. It measures X-rays emitted during solar flares in the energy range from 4 to 150 keV and takes X-ray images by using an indirect imaging technique, based on the Moiré effect. STIX instrument
consists of 32 pairs of tungsten grids and 32 pixelated CdTe detector units. Flare Images can be reconstructed on the ground using algorithms such as back-projection, forward-fit, and maximum-entropy after full pixel data are downloaded. Here we report a new image reconstruction and
classification model based on machine learning. Results will be discussed and compared with those from the traditional algorithms.

How to cite: Xiao, H., Krucker, S., Ryan, D., Battaglia, A., Lastufka, E., László, E., Dickson, E., and Wang, W.: STIX solar flare image reconstruction and classification using machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6371,, 2022.

EGU22-8940 | Presentations | ITS2.1/PS1.2

Mars events polyphonic detection, segmentation and classification with a hybrid recurrent scattering neural network using InSight mission data 

Salma Barkaoui, Angel Bueno Rodriguez, Philippe Lognonné, Maarten De Hoop, Grégory Sainton, Mathieu Plasman, and Taichi kawamura

Since deployed on the Martian surface, the seismometer SEIS (Seismic Experiment for Interior Structure) and the APSS (Auxiliary Payload Sensors Suite) of the InSight (Interior Exploration using Seismic Investigations, Geodesy and Heat Transport) mission have been recorded the daily Martian respectively ground acceleration and pressure. These data are essential to investigate the geophysical and atmospheric features of the red planet. So far, the InSight team were able to detect multiple Martian events. We distinguish two types: the artificial events like the lander modes or the micro-tilts known as glitches or the natural events like the pressure drops which are important to estimate the Martian subsurface and the seismic events used to study the interior structure of Mars. Despite the data complexity, the InSight team was able to catalog these events (Clinton et al 2020 for the seismic event catalog, Banfield et al., 2018, 2020 for the pressure drops catalog and Scholz et al. (2020) for the glitches catalog). However, despite all this effort, we are still in front of multiple challenges. In fact,  the seismic events' detection is limited  due to the SEIS sensitivity, which is the origin of a high noise level that may contaminate the seismic events. Thus, we can miss some of them, especially in the noisy period. Besides, their detection is very challenging and require multiple preprocessing task which is time-consuming. For the pressure drops, the detection method used in Banfield et al.  2020 is limited by a threshold equal to 0.3 Pa. Thus, the rest of pressure drops are not included. Plus, due to lack of energy, the pressure sensor was off for several days. As a result, many pressure drops were missed. As a result, being able to detect them directly on the SEIS data which are, in contrast,  provided continuously, is very important.

In this regard, the aim of this study is to overcome these challenges and thus improve the Martian events detection and provide an updated catalog automatically. For that, we were inspired of one of the main technics used today in data processing and analysis in a complete automatic way: it is the Machine Learning and particularly in our case is the Deep Learning. The architecture used for that is the “Hybrid Recurrent Scattering Neural Network” (Bueno et al 2021)  adapted for Mars

How to cite: Barkaoui, S., Bueno Rodriguez, A., Lognonné, P., De Hoop, M., Sainton, G., Plasman, M., and kawamura, T.: Mars events polyphonic detection, segmentation and classification with a hybrid recurrent scattering neural network using InSight mission data, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8940,, 2022.

EGU22-9077 | Presentations | ITS2.1/PS1.2

Automatic Detection of Interplanetary Coronal Mass Ejections 

Hannah Ruedisser, Andreas Windisch, Ute V. Amerstorfer, Tanja Amerstorfer, Christian Möstl, Martin A. Reiss, and Rachel L. Bailey

Interplanetary coronal mass ejections (ICMEs) are one of the main drivers for space weather disturbances. In the past,
different machine learning approaches have been used to automatically detect events in existing time series resulting from
solar wind in situ data. However, classification, early detection and ultimately forecasting still remain challenges when facing
the large amount of data from different instruments. We propose a pipeline using a Network similar to the ResUNet++ (Jha et al. (2019)), for the automatic detection of ICMEs. Comparing it to an existing method, we find that while achieving similar results, our model outperforms the baseline regarding GPU usage, training time and robustness to missing features, thus making it more usable for other datasets.
The method has been tested on in situ data from WIND. Additionally, it produced reasonable results on STEREO A and STEREO B datasets
with less input parameters. The relatively fast training allows straightforward tuning of hyperparameters and could therefore easily be used to detect other structures and phenomena in solar wind data, such as corotating interaction regions.

How to cite: Ruedisser, H., Windisch, A., Amerstorfer, U. V., Amerstorfer, T., Möstl, C., Reiss, M. A., and Bailey, R. L.: Automatic Detection of Interplanetary Coronal Mass Ejections, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9077,, 2022.

EGU22-9621 | Presentations | ITS2.1/PS1.2

Machine Learning Techniques for Automated ULF Wave Recognition in Swarm Time Series 

Georgios Balasis, Alexandra Antonopoulou, Constantinos Papadimitriou, Adamantia Zoe Boutsi, Omiros Giannakis, and Ioannis A. Daglis

Machine learning (ML) techniques have been successfully introduced in the fields of Space Physics and Space Weather, yielding highly promising results in modeling and predicting many disparate aspects of the geospace. Magnetospheric ultra-low frequency (ULF) waves play a key role in the dynamics of the near-Earth electromagnetic environment and, therefore, their importance in Space Weather studies is indisputable. Magnetic field measurements from recent multi-satellite missions are currently advancing our knowledge on the physics of ULF waves. In particular, Swarm satellites have contributed to the expansion of data availability in the topside ionosphere, stimulating much recent progress in this area. Coupled with the new successful developments in artificial intelligence, we are now able to use more robust approaches for automated ULF wave identification and classification. Here, we present results employing various neural networks (NNs) methods (e.g. Fuzzy Artificial Neural Networks, Convolutional Neural Networks) in order to detect ULF waves in the time series of low-Earth orbit (LEO) satellites. The outputs of the methods are compared against other ML classifiers (e.g. k-Nearest Neighbors (kNN), Support Vector Machines (SVM)), showing a clear dominance of the NNs in successfully classifying wave events.

How to cite: Balasis, G., Antonopoulou, A., Papadimitriou, C., Boutsi, A. Z., Giannakis, O., and Daglis, I. A.: Machine Learning Techniques for Automated ULF Wave Recognition in Swarm Time Series, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9621,, 2022.

The solar wind and its variability is well understood at Earth. However, at distances larger than 1AU the is less clear, mostly due to the lack of in-situ measurements. In this study we use transfer learning principles to infer solar wind conditions at Mars in periods where no measurements are available, with the aim of better illuminating the interaction between the partially magnetised Martian plasma environment and the upstream solar wind. Initially, a convolutional neural network (CNN) model for forecasting measurements of the interplanetary magnetic field, solar wind velocity, density and dynamic pressure is trained on terrestrial solar wind data. Afterwards, knowledge from this model is incorporated into a secondary CNN model which is used for predicting solar wind conditions upstream of Mars up to 5 hours in the future. We present the results of this study as well as the opportunities to expand this method for use at other planets.

How to cite: Durward, S.: Forecasting solar wind conditions at Mars using transfer learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10105,, 2022.

EGU22-11501 | Presentations | ITS2.1/PS1.2

Automatic detection of solar magnetic tornadoes based on computer vision methods. 

Dmitrii Vorobev, Mark Blumenau, Mikhail Fridman, Olga Khabarova, and Vladimir Obridko

We propose a new method for automatic detection of solar magnetic tornadoes based on computer vision methods. Magnetic tornadoes are magneto-plasma structures with a swirling magnetic field in the solar corona, and there is also evidence for the rotation of plasma in them. A theoretical description and numerical modeling of these objects are very difficult due to the three-dimensionality of the structures and peculiarities of their spatial and temporal dynamics [Wedemeyer-Böhm et al, 2012, Nature]. Typical sizes of magnetic tornadoes vary from 102 km up to 106 km, and their lifetime is from several minutes to many hours. So far, quite a few works are devoted to their study, and there are no accepted algorithms for detecting solar magnetic tornadoes by machine methods. An insufficient number of identified structures is one of many problems that do not allow studying physics of magnetic tornadoes and the processes associated with them. In particular, the filamentous rotating structures are well delectable only at the limb, while one can only make suppositions about their presence at the solar disk.
Our method is based on analyzing SDO/AIA images at wavelengths 171 Å, 193 Å, 211 Å and 304 Å, to which several different algorithms are applied, namely, the convolution with filters, convolutional neural network, and gradient boosting. The new technique is a combination of several approaches (transfer learning & stacking) that are widely used in various fields of data analysis. Such an approach allows detecting the structures in a short time with sufficient accuracy. As test objects, we used magnetic tornadoes previously described in the literature [e.g., Wedemeyer et al 2013, ApJ; Mghebrishvili et al. 2015 ApJ]. Our method made it possible to detect those structures, as well as to reveal previously unknown magnetic tornadoes.

How to cite: Vorobev, D., Blumenau, M., Fridman, M., Khabarova, O., and Obridko, V.: Automatic detection of solar magnetic tornadoes based on computer vision methods., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11501,, 2022.

EGU22-12480 | Presentations | ITS2.1/PS1.2

A versatile exploration method for simulated data based on Self Organizing Maps 

Maria Elena Innocenti, Sophia Köhne, Simon Hornisch, Rainer Grauer, Jorge Amaya, Jimmy Raeder, Banafsheh Ferdousi, James "Andy" Edmond, and Giovanni Lapenta

The large amount of data produced by measurements and simulations of space plasmas has made it fertile ground for the application of classification methods, that can support the scientist in preliminary data analysis. Among the different classification methods available, Self Organizing Maps, SOMs [Kohonen, 1982] offer the distinct advantage of producing an ordered, lower-dimensional representation of the input data that preserves their topographical relations. The 2D map obtained after training can then be explored to gather knowledge on the data it represents. The distance between nodes reflects the distance between the input data: one can then further cluster the map nodes to identify large scale regions in the data where plasma properties are expected to be similar.

In this work, we train SOMs using data from different simulations of different aspects of the heliospheric environment: a global magnetospheric simulation done with the OpenGGCM-CTIM-RCM code, a Particle In Cell simulation of plasmoid instability done with the semi-implicit code ECSIM, a fully kinetic simulation of single X point reconnection done with the Vlasov code implemented in MuPhy2.

We examine the SOM feature maps, unified distance matrix and SOM node weights to unlock information on the input data. We then classify the nodes of the different SOMs into a lower and automatically selected number of clusters, and we obtain, in all three cases, clusters that map well to our a priori knowledge on the three systems. Results for the magnetospheric simulations are described in Innocenti et al, 2021. 

This classification strategy then emerges as a useful, relatively cheap and versatile technique for the analysis of simulation, and possibly observational, plasma physics data.

Innocenti, M. E., Amaya, J., Raeder, J., Dupuis, R., Ferdousi, B., & Lapenta, G. (2021). Unsupervised classification of simulated magnetospheric regions. Annales Geophysicae Discussions, 1-28.

How to cite: Innocenti, M. E., Köhne, S., Hornisch, S., Grauer, R., Amaya, J., Raeder, J., Ferdousi, B., Edmond, J. "., and Lapenta, G.: A versatile exploration method for simulated data based on Self Organizing Maps, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12480,, 2022.

EGU22-12830 | Presentations | ITS2.1/PS1.2

Re-implementing and Extending the NURD Algorithm to the Full Duration of the Van Allen Probes Mission 

Matyas Szabo-Roberts, Karolina Kume, Artem Smirnov, Irina Zhelavskaya, and Yuri Shprits

Generating reliable databases of electron density measurements over a wide range of geomagnetic conditions is essential for improving empirical models of electron density. The Neural-network-based Upper hybrid Resonance Determination (NURD) algorithm has been developed for automated extraction of electron density from Van Allen Probes electric field measurements, and has been shown to be in good agreement with existing semi-automated methods and empirical models. The extracted electron density data has since then been used to develop the PINE (Plasma density in the Inner magnetosphere Neural network-based Empirical) model, an empirical model for reconstructing the global dynamics of the cold plasma density distribution based only on solar wind data and geomagnetic indices.
In this study we re-implement the NURD algorithm in both Python and Matlab, and compare the performance of these implementations to each other and previous NURD results. We take advantage of a labeled training data set now being available for the full duration of the Van Allen Probes mission to train the network and generate an electron density data set for a significantly longer time period. We perform detailed comparisons between this output, electron density produced from Van Allen Probes electric field measurements using the AURA semi-automated algorithm, and electron density obtained from existing empirical models. We also present preliminary results from the PINE plasmasphere model trained on this extended NURD electron density data set.

How to cite: Szabo-Roberts, M., Kume, K., Smirnov, A., Zhelavskaya, I., and Shprits, Y.: Re-implementing and Extending the NURD Algorithm to the Full Duration of the Van Allen Probes Mission, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12830,, 2022.

The ITU/WMO/UNEP Focus Group on AI for Natural Disaster Management (FG-AI4NDM) explores the potential of AI to support the monitoring and detection, forecasting, and communication of natural disasters. Building on the presentation at EGU2021, we will show how detailed analysis of real-life use cases by an interdisciplinary, multistakeholder, and international community of experts is leading to the development of three technical reports (dedicated to best practices in data collection and handling, AI-based algorithms, and AI-based communications technologies, respectively), a roadmap of ongoing pre-standardization and standardization activities in this domain, a glossary of relevant terms and definitions, and educational materials to support capacity building. It is hoped that these deliverables will form the foundation of internationally recognized standards.

How to cite: Kuglitsch, M.: Nature can be disruptive, so can technology: ITU/WMO/UNEP Focus Group on AI for Natural Disaster Management, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8,, 2022.

EGU22-79 | Presentations | ITS2.5/NH10.8

Assessing the impact of sea-level rise on future compound flooding hazards in the Kapuas River delta 

Joko Sampurno, Valentin Vallaeys, Randy Ardianto, and Emmanuel Hanert

Compound flooding hazard in estuarine delta is increasing due to mean sea-level rise (SLR) as the impact of climate change. Decision-makers need future hazard analysis to mitigate the event and design adaptation strategies. However, to date, no future hazard analysis has been made for the Kapuas River delta, a low-lying area on the west coast of the island of Borneo, Indonesia. Therefore, this study aims to assess future compound flooding hazards under SLR over the delta, particularly in Pontianak (the densest urban area over the region). Here we consider three SLR scenarios due to climate change, i.e., low emission scenario (RCP2.6), medium emission scenario (RCP4.5), and high emission scenario (RCP8.5). We implement a machine-learning technique, i.e., the multiple linear regression (MLR) algorithm, to model the river water level dynamics within the city. We then predict future extreme river water levels due to interactions of river discharges, rainfalls, winds, and tides. Furthermore, we create flood maps with a likelihood of areas to be flooded in 100 years return period (1% annual exceedance probability) due to the expected sea-level rise. We find that the extreme 1% return water level for the study area in 2100 is increased from about 2.80 m (current flood frequency state) to 3.03 m (under the RCP2.6), to 3.13 m (under the RCP4.5), and 3.38 m (under the RCP8.5).

How to cite: Sampurno, J., Vallaeys, V., Ardianto, R., and Hanert, E.: Assessing the impact of sea-level rise on future compound flooding hazards in the Kapuas River delta, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-79,, 2022.

According to UNDRR2021, there are 389 reported disasters in 2020. Disasters claim the lives of 15,080 people, 98.4 million people are affected globally, and US171.3 billion dollars are spent on economic damage. International agreements such as the Sendai framework for disaster risk reduction encourage the use of social media to strengthen disaster risk communication. With the advent of new technologies, social media has emerged out to be an important source of information in disaster management, and there is an increase in social media activity whilst disasters. Social media is the fourth most used platform for accessing emergency information. People seek to contact family, friends and search for food, water, transportation, and shelter. During cataclysmic events, the critical information posted on social media is immersed in irrelevant information. To assist and streamline emergency situations, staunch methodologies are required for extracting relevant information. The research study explores new-fangled deep learning methods for automatically identifying the relevancy of disaster-related social media messages. The contributions of this study are three-fold. Firstly, we present a hybrid deep learning-based framework to ameliorate the classification of disaster-related social media messages. The data is gathered from the Twitter platform, using the Search Application Programming Interface. The messages that contain information regarding the need, availability of vital resources like food, water, electricity, etc., and provide situational information are categorized into relevant messages. The rest of the messages are categorized into irrelevant messages. To demonstrate the applicability and effectiveness of the proposed approach, it is applied to the thunderstorm and cyclone Fani dataset. Both the disasters happened in India in 2019. Secondly, the performance of the proposed approach is compared with baseline methods, i.e., convolutional neural network, long short-term memory network, bidirectional long short-term memory network. The results of the proposed approach outperform the baseline methods. The performance of the proposed approach is evaluated using multiple metrics. The considered evaluation metrics are accuracy, precision, recall, f-score, area under receiver operating curve, area under precision-recall curve. The accurate and inaccurate classifications are shown on both the datasets. Thirdly, to incorporate our evaluated models into a working application, we extend an existing application DisDSS, which has been granted copyright invention award by Government of India. We call the newly extended system DisDSS 2.0, which integrates our framework to address the disaster relevancy identification issue. The output from the research study is helpful for disaster managers to make effective decisions on time. It bridges the gap between the decision-makers and citizens during disasters through the lens of deep learning.

How to cite: Singla, A., Agrawal, R., and Garg, A.: DisDSS 2.0: A Multi-Hazard Web-based Disaster Management System to Identify Disaster-Relevancy of a Social Media Message for Decision-Making Using Deep Learning Techniques, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-266,, 2022.

Background and objective: The fields of urban resilience to flooding and data science are on a collision course giving rise to the emerging field of smart resilience. The objective of this study is to propose and demonstrate a smart flood resilience framework that leverages various heterogeneous community-scale big data and infrastructure sensor data to enhance predictive risk monitoring and situational awareness.

Smart flood resilience framework: The smart flood resilience framework focuses on four core capabilities that could be augmented through the use of heterogeneous community-scale big data and analytics techniques: (1) predictive flood risk mapping: prediction capability of imminent flood risks (such as overflow of channels) to inform communities and emergency management agencies to take preparation and response actions; (2) automated rapid impact assessment: the ability to automatically and quickly evaluate the extent of flood impacts (i.e., physical, social, and economic impacts) to enable crisis responders and public officials to allocate relief and rescue resources on time; (3) predictive infrastructure failure prediction and monitoring: the ability to anticipate imminent failures in infrastructure systems as a flood event unfolds; and (4) smart situational awareness capabilities: the capability to derive proactive insights regarding the evolution of flood impacts (e.g., disrupted access to critical facilities and spatio-temporal patterns of recovery) on the communities.

Case study: We demonstrate the components of these core capabilities in the smart flood resilience framework in the context of the 2017 Hurricane Harvey in Harris. First, with Bayesian network modeling and deep learning methods, we reveal the use of flood sensor data for the prediction of floodwater overflow in channel networks and inundation of co-located road networks. Second, we discuss the use of social media data and machine learning techniques for assessing the impacts of floods on communities and sensing emotion signals to examine societal impacts. Third, we illustrate the use of high-resolution traffic data in network-theoretic models for now-casting of flood propagation on road networks and the disrupted access to critical facilities such as hospitals. Fourth, we leverage location-based and credit card transaction data in advanced spatial data analytics to proactively evaluate the recovery of communities and the impacts of floods on businesses.

Significances: This study shows that the significance of different core capabilities of the smart flood resilience framework in helping emergency managers, city planners, public officials, responders, and volunteers to better cope with the impacts of catastrophic flooding events.

How to cite: Mostafavi, A. and Yuan, F.: Smart Flood Resilience: Harnessing Community-Scale Big Data for Predictive Flood Risk Monitoring, Rapid Impact Assessment, and Situational Awareness, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-781,, 2022.


Operations Risk Insight (ORI) with Watson is an IBM AI application on the cloud.  ORI analyzes thousands of news sources and alert services daily.  There are too many data sources, warnings, watches and advisories for an individual to understand.  For example, during a week in 2021 with record wildfires, hurricanes and COVID hotspots across the US, thousands of impacting risk events hit key points of interest to IBM globally and were analyzed in real time.  

Which events impacted IBM’s business, and which didn’t? ORI has saved IBM millions of dollars annually for the past 5 years.  Our non-profit disaster relief partners have used ORI to respond more effectively to the needs of the vulnerable groups impacted by disasters.  Find out how disaster response leaders identify severe risks using Watson, the Hybrid Cloud, Big Data, Machine Learning and AI.

Presentation Objectives:

The objectives of this session are:

  • Educate the audience on a pragmatic and relevant IBM internal use case for an AI on the Cloud application, using many Watson and The Weather Company API's, plus machine learning running on IBM's cloud.
  • Obtain feedback and suggestions from the audience on how to expand and improve the machine learning and data analysis for this application to expanded the value for natural disaster response leaders. .
  • Inspire others to create their own grass roots cognitive project and learn more about AI and cloud technologies.
  • Discuss how this relates to the Call for Code and is used by Disaster Relief Agencies for free to assist the most vulnerable in society.

References Links:  

  • ORI has been featured in two Cloud Pak for Data (CP4D) workbooks:  CP4D Watson Studio Tutorial on Risk Analysis: and the Flood Risk Project:  Each demonstrate the application and methods for Machine Learning to be applied to AI for Natural Disaster Management (NDM). 
  • IBM use case for non-profit partners:
  • NC Tech article:
  • Supply Chain Management Review (SCMR) interview:
  • Supply Chain navigator article:

How to cite: Ward, T. and Kanwar, R.: IBM Operations Risk Insights with Watson:  a multi-hazard risk, AI for Natural Disaster Management use case, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1230,, 2022.

EGU22-1510 | Presentations | ITS2.5/NH10.8

From virtual environment to real observations: short-term hydrological forecasts with an Artificial Neural Network model. 

Renaud Jougla, Manon Ahlouche, Morgan Buire, and Robert Leconte

Machine learning model approaches for hydrological forecasts are nowadays common in research. Artificial Neural Network (ANN) is one of the most popular due to its good performance on watersheds with different hydrologic regimes and over several timescales. A short-term (1 to 7 days ahead) forecast model was explored to predict streamflow. This study focused on the summer season defined from May to October. Cross-validation was done over a period of 16 years, each time keeping a single year as a validation set.

The ANN model was parameterized with a single hidden layer of 6 neurons. It was developed in a virtual environment based on datasets generated by the physically based distributed hydrological model Hydrotel (Fortin et al., 2012). In a preliminary analysis, several combinations of inputs were assessed, the best combining precipitation and temperature with surface soil moisture and antecedent streamflow. Different spatial discretizations were compared. A semi-distributed discretization was selected to facilitate transferring the ANN model from a virtual environment to real observations such as remote sensing soil moisture products or ground station time series.

Four watersheds were under study: the Au Saumon and Magog watersheds located in south Québec (Canada); the Androscoggin watershed in Maine (USA); and the Susquehanna watershed located in New-York and Pennsylvania (USA). All but the Susquehanna watershed are mainly forested, while the latter has a 57% forest cover. To evaluate whether a model with a data-driven structure can mimic a deterministic model, ANN and Hydrotel simulated flows were compared. Results confirm that the ANN model can reproduce streamflow output from Hydrotel with confidence.

Soil moisture observation stations were deployed in the Au Saumon and Magog watersheds during the summers 2018 to 2021. Meteorological data were extracted from the ERA5-Land reanalysis dataset. As the period of availability of observed data is short, the ANN model was trained in a virtual environment. Two validations were done: one in the virtual environment and one using real soil moisture observations and flows. The number and locations of the soil moisture probes slightly differed during each of the four summers. Therefore, four models were trained depending on the number of probes and their location. Results highlight that location of the soil moisture probes has a large influence on the ANN streamflow outputs and identifies more representative sub-regions of the watershed.

The use of remote sensing data as inputs of the ANN model is promising. Soil moisture datasets from SMOS and SMAP missions are available for the four watersheds under study, although downscaling approaches should be applied to bring the spatial resolution of those products at the watershed scale. One other future lead could be the development of a semi-distributed ANN model in virtual environment based on a restricted selection of hydrological units based on physiographic characteristics. The future L-band NiSAR product could be relevant for this purpose, having a finer spatial resolution compared to SMAP and SMOS and a better penetration of the signal in forested areas than C-band SAR satellites such as Sentinel-1 and the Radarsat Constellation Mission.

How to cite: Jougla, R., Ahlouche, M., Buire, M., and Leconte, R.: From virtual environment to real observations: short-term hydrological forecasts with an Artificial Neural Network model., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1510,, 2022.

Tropical Cyclones (TCs) are deadly but rare events that cause considerable loss of life and property damage every year. Traditional TC forecasting and tracking methods focus on numerical forecasting models, synoptic forecasting and statistical methods. However, in recent years there have been several studies investigating applications of Deep Learning (DL) methods for weather forecasting with encouraging results.

We aim to test the efficacy of several DL methods for TC nowcasting, particularly focusing on Generative Adversarial Neural Networks (GANs) and Recurrent Neural Networks (RNNs). The strengths of these network types align well with the given problem: GANs are particularly apt to learn the form of a dataset, such as the typical shape and intensity of a TC, and RNNs are useful for learning timeseries data, enabling a prediction to be made based on the past several timesteps.

The goal is to produce a DL based pipeline to predict the future state of a developing cyclone with accuracy that measures up to current methods.  We demonstrate our approach based on learning from high-resolution numerical simulations of TCs from the Indian and Pacific oceans and discuss the challenges and advantages of applying these DL approaches to large high-resolution numerical weather data.

How to cite: Steptoe, H. and Xirouchaki, T.: Deep Learning for Tropical Cyclone Nowcasting: Experiments with Generative Adversarial and Recurrent Neural Networks, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1650,, 2022.

EGU22-1662 | Presentations | ITS2.5/NH10.8

Exploring the challenges of Digital Twins for weather & climate through an Atmospheric Dispersion modelling prototype 

Stephen Haddad, Peter Killick, Aaron Hopkinson, Tomasz Trzeciak, Mark Burgoyne, and Susan Leadbetter

Digital Twins present a new user-centric paradigm for developing and using weather & climate simulations that is currently being widely embraced, for example through large projects such as Destination Earth led by ECMWF.  In this project we have taken a smaller scale approach in understanding the opportunities and challenges in translating the Digital Twin concept from the original domain of manufacturing and the built environment to modelling of the earth’s atmosphere.

We describe our approach to creating a Digital Twin based on the Met Office’s Atmospheric Dispersion simulation package called NAME. We will discuss the advantages of doing this, such as the ability of nonexpert users to more easily produce scientifically valid simulations of dispersion events, such as industrial fires, and easily obtain results to feed into downstream analysis, for example of health impacts. We will describe the requirements of each of the key components of a digital twin and potential implementation approaches.

We will describe how a Digital Twin framework enables multiple models to be joined together to model complex systems as required for atmospheric concentrations around chemical spills or fires modelled by NAME. Overall, we outline a potential project blueprint for future work to improve usability and scientific throughput of existing modelling systems by creating a Digital Twins from core current modelling code and data gathering systems.

How to cite: Haddad, S., Killick, P., Hopkinson, A., Trzeciak, T., Burgoyne, M., and Leadbetter, S.: Exploring the challenges of Digital Twins for weather & climate through an Atmospheric Dispersion modelling prototype, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1662,, 2022.

Massive groundwater pumping for agricultural and industrial activities results in significant land subsidence in the arid world. In an acute water crisis, monitoring land subsidence and its key drivers is essential to assist groundwater depletion mitigation strategy. Physical models for aquifer simulation related to land deformation are computationally expensive. The interferometric synthetic aperture radar (InSAR) technique provides precise deformation mapping yet is affected by tropospheric and ionospheric errors. This study explores the capabilities of the deep learning approach coupled with satellite-derived variables in modeling subsidence, spatially and temporally, from 2016 to 2020 and predicting subsidence in the near future by using a recurrent neural network (RNN) in the Shabestar basin, Iran. The basin is part of the Urmia Lake River Basin, embracing 6.4 million people, yet has been primarily desiccated due to the over-usage of water resources in the basin. The deep learning model incorporates InSAR-derived land subsidence and its satellite-based key drivers such as actual evapotranspiration, Normalized Difference Vegetation Index (NDVI), land surface temperature, precipitation to yield the importance of critical drivers to inform groundwater governance. The land deformation in the area varied between -93.2 mm/year to 16 mm/year on average in 2016-2020. Our findings reveal that precipitation, evapotranspiration, and vegetation coverage primarily affected land subsidence; furthermore, the subsidence rate is predicted to increase rapidly. The phenomenon has the same trend with the variation of the Urmia Lake level. This study demonstrates the potential of artificial intelligence incorporating satellite-based ancillary data in land subsidence monitoring and prediction and contributes to future groundwater management.

How to cite: Zhang, Y. and Hashemi, H.: InSAR-Deep learning approach for simulation and prediction of land subsidence in arid regions, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2011,, 2022.

EGU22-2879 | Presentations | ITS2.5/NH10.8

Automatically detecting avalanches with machine learning in optical SPOT6/7 satellite imagery 

Elisabeth D. Hafner, Patrick Barton, Rodrigo Caye Daudt, Jan Dirk Wegner, Konrad Schindler, and Yves Bühler

Safety related applications like avalanche warning or risk management depend on timely information about avalanche occurrence. Knowledge on the locations and sizes of avalanches releasing is crucial for the responsible decision-makers. Such information is still collected today in a non-systematic way by observes in the field, for example from ski resort patrols or community avalanche services. Consequently, the existing avalanche mapping is, in particular in situations with high avalanche danger, strongly biased towards accessible terrain in proximity to (winter sport) infrastructure.

Recently, remote sensing has been shown to be capable of partly filling this gap, providing spatially continuous information on avalanche occurrences over large regions. In previous work we applied optical SPOT 6/7 satellite imagery to manually map two avalanche periods over a large part of the swiss Alps (2018: 12’500 and 2019: 9’500 km2). Subsequently, we investigated the reliability of this mapping and proved its suitability by identifying almost ¾ of all occurred avalanches (larger size 1) from SPOT 6/7 imagery. Therefore, optical SPOT data is an excellent source for continuous avalanche mapping, currently restricted by the time intensive manual mapping. To speed up this process we now propose a fully convolutional neural network (CNN) called AvaNet. AvaNet is based on a Deeplabv3+ architecture adapted to specifically learn how avalanches look like by explicitly including height information from a digital terrain model (DTM) for example. Relying on the manually mapped 24’737 avalanches for training, validation and testing, AvaNet achieves an F1 score of 62.5% when thresholding the probabilities from the network predictions at 0.5. In this study we present the results from our network in more detail, including different model variations and results of predictions on data from a third avalanche period we did not train on.

The ability to automate the mapping and therefor quickly identify avalanches from satellite imagery is an important step forward in regularly acquiring spatially continuous avalanche occurrence data. This enables the provision of essential information for the complementation of avalanche databases, making Alpine regions safer.

How to cite: Hafner, E. D., Barton, P., Caye Daudt, R., Wegner, J. D., Schindler, K., and Bühler, Y.: Automatically detecting avalanches with machine learning in optical SPOT6/7 satellite imagery, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2879,, 2022.

EGU22-3212 | Presentations | ITS2.5/NH10.8

Predicting Landslide Susceptibility in Cross River State of Nigeria using Machine Learning 

Joel Efiong, Devalsam Eni, Josiah Obiefuna, and Sylvia Etu

Landslides have continued to wreck its havoc in many parts of the globe; comprehensive studies of landslide susceptibilities of many of these areas are either lacking or inadequate. Hence, this study was aimed at predicting landslide susceptibility in Cross River State of Nigeria, using machine learning. Precisely, the frequency ratio (FR) model was adopted in this study. In adopting this approach, a landslide inventory map was developed using 72 landslide locations identified during fieldwork combined with other relevant data sources. Using appropriate geostatistical analyst tools within a geographical information environment, the landslide locations were randomly divided into two parts in the ratio of 7:3 for the training and validation processes respectively. A total of 12 landslide causing factors, such as; elevation, slope, aspect, profile curvature, plan curvature, topographic position index, topographic wetness index, stream power index, land use/land cover, geology, distance to waterbody and distance to major roads, were selected and used in the spatial relationship analysis of the factors influencing landslide occurrences in the study area. FR model was then developed using the training sample of the landslide to investigate landslide susceptibility in Cross River State which was subsequently validated. It was found out that the distribution of landslides in Cross River State of Nigeria was largely controlled by a combined effect of geo-environmental factors such as elevation of 250 – 500m, slope gradient of >35o, slopes facing the southwest direction, decreasing degree of both positive and negative curvatures, increasing values of topographic position index, fragile sands, sparse vegetation, especially in settlement and bare surfaces areas, distance to waterbody and major road of < 500m. About 46% of the mapped area was found to be at landslide susceptibility risk zones, ranging from moderate – very high levels. The susceptibility model was validated with 90.90% accuracy. This study has shown a comprehensive investigation of landslide susceptibility in Cross River State which will be useful in land use planning and mitigation measures against landslide induced vulnerability in the study area including extrapolation of the findings to proffer solutions to other areas with similar environmental conditions. This is a novel use of a machine learning technique in hazard susceptibility mapping.


Keywords: Landslide; Landslide Susceptibility mapping; Cross River State, Nigeria; Frequency ratio, Machine learning

How to cite: Efiong, J., Eni, D., Obiefuna, J., and Etu, S.: Predicting Landslide Susceptibility in Cross River State of Nigeria using Machine Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3212,, 2022.

EGU22-3283 | Presentations | ITS2.5/NH10.8

Assessment of Flood-Damaged Cropland Trends Under Future Climate Scenarios Using Convolutional Neural Network 

Rehenuma Lazin, Xinyi Shen, and Emmanouil Anagnostou

Every year flood causes severe damages in the cropland area leading to global food insecurity. As climate change continues, floods are predicted to be more frequent in the future. To cope with the future climate impacts, mitigate damages, and ensure food security, it is now imperative to study the future flood damage trends in the cropland area. In this study, we use a convolutional neural network (CNN) to estimate the damages (in acre) in the corn and soybean lands across the mid-western USA with projections from climate models. Here, we extend the application of the CNN model developed by Lazin et. al, (2021) that shows ~25% mean relative error for county-level flood-damaged crop loss estimation. The meteorological variables are derived from the reference gridMet datasets as predictors to train the model from 2008-2020. We then use downscaled climate projections from Multivariate Adaptive Constructed Analogs (MACA) dataset in the trained CNN model to assess future flood damage patterns in the cropland in the early (2011-2040), mid (2041-2070), and late (2071-2100) century, relative to the baseline historical period (1981-2010). Results derived from this study will help understand the crop loss trends due to floods under climate change scenarios and plan necessary arrangements to mitigate damages in the future.



[1] Lazin, R., Shen, X., & Anagnostou, E. (2021). Estimation of flood-damaged cropland area using a convolutional neural network. Environmental Research Letters16(5), 054011.

How to cite: Lazin, R., Shen, X., and Anagnostou, E.: Assessment of Flood-Damaged Cropland Trends Under Future Climate Scenarios Using Convolutional Neural Network, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3283,, 2022.

EGU22-3422 | Presentations | ITS2.5/NH10.8

Weather history encoding for machine learning-based snow avalanche detection 

Thomas Gölles, Kathrin Lisa Kapper, Stefan Muckenhuber, and Andreas Trügler

Since its start in 2014, the Copernicus Sentinel-1 programme has provided free of charge, weather independent, and high-resolution satellite Earth observations and has set major scientific advances in the detection of snow avalanches from satellite imagery in motion. Recently, operational avalanche detection from Sentinel-1 synthetic Aperture radar (SAR) images were successfully introduced for some test regions in Norway. However, current state of the art avalanche detection algorithms based on machine learning do not include weather history. We propose a novel way to encode weather data and include it into an automatic avalanche detection pipeline for the Austrian Alps. The approach consists of four steps. At first the raw data in netCDF format is downloaded, which consists of several meteorological parameters over several time steps. In the second step the weather data is downscaled onto the pixel locations of the SAR image. Then the data is aggregated over time, which produces a two-dimensional grid of one value per SAR pixel at the time when the SAR data was recorded. This aggregation function can range from simple averages to full snowpack models. In the final step, the grid is then converted to an image with greyscale values corresponding to the aggregated values. The resulting image is then ready to be fed into the machine learning pipeline. We will include this encoded weather history data to increase the avalanche detection performance, and investigate contributing factors with model interpretability tools and explainable artificial intelligence.

How to cite: Gölles, T., Kapper, K. L., Muckenhuber, S., and Trügler, A.: Weather history encoding for machine learning-based snow avalanche detection, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3422,, 2022.

EGU22-4250 | Presentations | ITS2.5/NH10.8

Landslide Susceptibility Modeling of an Escarpment in Southern Brazil using Artificial Neural Networks as a Baseline for Modeling Triggering Rainfall 

Luísa Vieira Lucchese, Guilherme Garcia de Oliveira, Alexander Brenning, and Olavo Correa Pedrollo

Landslide Susceptibility Mapping (LSM) and rainfall thresholds are well-documented tools used to model the occurrence of rainfall-induced landslides. In the case of locations where only rainfall can be considered a main landslide trigger, both methodologies apply essentially to the same locations, and a model that encompasses both would be an important step towards a better understanding and prediction of landslide-triggering rainfall events. In this research, we employ spatially cross-validated, hyperparameter tuned Artificial Neural Networks (ANNs) to predict the susceptibility to landslides of an area in southern Brazil. In a next step, we plan to add the triggering rainfall to this Artificial Intelligence model, which will concurrently model the susceptibility and the triggering rainfall event for a given area. The ANN is of type Multi-Layer Perceptron with three layers. The number of neurons in the hidden layer was tuned separately for each cross-validation fold, using a method described in previous work. The study area is the escarpment in the limits of the municipalities of Presidente Getúlio, Rio do Sul, and Ibirama, in southern Brazil. For this area, 82 landslides scars related to the event of December 17th, 2020, were mapped. The metrics for each fold are presented and the final susceptibility map for the area is shown and analyzed. The evaluation metrics attained are satisfactory and the resulting susceptibility map highlights the escarpment areas as most susceptible to landslides. The ANN-based susceptibility mapping in the area is considered successful and seen as a baseline for identifying rainfall thresholds in susceptible areas, which will be accomplished with a combined susceptibility and rainfall model in our future work.

How to cite: Vieira Lucchese, L., Garcia de Oliveira, G., Brenning, A., and Correa Pedrollo, O.: Landslide Susceptibility Modeling of an Escarpment in Southern Brazil using Artificial Neural Networks as a Baseline for Modeling Triggering Rainfall, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4250,, 2022.

EGU22-4266 | Presentations | ITS2.5/NH10.8

Camera Rain Gauge Based on Artificial Intelligence 

Raffaele Albano, Nicla Notarangelo, Kohin Hirano, and Aurelia Sole

Flood risk monitoring, alert and adaptation in urban areas require near-real-time fine-scale precipitation observations that are challenging to obtain from currently available measurement networks due to their costs and installation difficulties. In this sense, newly available data sources and computational techniques offer enormous potential, in particular, the exploiting of not-specific, widespread, and accessible devices.

This study proposes an unprecedented system for rainfall monitoring based on artificial intelligence, using deep learning for computer vision, applied to cameras images. As opposed to literature, the method is not device-specific and exploits general-purpose cameras (e.g., smartphones, surveillance cameras, dashboard cameras, etc.), in particular, low-cost device, without requiring parameter setting, timeline shots, or videos. Rainfall is measured directly from single photographs through Deep Learning models based on transfer learning with Convolutional Neural Networks. A binary classification algorithm is developed to detect the presence of rain. Moreover, a multi-class classification algorithm is used to estimate a quasi-instantaneous rainfall intensity range. Open data, dash-cams in Japan coupled with high precision multi-parameter radar XRAIN, and experiments in the NIED Large Scale Rainfall Simulator combined to form heterogeneous and verisimilar datasets for training, validation, and test. Finally, a case study over the Matera urban area (Italy) was used to illustrate the potential and limitations of rainfall monitoring using camera-based detectors.

The prototype was deployed in a real-world operational environment using a pre-existent 5G surveillance camera. The results of the binary classifier showed great robustness and portability: the accuracy and F1-score value were 85.28% and 85.13%, 0.86 and 0.85 for test and deployment, respectively, whereas the literature algorithms suffer from drastic accuracy drops changing the image source (e.g. from 91.92% to 18.82%). The 6-way classifier results reached test average accuracy and macro-averaged F1 values of 77.71% and 0.73, presenting the best performances with no-rain and heavy rainfall, which represents critical condition for flood risk. Thus, the results of the tests and the use-case demonstrate the model’s ability to detect a significant meteorological state for early warning systems. The classification can be performed on single pictures taken in disparate lighting conditions by common acquisition devices, i.e. by static or moving cameras without adjusted parameters. This system does not suit scenes that are also misleading for human visual perception. The proposed method features readiness level, cost-effectiveness, and limited operational requirements that allow an easy and quick implementation by exploiting pre-existent devices with a parsimonious use of economic and computational resources.

Altogether, this study corroborates the potential of non-traditional and opportunistic sensing networks for the development of hydrometeorological monitoring systems in urban areas, where traditional measurement methods encounter limitations, and in data-scarce contexts, e.g. where remote-sensed rainfall information is unavailable or has broad resolution respect with the scale of the proposed study. Future research will involve incremental learning algorithms and further data collection via experiments and crowdsourcing, to improve accuracy and at the same time promote public resilience from a smart city perspective.

How to cite: Albano, R., Notarangelo, N., Hirano, K., and Sole, A.: Camera Rain Gauge Based on Artificial Intelligence, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4266,, 2022.

EGU22-4730 | Presentations | ITS2.5/NH10.8

floodGAN – A deep learning-based model for rapid urban flood forecasting 

Julian Hofmann and Holger Schüttrumpf

Recent urban flood events revealed how severe and fast the impacts of heavy rainfall can be. Pluvial floods pose an increasing risk to communities worldwide due to ongoing urbanization and changes in climate patterns. Still, pluvial flood warnings are limited to meteorological forecasts or water level monitoring which are insufficient to warn people against the local and terrain-specific flood risks. Therefore, rapid flood models are essential to implement effective and robust early warning systems to mitigate the risk of pluvial flooding. Although hydrodynamic (HD) models are state-of-the-art for simulation pluvial flood hazards, the required computation times are too long for real-time applications.

In order to overcome the computation time bottleneck of HD models, the deep learning model floodGAN has been developed. FloodGAN combines two adversarial Convolutional Neural Networks (CNN) that are trained on high-resolution rainfall-flood data generated from rainfall generators and HD models. FloodGAN translates the flood forecasting problem into an image-to-image translation task whereby the model learns the non-linear spatial relationships of rainfall and hydraulic data. Thus, it directly translates spatially distributed rainfall forecasts into detailed hazard maps within seconds. Next to the inundation depth, the model can predict the velocities and time periods of hydraulic peaks of an upcoming rainfall event. Due to its image-translation approach, the floodGAN model can be applied for large areas and can be run on standard computer systems, fulfilling the tasks of fast and practical flood warning systems.

To evaluate the accuracy and generalization capabilities of the floodGAN model, numerous performance tests were performed using synthetic rainfall events as well as a past heavy rainfall event of 2018. Therefore, the city of Aachen was used as a case study. Performance tests demonstrated a speedup factor of 106 compared to HD models while maintaining high model quality and accuracy and good generalization capabilities for highly variable rainfall events. Improvements can be obtained by integrating recurrent neural network architectures and training with temporal rainfall series to forecast the dynamics of the flooding processes.

How to cite: Hofmann, J. and Schüttrumpf, H.: floodGAN – A deep learning-based model for rapid urban flood forecasting, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4730,, 2022.

EGU22-4900 | Presentations | ITS2.5/NH10.8

A modular and scalable workflow for data-driven modelling of shallow landslide susceptibility 

Ann-Kathrin Edrich, Anil Yildiz, Ribana Roscher, and Julia Kowalski

The spatial impact of a single shallow landslide is small compared to a deep-seated, impactful failure and hence its damage potential localized and limited. Yet, their higher frequency of occurrence and spatio-temporal correlation in response to external triggering events such as strong precipitation, nevertheless result in dramatic risks for population, infrastructure and environment. It is therefore essential to continuously investigate and analyze the spatial hazard that shallow landslides pose. Its visualisation through regularly-updated, dynamic hazard maps can be used by decision and policy makers. Even though a number of data-driven approaches for shallow landslide hazard mapping exist, a generic workflow has not yet been described. Therefore, we introduce a scalable and modular machine learning-based workflow for shallow landslide hazard prediction in this study. The scientific test case for the development of the workflow investigates the rainfall-triggered shallow landslide hazard in Switzerland. A benchmark dataset was compiled based on a historic landslide database as presence data, as well as a pseudo-random choice of absence locations, to train the data-driven model. Features included in this dataset comprise at the current stage 14 parameters from topography, soil type, land cover and hydrology. This work also focuses on the investigation of a suitable approach to choose absence locations and the influence of this choice on the predicted hazard as their influence is not comprehensively studied. We aim at enabling time-dependent and dynamic hazard mapping by incorporating time-dependent precipitation data into the training dataset with static features. Inclusion of temporal trigger factors, i.e. rainfall, enables a regularly-updated landslide hazard map based on the precipitation forecast. Our approach includes the investigation of a suitable precipitation metric for the occurrence of shallow landslides at the absence locations based on the statistical evaluation of the precipitation behavior at the presence locations. In this presentation, we will describe the modular workflow as well as the benchmark dataset and show preliminary results including above mentioned approaches to handle absence locations and time-dependent data.

How to cite: Edrich, A.-K., Yildiz, A., Roscher, R., and Kowalski, J.: A modular and scalable workflow for data-driven modelling of shallow landslide susceptibility, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4900,, 2022.

EGU22-6568 | Presentations | ITS2.5/NH10.8

Harnessing Machine Learning and Deep Learning applications for climate change risk assessment: a survey 

Davide Mauro Ferrario, Elisa Furlan, Silvia Torresan, Margherita Maraschini, and Andrea Critto

In the last years there has been a growing interest around Machine Learning (ML) in climate risk/ multi-risk assessment, steered mainly by the growing amount of data available and the reduction of associated computational costs. Extracting information from spatio-temporal data is critically important for problems such as extreme events forecasting and assessing risks and impacts from multiple hazards. Typical challenges in which AI and ML are now being applied require understanding the dynamics of complex systems, which involve many features with non-linear relations and feedback loops, analysing the effects of phenomena happening at different time scales, such as slow-onset events (sea level rise) and short-term episodic events (storm surges, floods) and estimating uncertainties of long-term predictions and scenarios. 
While in the last years there were many successful applications of AI/ML, such as Random Forest or Long-Short Term Memory (LSTM) in floods and storm surges risk assessment, there are still open questions and challenges that need to be addressed. In fact, there is a lack of data for extreme events and Deep Learning (DL) algorithms often need huge amounts of information to disentangle the relationships among hazard, exposure and vulnerability factors contributing to the occurrence of risks. Moreover, the spatio-temporal resolution can be highly irregular and need to be reconstructed to produce accurate and efficient models. For example, using data from meteorological ground stations can offer accurate datasets with fine temporal resolution, but with an irregular distribution in the spatial dimension; on the other hand, leveraging on satellite images can give access to more spatially refined data, but often lacking the temporal dimension (fewer events available to due atmospheric disturbances). 
Several techniques have been applied, ranging from classical multi-step forecasting, state-space and Hidden Markov models to DL techniques, such as Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). ANN and Deep Generative Models (DGM) have been used to reconstruct spatio-temporal grids and modelling continuous time-series, CNN to exploit spatial relations, Graph Neural Networks (GNN) to extract multi-scale localized spatial feature and RNN and LSTM for multi-scale time series prediction.  
To bridge these gaps, an in-depth state-of-the-art review of the mathematical and computer science innovations in ML/DL techniques that could be applied to climate /multi-risk assessment was undertaken. The review focuses on three possible ML/DL applications: analysis of spatio-temporal dynamics of risk factors, with particular attention on applications for irregular spatio-temporal grids; multivariate analysis for multi-hazard interactions and multiple risk assessment endpoints; analysis of future scenarios under climate change. We will present the main outcomes of the scientometric and systematic review of publications across the 2000- 2021 timeframe, which allowed us to: i) summarize keywords and word co-occurrence networks, ii) highlight linkages, working relations and co-citation clusters, iii) compare ML and DL approaches with classical statistical techniques and iv) explore applications at the forefront of the risk assessment community.

How to cite: Ferrario, D. M., Furlan, E., Torresan, S., Maraschini, M., and Critto, A.: Harnessing Machine Learning and Deep Learning applications for climate change risk assessment: a survey, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6568,, 2022.

EGU22-6576 | Presentations | ITS2.5/NH10.8

Swept Away: Flooding and landslides in Mexican poverty nodes 

Silvia García, Raul Aquino, and Walter Mata

Natural disasters should be examined within a risk-perspective framework where both natural threat and vulnerability are considered as intricate components of an extremely complex equation. The trend toward more frequent floods and landslides in Mexico in recent decades is not only the result of more intense rainfall, but also a consequence of increased vulnerability. As a multifactorial element, vulnerability is a low-frequency modulating factor of the risk dynamics to intense rainfall. It can be described in terms of physical, social, and economical factors. For instance, deforested or urbanized areas are the physical and social factors that lead to the deterioration of watersheds and an increased vulnerability to intense rains. Increased watershed vulnerability due to land-cover changes is the primary factor leading to more floods, particularly over pacific Mexico. ln some parts of the country, such as Colima, the increased frequency of intense rainfall (i.e., natural hazard) associated with high-intensity tropical cyclones and hurricanes is the leading cause of more frequent floods.


In this research an intelligent rain management-system is presented. The object is built to forecast and to simulate the components of risk, to stablish communication between rescue/aid teams and to help in preparedness activities (training). Detection, monitoring, analysis and forecasting of the hazards and scenarios that promote floods and landslides, is the main task. The developed methodology is based on a database that permits to relate heavy rainfall measurements with changes in land cover and use, terrain slope, basin compactness and communities’ resilience as key vulnerability factors. A neural procedure is used for the spatial definition of exposition and susceptibility (intrinsic and extrinsic parameters) and Machine Learning techniques are applied to find the If-Then relationships. The capability of the intelligent model for Colima, Mexico was tested by comparing the observed and modeled frequency of landslides and floods for ten years period. It was found that over most of the Mexican territory, more frequent floods are the result of a rapid deforestation process and that landslides and their impact on communities are directly related to the unauthorized growth of populations in high geo-risk areas (due to forced migration because of violence or extreme poverty) and the development of civil infrastructure (mainly roads) with a high impact on the natural environment. Consequently, the intelligent rain-management system offers the possibility to redesign and to plan the land use and the spatial distribution of poorest communities.

How to cite: García, S., Aquino, R., and Mata, W.: Swept Away: Flooding and landslides in Mexican poverty nodes, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6576,, 2022.

EGU22-6690 | Presentations | ITS2.5/NH10.8

A machine learning-based ensemble model for estimation of seawater quality parameters in coastal area 

Xiaotong Zhu, Jinhui Jeanne Huang, Hongwei Guo, Shang Tian, and Zijie Zhang

The precise estimation of seawater quality parameters is crucial for decision-makers to manage coastal water resources. Although various machine learning (ML)-based algorithms have been developed for seawater quality retrieval using remote sensing technology, the performance of these models in the application of specific regions remains significant uncertainty due to the different properties of coastal waters. Moreover, the prediction results of these ML models are unexplainable. To address these problems, an ML-based ensemble model was developed in this study. The model was applied to estimate chlorophyll-a (Chla), turbidity, and dissolved oxygen (DO) based on Sentinel-2 satellite imagery in Shenzhen Bay, China. The optimal input features for each seawater quality parameter were selected from the nine simulation scenarios which generated from eight spectral bands and six spectral indices. A local explanation method called SHapley Additive exPlanations (SHAP) was introduced to quantify the contributions of various features to the predictions of the seawater quality parameters. The results suggested that the ensemble model with feature selection enhanced the performance for three types of seawater quality parameters estimations (The errors were 1.7%, 1.5%, and 0.02% for Chla, turbidity, and DO, respectively). Furthermore, the reliability of the model performance was further verified for mapping the spatial distributions of water quality parameters during the model validation period. The spatial-temporal patterns of seawater quality parameters revealed that the distributions of seawater quality were mainly influenced by estuary input. Correlation analysis demonstrated that air temperature (Temp) and average air pressure (AAP) exhibited the closest relationship with Chla. The DO was most relevant with Temp, and turbidity was not sensitive to Temp, average wind speed (AWS), and AAP. This study enhanced the prediction capability of seawater quality parameters and provided a scientific coastal waters management approach for decision-makers.

How to cite: Zhu, X., Huang, J. J., Guo, H., Tian, S., and Zhang, Z.: A machine learning-based ensemble model for estimation of seawater quality parameters in coastal area, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6690,, 2022.

EGU22-6758 | Presentations | ITS2.5/NH10.8

AI-enhanced Integrated Alert System for effective Disaster Management 

Pankaj Kumar Dalela, Saurabh Basu, Sandeep Sharma, Anugandula Naveen Kumar, Suvam Suvabrata Behera, and Rajkumar Upadhyay

Effective communication systems supported by Information and Communication Technologies (ICTs) are integral and important components for ensuring comprehensive disaster management. Continuous warning monitoring, prediction, dissemination, and response coordination along with public engagement by utilizing the capabilities of emerging technologies including Artificial Intelligence (AI) can assist in building resilience and ensuring Disaster Risk Reduction. Thus, for effective disaster management, an Integrated Alert System is proposed which encapsulates all concerned disaster management authorities, alert forecasting and disseminating agencies under a single umbrella for alerting the targeted public through various communication channels. Enhancing the capabilities of the system through AI, its integral part includes the data-driven citizen-centric Decision Support System which can help disaster managers by performing complete impact assessment of disaster events through configuration of decision models developed by learning inter-relationships of different parameters. The system needs to be capable of identification of possible communication means to address community outreach, prediction of scope of alert, providing influence of alert message on targeted vulnerable population, performing crowdsourced data analysis, evaluating disaster impact through threat maps and dashboards, and thereby, providing complete analysis of the disaster event in all phases of disaster management. The system aims to address challenges including limited communication channels utilization and audience reach, language differences, and lack of ground information in decision making posed by current systems by utilizing the latest state of art technologies.

How to cite: Dalela, P. K., Basu, S., Sharma, S., Kumar, A. N., Behera, S. S., and Upadhyay, R.: AI-enhanced Integrated Alert System for effective Disaster Management, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6758,, 2022.

Main purpose of current research article is to present latest findings on automatic methods of manipulating social network data for developing seismic intensity maps. As case study the author selected the 2020 Samos earthquake event (Mw= 7, 30 October 2020, Greece). That earthquake event had significant consequences to the urban environment along with 2 deaths and 19 injuries. Initially an automatic approach, presented recently in the international literature was applied producing thus seismic intensity maps from tweets. Furthermore, some initial findings regarding the use of machine learning in various parts of the automatic methodology were presented along with potential of using photos posted in social networks. The data used were several thousands tweets and instagram posts.The results, provide vital findings in enriching data sources, data types, and effective rapid processing.

How to cite: Arapostathis, S. G.: The Samos earthquake event (Mw = 7, 30 October 2020, Greece) as case study for applying machine learning on texts and photos scraped from social networks for developing seismic intensity maps., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7129,, 2022.

EGU22-7308 | Presentations | ITS2.5/NH10.8

Building an InSAR-based database to support geohazard risk management by exploiting large ground deformation datasets 

Marta Béjar-Pizarro, Pablo Ezquerro, Carolina Guardiola-Albert, Héctor Aguilera Alonso, Margarita Patricia Sanabria Pabón, Oriol Monserrat, Anna Barra, Cristina Reyes-Carmona, Rosa Maria Mateos, Juan Carlos García López Davalillo, Juan López Vinielles, Guadalupe Bru, Roberto Sarro, Jorge Pedro Galve, Roberto Tomás, Virginia Rodríguez Gómez, Joaquín Mulas de la Peña, and Gerardo Herrera

The detection of areas of the Earth’s surface experiencing active deformation processes and the identification of the responsible phenomena (e.g. landslides activated after rainy events, subsidence due to groundwater extraction in agricultural areas, consolidation settlements, instabilities in active or abandoned mines) is critical for geohazard risk management and ultimately to mitigate the unwanted effects on the affected populations and the environment.

This will now be possible at European level thanks to the Copernicus European Ground Motion Service (EGMS), which will provide ground displacement measurements derived from time series analyses of Sentinel-1 data, using Interferometric Synthetic Aperture Radar (InSAR). The EGMS, which will be available to users in the first quarter of 2022 and will be updated annually, will be especially useful to identify displacements associated to landslides, subsidence and deformation of infrastructure.  To fully exploit the capabilities of this large InSAR datasets, it is fundamental to develop automatic analysis tools, such as machine learning algorithms, which require an InSAR-derived deformation database to train and improve them.  

Here we present the preliminary InSAR-derived deformation database developed in the framework of the SARAI project, which incorporates the previous InSAR results of the IGME-InSARlab and CTTC teams in Spain. The database contains classified points of measurement with the associated InSAR deformation and a set of environmental variables potentially correlated with the deformation phenomena, such as geology/lithology, land-surface slope, land cover, meteorological data, population density, and inventories such as the mining registry, the groundwater database, and the IGME’s land movements database (MOVES). We discuss the main strategies used to identify and classify pixels and areas that are moving, the covariables used and some ideas to improve the database in the future. This work has been developed in the framework of project PID2020-116540RB-C22 funded by MCIN/ AEI /10.13039/501100011033 and e-Shape project, with funding from the European Union’s Horizon 2020 research and innovation program under grant agreement 820852.

How to cite: Béjar-Pizarro, M., Ezquerro, P., Guardiola-Albert, C., Aguilera Alonso, H., Sanabria Pabón, M. P., Monserrat, O., Barra, A., Reyes-Carmona, C., Mateos, R. M., García López Davalillo, J. C., López Vinielles, J., Bru, G., Sarro, R., Galve, J. P., Tomás, R., Rodríguez Gómez, V., Mulas de la Peña, J., and Herrera, G.: Building an InSAR-based database to support geohazard risk management by exploiting large ground deformation datasets, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7308,, 2022.

EGU22-7313 | Presentations | ITS2.5/NH10.8

The potential of automated snow avalanche detection from SAR images for the Austrian Alpine region using a learning-based approach 

Kathrin Lisa Kapper, Stefan Muckenhuber, Thomas Goelles, Andreas Trügler, Muhamed Kuric, Jakob Abermann, Jakob Grahn, Eirik Malnes, and Wolfgang Schöner

Each year, snow avalanches cause many casualties and tremendous damage to infrastructure. Prevention and mitigation mechanisms for avalanches are established for specific regions only. However, the full extent of the overall avalanche activity is usually barely known as avalanches occur in remote areas making in-situ observations scarce. To overcome these challenges, an automated avalanche detection approach using the Copernicus Sentinel-1 synthetic aperture radar (SAR) data has recently been introduced for some test regions in Norway. This automated detection approach from SAR images is faster and gives more comprehensive results than field-based detection provided by avalanche experts. The Sentinel-1 programme has provided - and continues to provide - free of charge, weather-independent, and high-resolution satellite Earth observations since its start in 2014. Recent advances in avalanche detection use deep learning algorithms to improve the detection rates. Consequently, the performance potential and the availability of reliable training data make learning-based approaches an appealing option for avalanche detection.  

         In the framework of the exploratory project SnowAV_AT, we intend to build the basis for a state-of-the-art automated avalanche detection system for the Austrian Alps, including a "best practice" data processing pipeline and a learning-based approach applied to Sentinel-1 SAR images. As a first step towards this goal, we have compiled several labelled training datasets of previously detected avalanches that can be used for learning. Concretely, these datasets contain 19000 avalanches that occurred during a large event in Switzerland in January 2018, around 6000 avalanches that occurred in Switzerland in January 2019, and around 800 avalanches that occurred in Greenland in April 2016. The avalanche detection performance of our learning-based approach will be quantitatively evaluated against held-out test sets. Furthermore, we will provide qualitative evaluations using SAR images of the Austrian Alps to gauge how well our approach generalizes to unseen data that is potentially differently distributed than the training data. In addition, selected ground truth data from Switzerland, Greenland and Austria will allow us to validate the accuracy of the detection approach. As a particular novelty of our work, we will try to leverage high-resolution weather data and combine it with SAR images to improve the detection performance. Moreover, we will assess the possibilities of learning-based approaches in the context of the arguably more challenging avalanche forecasting problem.

How to cite: Kapper, K. L., Muckenhuber, S., Goelles, T., Trügler, A., Kuric, M., Abermann, J., Grahn, J., Malnes, E., and Schöner, W.: The potential of automated snow avalanche detection from SAR images for the Austrian Alpine region using a learning-based approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7313,, 2022.

Flood events cause substantial damage to infrastructure and disrupt livelihoods. There is a need for the development of an innovative, open-access and real-time disaster map pipeline which is automatically initiated at the time of a flood event to highlight flooded regions, potential damage and vulnerable communities. This can help in directing resources appropriately during and after a disaster to reduce disaster risk. To implement this pipeline, we explored the integration of three heterogeneous data sources which include remote sensing data, social sensing data and geospatial sensing data to guide disaster relief and response. Remote sensing through satellite imagery is an effective method to identify flooded areas where we utilized existing deep learning models to develop a pipeline to process both optical and radar imagery. Whilst this can offer situational awareness right after a disaster, satellite-based flood extent maps lack important contextual information about the severity of structural damage or urgent needs of affected population. This is where the potential of social sensing through microblogging sites comes into play as it provides insights directly from eyewitnesses and affected people in real-time. Whilst social sensing data is advantageous, these streams are usually extremely noisy where there is a need to build disaster relevant taxonomies for both text and images. To develop a disaster taxonomy for social media texts, we conducted literature review to better understand stakeholder information needs. The final taxonomy consisted of 30 categories organized among three high-level classes. This built taxonomy was then used to label a large number of tweet texts (~ 10,000) to train machine learning classifiers so that only relevant social media texts are visualized on the disaster map. Moreover, a disaster object taxonomy for social media images was developed in collaboration with a certified emergency manager and trained volunteers from Montgomery County, MD Community Emergency Response Team. In total, 106 object categories were identified and organized as a hierarchical  taxonomy with  three high-level classes and 10 sub-classes. This built taxonomy will be used to label a large set of disaster images for object detection so that machine learning classifiers can be trained to effectively detect disaster relevant objects in social media imagery. The wide perspective provided by the satellite view combined with the ground-level perspective from locally collected textual and visual information helped us in identifying three types of signals: (i) confirmatory signals from both sources, which puts greater confidence that a specific region is flooded, (ii) complementary signals that provide different contextual information including needs and requests, disaster impact or damage reports and situational information, and (iii) novel signals when both data sources do not overlap and provide unique information. We plan to fuse the third component, geospatial sensing, to perform flood vulnerability analysis to allow easy identification of areas/zones that are most vulnerable to flooding. Thus, the fusion of remote sensing, social sensing and geospatial sensing for rapid flood mapping can be a powerful tool for crisis responders.

How to cite: Ofli, F., Akhtar, Z., Sadiq, R., and Imran, M.: Triangulation of remote sensing, social sensing, and geospatial sensing for flood mapping, damage estimation, and vulnerability assessment, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7561,, 2022.

EGU22-7711 | Presentations | ITS2.5/NH10.8

Global sensitivity analyses to characterize the risk of earth fissures in subsiding basins 

Yueting Li, Claudia Zoccarato, Noemi Friedman, András Benczúr, and Pietro Teatini

Earth fissure associated with groundwater pumping is a severe geohazard jeopardizing several subsiding basins generally in arid countries (e.g., Mexico, Arizona, Iran, China, Pakistan). Up to 15 km long, 1–2 m wide, 15–20 m deep, and more than 2 m vertically dislocated fissures have been reported. A common geological condition favoring the occurrence of earth fissures is the presence of shallow bedrock ridge buried by compacting sedimentary deposits. This study aims to improve the understanding of this mechanism by evaluating the effects of various factors on the risk of fissure formation and development. Several parameters playing a role in the fissure occurrence have been considered, such as the shape of the bedrock ridge, the aquifer thickness, the pressure depletion in the aquifer system, and its compressibility. A realistic case is developed where the characteristics of fissure like displacements and stresses are quantified with aid of a numerical approach based on finite elements for the continuum and interface elements for the discretization of the fissures. Modelling results show that the presence of bedrock ridge causes tension accumulation around its tip and results in fissure opening from land surface downward after long term piezometry depletion. Different global sensitivity analysis methods are applied to measure the importance of each single factor (or group of them) on the quantity of interest, i.e., the fissure opening. A conventional variance-based method is first presented with Sobol indices computed from Monte Carlo simulations, although its accuracy is only guaranteed with a high number of forward simulations. As alternatives, generalized polynomial chaos expansion and gradient boosting tree are introduced to approximate the forward model and implement the corresponding sensitivity assessment at a significantly reduced computational cost. All the measures provide similar results that highlight the importance of bedrock ridge in earth fissuring. Generally, the steeper bedrock ridge the higher the risk of significant fissure opening. Pore pressure depletion is secondarily key factor which is essential for fissure formation.

How to cite: Li, Y., Zoccarato, C., Friedman, N., Benczúr, A., and Teatini, P.: Global sensitivity analyses to characterize the risk of earth fissures in subsiding basins, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7711,, 2022.

Induced subsidence and seismicity caused by the production of hydrocarbons in the Groningen gas field (the Netherlands) is a widely known issue facing this naturally aseismic region (Smith et al., 2019). Extraction reduces pore-fluid pressure leading to accumulation of small elastic and inelastic strains and an increase in effective vertical stress driving compaction of reservoir sandstones.

Recent studies (Pijnenburg et al., 2019a, b and Verberne et al., 2021) identify grain-scale deformation of intergranular and grain-coating clays as largely responsible for accommodating (permanent) inelastic deformation at small strains relevant to production (≤1.0%). However, their distribution, microstructure, abundance, and contribution to inelastic deformation remains unconstrained, presenting challenges when evaluating grain-scale deformation mechanisms within a natural system. Traditional methods of mineral identification are costly, labor-intensive, and time-consuming. Digital imaging coupled with machine-learning-driven segmentation is necessary to accelerate the identification of clay microstructures and distributions within reservoir sandstones for later large-scale analysis and geomechanical modeling.

We performed digital imaging on thin-sections taken from core recovered from the highly-depleted Zeerijp ZRP-3a well located at the most seismogenic part of the field. The core was kindly made available by the field operator, NAM. Optical digital images were acquired using the Zeiss AxioScan optical light microscope at 10x magnification with a resolution of 0.44µm and compared to backscattered electron (BSE) digital images from the Zeiss EVO 15 Scanning Electron Microscope (SEM) at varying magnifications with resolutions ranging from 0.09µm - 2.24 µm. Digital images were processed in ilastik, an interactive machine-learning-based toolkit for image segmentation that uses a Random Forest classifier to separate clays from a digital image (Berg et al., 2019).

Comparisons between segmented optical and BSE digital images indicate that image resolution is the main limiting factor for successful mineral identification and image segmentation, especially for clay minerals. Lower resolution digital images obtained using optical light microscopy may be sufficient to segment larger intergranular/pore-filling clays, but higher resolution BSE images are necessary to segment smaller micron to submicron-sized grain-coating clays. Comparing the same segmented optical image (~11.5% clay) versus BSE image (~16.3% clay) reveals an error of ~30%, illustrating the potential of underestimating the clay content necessary for geomechanical modeling.

Our analysis shows that coupled automated electron microscopy with machine-learning-driven image segmentation has the potential to provide statistically relevant and robust information to further constrain the role of clay films on the compaction behavior of reservoir rocks.



Berg, S. et al., Nat Methods 16, 1226–1232 (2019).

(NAM) Nederlandse Aardolie Maatschappij BV (2015).

Pijnenburg, R. P. J. et al., Journal of Geophysical Research: Solid Earth, 124 (2019a).

Pijnenburg, R. P. J. et al., Journal of Geophysical Research: Solid Earth, 124, 5254–5282. (2019b)

Smith, J. D. et al., Journal of Geophysical Research: Solid Earth, 124, 6165–6178. (2019)

Verberne, B. A. et al., Geology, 49 (5): 483–487. (2020)

How to cite: Vogel, H., Amiri, H., Plümper, O., Hangx, S., and Drury, M.: Applications of digital imaging coupled with machine-learning for aiding the identification, analysis, and quantification of intergranular and grain-coating clays within reservoirs rocks., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7915,, 2022.

EGU22-9406 | Presentations | ITS2.5/NH10.8

Building exposure datasets using street-level imagery and deep learning object detection models 

Luigi Cesarini, Rui Figueiredo, Xavier Romão, and Mario Martina

The built environment is constantly under the threat of natural hazards, and climate change will only exacerbate such perils. The assessment of natural hazard risk requires exposure models representing the characteristics of the assets at risk, which are crucial to subsequently estimate damage and impacts of a given hazard to such assets. Studies addressing exposure assessment are expanding, in particular due to technological progress. In fact, several works are introducing data collected from volunteered geographic information (VGI), user-generated content, and remote sensing data. Although these methods generate large amounts of data, they typically require a time-consuming extraction of the necessary information. Deep learning models are particularly well suited to perform this labour-intensive task due to their ability to handle massive amount of data.

In this context, this work proposes a methodology that connects VGI obtained from OpenStreetMap (OSM), street-level imagery from Google Street View (GSV) and deep learning object detection models to create an exposure dataset of electrical transmission towers, an asset particularly vulnerable to strong winds among other perils (i.e., ice loads and earthquakes). The main objective of the study is to establish and demonstrate a complete pipeline that first obtains the locations of transmission towers from the power grid layer of OSM’s world infrastructure, and subsequently assigns relevant features of each tower based on the classification returned from an object detection model over street-level imagery of the tower, obtained from GSV.

The study area for the initial application of the methodology is the Porto district (Portugal), which has an area of around 1360 km2 and 5789 transmission towers. The area was found to be representative given its diverse land use, containing both densely populated settlements and rural areas, and the different types of towers that can be found. A single-stage detector (YOLOv5) and a two-stage detector (Detectron2) were trained and used to perform identification and classification of towers. The first task was used to test the ability of a model to recognize whether a tower is present in an image, while the second task assigned a category to each tower based on a taxonomy derived from a compilation of the most used type of towers. Preliminary results on the test partition of the dataset are promising. For the identification task, YOLOv5 returned a mean average precision (mAP) of 87% for an intersection over union (IoU) of 50%, while Detectron2 reached a mAP of 91% for the same IoU. In the classification problem, the performances were also satisfactory, particularly when the models were trained on a sufficient number of images per class. 

Additional analyses of the results can provide insights on the types of areas for which the methodology is more reliable. For example, in remote areas, the long distance of a tower to the street might prevent the object to be identified in the image. Nevertheless, the proposed methodology can in principle be used to generate exposure models of transmission towers at large spatial scales in areas for which the necessary datasets are available.


How to cite: Cesarini, L., Figueiredo, R., Romão, X., and Martina, M.: Building exposure datasets using street-level imagery and deep learning object detection models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9406,, 2022.

EGU22-10276 | Presentations | ITS2.5/NH10.8

Weather and climate in the AI-supported early warning system DAKI-FWS 

Elena Xoplaki, Andrea Toreti, Florian Ellsäßer, Muralidhar Adakudlu, Eva Hartmann, Niklas Luther, Johannes Damster, Kim Giebenhain, Andrej Ceglar, and Jackie Ma

The project DAKI-FWS (BMWi joint-project “Data and AI-supported early warning system to stabilise the German economy”; German: “Daten- und KI-gestütztes Frühwarnsystem zur Stabilisierung der deutschen Wirtschaft”) develops an early warning system (EWS) to strengthen economic resilience in Germany. The EWS enables better characterization of the development and course of pandemics or hazardous climate extreme events and can thus protect and support lives, jobs, land and infrastructures.

The weather and climate modules of the DAKI-FWS use state-of-the-art seasonal forecasts for Germany and apply innovative AI-approaches to prepare very high spatial resolution simulations. These are used for the climate-related practical applications of the project, such as pandemics or subtropical/tropical diseases, and contribute to the estimation of the outbreak and evolution of health crises. Further, the weather modules of the EWS objectively identify weather and climate extremes, such as heat waves, storms and droughts, as well as compound extremes from a large pool of key data sets. The innovative project work is complemented by the development and AI-enhancement of the European Flood Awareness System model, LISFLOOD, and forecasting system for Germany at very high spatial resolution. The model combined with the high-end output of the seasonal forecast prepares high-resolution, accurate flood risk assessment. The final output of the EWS and hazard maps not only support adaptation, but they also increase preparedness providing a time horizon of several months ahead, thus increasing the resilience of economic sectors to impacts of the ongoing anthropogenic climate change. The weather and climate modules of the EWS provide economic, political, and administrative decision-makers and the general public with evidence on the probability of occurrence, intensity and spatial and temporal extent of extreme events as well as with critical information during a disaster.

How to cite: Xoplaki, E., Toreti, A., Ellsäßer, F., Adakudlu, M., Hartmann, E., Luther, N., Damster, J., Giebenhain, K., Ceglar, A., and Ma, J.: Weather and climate in the AI-supported early warning system DAKI-FWS, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10276,, 2022.

Landslide inventories are essential for landslide susceptibility mapping, hazard modelling, and further risk mitigation management. For decades, experts and organisations worldwide have preferred manual visual interpretation of satellite and aerial images. However, there are various problems associated with manual inventories, such as manual extraction of landslide borders and their representation with polygons, which is a subjective process.  Manual delineation is affected by the applied methodology, the preferences of the experts and interpreters, and how much time and effort are invested in the inventory generating process. In recent years, a vast amount of research related to semi-automated and automatic mapping of landslide inventories has been carried out to overcome these issues. The automatic generation of landslide inventories using Artificial Intelligence (AI) techniques is still in its early phase as currently there is no published research that can create a ground truth representation of landslide situation after a landslide triggering event. The evaluation metrics in recent literature show a range of 50-80% of F1-score in terms of landslide boundary delineation using AI-based models. However, very few studies claim to have achieved more than 80% F1 score with the exception of those employing the testing of their model evaluation in the same study area. Therefore, there is still a research gap between the generation of AI-based landslide inventories and their usability for landslide hazard and risk studies. In this study, we explore several inventories developed by AI and manual delineation and test their usability for assessing landslide hazard.

How to cite: Meena, S. R., Floris, M., and Catani, F.: Can landslide inventories developed by artificial intelligence substitute manually delineated inventories for landslide hazard and risk studies?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11422,, 2022.

EGU22-11787 | Presentations | ITS2.5/NH10.8

Explainable deep learning for wildfire danger estimation 

Michele Ronco, Ioannis Prapas, Spyros Kondylatos, Ioannis Papoutsis, Gustau Camps-Valls, Miguel-Ángel Fernández-Torres, Maria Piles Guillem, and Nuno Carvalhais

Deep learning models have been remarkably successful in a number of different fields, yet their application to disaster management is obstructed by the lack of transparency and trust which characterises artificial neural networks. This is particularly relevant in the field of Earth sciences where fitting is only a tiny part of the problem, and process understanding becomes more relevant [1,2]. In this regard, plenty of eXplainable Artificial Intelligence (XAI) algorithms have been proposed in the literature over the past few years [3]. We suggest that combining saliency maps with interpretable approximations, such as LIME, is useful to extract complementary insights and reach robust explanations. We address the problem of wildfire forecasting for which interpreting the model's predictions is of crucial importance to put into action effective mitigation strategies. Daily risk maps have been obtained by training a convolutional LSTM with ten years of data of spatio-temporal features, including weather variables, remote sensing indices and static layers for land characteristics [4]. We show how the usage of XAI allows us to interpret the predicted fire danger, thereby shortening the gap between black-box approaches and disaster management.


[1] Deep learning for the Earth Sciences: A comprehensive approach to remote sensing, climate science and geosciences

Gustau Camps-Valls, Devis Tuia, Xiao Xiang Zhu, Markus Reichstein (Editors)

Wiley \& Sons 2021

[2] Deep learning and process understanding for data-driven Earth System Science

Reichstein, M. and Camps-Valls, G. and Stevens, B. and Denzler, J. and Carvalhais, N. and Jung, M. and Prabhat

Nature 566 :195-204, 2019

[3] Explainable AI: Interpreting, Explaining and Visualizing Deep Learning

 Wojciech Samek, Grégoire Montavon, Andrea Vedaldi, Lars Kai Hansen, Klaus-Robert Müller (Editors)

LNCS, volume 11700, Springer 

[4] Deep Learning Methods for Daily Wildfire Danger Forecasting

Ioannis Prapas, Spyros Kondylatos, Ioannis Papoutsis, Gustau Camps-Valls, Michele Ronco, Miguel-Ángel Fernández-Torres, Maria Piles Guillem, Nuno Carvalhais

arXiv: 2111.02736


How to cite: Ronco, M., Prapas, I., Kondylatos, S., Papoutsis, I., Camps-Valls, G., Fernández-Torres, M.-Á., Piles Guillem, M., and Carvalhais, N.: Explainable deep learning for wildfire danger estimation, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11787,, 2022.

EGU22-11872 | Presentations | ITS2.5/NH10.8

Recent Advances in Deep Learning for Spatio-Temporal Drought Monitoring, Forecasting and Model Understanding 

María González-Calabuig, Jordi Cortés-Andrés, Miguel-Ángel Fernández-Torres, and Gustau Camps-Valls

Droughts constitute one of the costliest natural hazards and have seriously destructive effects on the ecological environment, agricultural production and socio-economic conditions. Their elusive and subjective definition, due to the complex physical, chemical and biological processes of the Earth system they involve, makes their management an arduous challenge to researchers, as well as decision and policy makers. We present here our most recent advances in machine learning models in three complementary lines of research about droughts: monitoring, forecasting and understanding. While monitoring or detection is about gaining the time series of drought maps and discovering underlying patterns and correlations, forecasting or prediction is to anticipate future droughts. Last but not least, understanding or explaining models by means of expert-comprehensible representations is equally important as accurately addressing these tasks, especially for their deployment in real scenarios. Thanks to the emergence and success of deep learning, all of these tasks can be tackled by the design of spatio-temporal data-driven approaches built on the basis of climate variables (soil moisture, precipitation, temperature, vegetation health, etc.) and/or satellite imagery. The possibilities are endless, from the design of convolutional architectures and attention mechanisms to the use of generative models such as Normalizing Flows (NFs) or Generative Adversarial Networks (GANs), trained both in a supervised and unsupervised manner, among others. Different application examples in Europe from 2003 onwards are provided, with the aim of reflecting on the possibilities of the strategies proposed, and also of foreseeing alternatives and future lines of development. For that purpose, we make use of several mesoscale (1 km) spatial and 8 days temporal resolution variables included in the Earth System Data Cube (ESDC) [Mahecha et al., 2020] for drought detection, while high resolution (20 m, 5 days) Sentinel-2 data cubes, extracted from the extreme summer track in EarthNet2021 [Requena-Mesa et al., 2021], are considered for forecasting.



Mahecha, M. D., Gans, F., Brandt, G., Christiansen, R., Cornell, S. E., Fomferra, N., ... & Reichstein, M. (2020). Earth system data cubes unravel global multivariate dynamics. Earth System Dynamics, 11(1), 201-234.

Requena-Mesa, C., Benson, V., Reichstein, M., Runge, J., & Denzler, J. (2021). EarthNet2021: A large-scale dataset and challenge for Earth surface forecasting as a guided video prediction task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1132-1142).

How to cite: González-Calabuig, M., Cortés-Andrés, J., Fernández-Torres, M.-Á., and Camps-Valls, G.: Recent Advances in Deep Learning for Spatio-Temporal Drought Monitoring, Forecasting and Model Understanding, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11872,, 2022.

EGU22-12432 | Presentations | ITS2.5/NH10.8

Building wildfire intelligence at the edge: bridging the gap from development to deployment 

Maria João Sousa, Alexandra Moutinho, and Miguel Almeida

The increased frequency, intensity, and severity of wildfire events in several regions across the world has highlighted several disaster response infrastructure hindrances that call for enhanced intelligence gathering pipelines. In this context, the interest in the use of unmanned aerial vehicles for surveillance and active fire monitoring has been growing in recent years. However, several roadblocks challenge the implementation of these solutions due to their high autonomy requirements and energy-constrained nature. For these reasons, the artificial intelligence development focus on large models hampers the development of models suitable for deployment onboard these platforms. In that sense, while artificial intelligence approaches can be an enabling technology that can effectively scale real-time monitoring services and optimize emergency response resources, the design of these systems imposes: (i) data requirements, (ii) computing constraints and (iii) communications limitations. Here, we propose a decentralized approach, reflecting upon these three vectors.

Data-driven artificial intelligence is central to both handle multimodal sensor data in real-time and to annotate large amounts of data collected, which are necessary to build robust safety-critical monitoring systems. Nevertheless, these two objectives have distinct implications computation-wise, because the first must happen on-board, whereas the second can leverage higher processing capabilities off-board. While autonomy of robotic platforms drives mission performance, being a key reason for the need for edge computing of onboard sensor data, the communications design is essential to mission endurance as relaying large amounts of data in real-time is unfeasible energy-wise. 

For these reasons, real-time processing and data annotation must be tackled in a complimentary manner, instead of the general practice of only targeting overall accuracy improvement. To build wildfire intelligence at the edge, we propose developments on two tracks of solutions: (i) data annotation and (ii) on the edge deployment. The need for considerable effort in these two avenues stems from both having very distinct development requirements and performance evaluation metrics. On the one hand, improving data annotation capacity is essential to build high quality databases that can provide better sources for machine learning. On the other hand, for on the edge deployment the development architectures need to compromise on robustness and architectural parsimony in order to be efficient for edge processing. Whereas the first objective is driven foremost by accuracy, the second goal must emphasize timeliness.

This work was supported by FCT – Fundação para a Ciência e a Tecnologia, I.P., through IDMEC, under project Eye in the Sky, PCIF/SSI/0103/2018, and through IDMEC, under LAETA, project UIDB/50022/2020. M. J. Sousa acknowledges the support from FCT, through the Ph.D. Scholarship SFRH/BD/145559/2019, co-funded by the European Social Fund (ESF).

How to cite: Sousa, M. J., Moutinho, A., and Almeida, M.: Building wildfire intelligence at the edge: bridging the gap from development to deployment, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12432,, 2022.

EGU22-20 | Presentations | ITS2.6/AS5.1

PRECISIONPOP: a multi-scale monitoring system for poplar plantations integrating field, aerial and satellite remote sensing 

Francesco Chianucci, Francesca Giannetti, Clara Tattoni, Nicola Puletti, Achille Giorcelli, Carlo Bisaglia, Elio Romano, Massimo Brambilla, Piermario Chiarabaglio, Massimo Gennaro, Giovanni d'Amico, Saverio Francini, Walter Mattioli, Domenico Coaloa, Piermaria Corona, and Gherardo Chirici

Poplar (Populus spp.) plantations are globally widespread in the Northern Hemisphere, and provide a wide range of benefits and products, including timber, carbon sequestration and phytoremediation. Because of poplar specific features (fast growth, short rotation) the information needs require frequent updates, which exceed the traditional scope of National Forest Inventories, implying the need for ad-hoc monitoring solutions.

Here we presented a regional-level multi-scale monitoring system developed for poplar plantations, which is based on the integration of different remotely-sensed informations at different spatial scales, developed in Lombardy (Northern Italy) region. The system is based on three levels of information: 1) At plot scale, terrestrial laser scanning (TLS) was used to develop non-destructive tree stem volume allometries in calibration sites; the produced allometries were then used to estimate plot-level stand parameters from field inventory; additional canopy structure attributes were derived using field digital cover photography. 2) At farm level, unmanned aerial vehicles (UAVs) equipped with multispectral sensors were used to upscale results obtained from field data. 3) Finally, both field and unmanned aerial estimates were used to calibrate a regional-scale supervised continuous monitoring system based on multispectral Sentinel-2 imagery, which was implemented and updated in a Google Earth Engine platform.

The combined use of multi-scale information allowed an effective management and monitoring of poplar plantations. From a top-down perspective, the continuous satellite monitoring system allowed the detection of early warning poplar stress, which are suitable for variable rate irrigation and fertilizing scheduling. From a bottom-up perspective, the spatially explicit nature of TLS measurements allows better integration with remotely sensed data, enabling a multiscale assessment of poplar plantation structure with different levels of detail, enhancing conventional tree inventories, and supporting effective management strategies. Finally, use of UAV is key in poplar plantations as their spatial resolution is suited for calibrating metrics from coarser remotely-sensed products, reducing or avoiding the need of ground measurements, with a significant reduction of time and costs.

How to cite: Chianucci, F., Giannetti, F., Tattoni, C., Puletti, N., Giorcelli, A., Bisaglia, C., Romano, E., Brambilla, M., Chiarabaglio, P., Gennaro, M., d'Amico, G., Francini, S., Mattioli, W., Coaloa, D., Corona, P., and Chirici, G.: PRECISIONPOP: a multi-scale monitoring system for poplar plantations integrating field, aerial and satellite remote sensing, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-20,, 2022.

EGU22-124 | Presentations | ITS2.6/AS5.1

Unsupervised machine learning driven Prospectivity analysis of REEs in NE India 

Malcolm Aranha and Alok Porwal

Traditional mineral prospectivity modelling for mineral exploration and targeting relies heavily on manual data filtering and processing to extract desirable geologic features based on expert knowledge. It involves the integration of geological predictor maps that are manually derived by time-consuming and labour-intensive pre-processing of primary geoscientific data to serve as spatial proxies of mineralisation processes. Moreover, the selection of these spatial proxies is guided by conceptual genetic modelling of the targeted deposit type, which may be biased by the subjective preference of an expert geologist. This study applies Self-Organising Maps (SOM), a neural network-based unsupervised machine learning clustering algorithm, to gridded geophysical and topographical datasets in order to identify and delineate regional-scale exploration targets for carbonatite-alkaline-complex-related REE deposits in northeast India. The study did not utilise interpreted and processed or manually generated data, such as surface or bed-rock geological maps, fault traces, etc., and relies on the algorithm to identify crucial features and delineate prospective areas. The obtained results were then compared with those obtained from a previous supervised knowledge-driven prospectivity analysis. The results were found to be comparable. Therefore, unsupervised machine learning algorithms are reliable tools to automate the manual process of mineral prospectivity modelling and are robust, time-saving alternatives to knowledge-driven or supervised data-driven prospectivity modelling. These methods would be instrumental in unexplored terrains for which there is little or no geological knowledge available. 

How to cite: Aranha, M. and Porwal, A.: Unsupervised machine learning driven Prospectivity analysis of REEs in NE India, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-124,, 2022.

EGU22-654 | Presentations | ITS2.6/AS5.1

On the derivation of data-driven models for partially observed systems 

Said Ouala, Bertrand Chapron, Fabrice Collard, Lucile Gaultier, and Ronan Fablet

When considering the modeling of dynamical systems, the increasing interest in machine learning, artificial intelligence and more generally, data-driven representations, as well as the increasing availability of data, motivated the exploration and definition of new identification techniques. These new data-driven representations aim at solving modern questions regarding the modeling, the prediction and ultimately, the understanding of complex systems such as the ocean, the atmosphere and the climate. 

In this work, we focus on one question regarding the ability to define a (deterministic) dynamical model from a sequence of observations. We focus on sea surface observations and show that these observations typically relate to some, but not all, components of the underlying state space, making the derivation of a deterministic model in the observation space impossible. In this context, we formulate the identification problem as the definition, from data, of an embedding of the observations, parameterized by a differential equation. When compared to state-of-the-art techniques based on delay embedding and linear decomposition of the underlying operators, the proposed approach benefits from all the advances in machine learning and dynamical systems theory in order to define, constrain and tune the reconstructed sate space and the approximate differential equation. Furthermore, the proposed embedding methodology naturally extends to cases in which a dynamical prior (derived for example using physical principals) is known, leading to relevant physics informed data-driven models. 

How to cite: Ouala, S., Chapron, B., Collard, F., Gaultier, L., and Fablet, R.: On the derivation of data-driven models for partially observed systems, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-654,, 2022.

EGU22-1255 | Presentations | ITS2.6/AS5.1

A Deep Learning approach to de-bias Air Quality forecasts, using heterogeneous Open Data sources as reference 

Antonio Pérez, Mario Santa Cruz, Johannes Flemming, and Miha Razinger

The degradation of air quality is a challenge that policy-makers face all over the world. According to the World Health Organisation, air pollution causes an estimate of 7 million premature deaths every year. In this context, air quality forecasts are crucial tools for decision- and policy-makers, to achieve data-informed decisions.

Global forecasts, such as the Copernicus Atmosphere monitoring service model (CAMS), usually exhibit biases: systematic deviations from observations. Adjusting these biases is typically the first step towards obtaining actionable air quality forecasts. It is especially relevant in health-related decisions, when the metrics of interest depend on specific thresholds.

AQ (Air quality) - Bias correction was a project funded by the ECMWF Summer of Weather Code (ESOWC) 2021 whose aim is to improve CAMS model forecasts for air quality variables (NO2, O3, PM2.5), using as a reference the in-situ observations provided by OpenAQ. The adjustment, based on machine learning methods, was performed over a set of specific interesting locations provided by the ECMWF, for the period June 2019 to March 2021.

The machine learning approach uses three different deep learning based models, and an extra neural network that gathers the output of the three previous models. From the three DL-based models, two of them are independent and follow the same structure built upon the InceptionTime module: they use both meteorological and air quality variables, to exploit the temporal variability and to extract the most meaningful features of the past [t-24h, t-23h, … t-1h] and future [t, t+1h, …, t+23h] CAMS predictions. The third model uses the station static attributes (longitude, latitude and elevation), and a multilayer perceptron interacts with the station attributes. The extracted features from these three models are fed into another multilayer perceptron, to predict the upcoming errors with hourly resolution [t, t+1h, …, t+23h]. As a final step, 5 different initializations are considered, assembling them with equal weights to have a more stable regressor.

Previous to the modelisation, CAMS forecasts of air quality variables were actually biassed independently from the location of interest and the variable (on average: biasNO2 = -22.76, biasO3 = 44.30, biasPM2.5 = 12.70). In addition, the skill of the model, measured by the Pearson correlation, did not reach 0.5 for any of the variables—with remarkable low values for NO2 and O3 (on average: pearsonNO2 = 0.10, pearsonO3 = 0.14).

AQ-BiasCorrection modelisation properly corrects these biases. Overall, the number of stations that improve the biases both in train and test sets are: 52 out of 61 (85%) for NO2, 62 out of 67 (92%) for O3, and 80 out of 102 (78%) for PM2.5. Furthermore, the bias improves with declines of -1.1%, -9.7% and -13.9% for NO2, O3 and PM2.5 respectively. In addition, there is an increase in the model skill measured through the Pearson correlation, reaching values in the range of 100-400% for the overall improvement of the variable skill.

How to cite: Pérez, A., Santa Cruz, M., Flemming, J., and Razinger, M.: A Deep Learning approach to de-bias Air Quality forecasts, using heterogeneous Open Data sources as reference, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1255,, 2022.

EGU22-1992 | Presentations | ITS2.6/AS5.1

Approximating downward short-wave radiation flux using all-sky optical imagery using machine learning trained on DASIO dataset. 

Vasilisa Koshkina, Mikhail Krinitskiy, Nikita Anikin, Mikhail Borisov, Natalia Stepanova, and Alexander Osadchiev

Solar radiation is the main source of energy on Earth. Cloud cover is the main physical factor limiting the downward short-wave radiation flux. In modern models of climate and weather forecasts, physical models describing the passage of radiation through clouds may be used. This is a computationally extremely expensive option for estimating downward radiation fluxes. Instead, one may use parameterizations which are simplified schemes for approximating environmental variables. The purpose of this work is to improve the accuracy of the existing parametrizations of downward shortwave radiation fluxes. We solve the problem using various machine learning (ML) models for approximating downward shortwave radiation flux using all-sky optical imagery. We assume that an all-sky photo contains complete information about the downward shortwave radiation. We examine several types of ML models that we trained on dataset of all-sky imagery accompanied by short-wave radiation flux measurements. The Dataset of All-Sky Imagery over the Ocean (DASIO) is collected in Indian, Atlantic and Arctic oceans during several oceanic expeditions from 2014 till 2021. The quality of the best classic ML model is better compared to existing parameterizations known from literature. We will show the results of our study regarding classic ML models as well as the results of an end-to-end ML approach involving convolutional neural networks. Our results allow us to assume one may acquire downward shortwave radiation fluxes directly from all-sky imagery. We will also cover some downsides and limitations of the presented approach.

How to cite: Koshkina, V., Krinitskiy, M., Anikin, N., Borisov, M., Stepanova, N., and Osadchiev, A.: Approximating downward short-wave radiation flux using all-sky optical imagery using machine learning trained on DASIO dataset., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1992,, 2022.

EGU22-2058 | Presentations | ITS2.6/AS5.1

Deep learning for ensemble forecasting 

Rüdiger Brecht and Alexander Bihlo
Ensemble prediction systems are an invaluable tool for weather prediction. Practically, ensemble predictions are obtained by running several perturbed numerical simulations. However, these systems are associated with a high computational cost and often involve statistical post-processing steps to improve their qualities.
Here we propose to use a deep-learning-based algorithm to learn the statistical properties of a given ensemble prediction system, such that this system will not be needed to simulate future ensemble forecasts. This way, the high computational costs of the ensemble prediction system can be avoided while still obtaining the statistical properties from a single deterministic forecast. We show preliminary results where we demonstrate the ensemble prediction properties for a shallow water unstable jet simulation on the sphere. 

How to cite: Brecht, R. and Bihlo, A.: Deep learning for ensemble forecasting, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2058,, 2022.

Numerical weather prediction (NWP) models are currently popularly used for operational weather forecast in meteorological centers. The NWP models describe the flow of fluids by employing a set of governing equations, physical parameterization schemes and initial and boundary conditions. Thus, it often face bias of prediction due to insufficient data assimilation, assumptions or approximations of dynamical and physical processes. To make gridded forecast of rainfall with high confidence, in this study, we present a data-driven deep learning model for correction of rainfall from NWP model, which mainly includes a confidence network and a combinatorial network. Meanwhile, a focal loss is introduced to deal with the characteristics of longtail-distribution of rainfall. It is expected to alleviate the impact of the large span of rainfall magnitude by transferring the regression problem into several binary classification problems. The deep learning model is used to correct the gridded forecasts of rainfall from the European Centre for Medium-Range Weather Forecast Integrated Forecasting System global model (ECMWF-IFS) with a forecast lead time of 24 h to 240 h in Eastern China. First, the rainfall forecast correction problem is treated as an image-to-image translation problem in deep learning under the neural networks. Second, the ECMWF-IFS forecasts and rainfall observations in recent years are used as training, validation, and testing datasets. Finally, the correction performance of the new machine learning model is evaluated and compared to several classical machine learning algorithms. By performing a set of experiments for rainfall forecast error correction, it is found that the new model can effectively forecast rainfall over East China region during the flood season of the year 2020. Experiments also demonstrate that the proposed approach generally performs better in bias correction of rainfall prediction than most of the classical machine learning approaches .

How to cite: Ma, L.: A Deep Learning Bias Correction Approach for Rainfall Numerical Prediction, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2095,, 2022.

EGU22-2893 | Presentations | ITS2.6/AS5.1 | Highlight

Bias Correction of Operational Storm Surge Forecasts Using Neural Networks 

Paulina Tedesco, Jean Rabault, Martin Lilleeng Sætra, Nils Melsom Kristensen, Ole Johan Aarnes, Øyvind Breivik, and Cecilie Mauritzen

Storm surges can give rise to extreme floods in coastal areas. The Norwegian Meteorological Institute (MET Norway) produces 120-hour regional operational storm surge forecasts along the coast of Norway based on the Regional Ocean Modeling System (ROMS). Despite advances in the development of models and computational capability, forecast errors remain large enough to impact response measures and issued alerts, in particular, during the strongest storm events. Reducing these errors will positively impact the efficiency of the warning systems while minimizing efforts and resources spent on mitigation.

Here, we investigate how forecasts can be improved with residual learning, i.e., training data-driven models to predict, and correct, the error in the ROMS output. For this purpose, sea surface height data from stations around Norway were collected and compared with the ROMS output.

We develop two different residual learning frameworks that can be applied on top of the ROMS output. In the first one, we perform binning of the model error, conditionalized by pressure, wind, and waves. Clear error patterns are visible when the error conditioned by the wind is plotted in a polar plot for each station. These error maps can be stored as correction lookup tables to be applied on the ROMS output. However, since wind, pressure, and waves are correlated, we cannot simultaneously correct the error associated with each variable using this method. To overcome this limitation, we develop a second method, which resorts to Neural Networks (NNs) to perform nonlinear modeling of the error pattern obtained at each station. 

The residual NN method strongly outperforms the error map method, and is a promising direction for correcting storm surge models operationally. Indeed, i) this method is applied on top of the existing model and requires no changes to it, ii) all predictors used for NN inference are available operationally, iii) prediction by the NN is very fast, typically a few seconds per station, and iv) the NN correction can be provided to a human expert who gets to inspect it, compare it with the ROMS output, and see how much correction is brought by the NN. Using this NN residual error correction method, the RMS error in the Oslofjord is reduced by typically 7% for lead times of 24 hours, 17% for 48 hours, and 35% for 96 hours.

How to cite: Tedesco, P., Rabault, J., Sætra, M. L., Kristensen, N. M., Aarnes, O. J., Breivik, Ø., and Mauritzen, C.: Bias Correction of Operational Storm Surge Forecasts Using Neural Networks, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2893,, 2022.

EGU22-3977 | Presentations | ITS2.6/AS5.1 | Highlight

Learning quasi-geostrophic turbulence parametrizations from a posteriori metrics 

Hugo Frezat, Julien Le Sommer, Ronan Fablet, Guillaume Balarac, and Redouane Lguensat

Machine learning techniques are now ubiquitous in the geophysical science community. They have been applied in particular to the prediction of subgrid-scale parametrizations using data that describes small scale dynamics from large scale states. However, these models are then used to predict temporal trajectories, which is not covered by this instantaneous mapping. Following the model trajectory during training can be done using an end-to-end approach, where temporal integration is performed using a neural network. As a consequence, the approach is shown to optimize a posteriori metrics, whereas the classical instantaneous training is limited to a priori ones. When applied on a specific energy backscatter problem, found in quasi-geostrophic turbulent flows, the strategy demonstrates long-term stability and high fidelity statistical performance, without any increase in computational complexity during rollout. These improvements may question the future development of realistic subgrid-scale parametrizations in favor of differentiable solvers, required by the a posteriori strategy.

How to cite: Frezat, H., Le Sommer, J., Fablet, R., Balarac, G., and Lguensat, R.: Learning quasi-geostrophic turbulence parametrizations from a posteriori metrics, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3977,, 2022.

EGU22-4062 | Presentations | ITS2.6/AS5.1

Climatological Ocean Surface Wave Projections using Deep Learning 

Peter Mlakar, Davide Bonaldo, Antonio Ricchi, Sandro Carniel, and Matjaž Ličer

We present a numerically cheap machine-learning model which accurately emulates the performances of the surface wave model Simulating WAves Near Shore (SWAN) in the Adriatic basin (north-east Mediterranean Sea).

A ResNet50 inspired deep network architecture with customized spatio-temporal attention layers was used, the network being trained on a 1970-1997 dataset of time-dependent features based on wind fields retrieved from the COSMO-CLM regional climate model (The authors acknowledge Dr. Edoardo Bucchignani (Meteorology Laboratory, Centro Italiano Ricerche Aerospaziali -CIRA-, Capua, Italy), for providing the COSMO-CLM wind fields). SWAN surface wave model outputs for the period of 1970-1997 are used as labels. The period 1998-2000 is used to cross-validate that the network very accurately reproduces SWAN surface wave features (i.e. significant wave height, mean wave period, mean wave direction) at several locations in the Adriatic basin. 

After successful cross validation, a series of projections of ocean surface wave properties based on climate model projections for the end of 21st century (under RCP 8.5 scenario) are performed, and shifts in the emulated wave field properties are discussed.

How to cite: Mlakar, P., Bonaldo, D., Ricchi, A., Carniel, S., and Ličer, M.: Climatological Ocean Surface Wave Projections using Deep Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4062,, 2022.

EGU22-4493 | Presentations | ITS2.6/AS5.1 | Highlight

Semi-automatic tuning procedure for a GCM targeting continental surfaces: a first experiment using in situ observations 

Maëlle Coulon--Decorzens, Frédérique Cheruy, and Frédéric Hourdin

The tuning or calibration of General Circulation Models (GCMs) is an essential stage for their proper behavior. The need to have the best climate projections in the regions where we live drives the need to tune the models in particular towards the land surface, bearing in mind that the interactions between the atmosphere and the land surface remain a key source of uncertainty in regional-scale climate projections [1].

For a long time, this tuning has been done by hand, based on scientific expertise and has not been sufficiently documented [2]. Recent tuning tools offer the possibility to accelerate climate model development, providing a real tuning formalism as well as a new way to understand climate models. High Tune explorer is one of these statistic tuning tool, involving machine learning and based on uncertainty quantification. It aims to reduce the range of free parameters that allow realistic model behaviour [3]. A new automatic tuning experiment was developed with this tool for the atmospheric component of the IPSL GCM model, LMDZ. It was first tuned at the process level, using several single column test cases compared to large eddies simulations; and then at the global level by targeting radiative metrics at the top of the atmosphere [4].

We propose to add a new step to this semi-automatic tuning procedure targeting atmosphere and land-surface interactions. The first aspect of the proposition is to compare coupled atmosphere-continent simulations (here running LMDZ-ORCHIDEE) with in situ observations from the SIRTA observatory located southwest of Paris. In situ observations provide hourly joint colocated data with a strong potential for the understanding of the processes at stake and their representation in the model. These data are also subject to much lower uncertainties than the satellite inversions with respect to the surface observations. In order to fully benefit from the site observations, the model winds are nudged toward reanalysis. This forces the simulations to follow the effective meteorological sequence, thus allowing the comparison between simulations and observations at the process time scale. The removal of the errors arising from the representation of large-scale dynamics makes the tuning focus on the representation of physical processes «at a given meteorological situation». Finally, the model grid is zoomed in on the SIRTA observatory in order to reduce the computational cost of the simulations while preserving a fine mesh around this observatory.

We show the results of this new tuning step, which succeeds in reducing the domain of acceptable free parameters as well as the dispersion of the simulations. This method, which is less computationally costly than global tuning, is therefore a good way to precondition the latter. It allows the joint tuning of atmospheric and land surface models, traditionally tuned separately [5], and has the advantage of remaining close to the processes and thus improving their understanding.


[1] Cheruy et al., 2014,

[2] Hourdin et al., 2017,

[3] Couvreux et al., 2021,

[4] Hourdin et al., 2021,

[5] Cheruy et al., 2020,

How to cite: Coulon--Decorzens, M., Cheruy, F., and Hourdin, F.: Semi-automatic tuning procedure for a GCM targeting continental surfaces: a first experiment using in situ observations, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4493,, 2022.

EGU22-4923 | Presentations | ITS2.6/AS5.1

Constrained Generative Adversarial Networks for Improving Earth System Model Precipitation 

Philipp Hess, Markus Drüke, Stefan Petri, Felix Strnad, and Niklas Boers

The simulation of precipitation in numerical Earth system models (ESMs) involves various processes on a wide range of scales, requiring high temporal and spatial resolution for realistic simulations. This can lead to biases in computationally efficient ESMs that have a coarse resolution and limited model complexity. Traditionally, these biases are corrected by relating the distributions of historical simulations with observations [1]. While these methods successfully improve the modelled statistics, unrealistic spatial features that require a larger spatial context are not addressed.

Here we apply generative adversarial networks (GANs) [2] to transform precipitation of the CM2Mc-LPJmL ESM [3] into a bias-corrected and more realistic output. Feature attribution shows that the GAN has correctly learned to identify spatial regions with the largest bias during training. Our method presents a general bias correction framework that can be extended to a wider range of ESM variables to create highly realistic but computationally inexpensive simulations of future climates. We also discuss the generalizability of our approach to projections from CMIP6, given that the GAN is only trained on historical data.

[1] A.J. Cannon et al. "Bias correction of GCM precipitation by quantile mapping: How well do methods preserve changes in quantiles and extremes?." Journal of Climate 28.17 (2015): 6938-6959.

[2] I. Goodfellow et al. "Generative adversarial nets." Advances in neural information processing systems 27 (2014).

[3] M. Drüke et al. "CM2Mc-LPJmL v1.0: Biophysical coupling of a process-based dynamic vegetation model with managed land to a general circulation model." Geoscientific Model Development 14.6 (2021): 4117--4141.

How to cite: Hess, P., Drüke, M., Petri, S., Strnad, F., and Boers, N.: Constrained Generative Adversarial Networks for Improving Earth System Model Precipitation, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4923,, 2022.

EGU22-5219 | Presentations | ITS2.6/AS5.1 | Highlight

Neural Partial Differential Equations for Atmospheric Dynamics 

Maximilian Gelbrecht and Niklas Boers

When predicting complex systems such as parts of the Earth system, one typically relies on differential equations which can often be incomplete, missing unknown influences or higher order effects. Using the universal differential equations framework, we can augment the equations with artificial neural networks that can compensate these deficiencies. We show that this can be used to predict the dynamics of high-dimensional spatiotemporally chaotic partial differential equations, such as the ones describing atmospheric dynamics. In a first step towards a hybrid atmospheric model, we investigate the Marshall Molteni Quasigeostrophic Model in the form of a Neural Partial Differential Equation. We use it in synthetic examples where parts of the governing equations are replaced with artificial neural networks (ANNs) and demonstrate how the ANNs can recover those terms.

How to cite: Gelbrecht, M. and Boers, N.: Neural Partial Differential Equations for Atmospheric Dynamics, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5219,, 2022.

EGU22-5631 | Presentations | ITS2.6/AS5.1

Autonomous Assessment of Source Area Distributions for Sections in Lagrangian Particle Release Experiments 

Carola Trahms, Patricia Handmann, Willi Rath, Matthias Renz, and Martin Visbeck

Lagrangian experiments for particle tracing in atmosphere or ocean models and their analysis are a cornerstone of earth-system studies. They cover diverse study objectives such as the identification of pathways or source regions. Data for Lagrangian studies are generated by releasing virtual particles in one or in multiple locations of interest and simulating their advective-diffusive behavior backwards or forwards in time. Identifying main pathways connecting two regions of interest is often done by counting the trajectories that reach both regions. Here, the exact source and target region must be defined manually by a researcher. Manually defining the importance and exact location of these regions introduces a highly subjective perspective into the analysis. Additionally, to investigate all major target regions, all of them must be defined manually and the data must be analyzed accordingly. This human element slows down and complicates large scale analyses with many different sections and possible source areas.

We propose to significantly reduce the manual aspect by automatizing this process. To this end, we combine methods from different areas of machine learning and pattern mining into a sequence of steps. First, unsupervised methods, i.e., clustering, identify possible source areas on a randomized subset of the data. In a successive second step, supervised learning, i.e., classification, labels the positions along the trajectories according to their most probable source area using the previously automatically identified clusters as labels. The results of this approach can then be compared quantitatively to the results of analyses with manual definition of source areas and border-hitting-based labeling of the trajectories. Preliminary findings suggest that this approach could indeed help greatly to objectify and fasten the analysis process for Lagrangian Particle Release Experiments.

How to cite: Trahms, C., Handmann, P., Rath, W., Renz, M., and Visbeck, M.: Autonomous Assessment of Source Area Distributions for Sections in Lagrangian Particle Release Experiments, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5631,, 2022.

EGU22-5632 | Presentations | ITS2.6/AS5.1

Data-Driven Sentinel-2 Based Deep Feature Extraction to Improve Insect Species Distribution Models 

Joe Phillips, Ce Zhang, Bryan Williams, and Susan Jarvis

Despite being a vital part of ecosystems, insects are dying out at unprecedented rates across the globe. To help address this in the UK, UK Centre for Ecology & Hydrology (UKCEH) are creating a tool to utilise insect species distribution models (SDMs) for better facilitating future conservation efforts via volunteer-led insect tracking procedures. Based on these SDM models, we explored the inclusion of additional covariate information via 10-20m2 bands of temporally-aggregated Sentinel-2 data taken over the North of England in 2017 to improve the predictive performance. Here, we matched the 10-20m2 resolution of the satellite data to the coarse 1002 insect observation data via four methodologies of increasing complexity. First, we considered standard pixel-based approaches, performing aggregation by taking both the mean and standard deviation over the 10m2 pixels. Second, we explored object-based approaches to address the modifiable areal unit problem by applying the SNIC superpixels algorithm over the extent, with the mean and standard deviation of the pixels taken within each segment. The resulting dataset was then re-projected to a resolution of 100m2 by taking the modal values of the 10m2 pixels, which were provided with the aggregated values of their parent segment. Third, we took the UKCEH-created 2017 Land Cover Map (LCM) dataset and sampled 42,000, random 100m2 areas, evenly distributed about their modal land cover classes. We trained the U-Net Deep Learning model using the Sentinel-2 satellite images and LCM classes, by which data-driven features were extracted from the network over each 100m2 extent. Finally, as with the second approach, we used the superpixels segments instead as the units of analysis, sampling 21,000 segments, and taking the smallest bounding box around each of them. An attention-based U-Net was then adopted to mask each of the segments from their background and extract deep features. In a similar fashion to the second approach, we then re-projected the resulting dataset to a resolution of 100m2, taking the modal segment values accordingly. Using cross-validated AUCs over various species of moths and butterflies, we found that the object-based deep learning approach achieved the best accuracy when used with the SDMs. As such, we conclude that the novel approach of spatially aggregating satellite data via object-based, deep feature extraction has the potential to benefit similar, model-based aggregation needs and catalyse a step-change in ecological and environmental applications in the future.

How to cite: Phillips, J., Zhang, C., Williams, B., and Jarvis, S.: Data-Driven Sentinel-2 Based Deep Feature Extraction to Improve Insect Species Distribution Models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5632,, 2022.

EGU22-5681 | Presentations | ITS2.6/AS5.1

AtmoDist as a new pathway towards quantifying and understanding atmospheric predictability 

Sebastian Hoffmann, Yi Deng, and Christian Lessig

The predictability of the atmosphere is a classical problem that has received much attention from both a theoretical and practical point of view. In this work, we propose to use a purely data-driven method based on a neural network to revisit the problem. The analysis is built upon the recently introduced AtmoDist network that has been trained on high-resolution reanalysis data to provide a probabilistic estimate of the temporal difference between given atmospheric fields, represented by vorticity and divergence. We define the skill of the network for this task as a new measure of atmospheric predictability, hypothesizing that the prediction of the temporal differences by the network will be more susceptible to errors when the atmospheric state is intrinsically less predictable. Preliminary results show that for short timescales (3-48 hours) one sees enhanced predictability in warm season compared to cool season over northern midlatitudes, and lower predictability over ocean compared to land. These findings support the hypothesis that across short timescales, AtmoDist relies on the recurrences of mesoscale convection with coherent spatiotemporal structures to connect spatial evolutions to temporal differences. For example, the prevalence of mesoscale convective systems (MCSs) over the central US in boreal warm season can explain the increase of mesoscale predictability there and oceanic zones marked by greater predictability corresponds well to regions of elevated convective activity such as the Pacific ITCZ. Given the dependence of atmospheric predictability on geographic location, season, and most importantly, timescales, we further apply the method to synoptic scales (2-10 days), where excitation and propagation of large-scale disturbances such as Rossby wave packets are expected to provide the connection between temporal and spatial differences. The design of the AtmoDist network is thereby adapted to the prediction range, for example, the size of the local patches that serve as input to AtmoDist is chosen based on the spatiotemporal atmospheric scales that provide the expected time and space connections.

By providing to the community a powerful, purely data-driven technique for quantifying, evaluating, and interpreting predictability, our work lays the foundation for efficiently detecting the existence of sub-seasonal to seasonal (S2S) predictability and, by further analyzing the mechanism of AtmoDist, understanding the physical origins, which bears major scientific and socioeconomic significances.

How to cite: Hoffmann, S., Deng, Y., and Lessig, C.: AtmoDist as a new pathway towards quantifying and understanding atmospheric predictability, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5681,, 2022.

EGU22-5746 | Presentations | ITS2.6/AS5.1

Model Output Statistics (MOS) and Machine Learning applied to CAMS O3 forecasts: trade-offs between continuous and categorical skill scores 

Hervé Petetin, Dene Bowdalo, Pierre-Antoine Bretonnière, Marc Guevara, Oriol Jorba, Jan Mateu armengol, Margarida Samso Cabre, Kim Serradell, Albert Soret, and Carlos Pérez García-Pando

Air quality (AQ) forecasting systems are usually built upon physics-based numerical models that are affected by a number of uncertainty sources. In order to reduce forecast errors, first and foremost the bias, they are often coupled with Model Output Statistics (MOS) modules. MOS methods are statistical techniques used to correct raw forecasts at surface monitoring station locations, where AQ observations are available. In this study, we investigate to what extent AQ forecasts can be improved using a variety of MOS methods, including persistence (PERS), moving average (MA), quantile mapping (QM), Kalman Filter (KF), analogs (AN), and gradient boosting machine (GBM). We apply our analysis to the Copernicus Atmospheric Monitoring Service (CAMS) regional ensemble median O3 forecasts over the Iberian Peninsula during 2018–2019. A key aspect of our study is the evaluation, which is performed using a very comprehensive set of continuous and categorical metrics at various time scales (hourly to daily), along different lead times (1 to 4 days), and using different meteorological input data (forecast vs reanalyzed).

Our results show that O3 forecasts can be substantially improved using such MOS corrections and that this improvement goes much beyond the correction of the systematic bias. Although it typically affects all lead times, some MOS methods appear more adversely impacted by the lead time. When considering MOS methods relying on meteorological information and comparing the results obtained with IFS forecasts and ERA5 reanalysis, the relative deterioration brought by the use of IFS is minor, which paves the way for their use in operational MOS applications. Importantly, our results also clearly show the trade-offs between continuous and categorical skills and their dependencies on the MOS method. The most sophisticated MOS methods better reproduce O3 mixing ratios overall, with lowest errors and highest correlations. However, they are not necessarily the best in predicting the highest O3 episodes, for which simpler MOS methods can give better results. Although the complex impact of MOS methods on the distribution and variability of raw forecasts can only be comprehended through an extended set of complementary statistical metrics, our study shows that optimally implementing MOS in AQ forecast systems crucially requires selecting the appropriate skill score to be optimized for the forecast application of interest.

Petetin, H., Bowdalo, D., Bretonnière, P.-A., Guevara, M., Jorba, O., Armengol, J. M., Samso Cabre, M., Serradell, K., Soret, A., and Pérez Garcia-Pando, C.: Model Output Statistics (MOS) applied to CAMS O3 forecasts: trade-offs between continuous and categorical skill scores, Atmos. Chem. Phys. Discuss. [preprint],, in review, 2021.

How to cite: Petetin, H., Bowdalo, D., Bretonnière, P.-A., Guevara, M., Jorba, O., Mateu armengol, J., Samso Cabre, M., Serradell, K., Soret, A., and Pérez García-Pando, C.: Model Output Statistics (MOS) and Machine Learning applied to CAMS O3 forecasts: trade-offs between continuous and categorical skill scores, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5746,, 2022.

With the goal of developing a data-driven parameterization of unresolved gravity waves (GW) momentum transport for use in general circulation models (GCMs), we investigate neural network architectures that emulate the Alexander-Dunkerton 1999 (AD99) scheme, an existing physics-based GW parameterization. We analyze the distribution of errors as functions of shear-related metrics in an effort to diagnose the disparity between online and offline performance of the trained emulators, and develop a sampling algorithm to treat biases on the tails of the distribution without adversely impacting mean performance. 

It has been shown in previous efforts [1] that stellar offline performance does not necessarily guarantee adequate online performance, or even stability. Error analysis reveals that the majority of the samples are learned quickly, while some stubborn samples remain poorly represented. We find that the more error-prone samples are those with wind profiles that have large shears– this is consistent with physical intuition as gravity waves encounter a wider range of critical levels when experiencing large shear;  therefore parameterizing gravity waves for these samples is a more difficult, complex task. To remedy this, we develop a sampling strategy that performs a parameterized histogram equalization, a concept borrowed from 1D optimal transport. 

The sampling algorithm uses a linear mapping from the original histogram to a more uniform histogram parameterized by $t \in [0,1]$, where $t=0$ recovers the original distribution and $t=1$ enforces a completely uniform distribution. A given value $t$ assigns each bin a new probability which we then use to sample from each bin. If the new probability is smaller than the original, then we invoke sampling without replacement, but limited to a reduced number consistent with the new probability. If the new probability is larger than the original, then we repeat all the samples in the bin up to some predetermined maximum repeat value (a threshold to avoid extreme oversampling at the tails). We optimize this sampling algorithm with respect to $t$, the maximum repeat value, and the number and distribution (uniform or not) of the histogram bins. The ideal combination of those parameters yields errors that are closer to a constant function of the shear metrics while maintaining high accuracy over the whole dataset. Although we study the performance of this algorithm in the context of training a gravity wave parameterization emulator, this strategy can be used for learning datasets with long tail distributions where the rare samples are associated with low accuracy. Instances of this type of datasets are prevalent in earth system dynamics: launching of gravity waves, and extreme events like hurricanes, heat waves are just a few examples. 

[1] Espinosa, Z. I., A. Sheshadri, G. R. Cain, E. P. Gerber, and K. J. DallaSanta, 2021: A Deep Learning Parameterization of Gravity Wave Drag Coupled to an Atmospheric Global Climate Model,Geophys. Res. Lett., in review. []

How to cite: Yang, L. and Gerber, E.: Sampling strategies for data-driven parameterization of gravity wave momentum transport, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5766,, 2022.

EGU22-5980 | Presentations | ITS2.6/AS5.1 | Highlight

Probabilistic forecasting of heat waves with deep learning 

George Miloshevich, Valerian Jacques-Dumas, Pierre Borgnat, Patrice Abry, and Freddy Bouchet
Extreme events such as storms, floods, cold spells and heat waves are expected to have an increasing societal impact with climate change. However the study of rare events is complicated due to computational costs of highly complex models and lack of observations. However, with the help of machine learning synthetic models for forecasting can be constructed and cheaper resampling techniques can be developed. Consequently, this may also clarify more regional impacts of climate change. .

In this work, we perform detailed analysis of how deep neural networks (DNNs) can be used in intermediate-range forecasting of prolonged heat waves of duration of several weeks over synoptic spatial scales. In particular, we train a convolutional neural network (CNN) on the 7200 years of a simulation of a climate model. As such, we are interested in probabilistic prediction (committor function in transition theory). Thus we discuss the proper forecasting scores such as Brier skill score, which is popular in weather prediction, and cross-entropy skill, which is based on information-theoretic considerations. They allow us to measure the success of various architectures and investigate more efficient pipelines to extract the predictions from physical observables such as geopotential, temperature and soil moisture. A priori, the committor is hard to visualize as it is a high dimensional function of its inputs, the grid points of the climate model for a given field. Fortunately, we can construct composite maps conditioned to its values which reveal that the CNN is likely relying on the global teleconnection patterns of geopotential. On the other hand, soil moisture signal is more localized with predictive capability over much longer times in future (at least a month). The latter fact relates to the soil-atmosphere interactions. One expects the performance of DNNs to greatly improve with more data. We provide quantitative assessment of this fact. In addition, we offer more details on how the undersampling of negative events affects the knowledge of the committor function. We show that transfer learning helps ensure that the committor is a smooth function along the trajectory. This will be an important quality when such a committor will be applied in rare event algorithms for importance sampling. 
While DNNs are universal function approximators the issue of extrapolation can be somewhat problematic. In addressing this question we train a CNN on a dataset generated from a simulation without a diurnal cycle, where the feedbacks between soil moisture and heat waves appear to be significantly stronger. Nevertheless, when the CNN with the given weights is validated on a dataset generated from a simulation with a daily cycle the predictions seem to generalize relatively well, despite a small reduction in skill. This generality validates the approach. 

How to cite: Miloshevich, G., Jacques-Dumas, V., Borgnat, P., Abry, P., and Bouchet, F.: Probabilistic forecasting of heat waves with deep learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5980,, 2022.

EGU22-6479 | Presentations | ITS2.6/AS5.1

Parameter inference and uncertainty quantification for an intermediate complexity climate model 

Benedict Roeder, Jakob Schloer, and Bedartha Goswami

Well-adapted parameters in climate models are essential to make accurate predictions
for future projections. In climate science, the record of precise and comprehensive obser-
vational data is rather short and parameters of climate models are often hand-tuned or
learned from artificially generated data. Due to limited and noisy data, one wants to use
Bayesian models to have access to uncertainties of the inferred parameters. Most popu-
lar algorithms for learning parameters from observational data like the Kalman inversion
approach only provide point estimates of parameters.
In this work, we compare two Bayesian parameter inference approaches applied to the
intermediate complexity model for the El Niño-Southern Oscillation by Zebiak & Cane. i)
The "Calibrate, Emulate, Sample" (CES) approach, an extension of the ensemble Kalman
inversion which allows posterior inference by emulating the model via Gaussian Processes
and thereby enables efficient sampling. ii) The simulation-based inference (SBI) approach
where the approximate posterior distribution is learned from simulated model data and
observational data using neural networks.
We evaluate the performance of both approaches by comparing their run times and the
number of required model evaluations, assess the scalability with respect to the number
of inference parameters, and examine their posterior distributions.

How to cite: Roeder, B., Schloer, J., and Goswami, B.: Parameter inference and uncertainty quantification for an intermediate complexity climate model, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6479,, 2022.

EGU22-6553 | Presentations | ITS2.6/AS5.1

Can simple machine learning methods predict concentrations of OH better than state of the art chemical mechanisms? 

Sebastian Hickman, Paul Griffiths, James Weber, and Alex Archibald

Concentrations of the hydroxyl radical, OH, control the lifetime of methane, carbon monoxide and other atmospheric constituents.  The short lifetime of OH, coupled with the spatial and temporal variability in its sources and sinks, makes accurate simulation of its concentration particularly challenging. To date, machine learning (ML) methods have been infrequently applied to global studies of atmospheric chemistry.

We present an assessment of the use of ML methods for the challenging case of simulation of the hydroxyl radical at the global scale, and show that several approaches are indeed viable.  We use observational data from the recent NASA Atmospheric Tomography Mission to show that machine learning methods are comparable in skill to state of the art forward chemical models and are capable, if appropriately applied, of simulating OH to within observational uncertainty.  

We show that a simple ridge regression model is a better predictor of OH concentrations in the remote atmosphere than a state of the art chemical mechanism implemented in a forward box model. Our work shows that machine learning may be an accurate emulator of chemical concentrations in atmospheric chemistry, which would allow a significant speed up in climate model runtime due to the speed and efficiency of simple machine learning methods. Furthermore, we show that relatively few predictors are required to simulate OH concentrations, suggesting that the variability in OH can be quantitatively accounted for by few observables with the potential to simplify the numerical simulation of atmospheric levels of key species such as methane. 

How to cite: Hickman, S., Griffiths, P., Weber, J., and Archibald, A.: Can simple machine learning methods predict concentrations of OH better than state of the art chemical mechanisms?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6553,, 2022.

EGU22-6674 | Presentations | ITS2.6/AS5.1

The gravity wave parameterization calibration problem: A 1D QBO model testbed 

Ofer Shamir, L. Minah Yang, David S. Connelly, and Edwin P. Gerber

An essential step in implementing any new parameterization is calibration, where the parameterization is adjusted to work with an existing model and yield some desired improvement. In the context of gravity wave (GW) momentum transport, calibration is necessitated by the facts that: (i) Some GWs are always at least partially resolved by the model, and hence a parameterization should only account for the missing waves. Worse, the parameterization may need to correct for the misrepresentation of under-resolved GWs, i.e., coarse vertical resolution can bias GW breaking level, leading to erroneous momentum forcing. (ii) The parameterized waves depend on the resolved solution for both their sources and dissipation, making them susceptible to model biases. Even a "perfect" parameterization could then yield an undesirable result, e.g., an unrealistic Quasi-Biennial Oscillation (QBO).  While model-specific calibration is required, one would like a general "recipe" suitable for most models. From a practical point of view, the adoption of a new parameterization will be hindered by a too-demanding calibration process. This issue is of particular concern in the context of data-driven methods, where the number of tunable degrees of freedom is large (possibly in the millions). Thus, more judicious ways for addressing the calibration step are required. 

To address the above issues, we develop a 1D QBO model, where the "true" gravity wave momentum deposition is determined from a source distribution and critical level breaking, akin to a traditional physics-based GW parameterization. The control parameters associated with the source consist of the total wave flux (related to the total precipitation for convectively generated waves) and the spectrum width (related to the depth of convection). These parameters can be varied to mimic the variability in GW sources between different models, i.e., biases in precipitation variability. In addition, the model’s explicit diffusivity and vertical advection can be varied to mimic biases in model numerics and circulation, respectively. The model thus allows us to assess the ability of a data-driven parameterization to (i) extrapolate, capturing the response of GW momentum transport to a change in the model parameters and (ii) be calibrated, adjusted to maintain the desired simulation of the QBO in response to a change in the model parameters. The first property is essential for a parameterization to be used for climate prediction, the second, for a parameterization to be used at all. We focus in particular on emulators of the GW momentum transport based on neural network and regression trees, contrasting their ability to satisfy both of these goals.  


How to cite: Shamir, O., Yang, L. M., Connelly, D. S., and Gerber, E. P.: The gravity wave parameterization calibration problem: A 1D QBO model testbed, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6674,, 2022.

All oceanic general circulation models (GCMs) include parametrizations of the unresolved subgrid-scale (eddy) effects on the large-scale motions, even at the (so-called) eddy-permitting resolutions. Among the many problems associated with the development of accurate and efficient eddy parametrizations, one problem is a reliable decomposition of a turbulent flow into resolved and unresolved (subgrid) scale components. Finding an objective way to separate eddies is a fundamental, critically important and unresolved problem. 
Here a statistically consistent correlation-based flow decomposition method (CBD) that employs the Gaussian filtering kernel with geographically varying topology – consistent with the observed local spatial correlations – achieves the desired scale separation. CBD is demonstrated for an eddy-resolving solution of the classical midlatitude double-gyre quasigeostrophic (QG) circulation, that possess two asymmetric gyres of opposite circulations and a strong meandering eastward jet, such as the Gulf Stream in the North Atlantic and Kuroshio in the North Pacific. CBD facilitates a comprehensive analysis of the feedbacks of eddies on the large-scale flow via the transient part of the eddy forcing. A  `product integral' based on time-lagged correlation between the diagnosed eddy forcing and the evolving large-scale flow, uncovers robust `eddy backscatter' mechanism. Data-driven augmentation of non-eddy-resolving ocean model by stochastically-emulated eddy fields allows to restore the missing eddy-driven features, such as the merging western boundary currents, their eastward extension and low-frequency variabilities of gyres.

  • N. Argawal, Ryzhov, E.A., Kondrashov, D., and P.S. Berloff, 2021: Correlation-based flow decomposition and statistical analysis of the eddy forcing, Journal of Fluid Mechanics, 924, A5. doi:10.1017/jfm.2021.604

  • N. Argawal, Kondrashov, D., Dueben, P., Ryzhov, E.A., and P.S. Berloff, 2021: A comparison of data-driven approaches to build low-dimensional ocean modelsJournal of Advances in Modelling Earth Systems, doi:10.1029/2021MS002537


How to cite: Kondrashov, D.: Towards physics-informed stochastic parametrizations of subgrid physics in ocean models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6859,, 2022.

EGU22-7044 | Presentations | ITS2.6/AS5.1

Seismic Event Characterization using Manifold Learning Methods 

Yuri Bregman, Yochai Ben Horin, Yael Radzyner, Itay Niv, Maayan Kahlon, and Neta Rabin

Manifold learning is a branch of machine learning that focuses on compactly representing complex data-sets based on their fundamental intrinsic parameters. One such method is diffusion maps, which reduces the dimension of the data while preserving its geometric structure. In this work, diffusion maps are applied to several seismic event characterization tasks. The first task is automatic earthquake-explosion discrimination, which is an essential component of nuclear test monitoring. We also use this technique to automatically identify mine explosions and aftershocks following large earthquakes. Identification of such events helps to lighten the analysts’ burden and allow for timely production of reviewed seismic bulletins.

The proposed methods begin with a pre-processing stage in which a time–frequency representation is extracted from each seismogram while capturing common properties of seismic events and overcoming magnitude differences. Then, diffusion maps are used in order to construct a low-dimensional model of the original data. In this new low-dimensional space, classification analysis is carried out.

The algorithm’s discrimination performance is demonstrated on several seismic data sets. For instance, using the seismograms from EIL station, we identify arrivals that were caused by explosions at the nearby Eshidiya mine in Jordan. The model provides a visualization of the data, organized by its intrinsic factors. Thus, along with the discrimination results, we provide a compact organization of the data that characterizes the activity patterns in the mine.

Our results demonstrate the potential and strength of the manifold learning based approach, which may be suitable to other in other geophysics domains.

How to cite: Bregman, Y., Ben Horin, Y., Radzyner, Y., Niv, I., Kahlon, M., and Rabin, N.: Seismic Event Characterization using Manifold Learning Methods, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7044,, 2022.

Accurate streamflow forecasts can provide guidance for reservoir managements, which can regulate river flows, manage water resources and mitigate flood damages. One popular way to forecast streamflow is to use bias-corrected meteorological forecasts to drive a calibrated hydrological model. But for cascade reservoirs, such approaches suffer significant deficiencies because of the difficulty to simulate reservoir operations by physical approach and the uncertainty of meteorological forecasts over small catchment. Another popular way is to forecast streamflow with machine learning method, which can fit a statistical model without inputs like reservoir operating rules. Thus, we integrate meteorological forecasts, land surface hydrological model and machine learning to forecast hourly streamflow over the Yantan catchment, which is one of the cascade reservoirs in the Hongshui River with streamflow influenced by both the upstream reservoir water release and the rainfall runoff process within the catchment.

Before evaluating the streamflow forecast system, it is necessary to investigate the skill by means of a series of specific hindcasts that isolate potential sources of predictability, like meteorological forcing and the initial condition (IC). Here, we use ensemble streamflow prediction (ESP)/reverse ESP (revESP) method to explore the impact of IC on hourly stream prediction. Results show that the effect of IC on runoff prediction is 16 hours. In the next step, we evaluate the hourly streamflow hindcasts during the rainy seasons of 2013-2017 performed by the forecast system. We use European Centre for Medium-Range Weather Forecasts perturbed forecast forcing from the THORPEX Interactive Grand Global Ensemble (TIGGE-ECMWF) as meteorological inputs to perform the hourly streamflow hindcasts. Compared with the ESP, the hydrometeorological ensemble forecast approach reduces probabilistic and deterministic forecast errors by 6% during the first 7 days. After integrated the long short-term memory (LSTM) deep learning method into the system, the deterministic forecast error can be further reduced by 6% in the first 72 hours. We also use historically observed streamflow to drive another LSTM model to perform an LSTM-only streamflow forecast. Results show that its skill sharply dropped after the first 24 hours, which indicates that the meteorology-hydrology modeling approach can improve the streamflow forecast.

How to cite: Liu, J. and Yuan, X.: Reservoir inflow forecast by combining meteorological ensemble forecast, physical hydrological simulation and machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7093,, 2022.

EGU22-7113 | Presentations | ITS2.6/AS5.1 | Highlight

Coupling regional air quality simulations of EURAD-IM with street canyon observations - a machine learning approach 

Charlotte Neubacher, Philipp Franke, Alexander Heinlein, Axel Klawonn, Astrid Kiendler-Scharr, and Anne-Caroline Lange

State of the art atmospheric chemistry transport models on regional scales as the EURAD-IM (EURopean Air pollution Dispersion-Inverse Model) simulate physical and chemical processes in the atmosphere to predict the dispersion of air pollutants. With EURAD-IM’s 4D-var data assimilation application, detailed analyses of the air quality can be conducted. These analyses allow for improvements of atmospheric chemistry forecast as well as emission source strength assessments. Simulations of EURAD-IM can be nested to a spatial resolution of 1 km, which does not correspond to the urban scale. Thus, inner city street canyon observations cannot be exploited since here, anthropogenic pollution vary vastly over scales of 100 m or less.

We address this issue by implementing a machine learning (ML) module into EURAD-IM, forming a hybrid model that enable bridging the representativeness gap between model resolution and inner-city observations. Thus, the data assimilation of EURAD-IM is strengthened by additional observations in urban regions. Our approach of the ML module is based on a neural network (NN) with relevant environmental information of street architecture, traffic density, meteorology, and atmospheric pollutant concentrations from EURAD-IM as well as the street canyon observation of pollutants as input features. The NN then maps the observed concentration from street canyon scale to larger spatial scales.

We are currently working with a fully controllable test environment created from EURAD-IM forecasts of the years 2020 and 2021 at different spatial resolutions. Here, the ML model maps the high-resolution hourly NO2 concentration to the concentration of the low resolution model grid. It turns out that it is very difficult for NNs to learn the hourly concentrations with equal accuracy using diurnal cycles of pollutant concentrations. Thus, we develop a model that uses an independent NN for each hour to support time-of-day learning. This allows to reduce the training error by a factor of 102. As a proof of concept, we trained the ML model in an overfitting regime where the mean squared training error reduce to 0.001% for each hour. Furthermore, by optimizing the hyperparameters and introducing regularization terms to reduce the overfitting, we achieved a validation error of 9−12% during night and 9−16% during day.

How to cite: Neubacher, C., Franke, P., Heinlein, A., Klawonn, A., Kiendler-Scharr, A., and Lange, A.-C.: Coupling regional air quality simulations of EURAD-IM with street canyon observations - a machine learning approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7113,, 2022.

EGU22-7135 | Presentations | ITS2.6/AS5.1 | Highlight

How to calibrate a climate model with neural network based physics? 

Blanka Balogh, David Saint-Martin, and Aurélien Ribes

Unlike the traditional subgrid scale parameterizations used in climate models, current neural network (NN) parameterizations are only tuned offline, by minimizing a loss function on outputs from high resolution models. This approach often leads to numerical instabilities and long-term biases. Here, we propose a method to design tunable NN parameterizations and calibrate them online. The calibration of the NN parameterization is achieved in two steps. First, some model parameters are included within the NN model input. This NN model is fitted at once for a range of values of the parameters, using an offline metric. Second, once the NN parameterization has been plugged into the climate model, the parameters included among the NN inputs are optimized with respect to an online metric quantifying errors on long-term statistics. We illustrate our method with two simple dynamical systems. Our approach significantly reduces long-term biases of the climate model with NN based physics.

How to cite: Balogh, B., Saint-Martin, D., and Ribes, A.: How to calibrate a climate model with neural network based physics?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7135,, 2022.

EGU22-8279 | Presentations | ITS2.6/AS5.1

Using deep learning to improve the spatial resolution of the ocean model 

Ihor Hromov, Georgy Shapiro, Jose Ondina, Sanjay Sharma, and Diego Bruciaferri

For the ocean models, the increase of spatial resolution is a matter of significant importance and thorough research. Computational resources limit our capabilities of the increase in model resolution. This constraint is especially true for the traditional dynamical models, for which an increase of a factor of two in the horizontal resolution results in simulation times increased approximately tenfold. One of the potential methods to relax this limitation is to use Artificial Intelligence methods, such as Neural Networks (NN). In this research, NN is applied to ocean circulation modelling. More specifically, NN is used on data output from the dynamical model to increase the spatial resolution of the model output. The main dataset being used is Sea Surface Temperature data in 0.05- and 0.02-degree horizontal resolutions for Irish Sea. 

Several NN architectures were applied to address the task. Generative Adversarial Networks (GAN), Convolutional Neural Networks (CNN) and Multi-level Wavelet CNN. They are used in other areas of knowledge in problems related to the increase of resolution. The work will contrast and compare the efficiency of and present a provisional assessment of the efficiency of each of the methods. 

How to cite: Hromov, I., Shapiro, G., Ondina, J., Sharma, S., and Bruciaferri, D.: Using deep learning to improve the spatial resolution of the ocean model, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8279,, 2022.

EGU22-8334 | Presentations | ITS2.6/AS5.1

Information theory solution approach for air-pollution sensors' location-allocation problem 

Barak Fishbain, Ziv Mano, and Shai Kendler

Urbanization and industrialization processes are accompanied by adverse environmental effects, such as air pollution. The first action in reducing air pollution is the detection of its source(s). This is achievable through monitoring. When deploying a sensor array, one must balance between the array's cost and performance. This optimization problem is known as the location-allocation problem. Here, a new solution approach, which draws its foundation from information theory is presented. The core of the method is air-pollution levels computed by a dispersion model in various meteorological conditions. The sensors are then placed in the locations which information theory identifies as the most uncertain. The method is compared with two other heuristics typically applied for solving the location-allocation problem. In the first, sensors are randomly deployed, in the second, the sensors are placed according to the maximal cumulative pollution levels (i.e., hot spot). For the comparison two simulated scenes were evaluated, one contains point sources and buildings, and the other also contains line sources (i.e., roads). It shows that the Entropy method resulted in a superior sensors' deployment compared to the other two approaches in terms of source apportionment and dense pollution field reconstruction from the sensors' network measurements.

How to cite: Fishbain, B., Mano, Z., and Kendler, S.: Information theory solution approach for air-pollution sensors' location-allocation problem, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8334,, 2022.

EGU22-8719 | Presentations | ITS2.6/AS5.1

Multi-station Multivariate Multi-step Convection Nowcasting with Deep Neural Networks 

Sandy Chkeir, Aikaterini Anesiadou, and Riccardo Biondi

Extreme weather nowcasting has always been a challenging task in meteorology. Many research studies have been conducted to accurately forecast extreme weather events, related to rain rates and/or wind speed thresholds, in spatio-temporal scales. Over decades, this field gained attention in the artificial intelligence community which is aiming towards creating more accurate models using the latest algorithms and methods.  

In this work, within the H2020 SESAR ALARM project, we aim to nowcast rain and wind speed as target features using different input configurations of the available sources such as weather stations, lightning detectors, radar, GNSS receivers, radiosonde and radio occultations data. This nowcasting task has been firstly conducted at 14 local stations around Milano Malpensa Airport as a short-term temporal multi-step forecasting. At a second step, all stations will be combined, meaning that the forecasting becomes a spatio-temporal problem. Concretely, we want to investigate the predicted rain and wind speed values using the different inputs for two case scenarios: for each station, and joining all stations together. 

The chaotic nature of the atmosphere, e.g. non-stationarity of the driving series of each weather feature, makes the predictions unreliable and inaccurate and thus dealing with these data is a very delicate task. For this reason, we have devoted some work to cleaning, feature engineering and preparing the raw data before feeding them into the model architectures. We have managed to preprocess large amounts of data for local stations around the airport, and studied the feasibility of nowcasting rain and wind speed targets using different data sources altogether. The temporal multivariate driving series have high dimensionality and we’ve  made multi-step predictions for the defined target functions.

We study and test different machine learning architectures starting from simple multi-layer perceptrons to convolutional models, and Recurrent Neural Networks (RNN) for temporal and spatio-temporal nowcasting. The Long Short-Term Memory (LSTM) encoder decoder architecture outperforms other models achieving more accurate predictions for each station separately.  Furthermore, to predict the targets in a spatio-temporal scale, we will deploy a 2-layer spatio-temporal stacked LSTM model consisting of independent LSTM models per location in the first LSTM layer, and another LSTM layer to finally predict targets for multi-steps ahead. And the results obtained with different algorithm architectures applied to a dense network of sensors are to be reported.

How to cite: Chkeir, S., Anesiadou, A., and Biondi, R.: Multi-station Multivariate Multi-step Convection Nowcasting with Deep Neural Networks, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8719,, 2022.

EGU22-8852 | Presentations | ITS2.6/AS5.1

Time-dependent Hillshades: Dispelling the Shadow Curse of Machine Learning Applications in Earth Observation 

Freddie Kalaitzis, Gonzalo Mateo-Garcia, Kevin Dobbs, Dolores Garcia, Jason Stoker, and Giovanni Marchisio

We show that machine learning models learn and perform better when they know where to expect shadows, through hillshades modeled to the time of imagery acquisition.

Shadows are detrimental to all machine learning applications on satellite imagery. Prediction tasks like semantic / instance segmentation, object detection, counting of rivers, roads, buildings, trees, all rely on crisp edges and colour gradients that are confounded by the presence of shadows in passive optical imagery, which rely on the sun’s illumination for reflectance values.

Hillshading is a standard technique for enriching a mapped terrain with relief effects, which is done by emulating the shadow caused by steep terrain and/or tall vegetation. A hillshade that is modeled to the time of day and year can be easily derived through a basic form of ray tracing on a Digital Terrain Model (DTM) (also known as a bare-earth DEM) or Digital Surface Model (DSM) given the sun's altitude and azimuth angles. In this work, we use lidar-derived DSMs. A DSM-based hillshade conveys a lot more information on shadows than a bare-earth DEM alone, namely any non-terrain vertical features (e.g. vegetation, buildings) resolvable at a 1-m resolution. The use of this level of fidelity of DSM for hillshading and its input to a machine learning model is novel and the main contribution of our work. Any uncertainty over the angles can be captured through a composite multi-angle hillshade, which shows the range where shadows can appear throughout the day.

We show the utility of time-dependent hillshades in the daily mapping of rivers from Very High Resolution (VHR) passive optical and lidar-derived terrain data [1]. Specifically, we leverage the acquisition timestamps within a daily 3m PlanetScope product over a 2-year period. Given a datetime and geolocation, we model the sun’s azimuth and elevation relative to that geolocation at that time of day and year. We can then generate a time-dependent hillshade and therefore locate shadows in any given time within that 2-year period. In our ablation study we show that, out of all the lidar-derived products, the time-dependent hillshades contribute a 8-9% accuracy improvement in the semantic segmentation of rivers. This indicates that a semantic segmentation machine learning model is less prone to errors of commission (false positives), by better disambiguating shadows from dark water.

Time-dependent hillshades are not currently used in ML for EO use-cases, yet they can be useful. All that is needed to produce them is access to high-resolution bare-earth DEMs, like that of the US National 3D Elevation Program covering the entire continental U.S at 1-meter resolution, or creation of DSMs from the lidar point cloud data itself. As the coverage of DSM and/or DEM products expands to more parts of the world, time-dependent hillshades could become as commonplace as cloud masks in EO use cases.

[1] Dolores Garcia, Gonzalo Mateo-Garcia, Hannes Bernhardt, Ron Hagensieker, Ignacio G. Lopez-Francos, Jonathan Stock, Guy Schumann, Kevin Dobbs and Freddie Kalaitzis Pix2Streams: Dynamic Hydrology Maps from Satellite-LiDAR Fusion. AI for Earth Sciences Workshop, NeurIPS 2020

How to cite: Kalaitzis, F., Mateo-Garcia, G., Dobbs, K., Garcia, D., Stoker, J., and Marchisio, G.: Time-dependent Hillshades: Dispelling the Shadow Curse of Machine Learning Applications in Earth Observation, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8852,, 2022.

EGU22-9348 | Presentations | ITS2.6/AS5.1

Data-driven modelling of soil moisture: mapping organic soils 

Doran Khamis, Matt Fry, Hollie Cooper, Ross Morrison, and Eleanor Blyth

Improving our understanding of soil moisture and hydraulics is crucial for flood prediction, smart agriculture, modelling nutrient and pollutant spread and evaluating the role of land as a sink or source of carbon and other greenhouse gases. State of the art land surface models rely on poorly-resolved soil textural information to parametrise arbitrarily layered soil models; soils rich in organic matter – key to understanding the role of the land in achieving net zero carbon – are not well modelled. Here, we build a predictive data-driven model of soil moisture using a neural network composed of transformer layers to process time series data from point-sensors (precipitation gauges and sensor-derived estimates of potential evaporation) and convolutional layers to process spatial atmospheric driving data and contextual information (topography, land cover and use, location and catchment behaviour of water bodies). We train the model using data from the COSMOS-UK sensor network and soil moisture satellite products and compare the outputs with JULES to investigate where and why the models diverge. Finally, we predict regions of high peat content and propose a way to combine theory with our data-driven approach to move beyond the sand-silt-clay modelling framework.

How to cite: Khamis, D., Fry, M., Cooper, H., Morrison, R., and Blyth, E.: Data-driven modelling of soil moisture: mapping organic soils, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9348,, 2022.

EGU22-9452 | Presentations | ITS2.6/AS5.1

Eddy identification from along track altimeter data using deep learning: EDDY project 

Adili Abulaitijiang, Eike Bolmer, Ribana Roscher, Jürgen Kusche, Luciana Fenoglio, and Sophie Stolzenberger

Eddies are circular rotating water masses, which are usually generated near the large ocean currents, e.g., Gulf Stream. Monitoring eddies and gaining knowledge on eddy statistics over a large region are important for fishery, marine biology studies, and testing ocean models.

At mesoscale, eddies are observed in radar altimetry, and methods have been developed to identify, track and classify them in gridded maps of sea surface height derived from multi-mission data sets. However, this procedure has drawbacks since much information is lost in the gridded maps. Inevitably, the spatial and temporal resolution of the original altimetry data degrades during the gridding process. On the other hand, the task of identifying eddies has been a post-analysis process on the gridded dataset, which is, by far, not meaningful for near-real time applications or forecasts. In the EDDY project at the University of Bonn, we aim to develop methods for identifying eddies directly from along track altimetry data via a machine (deep) learning approach.

At the early stage of the project, we started with gridded altimetry maps to set up and test the machine learning algorithm. The gridded datasets are not limited to multi-mission gridded maps from AVISO, but also include the high resolution (~6 km) ocean modeling simulation dataset (e.g., FESOM, Finite Element Sea ice Ocean Model). Later, the gridded maps are sampled along the real altimetry ground tracks to obtain the single-track altimetry data. Reference data, as the training set for machine learning, will be produced by open-source geometry-based approach (e.g., py-eddy-tracker, Mason et al., 2014) with additional constraints like Okubo-Weiss parameter and Sea Surface Temperature (SST) profile signatures.

In this presentation, we introduce the EDDY project and show the results from the machine learning approach based on gridded datasets for the Gulf stream area for the period 2017, and first results of single-track eddy identification in the region.

How to cite: Abulaitijiang, A., Bolmer, E., Roscher, R., Kusche, J., Fenoglio, L., and Stolzenberger, S.: Eddy identification from along track altimeter data using deep learning: EDDY project, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9452,, 2022.

DINCAE (Data INterpolating Convolutional Auto-Encoder) is a neural network to reconstruct missing data (e.g. obscured by clouds or gaps between tracks) in satellite data. Contrary to standard image reconstruction (in-painting) with neural networks, this application requires a method to handle missing data (or data with variable accuracy) already in the training phase. Instead of using a cost function based on the mean square error, the neural network (U-Net type of network) is optimized by minimizing the negative log likelihood assuming a Gaussian distribution (characterized by a mean and a variance). As a consequence, the neural network also provides an expected error variance of the reconstructed field (per pixel and per time instance).


In this updated version DINCAE 2.0, the code was rewritten in Julia and a new type of skip connection has been implemented which showed superior performance with respect to the previous version. The method has also been extended to handle multivariate data (an example will be shown with sea-surface temperature, chlorophyll concentration and wind fields). The improvement of this network is demonstrated in the Adriatic Sea. 


Convolutional networks work usually with gridded data as input. This is however a limitation for some data types used in oceanography and in Earth Sciences in general, where observations are often irregularly sampled.  The first layer of the neural network and the cost function have been modified so that unstructured data can also be used as inputs to obtain gridded fields as output. To demonstrate this, the neural network is applied to along-track altimetry data in the Mediterranean Sea. Results from a 20-year reconstruction are presented and validated. Hyperparameters are determined using Bayesian optimization and minimizing the error relative to a development dataset.

How to cite: Barth, A., Alvera-Azcárate, A., Troupin, C., and Beckers, J.-M.: A multivariate convolutional autoencoder to reconstruct satellite data with an error estimate based on non-gridded observations: application to sea surface height, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9578,, 2022.

EGU22-9734 | Presentations | ITS2.6/AS5.1

High Impact Weather Forecasts in Southern Brazil using Ensemble Precipitation Forecasts and Machine Learning 

Cesar Beneti, Jaqueline Silveira, Leonardo Calvetti, Rafael Inouye, Lissette Guzman, Gustavo Razera, and Sheila Paz

In South America, southern parts of Brazil, Paraguay and northeast Argentina are regions particularly prone to high impact weather (intensive lightning activity, high precipitation, hail, flash floods and occasional tornadoes), mostly associated with extra-tropical cyclones, frontal systems and Mesoscale Convective Systems. In the south of Brazil, agricultural industry and electrical power generation are the main economic activities. This region is responsible for 35% of all hydro-power energy production in the country, with long transmission lines to the main consumer regions, which are severely affected by these extreme weather conditions. Intense precipitation events are a common cause of electricity outages in southern Brazil, which ranks as one of the regions in Brazil with the highest annual lightning incidence, as well. Accurate precipitation forecasts can mitigate this kind of problem. Despite improvements in the precipitation estimates and forecasts, some difficulties remain to increase the accuracy, mainly related to the temporal and spatial location of the events. Although several options are available, it is difficult to identify which deterministic forecast is the best or the most reliable forecast. Probabilistic products from large ensemble prediction systems provide a guide to forecasters on how confident they should be about the deterministic forecast, and one approach is using post processing methods such as machine learning (ML), which has been used to identify patterns in historical data to correct for systematic ensemble biases.

In this paper, we present a study, in which we used 20 members from the Global Ensemble Forecast System (GEFS) and 50 members from European Centre for Medium-Range Weather Forecasts (ECMWF)  during 2019-2021,  for seven daily precipitation thresholds: 0-1.0mm, 1.0mm-15mm, 15mm-40mm, 40mm-55mm, 55mm-105mm, 105mm-155mm and over 155mm. A ML algorithm was developed for each day, up to 15 days of forecasts, and several skill scores were calculated, for these daily precipitation thresholds. Initially, to select the best members of the ensembles, a gradient boosting algorithm was applied, in order to improve the skill of the model and reduce processing time. After preprocessing the data, a random forest classifier was used to train the model. Based on hyperparameter sensitivity tests, the random forest required 500 trees, a maximum tree depth of 12 levels, at least 20 samples per leaf node, and the minimization of entropy for splits. In order to evaluate the models, we used a cross-validation on a limited data sample. The procedure has a single parameter that refers to the number of groups that a given data sample is to be split into. In our work we created a twenty-six fold cross validation with 30 days per fold to verify the forecasts. The results obtained by the RF were evaluated through estimated value versus observed value. For the forecast range, we found values above 75% for the precision metrics in the first 3 days, and around 68% in the next days. The recall was also around 80% throughout the entire forecast range,  with promising results to apply this technique operationally, which is our intent in the near future. 

How to cite: Beneti, C., Silveira, J., Calvetti, L., Inouye, R., Guzman, L., Razera, G., and Paz, S.: High Impact Weather Forecasts in Southern Brazil using Ensemble Precipitation Forecasts and Machine Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9734,, 2022.

EGU22-9833 | Presentations | ITS2.6/AS5.1

Deep learning for laboratory earthquake prediction and autoregressive forecasting of fault zone stress 

Laura Laurenti, Elisa Tinti, Fabio Galasso, Luca Franco, and Chris Marone

Earthquakes forecasting and prediction have long, and in some cases sordid, histories but recent work has rekindled interest in this area based on advances in short-term early warning, hazard assessment for human induced seismicity and successful prediction of laboratory earthquakes.

In the lab, frictional stick-slip events provide an analog for the full seismic cycle and such experiments have played a central role in understanding the onset of failure and the dynamics of earthquake rupture. Lab earthquakes are also ideal targets for machine learning (ML) techniques because they can be produced in long sequences under a wide range of controlled conditions. Indeed, recent work shows that labquakes can be predicted from fault zone acoustic emissions (AE). Here, we generalize these results and explore additional ML and deep learning (DL) methods for labquake prediction. Key questions include whether improved ML/DL methods can outperform existing models, including prediction based on limited training, or if such methods can successfully forecast beyond a single seismic cycle for aperiodic failure. We describe significant improvements to existing methods of labquake prediction using simple AE statistics (variance) and DL models such as Long-Short Term Memory (LSTM) and Convolution Neural Network (CNN). We demonstrate: 1) that LSTMs and CNNs predict labquakes under a variety of conditions, including pre-seismic creep, aperiodic events and alternating slow and fast events and 2) that fault zone stress can be predicted with fidelity (accuracy in terms of R2 > 0.92), confirming that acoustic energy is a fingerprint of the fault zone stress. We predict also time to start of failure (TTsF) and time to the end of Failure (TTeF). Interestingly, TTeF is successfully predicted in all seismic cycles, while the TTsF prediction varies with the amount of fault creep before an event. We also report on a novel autoregressive forecasting method to predict future fault zone states, focusing on shear stress. This forecasting model is distinct from existing predictive models, which predict only the current state. We compare three modern approaches in sequence modeling framework: LSTM, Temporal Convolution Network (TCN) and Transformer Network (TF). Results are encouraging in forecasting the shear stress at long-term future horizons, autoregressively. Our ML/DL prediction models outperform the state of the art and our autoregressive model represents a novel forecasting framework that could enhance current methods of earthquake forecasting.

How to cite: Laurenti, L., Tinti, E., Galasso, F., Franco, L., and Marone, C.: Deep learning for laboratory earthquake prediction and autoregressive forecasting of fault zone stress, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9833,, 2022.

EGU22-10157 | Presentations | ITS2.6/AS5.1

How land cover changes affect ecosystem productivity 

Andreas Krause, Phillip Papastefanou, Konstantin Gregor, Lucia Layritz, Christian S. Zang, Allan Buras, Xing Li, Jingfeng Xiao, and Anja Rammig

Historically, many forests worldwide were cut down and replaced by agriculture. While this substantially reduced terrestrial carbon storage, the impacts of land-use change on ecosystem productivity have not been adequately resolved yet.

Here, we apply the machine learning algorithm Random Forests to predict the potential gross primary productivity (GPP) of forests, grasslands, and croplands around the globe using high-resolution datasets of satellite-derived GPP, land cover, and 20 environmental predictor variables.

With a mean potential GPP of around 2.0 kg C m-2 yr-1 forests are the most productive land cover on two thirds of the global suitable area, while grasslands and croplands are on average 23 and 9% less productive, respectively. These findings are robust against alternative input datasets and algorithms, even though results are somewhat sensitive to the underlying land cover map.

Combining our potential GPP maps with a land-use reconstruction from the Land-Use Harmonization project (LUH2) we estimate that historical agricultural expansion reduced global GPP by around 6.3 Gt C yr-1 (4.4%). This reduction in GPP induced by land cover changes is amplified in some future scenarios as a result of ongoing deforestation but partly reversed in other scenarios due to agricultural abandonment.

Finally, we compare our potential GPP maps to simulations from eight CMIP6 Earth System Models with an explicit representation of land management. While the mean GPP values of the ESM ensemble show reasonable agreement with our estimates, individual Earth System Models simulate large deviations both in terms of mean GPP values of different land cover types as well as in their spatial variations. Reducing these model biases would lead to more reliable simulations concerning the potential of land-based mitigation policies.

How to cite: Krause, A., Papastefanou, P., Gregor, K., Layritz, L., Zang, C. S., Buras, A., Li, X., Xiao, J., and Rammig, A.: How land cover changes affect ecosystem productivity, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10157,, 2022.

EGU22-10519 | Presentations | ITS2.6/AS5.1 | Highlight

Adaptive Bias Correction for Improved Subseasonal Forecasting 

Soukayna Mouatadid, Paulo Orenstein, Genevieve Flaspohler, Miruna Oprescu, Judah Cohen, Franklyn Wang, Sean Knight, Maria Geogdzhayeva, Sam Levang, Ernest Fraenkel, and Lester Mackey

Improving our ability to forecast the weather and climate is of interest to all sectors of the economy and government agencies from the local to the national level. In fact, weather forecasts 0-10 days ahead and climate forecasts seasons to decades ahead are currently used operationally in decision-making, and the accuracy and reliability of these forecasts has improved consistently in recent decades. However, many critical applications require subseasonal forecasts with lead times in between these two timescales. Subseasonal forecasting—predicting temperature and precipitation 2-6 weeks ahead—is indeed critical for effective water allocation, wildfire management, and drought and flood mitigation. Yet, accurate forecasts for the subseasonal regime are still lacking due to the chaotic nature of weather.

While short-term forecasting accuracy is largely sustained by physics-based dynamical models, these deterministic methods have limited subseasonal accuracy due to chaos. Indeed, subseasonal forecasting has long been considered a “predictability desert” due to its complex dependence on both local weather and global climate variables. Nevertheless, recent large-scale research efforts have advanced the subseasonal capabilities of operational physics-based models, while parallel efforts have demonstrated the value of machine learning and deep learning methods in improving subseasonal forecasting.

To counter the systematic errors of dynamical models at longer lead times, we introduce an adaptive bias correction (ABC) method that combines state-of-the-art dynamical forecasts with observations using machine learning. We evaluate our adaptive bias correction method in the contiguous U.S. over the years 2011-2020 and demonstrate consistent improvement over standard meteorological baselines, state-of-the-art learning models, and the leading subseasonal dynamical models, as measured by root mean squared error and uncentered anomaly correlation skill. When applied to the United States’ operational climate forecast system (CFSv2), ABC improves temperature forecasting skill by 20-47% and precipitation forecasting skill by 200-350%. When applied to the leading subseasonal model from the European Centre for Medium-Range Weather Forecasts (ECMWF), ABC improves temperature forecasting skill by 8-38% and precipitation forecasting skill by 40-80%.

Overall, we find that de-biasing dynamical forecasts with our learned adaptive bias correction method yields an effective and computationally inexpensive strategy for generating improved subseasonal forecasts and building the next generation of subseasonal forecasting benchmarks. To facilitate future subseasonal benchmarking and development, we release our model code through the subseasonal_toolkit Python package and our routinely updated SubseasonalClimateUSA dataset through the subseasonal_data Python package.

How to cite: Mouatadid, S., Orenstein, P., Flaspohler, G., Oprescu, M., Cohen, J., Wang, F., Knight, S., Geogdzhayeva, M., Levang, S., Fraenkel, E., and Mackey, L.: Adaptive Bias Correction for Improved Subseasonal Forecasting, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10519,, 2022.

EGU22-10711 | Presentations | ITS2.6/AS5.1

A new approach toward integrated inversion of reflection seismic and gravity datasets using deep learning 

Mahtab Rashidifard, Jeremie Giraud, Mark Jessell, and Mark Lindsay

Reflection seismic data, although sparsely distributed due to the high cost of acquisition, is the only type of data that can provide high-resolution images of the crust to reveal deep subsurface structures and the architectural complexity that may vector attention to minerally prospective regions. However, these datasets are not commonly considered in integrated geophysical inversion approaches due to computationally expensive forward modeling and inversion. Common inversion techniques on reflection seismic images are mostly utilized and developed for basin studies and have very limited application for hard-rock studies. Post-stack acoustic impedance inversions, for example, rely a lot on extracted petrophysical information along drilling borehole for depth correction purposes which are not necessarily available. Furthermore, the available techniques do not allow simple, automatic integration of seismic inversion with other geophysical datasets. 


 We introduce a new methodology that allows the utilization of the seismic images within the gravity inversion technique with the purpose of 3D boundary parametrization of the subsurface. The proposed workflow is a novel approach for incorporating seismic images into the integrated inversion techniques which relies on the image-ray method for depth-to-time domain conversion of seismic datasets. This algorithm uses a convolutional neural network to iterate over seismic images in time and depth domains. This iterative process is functional to compensate for the low depth resolution of the gravity datasets. We use a generalized level-set technique for gravity inversion to link the interfaces of the units with the depth-converted seismic images. The algorithm has been tested on realistic synthetic datasets generated from scenarios corresponding to different deformation histories. The preliminary results of this study suggest that post-stack seismic images can be utilized in integrated geophysical inversion algorithms without the need to run computationally expensive full wave-form inversions.  

How to cite: Rashidifard, M., Giraud, J., Jessell, M., and Lindsay, M.: A new approach toward integrated inversion of reflection seismic and gravity datasets using deep learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10711,, 2022.

EGU22-11043 | Presentations | ITS2.6/AS5.1

Framework for the deployment of DNNs in remote sensing inversion algorithms applied to Copernicus Sentinel-4 (S4) and TROPOMI/Sentinel-5 Precursor (S5P) 

Fabian Romahn, Victor Molina Garcia, Ana del Aguila, Ronny Lutz, and Diego Loyola

In remote sensing, the quantities of interest (e.g. the composition of the atmosphere) are usually not directly observable but can only be inferred indirectly via the measured spectra. To solve these inverse problems, retrieval algorithms are applied that usually depend on complex physical models, so-called radiative transfer models (RTMs). RTMs are very accurate, however also computationally very expensive and therefore often not feasible in combination with the strict time requirements of operational processing of satellite measurements. With the advances in machine learning, the methods of this field, especially deep neural networks (DNN), have become very promising for accelerating and improving the classical remote sensing retrieval algorithms. However, their application is not straightforward but instead quite challenging as there are many aspects to consider and parameters to optimize in order to achieve satisfying results.

In this presentation we show a general framework for replacing the RTM, used in an inversion algorithm, with a DNN that offers sufficient accuracy while at the same time increases the processing performance by several orders of magnitude. The different steps, sampling and generation of the training data, the selection of the DNN hyperparameters, the training and finally the integration of the DNN into an operational environment are explained in detail. We will also focus on optimizing the efficiency of each step: optimizing the generation of training samples through smart sampling techniques, accelerating the training data generation through parallelization and other optimizations of the RTM, application of tools for the DNN hyperparameter optimization as well as the use of automation tools (source code generation) and appropriate interfaces for the efficient integration in operational processing systems.

This procedure has been continuously developed throughout the last years and as a use case, it will be shown how it has been applied in the operational retrieval of cloud properties for the Copernicus satellite sensors Sentinel-4 (S4) and TROPOMI/Sentinel-5 Precursor (S5P).

How to cite: Romahn, F., Molina Garcia, V., del Aguila, A., Lutz, R., and Loyola, D.: Framework for the deployment of DNNs in remote sensing inversion algorithms applied to Copernicus Sentinel-4 (S4) and TROPOMI/Sentinel-5 Precursor (S5P), EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11043,, 2022.

EGU22-11420 | Presentations | ITS2.6/AS5.1

Histroy Matching for the tuning of coupled models: experiments on the Lorenz 96 model 

Redouane Lguensat, Julie Deshayes, and Venkatramani Balaji

The process of relying on experience and intuition to find good sets of parameters, commonly referred to as "parameter tuning" keeps having a central role in the roadmaps followed by dozens of modeling groups involved in community efforts such as the Coupled Model Intercomparison Project (CMIP). 

In this work, we study a tool from the Uncertainty Quantification community that started recently to draw attention in climate modeling: History Matching also referred to as "Iterative Refocussing". The core idea of History Matching is to run several simulations with different set of parameters and then use observed data to rule-out any parameter settings which are "implausible". Since climate simulation models are computationally heavy and do not allow testing every possible parameter setting, we employ an emulator that can be a cheap and accurate replacement. Here a machine learning algorithm, namely, Gaussian Process Regression is used for the emulating step. History Matching is then a good example where the recent advances in machine learning can be of high interest to climate modeling.

One objective of this study is to evaluate the potential for history matching to tune a climate system with multi-scale dynamics. By using a toy climate model, namely, the Lorenz 96 model, and producing experiments in perfect-model setting, we explore different types of applications of HM and highlight the strenghts and challenges of using such a technique. 

How to cite: Lguensat, R., Deshayes, J., and Balaji, V.: Histroy Matching for the tuning of coupled models: experiments on the Lorenz 96 model, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11420,, 2022.

EGU22-11465 | Presentations | ITS2.6/AS5.1

Quantile machine learning models for predicting European-wide, high resolution fine-mode Aerosol Optical Depth (AOD) based on ground-based AERONET and satellite AOD data 

Zhao-Yue Chen, Raul Méndez-Turrubiates, Hervé Petetin, Aleks Lacima, Albert Soret Miravet, Carlos Pérez García-Pando, and Joan Ballester

Air pollution is a major environmental risk factor for human health. Among the different air pollutants, Particulate Matter (PM) arises as the most prominent one, with increasing health effects over the last decades. According to the Global Burden of Disease, PM contributed to 4.14 million premature deaths globally in 2019, over twice as much as in 1990 (2.04 million). With these numbers in mind, the assessment of ambient PM exposure becomes a key issue in environmental epidemiology. However, the limited number of ground-level sites measuring daily PM values is a major constraint for the development of large-scale, high-resolution epidemiological studies.

In the last five years, there has been a growing number of initiatives estimating ground-level PM concentrations based on satellite Aerosol Optical Depth (AOD) data, representing a low-cost alternative with higher spatial coverage compared to ground-level measurements. At present, the most popular AOD product is NASA’s MODIS (Moderate Resolution Imaging Spectroradiometer), but the data that it provides is restricted to Total Aerosol Optical Depth (TAOD). Compared with TAOD, Fine-mode Aerosol Optical Depth (FAOD) better describes the distribution of small-diameter particles (e.g. PM10 and PM2.5), which are generally those associated with anthropogenic activity. Complementarily, AERONET (AErosol RObotic NETwork, which is the network of ground-based sun photometers), additionally provide Fine- and Coarse-mode Aerosol Optical Depth (FAOD and CAOD) products based on Spectral Deconvolution Algorithms (SDA).

Within the framework of the ERC project EARLY-ADAPT (, which aims to disentangle the association between human health, climate variability and air pollution to better estimate the early adaptation response to climate change, here we develop quantile machine learning models to further advance in the association between AERONET FAOD and satellite AOD over Europe during the last two decades. Due to large missing data form satellite estimations, we also included the AOD estimates from ECMWF’s Copernicus Atmosphere Monitoring Service Global Reanalysis (CAMSRA) and NASA’s Modern-Era Retrospective Analysis for Research and Applications v2 (MERRA-2), together with atmosphere, land and ocean variables such as boundary layer height, downward UV radiation and cloud cover from ECMWF’s ERA5-Land.

The models were thoroughly validated with spatial cross-validation. Preliminary results show that the R2 of the three AOD estimates (TAOD, FAOD and CAOD) predicted with quantile machine learning models range between 0.61 and 0.78, and the RMSE between 0.02 and 0.03. For the Pearson correlation with ground-level PM2.5, the predicted FAOD is highest (0.38), while 0.18, 0.11 and 0.09 are for Satellite, MERRA-2, CAMSRA AOD, respectively. This study provides three useful indicators for further estimating PM, which could improve our understanding of air pollution in Europe and open new avenues for large-scale, high-resolution environmental epidemiology studies.

How to cite: Chen, Z.-Y., Méndez-Turrubiates, R., Petetin, H., Lacima, A., Soret Miravet, A., Pérez García-Pando, C., and Ballester, J.: Quantile machine learning models for predicting European-wide, high resolution fine-mode Aerosol Optical Depth (AOD) based on ground-based AERONET and satellite AOD data, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11465,, 2022.

EGU22-11924 | Presentations | ITS2.6/AS5.1

Automated detection and classification of synoptic scale fronts from atmospheric data grids 

Stefan Niebler, Peter Spichtinger, Annette Miltenberger, and Bertil Schmidt

Automatic determination of fronts from atmospheric data is an important task for weather prediction as well as for research of synoptic scale phenomena. We developed a deep neural network to detect and classify fronts from multi-level ERA5 reanalysis data. Model training and prediction is evaluated using two different regions covering Europe and North America with data from two weather services. Due to a label deformation step performed during training we are able to directly generate frontal lines with no further thinning during post processing. Our network compares well against the weather service labels with a Critical Success Index higher than 66.9% and a Object Detection Rate of more than 77.3%. Additionally the frontal climatologies generated from our networks ouput are highly correlated (greater than 77.2%) to climatologies created from weather service data. Evaluation of cross sections of our detection results provide further insight in the characteristics of our predicted fronts and show that our networks classification is physically plausible.

How to cite: Niebler, S., Spichtinger, P., Miltenberger, A., and Schmidt, B.: Automated detection and classification of synoptic scale fronts from atmospheric data grids, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11924,, 2022.

EGU22-12043 | Presentations | ITS2.6/AS5.1

A Domain-Change Approach to the Semantic Labelling of Remote Sensing Images 

Chandrabali Karmakar, Gottfried Schwartz, Corneliu Octavian Dumitru, and Mihai Datcu

For many years, image classification – mainly based on pixel brightness statistics – has been among the most popular remote sensing applications. However, during recent years, many users were more and more interested in the application-oriented semantic labelling of remotely sensed image objects being depicted in given images.

In parallel, the development of deep learning algorithms has led to several powerful image classification and annotation tools that became popular in the remote sensing community. In most cases, these publicly available tools combine efficient algorithms with expert knowledge and/or external information ingested during an initial training phase, and we often encounter two alternative types of deep learning approaches, namely Autoencoders (AEs) and Convolutional Neural Networks (CNNs). Both approaches try to convert the pixel data of remote sensing images into semantic maps of the imaged areas. In our case, we made an attempt to provide an efficient new semantic annotation tool that helps in the semantic interpretation of newly recorded images with known and/or possibly unknown content.

Typical cases are remote sensing images depicting unexpected and hitherto uncharted phenomena such as flooding events or destroyed infrastructure. When we resort to the commonly applied AE or CNN software packages we cannot expect that existing statistics, or a few initial ground-truth annotations made by an image interpreter, will automatically lead to a perfect understanding of the image content. Instead, we have to discover and combine a number of additional relationships that define the actual content of a selected image and many of its characteristics.

Our approach consists of a two-stage domain-change approach where we first convert an image into a purely mathematical ‘topic representation’ initially introduced by Blei [1]. This representation provides statistics-based topics that do not yet require final application-oriented labelling describing physical categories or phenomena and support the idea of explainable machine learning [2]. Then, during a second stage, we try to derive physical image content categories by exploiting a weighted multi-level neural network approach that converts weighted topics into individual application-oriented labels. This domain-changing learning stage limits label noise and is initially supported by an image interpreter allowing the joint use of pixel statistics and expert knowledge [3]. The activity of the image interpreter can be limited to a few image patches. We tested our approach on a number of different use cases (e.g., polar ice, agriculture, natural disasters) and found that our concept provides promising results.  

[1] D.M. Blei, A.Y. Ng, and M.I. Jordan, (2003). Latent Dirichlet Allocation, Journal of Machine Learning Research, Vol. 3, pp. 993-1022.
[2] C. Karmakar, C.O. Dumitru, G. Schwarz, and M. Datcu (2020). Feature-free explainable data mining in SAR images using latent Dirichlet allocation, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 14, pp. 676-689.
[3] C.O. Dumitru, G. Schwarz, and M. Datcu (2021). Semantic Labelling of Globally Distributed Urban and Non-Urban Satellite Images Using High-Resolution SAR Data, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 15, pp. 6009-6068.

How to cite: Karmakar, C., Schwartz, G., Dumitru, C. O., and Datcu, M.: A Domain-Change Approach to the Semantic Labelling of Remote Sensing Images, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12043,, 2022.

EGU22-12489 | Presentations | ITS2.6/AS5.1

“Fully-automated” clustering method for stress inversions (CluStress) 

Lukács Kuslits, Lili Czirok, and István Bozsó

As it is well-known, stress fields are responsible for earthquake formation. In order to analyse stress relations in a study area using focal mechanisms’ (FMS) inversions, it is vital to consider three fundamental criteria:

(1)       The investigated area is characterized by a homogeneous stress field.

(2)       The earthquakes occur with variable directions on pre-existing faults.

(3)       The deviation of the fault slip vector from the shear stress vector is minimal (Wallace-Bott hypothesis).

The authors have attempted to develop a “fully-automated” algorithm to carry out the classification of the earthquakes as a prerequisite of stress estimations. This algorithm does not call for the setting of hyper-parameters, thus subjectivity can be reduced significantly and the running time can also decrease. Nevertheless, there is an optional hyper-parameter that is eligible to filter outliers, isolated points (earthquakes) in the input dataset.

In this presentation, they show the operation of this algorithm in case of synthetic datasets consisting of different groups of FMS and a real seismic dataset. The latter come from a survey area in the earthquake-prone Vrancea-zone (Romania). This is a relatively small region (around 30*70 km) in the external part of SE-Carpathians where the distribution of the seismic events is quite dense and heterogeneous.

It shall be noted that though the initial results are promising, further developments are still necessary. The source codes are soon to be uploaded to a public GitHub repository which will be available for the whole scientific community.

How to cite: Kuslits, L., Czirok, L., and Bozsó, I.: “Fully-automated” clustering method for stress inversions (CluStress), EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12489,, 2022.

EGU22-12549 | Presentations | ITS2.6/AS5.1

Joint calibration and mapping of satellite altimetry data using trainable variaitional models 

Quentin Febvre, Ronan Fablet, Julien Le Sommer, and Clément Ubelmann

Satellite radar altimeters are a key source of observation of ocean surface dynamics. However, current sensor technology and mapping techniques do not yet allow to systematically resolve scales smaller than 100km. With their new sensors, upcoming wide-swath altimeter missions such as SWOT should help resolve finer scales. Current mapping techniques rely on the quality of the input data, which is why the raw data go through multiple preprocessing stages before being used. Those calibration stages are improved and refined over many years and represent a challenge when a new type of sensor start acquiring data.

We show how a data-driven variational data assimilation framework could be used to jointly learn a calibration operator and an interpolator from non-calibrated data . The proposed framework significantly outperforms the operational state-of-the-art mapping pipeline and truly benefits from wide-swath data to resolve finer scales on the global map as well as in the SWOT sensor geometry.


How to cite: Febvre, Q., Fablet, R., Le Sommer, J., and Ubelmann, C.: Joint calibration and mapping of satellite altimetry data using trainable variaitional models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12549,, 2022.

EGU22-12574 | Presentations | ITS2.6/AS5.1 | Highlight

SWIFT-AI: Significant Speed-up in Modelling the Stratospheric Ozone Layer 

Helge Mohn, Daniel Kreyling, Ingo Wohltmann, Ralph Lehmann, Peter Maass, and Markus Rex

Common representations of the stratospheric ozone layer in climate modeling are widely considered only in a very simplified way. Neglecting the mutual interactions of ozone with atmospheric temperature and dynamics has the effect of making climate projections less accurate. Although, more elaborate and interactive models of the stratospheric ozone layer are available, they require far too much computation time to be coupled with climate models. Our aim with this project was to break new ground and pursue an interdisciplinary strategy that spans the fields of machine learning, atmospheric physics and climate modelling.

In this work, we present an implicit neural representation of the extrapolar stratospheric ozone chemistry (SWIFT-AI). An implicitly defined hyperspace of the stratospheric ozone chemistry offers a continuous and even differentiable representation that can be parameterized by artificial neural networks. We analysed different parameter-efficient variants of multilayer perceptrons. This was followed by an intensive, as far as possible energy-efficient search for hyperparameters involving Bayesian optimisation and early stopping techniques.

Our data source is the Lagrangian chemistry and transport model ATLAS. Using its full model of stratospheric ozone chemistry, we focused on simulating a wide range of stratospheric variability that will occur in future climate (e.g. temperature and meridional circulation changes). We conducted a simulation for several years and created a data-set with over 200E+6 input and output pairs. Each output is the 24h ozone tendency of a trajectory. We performed a dimensionality reduction of the input parameters by using the concept of chemical families and by performing a sensitivity analysis to choose a set of robust input parameters.

We coupled the resulting machine learning models with the Lagrangian chemistry and transport model ATLAS, substituting the full stratospheric chemistry model. We validated a two-year simulation run by comparing to the differences in accuracy and computation time from both the full stratospheric chemistry model and the previous polynomial approach of extrapolar SWIFT. We found that SWIFT-AI consistently outperforms the previous polynomial approach of SWIFT, both in terms of test data and simulation results. We discovered that the computation time of SWIFT-AI is more than twice as fast as the previous polynomial approach SWIFT and 700 times faster than the full stratospheric chemistry scheme of ATLAS, resulting in minutes instead of weeks of computation time per model year – a speed-up of several orders of magnitude.

To ensure reproducibility and transparency, we developed a machine learning pipeline, published a benchmark dataset and made our repository open to the public.

In summary, we could show that the application of state-of-the-art machine learning methods to the field of atmospheric physics holds great potential. The achieved speed-up of an interactive and very precise ozone layer enables a novel way of representing the ozone layer in climate models. This in turn will increase the quality of climate projections, which are crucial for policy makers and of great importance for our planet.

How to cite: Mohn, H., Kreyling, D., Wohltmann, I., Lehmann, R., Maass, P., and Rex, M.: SWIFT-AI: Significant Speed-up in Modelling the Stratospheric Ozone Layer, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12574,, 2022.

Recently, an increase in forecast skill of the seasonal climate forecast for winter in Europe has been achieved through an ensemble subsampling approach by way of predicting the mean winter North Atlantic Oscillation (NAO) index through linear regression (based on the autumn state of the four predictors sea surface temperature, Arctic sea ice volume, Eurasian snow depth and stratospheric temperature) and the sampling of the ensemble members which are able to reproduce this NAO state. This thesis shows that the statistical prediction of the NAO index can be further improved via nonlinear methods using the same predictor variables as in the linear approach. This likely also leads to an increase in seasonal climate forecast skill. The data used for the calculations stems from the global reanalysis by the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5. The available time span for use in this thesis covered only 40 years from 1980 till 2020, hence it was important to use a method that still yields statistically significant and meaningful results under those circumstances. The nonlinear method chosen was k-nearest neighbor, which is a simple, yet powerful algorithm when there is not a lot of data available. Compared to other methods like neural networks it is easy to interpret. The resulting method has been developed and tested in a double cross-validation setting. While sea ice in the Barents-Kara sea in September-October shows the most predictive capability for the NAO index in the subsequent winter as a single predictor, the highest forecast skill is achieved through a combination of different predictor variables.

How to cite: Hauke, C., Ahrens, B., and Dalelane, C.: Prediction of the North Atlantic Oscillation index for the winter months December-January-February via nonlinear methods, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12628,, 2022.

EGU22-12765 | Presentations | ITS2.6/AS5.1

Supervised machine learning to estimate instabilities in chaotic systems: computation of local Lyapunov exponents 

Daniel Ayers, Jack Lau, Javier Amezcua, Alberto Carrassi, and Varun Ojha

Weather and climate are well known exemplars of chaotic systems exhibiting extreme sensitivity to initial conditions. Initial condition errors are subject to exponential growth on average, but the rate and the characteristic of such growth is highly state dependent. In an ideal setting where the degree of predictability of the system is known in real-time, it may be possible and beneficial to take adaptive measures. For instance a local decrease of predictability may be counteracted by increasing the time- or space-resolution of the model computation or the ensemble size in the context of ensemble-based data assimilation or probabilistic forecasting.

Local Lyapunov exponents (LLEs) describe growth rates along a finite-time section of a system trajectory. This makes the LLEs the ideal quantities to measure the local degree of predictability, yet a main bottleneck for their real-time use in  operational scenarios is the huge computational cost. Calculating LLEs involves computing a long trajectory of the system, propagating perturbations with the tangent linear model, and repeatedly orthogonalising them. We investigate if machine learning (ML) methods can estimate the LLEs based only on information from the system’s solution, thus avoiding the need to evolve perturbations via the tangent linear model. We test the ability of four algorithms (regression tree, multilayer perceptron, convolutional neural network and long short-term memory network) to perform this task in two prototypical low dimensional chaotic dynamical systems. Our results suggest that the accuracy of the ML predictions is highly dependent upon the nature of the distribution of the LLE values in phase space: large prediction errors occur in regions of the attractor where the LLE values are highly non-smooth.  In line with classical dynamical systems studies, the neutral LLE is more difficult to predict. We show that a comparatively simple regression tree can achieve performance that is similar to sophisticated neural networks, and that the success of ML strategies for exploiting the temporal structure of data depends on the system dynamics.

How to cite: Ayers, D., Lau, J., Amezcua, J., Carrassi, A., and Ojha, V.: Supervised machine learning to estimate instabilities in chaotic systems: computation of local Lyapunov exponents, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12765,, 2022.

EGU22-13228 | Presentations | ITS2.6/AS5.1 | Highlight

Developing a data-driven ocean forecast system 

Rachel Furner, Peter Haynes, Dan Jones, Dave Munday, Brooks Paige, and Emily Shuckburgh

The recent boom in machine learning and data science has led to a number of new opportunities in the environmental sciences. In particular, process-based weather and climate models (simulators) represent the best tools we have to predict, understand and potentially mitigate the impacts of climate change and extreme weather. However, these models are incredibly complex and require huge amounts of High Performance Computing resources. Machine learning offers opportunities to greatly improve the computational efficiency of these models by developing data-driven emulators.

Here I discuss recent work to develop a data-driven model of the ocean, an integral part of the weather and climate system. Much recent progress has been made with developing data-driven forecast systems of atmospheric weather, highlighting the promise of these systems. These techniques can also be applied to the ocean, however modelling of the ocean poses some fundamental differences and challenges in comparison to modelling the atmosphere, for example, oceanic flow is bathymetrically constrained across a wide range of spatial and temporal scales.

We train a neural network on the output from an expensive process-based simulator of an idealised channel configuration of oceanic flow. We show the model is able to learn well the complex dynamics of the system, replicating the mean flow and details within the flow over single prediction steps. We also see that when iterating the model, predictions remain stable, and continue to match the ‘truth’ over a short-term forecast period, here around a week.


How to cite: Furner, R., Haynes, P., Jones, D., Munday, D., Paige, B., and Shuckburgh, E.: Developing a data-driven ocean forecast system, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-13228,, 2022.

EGU22-591 | Presentations | ITS2.7/AS5.2

Identifying precursors for extreme stratospheric polar vortex events  using an explainable neural network 

Zheng Wu, Tom Beucler, Raphaël de Fondeville, Eniko Székely, Guillaume Obozinski, William Ball, and Daniela Domeisen

The winter stratospheric polar vortex exhibits considerable variability in both magnitude and zonal wave structure, which arises in part from stratosphere-troposphere coupling associated with tropospheric precursors and can result in extreme polar vortex events. These extremes can subsequently influence weather in the troposphere and thus are important sources of surface prediction. However, the predictability limit of these extreme events is around 1-2 weeks in the state-of-the-art prediction system. In order to explore and improve the predictability limit of the extreme vortex events, in this study, we train an artificial neural network (ANN) to model stratospheric polar vortex anomalies and to identify strong and weak stratospheric vortex events. To pinpoint the origins of the stratospheric anomalies, we then employ two neural network visualization methods, SHapley Additive exPlanations (SHAP) and Layerwise Relevance Propagation (LRP), to uncover feature importance in the input variables (e.g., geopotential height and background zonal wind). The extreme vortex events can be identified by the ANN with an averaged accuracy of 60-80%. For the correctly identified extreme events, the composite of the feature importance of the input variables shows spatial patterns consistent with the precursors found for extreme stratospheric events in previous studies. This consistency provides confidence that the ANN is able to identify reliable indicators for extreme stratospheric vortex events and that it could help to identify the role of the previously found precursors, such as the sea level pressure anomalies associated with the Siberian high. In addition to the composite of all the events, the feature importance for each of the individual events further reveals the physical structures in the input variables (such as the locations of the geopotential height anomalies) that are specific to that event. Our results show the potential of explainable neural networks techniques in understanding and predicting the stratospheric variability and extreme events, and in searching for potential precursors for these events on subseasonal time scales. 

How to cite: Wu, Z., Beucler, T., de Fondeville, R., Székely, E., Obozinski, G., Ball, W., and Domeisen, D.: Identifying precursors for extreme stratospheric polar vortex events  using an explainable neural network, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-591,, 2022.

EGU22-676 | Presentations | ITS2.7/AS5.2

A two-stage machine learning framework using global satellite data of cloud classes for process-oriented model evaluation 

Arndt Kaps, Axel Lauer, Gustau Camps-Valls, Pierre Gentine, Luis Gómez-Chova, and Veronika Eyring

Clouds play a key role in weather and climate but are quite challenging to simulate with global climate models as the relevant physics include non-linear processes on scales covering several orders of magnitude in both the temporal and spatial dimensions. The numerical representation of clouds in global climate models therefore requires a high degree of parameterization, which makes a careful evaluation a prerequisite not only for assessing the skill in reproducing observed climate but also for building confidence in projections of future climate change. Current methods to achieve this usually involve the comparison of multiple large-scale physical properties in the model output to observational data. Here, we introduce a two-stage data-driven machine learning framework for process-oriented evaluation of clouds in climate models based directly on widely known cloud types. The first step relies on CloudSat satellite data to assign cloud labels in line with cloud types defined by the World Meteorological Organization (WMO) to MODIS pixels using deep neural networks. Since the method is supervised and trained on labels provided by CloudSat, the predicted cloud types remain objective and do not require a posteriori labeling. The second step consists of a regression algorithm that predicts fractional cloud types from retrieved cloud physical variables. This step aims to ensure that the method can be used with any data set providing physical variables comparable to MODIS. In particular, we use a Random Forest regression that acts as a transfer model to evaluate the spatially relatively coarse output of climate models and allows the use of varying input features. As a proof of concept, the method is applied to coarse grained ESA Cloud CCI data. The predicted cloud type distributions are physically consistent and show the expected features of the different cloud types. This demonstrates how advanced observational products can be used with this method to obtain cloud type distributions from coarse data, allowing for a process-based evaluation of clouds in climate models.

How to cite: Kaps, A., Lauer, A., Camps-Valls, G., Gentine, P., Gómez-Chova, L., and Eyring, V.: A two-stage machine learning framework using global satellite data of cloud classes for process-oriented model evaluation, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-676,, 2022.

EGU22-696 | Presentations | ITS2.7/AS5.2 | Highlight

Latent Linear Adjustment Autoencoder: a novel method for estimating dynamic precipitation at high resolution 

Christina Heinze-Deml, Sebastian Sippel, Angeline G. Pendergrass, Flavio Lehner, and Nicolai Meinshausen

A key challenge in climate science is to quantify the forced response in impact-relevant variables such as precipitation against the background of internal variability, both in models and observations. Dynamical adjustment techniques aim to remove unforced variability from a target variable by identifying patterns associated with circulation, thus effectively acting as a filter for dynamically induced variability. The forced contributions are interpreted as the variation that is unexplained by circulation. However, dynamical adjustment of precipitation at local scales remains challenging because of large natural variability and the complex, nonlinear relationship between precipitation and circulation particularly in heterogeneous terrain. 

In this talk, I will present the Latent Linear Adjustment Autoencoder (LLAAE), a novel statistical model that builds on variational autoencoders. The Latent Linear Adjustment Autoencoder enables estimation of the contribution of a coarse-scale atmospheric circulation proxy to daily precipitation at high resolution and in a spatially coherent manner. To predict circulation-induced precipitation, the LLAAE combines a linear component, which models the relationship between circulation and the latent space of an autoencoder, with the autoencoder's nonlinear decoder. The combination is achieved by imposing an additional penalty in the cost function that encourages linearity between the circulation field and the autoencoder's latent space, hence leveraging robustness advantages of linear models as well as the flexibility of deep neural networks. 

We show that our model predicts realistic daily winter precipitation fields at high resolution based on a 50-member ensemble of the Canadian Regional Climate Model at 12 km resolution over Europe, capturing, for instance, key orographic features and geographical gradients. Using the Latent Linear Adjustment Autoencoder to remove the dynamic component of precipitation variability, forced thermodynamic components are expected to remain in the residual, which enables the uncovering of forced precipitation patterns of change from just a few ensemble members. We extend this to quantify the forced pattern of change conditional on specific circulation regimes. 

Future applications could include, for instance, weather generators emulating climate model simulations of regional precipitation, detection and attribution at subcontinental scales, or statistical downscaling and transfer learning between models and observations to exploit the typically much larger sample size in models compared to observations.

How to cite: Heinze-Deml, C., Sippel, S., Pendergrass, A. G., Lehner, F., and Meinshausen, N.: Latent Linear Adjustment Autoencoder: a novel method for estimating dynamic precipitation at high resolution, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-696,, 2022.

EGU22-722 | Presentations | ITS2.7/AS5.2 | Highlight

Climate-Invariant, Causally Consistent Neural Networks as Robust Emulators of Subgrid Processes across Climates 

Tom Beucler, Fernando Iglesias-Suarez, Veronika Eyring, Michael Pritchard, Jakob Runge, and Pierre Gentine

Data-driven algorithms, in particular neural networks, can emulate the effects of unresolved processes in coarse-resolution Earth system models (ESMs) if trained on high-resolution simulation or observational data. However, they can (1) make large generalization errors when evaluated in conditions they were not trained on; and (2) trigger instabilities when coupled back to ESMs.

First, we propose to physically rescale the inputs and outputs of neural networks to help them generalize to unseen climates. Applied to the offline parameterization of subgrid-scale thermodynamics (convection and radiation) in three distinct climate models, we show that rescaled or "climate-invariant" neural networks make accurate predictions in test climates that are 8K warmer than their training climates. Second, we propose to eliminate spurious causal relations between inputs and outputs by using a recently developed causal discovery framework (PCMCI). For each output, we run PCMCI on the inputs time series to identify the reduced set of inputs that have the strongest causal relationship with the output. Preliminary results show that we can reach similar levels of accuracy by training one neural network per output with the reduced set of inputs; stability implications when coupled back to the ESM are explored.

Overall, our results suggest that explicitly incorporating physical knowledge into data-driven models of Earth system processes may improve their ability to generalize across climate regimes, while quantifying causal associations to select the optimal set of inputs may improve their consistency and stability.

How to cite: Beucler, T., Iglesias-Suarez, F., Eyring, V., Pritchard, M., Runge, J., and Gentine, P.: Climate-Invariant, Causally Consistent Neural Networks as Robust Emulators of Subgrid Processes across Climates, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-722,, 2022.

EGU22-1065 | Presentations | ITS2.7/AS5.2 | Highlight

Skilful US Soy-yield forecasts at pre-sowing lead-times 

Sem Vijverberg, Dim Coumou, and Raed Hamed

Soy harvest failure events can severely impact farmers, insurance companies and raise global prices. Reliable seasonal forecasts of mis-harvests would allow stakeholders to prepare and take appropriate early action. However, especially for farmers, the reliability and lead-time of current prediction systems provide insufficient information to justify for within-season adaptation measures. Recent innovations increased our ability to generate reliable statistical seasonal forecasts. Here, we combine these innovations to predict the 1-3 poor soy harvest years in eastern US. We first use a clustering algorithm to spatially aggregate crop producing regions within the eastern US that are particularly sensitive to hot-dry weather conditions. Next, we use observational climate variables (sea surface temperature (SST) and soil moisture) to extract precursor timeseries at multiple lags. This allows the machine learning model to learn the low-frequency evolution, which carries important information for predictability. A selection based on causal inference allows for physically interpretable precursors. We show that the robust selected predictors are associated with the evolution of the horseshoe Pacific SST pattern, in line with previous research. We use the state of the horseshoe Pacific to identify years with enhanced predictability. We achieve very high forecast skill of poor harvests events, even 3 months prior to sowing, using a strict one-step-ahead train-test splitting. Over the last 25 years, 90% of the predicted events in February were correct. When operational, this forecast would enable farmers (and insurance/trading companies) to make informed decisions on adaption measures, e.g., selecting more drought-resistant cultivars, invest in insurance, change planting management.

How to cite: Vijverberg, S., Coumou, D., and Hamed, R.: Skilful US Soy-yield forecasts at pre-sowing lead-times, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1065,, 2022.

EGU22-1835 | Presentations | ITS2.7/AS5.2

Using Deep Learning for a High-Precision Analysis of Atmospheric Rivers in a High-Resolution Large Ensemble Climate Dataset 

Timothy Higgins, Aneesh Subramanian, Andre Graubner, Lukas Kapp-Schwoerer, Karthik Kashinath, Sol Kim, Peter Watson, Will Chapman, and Luca Delle Monache

Atmospheric rivers (ARs) are elongated corridors of water vapor in the lower Troposphere that cause extreme precipitation over many coastal regions around the globe. They play a vital role in the water cycle in the western US, fueling most extreme west coast precipitation and sometimes accounting for more than 50% of total annual west coast precipitation (Gershunov et al. 2017). Severe ARs are associated with extreme flooding and damages while weak ARs are typically more beneficial to our society as they bring much needed drought relief.

Precipitation is particularly difficult to predict in traditional climate models.  Predicting water vapor is more reliable (Lavers et al. 2016), allowing IVT (integrated vapor transport) and ARs to be a favorable method for understanding changing patterns in precipitation (Johnson et al. 2009).  There are a variety of different algorithms used to track ARs due to their relatively diverse definitions (Shields et al. 2018). The Atmospheric River Tracking Intercomparison Project (ARTMIP) organizes and provides information on all of the widely accepted algorithms that exist. Nearly all of the algorithms included in ARTMIP rely on absolute and relative numerical thresholds, which can often be computationally expensive and have a large memory footprint. This can be particularly problematic in large climate datasets. The vast majority of algorithms also heavily factor in wind velocity at multiple vertical levels to track ARs, which is especially difficult to store in climate models and is typically not output at the temporal resolution that ARs occur.

A recent alternative way of tracking ARs is through the use of machine learning. There are a variety of neural networks that are commonly applied towards identifying objects in cityscapes via semantic segmentation. The first of these neural networks that was applied towards detecting ARs is DeepLabv3+ (Prabhat et al. 2020). DeepLabv3+ is a state of the art model that demonstrates one of the highest performances of any present day neural network when tasked with the objective of identifying objects in cityscapes (Wu et al. 2019). We employ a light-weight convolutional neural network adapted from CGNet (Kapp-Schwoerer et al. 2020) to efficiently track these severe events without using wind velocity at all vertical levels as a predictor variable. When applied to cityscapes, CGNet's greatest advantage is its performance relative to its memory footprint (Wu et al. 2019). It has two orders of magnitude less parameters than DeepLabv3+ and is computationally less expensive. This can be especially useful when identifying ARs in large datasets. Convolutional neural networks have not been used to track ARs in a regional domain. This will also be the first study to demonstrate the performance of this neural network on a regional domain by providing an objective analysis of its consistency with eight different ARTMIP algorithms.

How to cite: Higgins, T., Subramanian, A., Graubner, A., Kapp-Schwoerer, L., Kashinath, K., Kim, S., Watson, P., Chapman, W., and Delle Monache, L.: Using Deep Learning for a High-Precision Analysis of Atmospheric Rivers in a High-Resolution Large Ensemble Climate Dataset, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1835,, 2022.

EGU22-2012 | Presentations | ITS2.7/AS5.2

Gap filling in air temperature series by matrix completion methods 

Benoît Loucheur, Pierre-Antoine Absil, and Michel Journée

Quality control of meteorological data is an important part of atmospheric analysis and prediction, as missing or erroneous data can have a negative impact on the accuracy of these environmental products.

In Belgium, the Royal Meteorological Institute (RMI) is the national meteorological service that provide weather and climate services based on observations and scientific research. RMI collects and archives meteorological observations in Belgium since the 19th century. Currently, air temperature is monitored in Belgium in about 30 synoptic automatic weather stations (AWS) as well as in 110 manual climatological stations. In the latter stations, a volunteer observer records every morning at 8 o'clock the daily extreme air temperatures. All observations are routinely checked for errors, inconsistencies and missing values by the RMI staff. Misleading data are corrected and gaps are filled by estimations. This quality control tasks require a lot of human intervention. With the forthcoming deployment of low-cost weather stations and the subsequent increase in the volume of data to verify, the process of data quality control and completion should become as automated as much as possible.

In this work, the quality control process is fully automated by using mathematical tools. We present low-rank matrix completion methods (LRMC) that we used to solve the problem of completing missing data in daily minimum and maximum temperature series. We used a machine learning technique called Monte Carlo cross-validation to train our algorithms and then test them in a real case.

Among the matrix completion methods, some are regularised by graphs. In our case, it is then possible to represent the spatial and temporal component via graphs. By manipulating the construction of these graphs, we hope to improve the completion results. We were then able to compare our methods with what is done in the state of the art, such as the inverse distance weighting (IDW) method.

All our experiments were performed with a dataset provided by the RMI, including daily minimum and maximum temperature measurements from 100 stations over the period 2005-2019.

How to cite: Loucheur, B., Absil, P.-A., and Journée, M.: Gap filling in air temperature series by matrix completion methods, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2012,, 2022.

EGU22-2248 | Presentations | ITS2.7/AS5.2

Exploring flooding mechanisms and their trends in Europe through explainable AI 

Shijie Jiang, Yi Zheng, and Jakob Zscheischler

Understanding the mechanisms causing river flooding and their trends is important to interpret past flood changes and make better predictions of future flood conditions. However,  there is still a lack of quantitative assessment of trends in flooding mechanisms based on observations. Recent years have witnessed the increasing prevalence of machine learning in hydrological modeling and its predictive power has been demonstrated in numerous studies. Machine learning makes hydrological predictions by recognizing generalizable relationships between inputs and outputs, which, if properly interpreted, may provide us further scientific insights into hydrological processes. In this study, we propose a new method using interpretive machine learning to identify flooding mechanisms based on the predictive relationship between precipitation and temperature and flow peaks. Applying this method to more than a thousand catchments in Europe reveals three primary input-output patterns within flow predictions, which can be associated with three catchment-wide flooding mechanisms: extreme precipitation, soil moisture excess, and snowmelt. The results indicate that approximately one-third of the studied catchments are controlled by a combination of the above mechanisms, while others are mostly dominated by one single mechanism. Although no significant shifts from one dominant mechanism to another are observed for the catchments over the past seven decades overall, some catchments with single mechanisms have become dominated by mixed mechanisms and vice versa. In particular, snowmelt-induced floods have decreased significantly in general, whereas rainfall has become more dominant in causing floods, and their effects on flooding seasonality and magnitude are crucial. ​Overall, this study provides a new perspective for understanding climatic extremes and demonstrates the prospect of artificial intelligence(AI)-assisted scientific discovery in the future.

How to cite: Jiang, S., Zheng, Y., and Zscheischler, J.: Exploring flooding mechanisms and their trends in Europe through explainable AI, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2248,, 2022.

EGU22-2391 | Presentations | ITS2.7/AS5.2

Exploring cirrus cloud microphysical properties using explainable machine learning 

Kai Jeggle, David Neubauer, Gustau Camps-Valls, Hanin Binder, Michael Sprenger, and Ulrike Lohmann

Cirrus cloud microphysics and their interactions with aerosols remain one of the largest uncertainties in global climate models and climate change projections. The uncertainty originates from the high spatio-temporal variability and their non-linear dependence on meteorological drivers like temperature, updraft velocities, and aerosol environment. We combine ten years of CALIPSO/CloudSat satellite observations of cirrus clouds with ERA5 and MERRA-2 reanalysis data of meteorological and aerosol variables to create a spatial data cube. Lagrangian back trajectories are calculated for each cirrus cloud observation to add a temporal dimension to the data cube. We then train a gradient boosted tree machine learning (ML) model to predict vertically resolved cirrus cloud microphysical properties (i.e. observed ice crystal number concentration and ice water content). The explainable machine learning method of SHAP values is applied to assess the impact of individual cirrus drivers as well as combinations of drivers on cirrus cloud microphysical properties in varying meteorological conditions. In addition, we analyze how the impact of the drivers differs regionally, vertically, and temporally.

We find that the tree-based ML model is able to create a good mapping between cirrus drivers and microphysical properties (R² ~0.75) and the SHAP value analysis provides detailed insights in how different drivers impact the prediction of the microphysical cirrus cloud properties. These findings can be used to improve global climate model parameterizations of cirrus cloud formation in future works. Our approach is a good example for exploring unsolved scientific questions using explainable machine learning and feeding back insights to the domain science.

How to cite: Jeggle, K., Neubauer, D., Camps-Valls, G., Binder, H., Sprenger, M., and Lohmann, U.: Exploring cirrus cloud microphysical properties using explainable machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2391,, 2022.

Global circulation models (GCMs) form the basis of a vast portion of earth system research and inform our climate policy. However, our climate system is complex and connected across scales. To simulate it, we must use parameterisations. These parameterisations, which are present in all models, can have a detectable influence on the GCM outputs.

GCMs are improving, but we need to use their current output to optimally estimate the risks of extreme weather. Therefore, we must debias GCM outputs with respect to observations. Current debiasing methods cannot correct both spatial correlations and cross-variable correlations. This limitation means current methods can produce physically implausible weather events - even when the single-location, single-variable distributions match the observations. This limitation is very important for extreme event research. Compound events like heat and drought, which drastically increase wildfire risk, and spatially co-occurring events like multiple bread-basket failures, are not well corrected by these current methods.

We propose using unsupervised image-to-image translations networks to perform bias correction of GCMs. These neural network architectures are used to translate (perform bias correction) between different image domains. For example, they have been used to translate computer-generated city scenes into real-world photos, which requires spatial and cross-variable correlations to be translated. Crucially, these networks learn to translate between image domains without requiring corresponding pairs of images. Such pairs cannot be generated between climate simulations and observations due to the inherent chaos of weather.

In this work, we use these networks to bias correct historical recreation simulations from the HadGEM3-A-N216 atmosphere-only GCM with respect to the ERA5 reanalysis dataset. This GCM has a known bias in simulating the South Asian monsoon, and so we focus on this region. We show the ability of neural networks to correct this bias, and show how combining the neural network with classical techniques produces a better bias correction than either method alone. 

How to cite: Fulton, J. and Clarke, B.: Correcting biases in climate simulations using unsupervised image-to-image-translation networks, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2988,, 2022.

EGU22-3009 | Presentations | ITS2.7/AS5.2

Application of Machine Learning for spatio-temporal mapping of the air temperature in Warsaw 

Amirhossein Hassani, Núria Castell, and Philipp Schneider

Mapping the spatio-temporal distribution of near-surface urban air temperature is crucial to our understanding of climate-sensitive epidemiology, indoor-outdoor thermal comfort, urban biodiversity, and interactive impacts of climate change and urbanity. Urban-scale decision-making in face of future climatic uncertainties requires detailed information on near-surface air temperature at high spatio-temporal resolutions. However, reaching such fine resolutions cannot be currently realised by traditional observation networks, or even by regional or global climate models (Hamdi et al. 2020). Given the complexity of the processes affecting air temperature at the urban scale to the regional scale, here we apply Machine Learning (ML) algorithms, in particular, XGBoost gradient boosting method to build predictive models of near surface air temperature (Ta at 2-meter height). These predictive models establish data-driven relations between crowd-sourced measured Ta (data produced by citizens’ sensors) and a set of spatial and spatio-temporal predictors, primarily derived from Earth Observation satellite data including Modis Aqua/Landsat 8 Land Surface Temperature (LST), Modis Terra vegetative indices, and Sentinel-2 water vapour product. We use our models to predict sub-daily (at Modis Aqua satellite passing times) variation in urban scale Ta in city of Warsaw, Poland at spatial resolution of 1 km for the months July-September and the years 2016 to 2021. A 10-fold cross-validation of the developed models shows a root mean square error between 0.97 and 1.02 °C and a coefficient of determination between 0.96 and 0.98, which are satisfactory according to the literature (Taheri-Shahraiyni and Sodoudi 2017). The resulting maps allow us to identify regions of Warsaw that are vulnerable to heat stress. The strength of the method used here is that it can be easily replicated in other EU cities to achieve high resolution maps due to the accessibility and open-sourced nature of the training and predictor data. Contingent on data availability, the predictive framework developed also can be used for monitoring and downscaling of other urban governing climatic parameters such as relative humidity in the context of future climate uncertainties.

Hamdi, R., H. Kusaka, Q.-V. Doan, P. Cai, H. He, G. Luo, W. Kuang, S. Caluwaerts, F. Duchêne, B. J. E. S. Van Schaeybroek and Environment (2020). "The state-of-the-art of urban climate change modeling and observations." 1-16.

Taheri-Shahraiyni, H. and S. J. T. S. Sodoudi (2017). "High-resolution air temperature mapping in urban areas: A review on different modelling techniques."  21(6 Part A): 2267-2286.

How to cite: Hassani, A., Castell, N., and Schneider, P.: Application of Machine Learning for spatio-temporal mapping of the air temperature in Warsaw, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3009,, 2022.

The interdisciplinary research project "BayTreeNet" investigates the reactions of forest ecosystems to current climate dynamics. In the mid-latitudes, local climatic phenomena often show a strong dependence on the large-scale climate dynamics, the weather types (WT), which significantly determine the climate of a region through frequency and intensity. In the topographically diverse region of Bavaria, different WT show various weather conditions at different locations.

The meaning of every WT is explained for the different forest regions in Bavaria and the results of the climate dynamics sub-project provide the physical basis for the "BayTreeNet" project. Subsequently, climate-growth relationships are established in the dendroecology sub-project to investigate the response of forests to individual WT at different forest sites. Complementary steps allow interpretation of results for the past (20th century) and projection into the future (21st century). One hypothesis to be investigated is that forest sites in Bavaria are affected by a significant influence of climate change in the 21st century and the associated change in WT.

The automated classification of large-scale weather patterns is presented by Self-Organizing-Maps (SOM) developed by Kohonen, which enables visualization and reduction of high-dimensional data. The poster presents the evaluation and selection of an appropriate SOM-setting and its first results. Besides, it is planned to show first analyses of the environmental conditions of the different WT and how these are represented in global climate models (GCMs) in the past and future.

How to cite: Wehrmann, S. and Mölg, T.: Classifying weather types in Europe by Self-Organizing-Maps (SOM) with regard to GCM-based future projections, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3105,, 2022.

EGU22-3482 | Presentations | ITS2.7/AS5.2

Public perception assessment on climate change and natural disaster influence using social media big-data: A case study of USA 

SungKu Heo, Pouya Ifaei, Mohammad Moosazadeh, and ChangKyoo Yoo

Climate change is a global crisis to the world which influences the human race and society's development. Threatens of climate change have become increasingly recognized to the public and government in both environments, society, and economy across the globe; because the consequence of climate change is not only shown up as the increasing of global temperature, also expressed in an intensive natural hazard, such as floods, droughts, wildfires, and hurricanes. For the sustainability development in the globe, it is crucial to provide a response to mitigating climate change through the government’s policy and decision-making; however, the public's engagement in the actions towards the critical environmental crisis still needs to be largely promoted. Analyzing the relationship between the public awareness of climate change and natural disasters is an essential aspect in climate change mitigation and policymaking. In this study, based on the abundance of the text message in social media, especially Twitter, the public understanding and discussions upon climate change from the surrounding environment was recognized and analyzed through the human as the sensor which receiving information of climate change. Twitter content analysis and filed data impact analysis were conducted; text mining algorithms are implemented in the Twitter big-data information to find the similarity based on a cosine similarity score (CSS) between the climate change corpus and the natural events corpora. Then, the factors of nature disaster influence were predicted utilizing a multiple linear regression model and climate change tweets dataset. This research shows that the public is more pretend to link the natural events with climate change when they tweeting when serious natural disasters happened. The developed regression model indicated that natural events caused by the consequence of climate change influenced the people’s social media activity through messages on Twitter with having the awareness of climate change. From this study, the results indicated that the public experience of natural events including intensive disasters can lead them to link the climate change with the natural events easily; when compared with the people who rarely experience natural events.


This research was supported by the project (NRF-2021R1A2C2007838) through the National Research Foundation of Korea (NRF) and the Korea Ministry of Environment (MOE) as Graduate school specialized in Climate Change.

How to cite: Heo, S., Ifaei, P., Moosazadeh, M., and Yoo, C.: Public perception assessment on climate change and natural disaster influence using social media big-data: A case study of USA, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3482,, 2022.

EGU22-4431 | Presentations | ITS2.7/AS5.2

Identification of Global Drivers of Indian Summer Monsoon using Causal Inference and Interpretable AI 

Deepayan Chakraborty, Adway Mitra, Bhupendranath Goswami, and Pv Rajesh

Indian Summer Monsoon Rainfall (ISMR) is a complex phenomenon that depends on several climatic phenomena at different parts of the word through teleconnections. Each season is characterized by extended periods of wet and dry spells (which may cause floods or droughts) which contribute to intra-seasonal variability. Tropical and extra-tropical drivers jointly influence the intra-seasonal variability. Although El Nino and Southern Oscillation (ENSO) is known to be a driver of ISMR, researchers have also found its relation with Indian Ocean Dipole (IOD), North Atlantic Oscillations (NAO), Atlantic Multi-decadal Oscillation (AMO). In this work, we use ideas from Causality Theory and Explainable Machine Learning to quantify the influence of different climatic phenomena on the intraseasonal variation of ISMR.

To identify such causal relations, we applied two statistically sound causal inference approaches, i.e., PCMCI+ Algorithm (Conditional Independence based) and Granger Causal test (Regression-based).  For the Granger causality test, we have examined separately for both linear and non-linear regression. In case of PCMCI+, conditional independence tests were used between pairs of variables at different "lag periods". It is worth pointing out that, till now “causality” is not properly quantified in the Climate Science community and only linear correlations are used as a basis to identify relationships like ENSO-ISMR and AMO-ISMR. We performed experiments on mean monthly rainfall anomaly data (during the monsoon months of June-September over India) along with six probable drivers (ENSO, AMO, North Atlantic Oscillation, Pacific Decadal Oscillation, Atlantic Nino, and Indian Ocean Dipole) for May, June, July, August, September months during the period 1861-2016. While the two approaches produced some contradictions, they also produced a common conclusion that ENSO and AMO are equally important and independent drivers of ISMR. 

Additionally, we have studied the contribution of the drivers on annual extremes of ISMR (years of deficient and excess rainfall) using Shapley values based on the concept of Game Theory to quantify the contributions of different predictors in a model. In this work, we train a XGBoost model to predict the ISMR anomaly from any values of the predictor variables. The experiment is carried out in two approaches. One approach involves analyzing the contribution of each driver for each of the ISMR months of any year on the mean seasonal rainfall anomaly of that year. Another approach focuses on the contribution of the seasonal mean value of each driver on the same. In both approaches, we analyze the distribution of each driver’s Shapley values for excess and deficient monsoon years for contrast. We find that while ENSO is indeed the dominant driving factor for a majority of these years, AMO is another major factor which frequently contributes to such deficiencies, while Atlantic Nino and Indian Ocean Dipole too sometimes contribute. On the other hand, Indian Ocean Dipole seems to be a major contributor for several years of excess rainfall. As future work, we plan to carry out a robustness analysis of these results, and also examine the drivers of regional extremes.

How to cite: Chakraborty, D., Mitra, A., Goswami, B., and Rajesh, P.: Identification of Global Drivers of Indian Summer Monsoon using Causal Inference and Interpretable AI, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4431,, 2022.

EGU22-4534 | Presentations | ITS2.7/AS5.2

Spatial multi-modality as a way to improve both performance and interpretability of deep learning models to reconstruct phytoplankton time-series in the global ocean 

Joana Roussillon, Jean Littaye, Ronan Fablet, Lucas Drumetz, Thomas Gorgues, and Elodie Martinez

Phytoplankton plays a key role in the carbon cycle and fuels marine food webs. Its seasonal and interannual variations are relatively well-known at global scale thanks to satellite ocean color observations that have been continuously acquired since 1997. However, the satellite-derived chlorophyll-a concentrations (Chl-a, a proxy of phytoplankton biomass) time series are still too short to investigate phytoplankton biomass low-frequency variability. Machine learning models such as support vector regression (SVR) or multi-layer perceptron (MLP) have recently proven to be an alternative approach to mechanistic ones to reconstruct Chl-a past signals (including periods before the satellite era) from physical predictors, but they remain unsatisfactory. In particular, the relationships between phytoplankton and its physical surrounding environment are not homogeneous in space, and training such models over the entire globe does not allow them to capture these regional specificities. Moreover, if the global ocean is commonly partitioned into biogeochemical provinces into which phytoplankton growth is supposed to be governed by similar processes, their time-evolving nature makes it difficult to impose a priori spatial constraints to restrict the learning phase on specific areas. Here, we propose to overcome this limitation by introducing spatial multi-modalities into a convolutional neural network (CNN). The latter can learn with no particular supervision several spatially weighted modes of variability. Each of them is associated with a CNN submodel trained in parallel, standing for a mode-specific response of phytoplankton biomass to the physical forcing. Beyond improving performance reconstruction, we will show that the learned spatial modes appear physically consistent and may help to get new insights into physical-biogeochemical processes controlling phytoplankton repartition at global scale.

How to cite: Roussillon, J., Littaye, J., Fablet, R., Drumetz, L., Gorgues, T., and Martinez, E.: Spatial multi-modality as a way to improve both performance and interpretability of deep learning models to reconstruct phytoplankton time-series in the global ocean, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4534,, 2022.

EGU22-4584 | Presentations | ITS2.7/AS5.2

Super-Resolution based Deep Downscaling of Precipitation 

Sumanta Chandra Mishra Sharma and Adway Mitra

Downscaling is widely used to improve spatial resolution of meteorological variables. Broadly there are two classes of techniques used for downscaling i.e. dynamical downscaling and statistical downscaling. Dynamical downscaling depends on the boundary conditions of coarse resolution global models like General Circulation Models (GCMs) for its operation whereas the statistical model tries to interpret the statistical relationship between the high-resolution and low-resolution data (Kumar et. al. 2021). With the rapid development of deep learning techniques in recent years, deep learning based super-resolution (SR) models have been designed for image processing and computer vision, for increasing the resolution of a given image. But many researchers from other fields have also adapted these techniques and achieved state-of-the-art performance in various domains. To the best of our knowledge, only a few works exist that have used the super-resolution methods in climate domain, for deep downscaling of precipitation data.

These super-resolution approaches mostly use convolutional neural networks (CNN) to accomplish their task. In CNN when we increase the depth of the model then there is a chance of information loss and error propagation (Vandal To reduce this information loss, we have introduced residual-based deep downscaling models. These models have multiple residual blocks and skip connections between similar types of convolutional layers. The long skip connections in the model helps to reduce information loss in the network. These models take as input, data that is pre-upsampled by linear interpolation, and then improve the estimates of the pixel values.

In our experiments, we have focused on downscaling of rainfall over Indian landmass (for Indian summer monsoon rainfall) and for a region in the USA spanning the southeast CONUS and parts of its neighboring states that are present between the longitude 700 W to 1000 W and latitude 240 N to 400 N. The precipitation data for this task is collected from the India Meteorological Department (IMD), Pune, India, and NOAA Physical Science Laboratory. We have examined our model's predictive behavior and compared it with the existing super-resolution models like SRCNN and DeepSD, which have been earlier used for precipitation downscaling. In the DeepSD model, we have used the GTOPO30 land elevation data provided by USGS along with the precipitation data as input. All these models are trained and tested in both the geographical regions separately and it is found that the proposed model performs better than the existing models on multiple accuracy measures like PSNR, Correlation Coefficient, etc. for the specific region and scaling factor.

How to cite: Mishra Sharma, S. C. and Mitra, A.: Super-Resolution based Deep Downscaling of Precipitation, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4584,, 2022.

EGU22-4853 | Presentations | ITS2.7/AS5.2

Can cloud properties provide information on surface wind variations using deep learning? 

Sebastiaan Jamaer, Jérôme Neirynck, and Nicole van Lipzig

Recent studies have shown that the increasing sizes of offshore wind farms can cause a reduced energy production through mesoscale interactions with the atmosphere. Therefore, accurate nowcasting of the energy yields of large offshore wind farms depend on accurate predictions of the large synoptic weather systems as well as accurate predictions of the smaller mesoscale weather systems. In general, global or regional forecasting models are very well suited to predict synoptic-scale weather systems. However, satellite or radar data can support the nowcasting of shorter, smaller-scale systems. 

In this work, a first step towards nowcasting of the mesoscale wind using satellite images has been taken, namely the coupling of the mesoscale wind component to cloud properties that are available from satellite images using a deep learning framework. To achieve this, a high-resolution regional atmospheric model (COSMO-CLM) was used to generate one year of high resolution cloud en hub-height wind data. From this wind data the mesoscale component was filtered out and used as target images for the deep learning model. The input images of the model were several cloud-related fields from the atmospheric model. The model itself was a Deep Convolutional Neural Network (a U-Net) which was trained to minimize the mean squared error. 

This analysis indicates that cloud information can be used to extract information about the mesoscale weather systems and could be used for nowcasting by using the trained U-Net as a basis for a temporal deep learning model. However, future validation with real-world data is still needed to determine the added value of such an approach.

How to cite: Jamaer, S., Neirynck, J., and van Lipzig, N.: Can cloud properties provide information on surface wind variations using deep learning?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4853,, 2022.

EGU22-5058 | Presentations | ITS2.7/AS5.2

Can satellite images provide supervision for cloud systems characterization? 

Dwaipayan Chatterjee, Hartwig Deneke, and Susanne Crewell

With ever-increasing resolution, geostationary satellites are able to reveal the complex structure and organization of clouds. How cloud systems organize is important for the local climate and strongly connects to the Earth's response to warming through cloud system feedback.

Motivated by recent developments in computer vision for pattern analysis of uncurated images, our work aims to understand the organization of cloud systems based on high-resolution cloud optical depth images. We are exploiting the self-learning capability of a deep neural network to classify satellite images into different subgroups based on the distribution pattern of the cloud systems.

Unlike most studies, our neural network is trained over the central European domain, which is characterized by strong land surface type and topography variations. The satellite data is post-processed and retrieved at a higher spatio-temporal resolution (2 km, 5 min), enhanced by 66% compared to the current standard, equivalent to the future Meteosat third-generation satellite, which will be launched soon.

We show how recent advances in deep learning networks are used to understand clouds' physical properties in temporal and spatial scales. In a purely data-driven approach, we avoid the noise and bias obtained from human labeling, and with proper scalable techniques, it takes 0.86 ms and 2.13 ms to label an image at two different spatial configurations. We demonstrate explainable artificial intelligence (XAI), which helps gain trust for the neural network's performance.

To generalize the results, a thorough quantified evaluation is done on two spatial domains and two-pixel configurations (128x128, 64x64). We examine the uncertainty associated with distinct machine-detected cloud-pattern categories. For this, the learned features of the satellite images are extracted from the trained neural network and fed to an independent hierarchical - agglomerative algorithm. Therefore the work also explores the uncertainties associated with the automatic machine-detected patterns and how they vary with different cloud classification types.

How to cite: Chatterjee, D., Deneke, H., and Crewell, S.: Can satellite images provide supervision for cloud systems characterization?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5058,, 2022.

Extreme weather events, such as droughts, floods or heatwaves, severely impact agricultural yield. However, crop yield failure may also be caused by the temporal or multivariate compounding of more moderate weather events. An example of such an occurrence is the phenomenon of 'false spring', where the combined effects of a warm interval in late winter followed by a period of freezing temperatures can result in severe damage to vegetation. Alternatively, multiple weather events may impact crops simultaneously, as with compound hot and dry weather conditions.

Machine learning techniques are able to learn highly complex and nonlinear relationships between predictors. Such methods have previously been used to explore the influence of monthly- or seasonally-aggregated weather data as well as predefined extreme event indicators on crop yield. However, as crop yield may be impacted by climatic variables at different temporal scales, interpretable machine learning methods that can extract relevant meteorological features from higher-resolution time series data are desirable.

In this study we test the ability of adaptations of random forest models to identify compound meteorological drivers of crop failure from simulated data. In particular, adaptations of random forest models capable of ingesting daily multivariate time series data and spatial information are used. First, we train models to extract useful features from daily climatic data and predict crop yield failure probabilities. Second, we use permutation feature importances and sequential feature selection to investigate weather events and time periods identified by the models as most relevant for crop yield failure prediction. Finally, we explore the interactions learned by the models between these selected meteorological drivers, and compare the outcomes for several global crop models. Ultimately, our goal is to present a robust and highly interpretable machine learning method that can identify critical weather conditions from datasets with high temporal and spatial resolution, and is therefore able to identify drivers of crop failure using relatively few years of data.

How to cite: Sweet, L. and Zscheischler, J.: Using interpretable machine learning to identify compound meteorological drivers of crop yield failure, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5464,, 2022.

EGU22-5756 | Presentations | ITS2.7/AS5.2

The influence of meteorological parameters on wind speed extreme events:  A causal inference approach 

Katerina Hlavackova-Schindler (Schindlerova), Andreas Fuchs, Claudia Plant, Irene Schicker, and Rosmarie DeWit

Based on the ERA5  data of hourly  meteorological parameters [1], we investigate temporal effects of  12 meteorological parameters on  the extreme values occurring in  wind speed.  We approach the problem by using the Granger causal inference, namely by the heterogeneous graphical Granger model (HGGM) [2]. In contrary to the classical Granger model proposed for causal inference among Gaussian processes, the HGGM detects causal relations among time series with distributions from the exponential family, which includes a wider class of common distributions. In previous synthetic experiments, HGGM combined with the genetic algorithm search based on the minimum message length principle has been shown superior in precision over the baseline causal methods [2].  We investigate various experimental settings of all 12 parameters with respect to the wind extremes in various time intervals. Moreover, we compare the influence of various data preprocessing methods and evaluate the interpretability of the discovered causal connections based on meteorological knowledge.


[2] Behzadi, S, Hlaváčková-Schindler, K., Plant, C. (2019) Granger causality for heterogeneous processes, In: Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer, pp. 463-475.

[3] Hlaváčková-Schindler, K., Plant, C. (2020) Heterogeneous graphical Granger causality by minimum message length, Entropy, 22(1400). pp. 1-21 ISSN 1099-4300 MDPI (2020).

How to cite: Hlavackova-Schindler (Schindlerova), K., Fuchs, A., Plant, C., Schicker, I., and DeWit, R.: The influence of meteorological parameters on wind speed extreme events:  A causal inference approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5756,, 2022.

EGU22-6093 | Presentations | ITS2.7/AS5.2

Machine learning to quantify cloud responses to aerosols from satellite data 

Jessenia Gonzalez, Odran Sourdeval, Gustau Camps-Valls, and Johannes Quaas

The Earth's radiation budget may be altered by changes in atmospheric composition or land use. This is called radiative forcing. Among the human-generated influences in radiative forcing, aerosol-cloud interactions are the least understood. A way to quantify a key uncertainty in this regard, the adjustment of cloud liquid water path (LWP), is by the ratio (sensitivity) of LWP to changes in cloud droplet number concentration (Nd). A key problem in quantifying this sensitivity from large-scale observations is that these two quantities are not retrieved by operational satellite products and are subject to large uncertainties. 

In this work, we use machine learning techniques to show that inferring LWP and Nd directly from satellite observation data may yield a better understanding of this relationship without using retrievals, which may lead to large and systematic uncertainties. In particular, we use supervised learning on the basis of available high-resolution ICON-LEM (ICOsahedral Non-hydrostatic Large Eddy Model) simulations from the HD(CP)² project (High Definition Clouds and Precipitation for advancing Climate Prediction) and forward-simulated radiances obtained from the radiative transfer modeling (RTTOV, Radiative Transfer for TOVS) which uses MODIS (Moderate Resolution Imaging Spectroradiometer) data as a reference. Usually, only two channels from the reflectance of MODIS can be used to estimate the LWP and Nd. However, having access to 36 bands allows us to exploit data and find other patterns to get these parameters directly from the observation space rather than from the retrievals. A machine learning model is used to create an emulator which approximates the Radiative Transfer Model, and another machine learning model to directly predict the sensitivity of LWP - Nd from the satellite observation data.

How to cite: Gonzalez, J., Sourdeval, O., Camps-Valls, G., and Quaas, J.: Machine learning to quantify cloud responses to aerosols from satellite data, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6093,, 2022.

Microclimate is a relatively recent concept in atmospheric sciences, which started drawing attention of engineers and climatologists after proliferation of the open thermal (infrared, middle- and near-infrared) remote sensing instruments and high-resolution emissivity datasets. Rarely mentioned in the context of global climate change reversing, efficient management of microclimates nevertheless can be considered as a possible solution. Their function is bi-directional; On one hand, they can perform as ‘buffers’ by smoothing out effects of the already altered global climate on people and ecosystems, whilst also acting as the structural contributors to perturbations in the higher layers of the atmosphere. 

In the most abstract terms, microclimates tend to manifest themselves via land surface temperature conditions, which in turn are highly sensitive to the underlying land cover and use decisions. Forests are considered as the most efficient terrestrial carbon sinks and climate regulators, and various forms, configurations and continuity of logging can substantially alter the patterns of local temperature fluxes, precipitation and ecosystems. In this study we propose a novel heteroskedastic machine learning method, which can attribute localised forest loss patches due to industrial mining activity and estimate the resulting change in dynamics of the surrounding microclimate(s). 

How to cite: Tkachenko, N. and Garcia Velez, L.: Global attribution of microclimate dynamics to industrial deforestation sites using thermal remote sensing and machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6466,, 2022.

EGU22-6543 | Presentations | ITS2.7/AS5.2

High-resolution hybrid spatiotemporal modeling of daily relative humidity across Germany for epidemiological research: a Random Forest approach 

Nikolaos Nikolaou, Laurens Bouwer, Mahyar Valizadeh, Marco Dallavalle, Kathrin Wolf, Massimo Stafoggia, Annette Peters, and Alexandra Schneider

Introduction: Relative humidity (RH) is a meteorological variable of great importance as it affects other climatic variables and plays a role in plant and animal life as well as in human comfort and well-being. However, the commonly used weather station observations are inefficient to represent the great spatiotemporal RH variability, leading to exposure misclassification and difficulties to assess local RH health effects. There is also a lack of high-resolution RH spatial datasets and no readily available methods for modeling humidity across space and time. To tackle these issues, we aimed to improve the spatiotemporal coverage of RH data in Germany, using remote sensing and machine learning (ML) modeling.

Methods: In this study, we estimated German-wide daily mean RH at 1km2 resolution over the period 2000-2020. We used several predictors from multiple sources, including DWD RH observations, Ta predictions as well as satellite-derived DEM, NDVI and the True Color band composition (bands 1, 4 and 3: red, green and blue). Our main predictor for estimating the daily mean RH was the daily mean Ta. We had already mapped daily mean Ta in 1km2 across Germany through a regression-based hybrid approach of two linear mixed models using land surface temperature. Additionally, a very important predictor was the date, capturing the day-to-day variation of the response-explanatory variables relationship. All these variables were included in a Random Forest (RF) model, applied for each year separately. We assessed the model’s accuracy via 10-fold cross-validation (CV). First internally, using station observations that were not used for the model training, and then externally in the Augsburg metropolitan area using the REKLIM monitoring network over the period 2015-2019.

Results: Regarding the internal validation, the 21-year overall mean CV-R2 was 0.76 and the CV-RMSE was 6.084%. For the model’s external performance, at the same day, we found CV-R2=0.75 and CV-RMSE=7.051% and for the 7-day average, CV-R2=0.81 and CV-RMSE=5.420%. Germany is characterized by high relative humidity values, having a 20-year average RH of 78.4%. Even if the annual country-wide averages were quite stable, ranging from 81.2% for 2001 to 75.3% for 2020, the spatial variability exceeded 15% annually on average. Generally, winter was the most humid period and especially December was the most humid month. Extended urban cores (e.g., from Stuttgart to Frankfurt) or individual cities as Munich were less humid than the surrounding rural areas. There are also specific spatial patterns for RH distribution, including mountains, rivers and coastlines. For instance, the Alps and the North Sea coast are areas with elevated RH.

Conclusion: Our results indicate that the applied hybrid RF model is suitable for estimating nationwide RH at high spatiotemporal resolution, achieving a strong performance with low errors. Our method contributes to an improved spatial estimation of RH and the output product will help us understand better the spatiotemporal patterns of RH in Germany. We also plan to apply other ML techniques and compare the findings. Finally, our dataset will be used for epidemiological analyses, but could also be used for other research questions.

How to cite: Nikolaou, N., Bouwer, L., Valizadeh, M., Dallavalle, M., Wolf, K., Stafoggia, M., Peters, A., and Schneider, A.: High-resolution hybrid spatiotemporal modeling of daily relative humidity across Germany for epidemiological research: a Random Forest approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6543,, 2022.

EGU22-6958 | Presentations | ITS2.7/AS5.2

Causal Discovery in Ensembles of Climate Time Series 

Andreas Gerhardus and Jakob Runge

Understanding the cause and effect relationships that govern natural phenomena is central to the scientific inquiry. While being the gold standard for inferring causal relationships, there are many scenarios in which controlled experiments are not possible. This is for example the case for most aspects of Earth's complex climate system. Causal relationships then have to be learned from statistical dependencies in observational data, a task that is commonly referred to as (observational) causal discovery.

When applied to time series data for learning causal relationships in dynamical systems, methods for causal discovery face additional statistical challenges. This is so because, as licensed by an assumption of stationarity, samples are taken in a sliding window fashion and hence autocorrelated rather than iid. Moreover, strong autocorrelations also often occlude other relevant causal links. The recent PCMCI algorithm (Runge et al., 2019) and its variants PCMCI+ (Runge, 2020) and LPCMCI (Gerhardus and Runge, 2020) address and to some extent alleviate theses issues.

In this contribution we present the Ensemble-PCMCI method, an adaption of PCMCI (and its variants PCMCI+ and LPCMCI) to cases in which the data comprises several time series, i.e., measurements of several instances of the same underlying dynamical system. Samples can then be taken from these different time series instead of a in a sliding window fashion, thus avoiding the issue of autocorrelation and also allowing to relax the stationarity assumption. In particular, this opens the possibility to analyze temporal changes in the underlying causal mechanisms. A potential domain of application are ensemble forecasts.

Related references:
Jakob Runge et al. (2019). Detecting and quantifying causal associations in large nonlinear time series datasets. Science Advances 5 eaau4996.

Jakob Runge (2020). Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI). Proceedings of Machine Learning Research 124 1388–1397. PMLR.

Andreas Gerhardus and Jakob Runge (2020). High-recall causal discovery for autocorrelated time series with latent confounders. In Advances in Neural Information Processing Systems 33 12615–12625. Curran Associates, Inc.

How to cite: Gerhardus, A. and Runge, J.: Causal Discovery in Ensembles of Climate Time Series, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6958,, 2022.

EGU22-6998 | Presentations | ITS2.7/AS5.2

Inferring the Cloud Vertical Distribution from Geostationary Satellite Data 

Sarah Brüning, Holger Tost, and Stefan Niebler

Clouds and their radiative feedback mechanisms are of vital importance for the atmospheric cycle of the Earth regarding global weather today as well as climate changes in the future. Climate models and simulations are sensitive to the vertical distribution of clouds, emphasizing the need for broadly accessible fine resolution data. Although passive satellite sensors provide continuous cloud monitoring on a global scale, they miss the ability to infer physical properties below the cloud top. Active instruments like radar are particularly suitable for this task but lack an adequate spatio-temporal resolution. Here, recent advances in Deep-Learning models open up the possibility to transfer spatial information from a 2D towards a 3D perspective on a large-scale.

By an example period in 2017, this study aims to explore the feasibility and potential of neural networks to reconstruct the vertical distribution of volumetric radar data along a cloud’s column. For this purpose, the network has been tested on the Full Disk domain of a geostationary satellite with high spatio-temporal resolution data. Using raw satellite channels, spectral indices, and topographic data, we infer the 3D radar reflectivity from these physical predictors. First results demonstrate the network’s capability to reconstruct the cloud vertical distribution. Finally, the ultimate goal of interpolating the cloud column for the whole domain is supported by a considerably high accuracy in predicting the radar reflectivity. The resulting product can open up the opportunity to enhance climate models by an increased spatio-temporal resolution of 3D cloud structures.

How to cite: Brüning, S., Tost, H., and Niebler, S.: Inferring the Cloud Vertical Distribution from Geostationary Satellite Data, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6998,, 2022.

EGU22-7011 | Presentations | ITS2.7/AS5.2

Unlocking the potential of ML for Earth and Environment researchers 

Tobias Weigel, Frauke Albrecht, Caroline Arnold, Danu Caus, Harsh Grover, and Andrey Vlasenko

This presentation reports on support done under the aegis of Helmholtz AI for a wide range of machine learning based solutions for research questions related to Earth and Environmental sciences. We will give insight into typical problem statements from Earth observation and Earth system modeling that are good candidates for experimentation with ML methods and report on our accumulated experience tackling such challenges with individual support projects. We address these projects in an agile, iterative manner and during the definition phase, we direct special attention towards assembling practically meaningful demonstrators within a couple of months. A recent focus of our work lies on tackling software engineering concerns for building ML-ESM hybrids.

Our implementation workflow covers stages from data exploration to model tuning. A project may often start with evaluating available data and deciding on basic feasibility, apparent limitations such as biases or a lack of labels, and splitting into training and test data. Setting up a data processing workflow to subselect and compile training data is often the next step, followed by setting up a model architecture. We have made good experience with automatic tooling to tune hyperparameters and test and optimize network architectures. In typical implementation projects, these stages may repeat many times to improve results and cover aspects such as errors due to confusing samples, incorporating domain model knowledge, testing alternative architectures and ML approaches, and dealing with memory limitations and performance optimization.

Over the past two years, we have supported Helmholtz-based researchers from many subdisciplines on making the best use of ML methods along with these steps. Example projects include wind speed regression on GNSS-R data, emulation of atmospheric chemistry modeling, Earth System model parameterizations with ML, marine litter detection, and rogue waves prediction. The poster presentation will highlight selected best practices across these projects. We are happy to share our experience as it may prove useful to applications in wider Earth System modeling. If you are interested in discussing your challenge with us, please feel free to chat with us.

How to cite: Weigel, T., Albrecht, F., Arnold, C., Caus, D., Grover, H., and Vlasenko, A.: Unlocking the potential of ML for Earth and Environment researchers, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7011,, 2022.

EGU22-7034 | Presentations | ITS2.7/AS5.2

Developing a new emergent constraint through network analysis 

Lucile Ricard, Athanasios Nenes, Jakob Runge, and Fabrizio Falasca

Climate sensitivity expresses how average global temperature responds to an increase in greenhouse gas concentration. It is a key metric to assess climate change, and to formulate policy decisions, but its estimation from the Earth System Models (ESM) provides a wide range: between 2.5 and 4.0 K based on the sixth assessment report (AR6) of the Intergovernmental Panel on Climate Change (IPCC). To narrow down this spread, a number of observable metrics, called “emergent constraints” have been proposed, but often are based on relatively few parameters from a simulation – thought to express the “essence” of the climate simulation and its relationship with climate sensitivity. Many of the constraints to date however are model-dependent, therefore questionable in terms of their robustness.

We postulate that methods based on “holistic” consideration of the simulations and observations may provide more robust constraints; we also focus on Sea Surface Temperature (SST) ensembles as SST is a major driver of climate variability. To extract the essential patterns of SST variability, we use a knowledge discovery and network inference method, δ-Maps (Fountalis et al., 2016, Falasca et al, 2019), expanded to include a causal discovery algorithm (PCMCI) that relies on conditional independence testing, to capture the essential dynamics of the climate simulation on a functional graph and explore the true causal effects of the underlying dynamical system (Runge et al., 2019). The resulting networks are then quantitatively compared using network “metrics” that capture different aspects, including the regions of uniform behavior, how they alternate over time and the strength of association. These metrics are then compared between simulations, and observations and used as emergent constraints, called Causal Model Evaluation (CME).

We apply δ-Maps and CME to CMIP6 model SST outputs and demonstrate how the networks and related metrics can be used to assess the historical performance of CMIP models, and climate sensitivity. We start by comparing the CMIP6 simulations against CMIP5 models, by using the reanalysis dataset HadISST (Met Office Hadley Centre) as a proxy for observations. Each field is reduced to a network and then how similar they are with reanalysis SST. The CMIP6 historical networks are then compared against CMIP6 projected networks, build from the Shared Socio-Economic Pathway ssp245 (“Middle of the road”) scenario. Comparing past and future SST networks help us to evaluate the extent to which climate warming is encompassed in the change overlying dynamical system of our networks. A large distance from network build over the past period to network build over a future scenario could be tightly related to a large temperature response to an increase of greenhouse gas emission, that is the way we define climate sensitivity. We finally give a new estimation of the climate sensitivity with a weighting scheme approach, derived from a combination of its performance metrics.

How to cite: Ricard, L., Nenes, A., Runge, J., and Falasca, F.: Developing a new emergent constraint through network analysis, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7034,, 2022.

EGU22-7355 | Presentations | ITS2.7/AS5.2

Combining cloud properties and synoptic observations to predict cloud base height using Machine Learning 

Julien Lenhardt, Johannes Quaas, and Dino Sejdinovic

Cloud base height (CBH) is an important geometric parameter of a cloud and shapes its radiative properties. The CBH is also further of practical interest in the aviation community regarding pilot visibility and aircraft icing hazards. While the cloud-top height has been successfully derived from passive imaging radiometers on satellites during recent years, the derivation of the CBH remains a more difficult challenge with these same retrievals.

In our study we combine surface observations and passive satellite remote-sensing retrievals to create a database of CBH labels and cloud properties to ultimately train a machine learning model predicting CBH. The labels come from the global marine meteorological observations dataset (UK Met Office, 2006) which consists of near-global synoptic observations made on sea. This data set provides information about CBH, cloud type, cloud cover and other meteorological surface quantities with CBH being the main interest here. The features based upon which the machine learning model is trained consist in different cloud-top and cloud optical properties (Level 2 products MOD06/MYD06 from the MODIS sensor) extracted on a 127km x 127km grid around the synoptic observation point. To study the large diversity in cloud scenes, an auto-encoder architecture is chosen. The regression task is then carried out in the modelled latent space which is output by the encoder part of the model. To account for the spatial relationships in our input data the model architecture is based on Convolutional Neural Networks. We define a study domain in the Atlantic ocean, around the equator. The combination of information from below and over the cloud could allow us to build a robust model to predict CBH and then extend predictions to regions where surface measurements are not available.

How to cite: Lenhardt, J., Quaas, J., and Sejdinovic, D.: Combining cloud properties and synoptic observations to predict cloud base height using Machine Learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-7355,, 2022.

EGU22-8068 | Presentations | ITS2.7/AS5.2

Generative Adversarial Modeling of Tropical Precipitation and the Intertropical Convergence Zone 

Cody Nash, Balasubramanya Nadiga, and Xiaoming Sun

In this study we evaluate the use of generative adversarial networks (GANs) to model satellite-based estimates of precipitation conditioned on reanalysis temperature, humidity, wind, and surface latent heat flux.  We are interested in the climatology of precipitation and modeling it in terms of atmospheric state variables, in contrast to a weather forecast or precipitation nowcast perspective.  We consider a hierarchy of models in terms of complexity, including simple baselines, generalized linear models, gradient boosted decision trees, pointwise GANs and deep convolutional GANs. To gain further insight into the models we apply methods for analyzing machine learning models, including model explainability, ablation studies, and a diverse set of metrics for pointwise and distributional differences, including information theory based metrics.  We find that generative models significantly outperform baseline models on metrics based on the distribution of predictions, particularly in capturing the extremes of the distributions.  Overall, a deep convolutional model achieves the highest accuracy.  We also find that the relative importance of atmospheric variables and of their interactions vary considerably among the different models considered. 

How to cite: Nash, C., Nadiga, B., and Sun, X.: Generative Adversarial Modeling of Tropical Precipitation and the Intertropical Convergence Zone, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8068,, 2022.

EGU22-8130 | Presentations | ITS2.7/AS5.2

A comparison of explainable AI solutions to a climate change prediction task 

Philine Lou Bommer, Marlene Kretschmer, Dilyara Bareeva, Kadircan Aksoy, and Marina Höhne

In climate change research we are dealing with a chaotic system, usually leading to huge computational efforts in order to make faithful predictions. Deep neural networks (DNNs) offer promising new approaches due to their computational efficiency and universal solution properties. However, despite the increase in successful application cases with DNNs, the black-box nature of such purely data-driven approaches limits their trustworthiness and therefore the useability of deep learning in the context of climate science.

The field of explainable artificial intelligence (XAI) has been established to enable a deeper understanding of the complex, highly-nonlinear methods and their predictions. By shedding light onto the reasons behind the predictions made by DNNs, XAI methods can serve as a support for researchers to reveal the underlying physical mechanisms and properties inherent in the studied data. Some XAI methods have already been successfully applied to climate science, however, no detailed comparison of their performances is available. As the number of XAI methods on the one hand, and DNN applications on the other hand are growing, a comprehensive evaluation is necessary in order to understand the different XAI methods in the climate context.

In this work we provide an overview of different available XAI methods and their potential applications for climate science. Based on a previously published climate change prediction task, we compare several explanation approaches, including model-aware (e.g. Saliency, IntGrad, LRP) and model-agnostic methods (e.g. SHAP). We analyse their ability to verify the physical soundness of the DNN predictions as well as their ability to uncover new insights into the underlying climate phenomena. Another important aspect we address in our work is the possibility to assess the underlying uncertainties of DNN predictions using XAI methods. This is especially crucial in climate science applications where uncertainty due to natural variability is usually large. To this end, we investigate the potential of two recently introduced XAI methods —UAI+ and NoiseGrad, which have been designed to include uncertainty information of the predictions into the explanations. We demonstrate that those XAI methods enable more stable explanations with respect to model noise and can further deal with uncertainties of network information. We argue that these methods are therefore particularly suitable for climate science application cases.

How to cite: Bommer, P. L., Kretschmer, M., Bareeva, D., Aksoy, K., and Höhne, M.: A comparison of explainable AI solutions to a climate change prediction task, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8130,, 2022.

Despite the importance of the Atlantic Meridional Overturning Circulation (AMOC) to the climate on decadal and multidecadal timescales, Earth System Models (ESM) exhibit large differences in their estimation of the amplitude and spectrum of its variability. In addition, observational data is sparse and before the onset of the current century, many reconstructions of the AMOC rely on linear relationships to the more readily observed surface properties of the Atlantic rather than the less explored deeper ocean. Yet, it is conceptually well established that the density distribution is dynamically closely related to the AMOC, and in this contribution, we investigate this connection in model simulations to identify which density information is necessary to reconstruct the AMOC. We chose to establish these links in a data-driven approach. 

We use simulations from a historically forced large ensemble as well as abruptly forced long term simulations with varying strength of forcing and therefore comprising vastly different states of the AMOC. In a first step, we train uncertainty-aware neural networks to infer the state of the AMOC from the density information at different layers in the North Atlantic. In a second step, we compare the performance of the trained neural networks across depth and with their linear counterparts in simulations that were not part of the training process. Finally, we investigate how the networks arrived at their specific prediction using Layer-Wise-Relevance Propagation (LRP), a recently developed technique that propagates relevance backwards through the network to the input density field, effectively filtering out important from unimportant information and identifying regions of high relevance for the reconstruction of the AMOC.

Our preliminary results show that in general, the information provided by only one density layer between the surface and 1100 m is sufficient to reconstruct the AMOC with high precision, and neural networks are capable of generalizing to unseen simulations. From the set of these neural networks trained on different layers, we choose the surface layer as well as one subsurface layer close to 1000 m for further investigation of their decision-making process using LRP. Our preliminary investigation reveals that the LRP in the subsurface layer identifies regions of potentially high physical relevance for the AMOC. By contrast, the regions identified in the surface layer show little physical relevance for the AMOC.

How to cite: Mayer, B., Barnes, E., Marotzke, J., and Baehr, J.: Reconstructing the Atlantic Meridional Overturning Circulation in Earth System Model simulations from density information using explainable machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8411,, 2022.

EGU22-8454 | Presentations | ITS2.7/AS5.2

Using Generative Adversarial Networks (GANs) to downscale tropical cyclone precipitation. 

Emily Vosper, Dann Mitchell, Peter Watson, Laurence Aitchison, and Raul Santos-Rodriguez

Fluvial flood hazards from tropical cyclones (TCs) are frequently the leading cause of mortality and damages (Rezapour and Baldock, 2014). Accurately modeling TC precipitation is vital for studying the current and future impacts of TCs. However, general circulation models at typical resolution struggle to accurately reproduce TC rainfall, especially for the most extreme storms (Murakami et al., 2015). Increasing horizontal resolution can improve precipitation estimates (Roberts et al., 2020; Zhang et al., 2021), but as these methods are computationally expensive there is a trade-off between accuracy and generating enough ensemble members to generate sufficient high impact, low probability events. Often, downscaling models are used as a computationally cheaper alternative. 

Here, we downscale TC precipitation data from 100 km to 10 km resolution using a generative adversarial network (GAN). Generative approaches have the potential to well reproduce the fine spatial detail and stochastic nature of precipitation (Ravuri et al., 2021). Using observational products for tracking (IBTrACS) and rainfall (MSWEP), we train our GAN over the historical period 1979 - 2020. We are interested in how well our model reproduces precipitation intensity and structure with a focus on the most extreme events, where models have traditionally struggled. 


Murakami, H., et al., 2015. Simulation and Prediction of Category 4 and 5 Hurricanes in the High-Resolution GFDL HiFLOR Coupled Climate Model*. Journal of Climate, 28(23), pp.9058-9079. 

Ravuri, S., et al., 2021. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597(7878), pp.672-677. 

Rezapour, M. and Baldock, T., 2014. Classification of Hurricane Hazards: The Importance of Rainfall. Weather and Forecasting, 29(6), pp.1319-1331. 

Roberts, M., et al., 2020. Impact of Model Resolution on Tropical Cyclone Simulation Using the HighResMIP–PRIMAVERA Multimodel Ensemble. Journal of Climate, 33(7), pp.2557-2583. 

Zhang, W., et al., 2021. Tropical cyclone precipitation in the HighResMIP atmosphere-only experiments of the PRIMAVERA Project. Climate Dynamics, 57(1-2), pp.253-273. 

How to cite: Vosper, E., Mitchell, D., Watson, P., Aitchison, L., and Santos-Rodriguez, R.: Using Generative Adversarial Networks (GANs) to downscale tropical cyclone precipitation., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8454,, 2022.

EGU22-8499 | Presentations | ITS2.7/AS5.2 | Highlight

Matryoshka Neural Operators: Learning Fast PDE Solvers for Multiscale Physics 

Björn Lütjens, Catherine H. Crawford, Campbell Watson, Chris Hill, and Dava Newman

Running a high-resolution global climate model can take multiple days on the world's largest supercomputers. Due to the long runtimes that are caused by solving the underlying partial differential equations (PDEs), climate researchers struggle to generate ensemble runs that are necessary for uncertainty quantification or exploring climate policy decisions.


Physics-informed neural networks (PINNs) promise a solution: they can solve single instances of PDEs up to three orders of magnitude faster than traditional finite difference numerical solvers. However, most approaches in physics-informed machine learning learn the solution of PDEs over the full spatio-temporal domain, which requires infeasible amounts of training data, does not exploit knowledge of the underlying large-scale physics, and reduces model trust. Our philosophy is to limit learning to the hard-to-model parts. Hence, we are proposing a novel method called \textit{matryoshka neural operator} that leverages an old scheme called super-parametrizations developed in geophysical fluid dynamics. Using this scheme our proposed physics-informed architecture exploits knowledge of approximate large-scale dynamics and only learns the influence of small-scale dynamics onto large-scale dynamics, also called subgrid parametrizations.


Some work in geophysical fluid dynamics is conceptually similar, but fully relies on neural networks which can only operate on fixed grids (Gentine et al., 2018). We are the first to learn grid-independent subgrid parametrizations by leveraging neural operators that learn the dynamics in a grid-independent latent space. Neural operators can be seen as an extension of neural networks to infinite-dimensions: They encode infinite-dimensional inputs into a finite-dimensional representations, such as Eigen or Fourier modes, and learn the nonlinear temporal dynamics in the encoded state.


We demonstrate the neural operators for learning non-local subgrid parametrizations over the full large-scale domain of the two-scale Lorenz96 equation. We show that the proposed learning-based PDE solver is grid-independent, has quasilinear instead of quadratic complexity in comparison to a fully-resolving numerical solver, is more accurate than current neural network or polynomial-based parametrizations, and offers interpretability through Fourier modes.


Gentine, P., Pritchard, M., Rasp, S., Reinaudi, G., and Yacalis, G. (2018). Could machine learning break the convection parameterization deadlock? Geophysical Research Letters, 45, 5742– 5751.

How to cite: Lütjens, B., Crawford, C. H., Watson, C., Hill, C., and Newman, D.: Matryoshka Neural Operators: Learning Fast PDE Solvers for Multiscale Physics, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8499,, 2022.

EGU22-8649 | Presentations | ITS2.7/AS5.2

Physically Based Deep Learning Framework to Model Intense Precipitation Events at Engineering Scales 

Bernardo Teufel, Fernanda Carmo, Laxmi Sushama, Lijun Sun, Naveed Khaliq, Stephane Belair, Asaad Yahia Shamseldin, Dasika Nagesh Kumar, and Jai Vaze

The high computational cost of super-resolution (< 250 m) climate simulations is a major barrier for generating climate change information at such high spatial and temporal resolutions required by many sectors for planning local and asset-specific climate change adaptation strategies. This study couples machine learning and physical modelling paradigms to develop a computationally efficient simulator-emulator framework for generating super-resolution climate information. To this end, a regional climate model (RCM) is applied over the city of Montreal, for the summers of 2015 to 2020, at 2.5 km (i.e., low resolution – LR) and 250 m (i.e., high resolution – HR), which is used to train and validate the proposed super-resolution deep learning (DL) model. In the field of video super-resolution, convolutional neural networks combined with motion compensation have been used to merge information from multiple LR frames to generate high-quality HR images. In this study, a recurrent DL approach based on passing the generated HR estimate through time helps the DL model to recreate fine details and produce temporally consistent fields, resembling the data assimilation process commonly used in numerical weather prediction. Time-invariant HR surface fields and storm motion (approximated by RCM-simulated wind) are also considered in the DL model, which helps further improve output realism. Results suggest that the DL model is able to generate HR precipitation estimates with significantly lower errors than other methods used, especially for intense short-duration precipitation events, which often occur during the warm season and are required to evaluate climate resiliency of urban storm drainage systems. The generic and flexible nature of the developed framework makes it even more promising as it can be applied to other climate variables, periods and regions.

How to cite: Teufel, B., Carmo, F., Sushama, L., Sun, L., Khaliq, N., Belair, S., Shamseldin, A. Y., Nagesh Kumar, D., and Vaze, J.: Physically Based Deep Learning Framework to Model Intense Precipitation Events at Engineering Scales, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8649,, 2022.

EGU22-8656 | Presentations | ITS2.7/AS5.2 | Highlight

Conditional normalizing flow for predicting the occurrence of rare extreme events on long time scales 

Jakob Kruse, Beatrice Ellerhoff, Ullrich Köthe, and Kira Rehfeld

The socio-economic impacts of rare extreme events, such as droughts, are one of the main ways in which climate affects humanity. A key challenge is to quantify the changing risk of once-in-a-decade or even once-in-a-century events under global warming, while leaning heavily on comparatively short observation spans. The predictive power of classical statistical methods from extreme value theory (EVT) often remains limited to uncorrelated events with short return periods. This is mainly due to their strong assumption of an underlying exponential family distribution of the variable in question. Standard EVT is therefore at odds with the rich and large-scale correlations found in various surface climate parameters such as local temperatures, as well as the more complex shape of empirical distributions. Here, we turn to recent developments in machine learning, namely to conditional normalizing flows, which are flexible neural networks for modeling highly-correlated unknown distributions. Given a short time series, we show how such networks can model the posterior probability of events whose return periods are much longer than the observation span. The necessary correlations and patterns can be extracted from a paired set of inputs, i.e. time series, and outputs, i.e. return periods. To evaluate this approach in a controlled setting, we generate synthetic training data by sampling temporally autoregressive processes with a non-trivial covariance structure. We compare the results to a baseline analysis using EVT. In this work, we focus on the prediction of return periods of rare statistical events. However, we expect the same potential for a wide range of statistical measures, such as the power spectrum and rate functions. Future work should also investigate its applicability to compound and spatially extended events, as well as changing conditions under warming scenarios.

How to cite: Kruse, J., Ellerhoff, B., Köthe, U., and Rehfeld, K.: Conditional normalizing flow for predicting the occurrence of rare extreme events on long time scales, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8656,, 2022.

EGU22-8848 | Presentations | ITS2.7/AS5.2 | Highlight

Defining regime specific cloud sensitivities using the learnings from machine learning 

Alyson Douglas and Philip Stier

Clouds remain a core uncertainty in quantifying Earth’s climate sensitivity due to their complex dynamical and microphysical  interactions with multiple components of the Earth system. Therefore it is pivotal to observationally constrain possible cloud changes in a changing climate in order to evaluate our current generation of Earth system models by a set of physically realistic sensitivities. We developed a novel observational regime framework from over 15 years of MODIS satellite observations, from which we have derived a set of regimes of cloud controlling factors. These regimes were established using the relationship strength, as measured by using the weights of a trained, simple machine learning model. We apply these as observational constraints on the ​​r1i1p1f1 and r1i1p1f3 historical runs from various CMIP6 models to test if CMIP6 climate models can accurately represent key cloud controlling factors.. Within our regime framework, we can compare the observed environmental drivers and sensitivities of each regime against the parameterization-driven, modeled outcomes. We find that, for almost every regime, CMIP6 models do not properly represent the global distribution of occurrence, raising into question how much we can trust our range of climate sensitivities when specific cloud controlling factors are so badly represented by these models. This is especially pertinent in southern ocean and marine stratocumulus regimes, as the changes in these clouds’ optical depths and cloud amount have increased the ECS from CMIP5 to CMIP6. Our results suggest that these uncertainties in CMIP6 cloud parameterizations propagate into derived cloud feedbacks and ultimately climate sensitivity, which is evident from a regimed based analysis of cloud controlling factors.

How to cite: Douglas, A. and Stier, P.: Defining regime specific cloud sensitivities using the learnings from machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8848,, 2022.

EGU22-9112 | Presentations | ITS2.7/AS5.2

Causal Orthogonal Functions: A Causal Inference approach to temporal feature extraction 

Nicolas-Domenic Reiter, Jakob Runge, and Andreas Gerhardus

Understanding complex dynamical systems is a major challenge in many scientific disciplines. There are two aspects which are of particular interest when analyzing complex dynamical systems: 1) the temporal patterns along which they evolve and 2) the governing causal mechanisms.

Temporal patterns in a time-series can be extracted and analyzed through a variety of time-series representations, that is a collection of filters. Discrete Wavelet and Fourier Transforms are prominent examples and have been widely applied to investigate the temporal structure of dynamical systems.

Causal Inference is a framework formalizing questions of cause and effect. In this work we propose an elementary and systematic approach to combine time-series representations with Causal Inference. Hereby we introduce a notion of cause and effect with respect to a pair of arbitrary time-series filters. Using a Singular Value Decomposition we derive an alternative representation of how one process drives another over a specified time-period. We call the building blocks of this representation Causal Orthogonal Functions. Combining the notion of Causal Orthogonal Functions with a Wavelet or Fourier decomposition of a time-series yields time-scale specific Causal Orthogonal Functions. As a result we obtain a time-scale specific representation of the causal influence one process has on another over some fixed time-period. This allows to conduct causal effect analysis in discrete-time stochastic dynamical systems at multiple time-scales. We illustrate our approach by examining linear VAR processes.

How to cite: Reiter, N.-D., Runge, J., and Gerhardus, A.: Causal Orthogonal Functions: A Causal Inference approach to temporal feature extraction, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9112,, 2022.

Outliers detection generally aims at identifying extreme events and insightful changes in climate behavior. One important type of outlier is pattern outlier also called discord, where the outlier pattern detected covers a time interval instead of a single point in the time series. Machine learning contributes many algorithms and methods in this field especially unsupervised algorithms for different types of data time series. In a first submitted paper, we have investigated discord detection applied to climate-related impact observations. We have introduced the prominent discord notion, a contextual concept that derives a set of insightful discords by identifying dependencies among variable length discords, and which is ordered based on the number of discords they subsume. 

Following this study, here we propose a ranking function based on the length of the first subsumed discord and the total length of the prominent discord, and make use of the powerful matrix profile technique. Preliminary results show that our approach, applied to monthly runoff timeseries between 1902 and 2005 over West Africa, detects both the emergence of long term change with the associated former climate regime, and the regional driest decade (1982-1992) of the 20th century (i.e. climate extreme event). In order to demonstrate the genericity and multiple insights gained by our method, we go further by evaluating the approach on other impact (e.g. crop data, fires, water storage) and climate (precipitation and temperature) observations, to provide similar results on different variables, extract relationships among them and identify what constitutes a prominent discord in such cases. A further step will consist in evaluating our methodology on climate and impact historical simulations, to determine if prominent discords highlighted in observations can be captured in climate and impact models.

How to cite: El Khansa, H., Gervet, C., and Brouillet, A.: Prominent discords in climate data through matrix profile techniques: detecting emerging long term pattern changes and anomalous events , EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9250,, 2022.

EGU22-9281 | Presentations | ITS2.7/AS5.2

Machine learning-based identification and classification of ocean eddies 

Eike Bolmer, Adili Abulaitijiang, Jürgen Kusche, Luciana Fenoglio-Marc, Sophie Stolzenberger, and Ribana Roscher

The automatic detection and tracking of mesoscale ocean eddies, the ‘weather of the ocean’, is a well-known task in oceanography. These eddies have horizontal scales from 10 km up to 100 km and above. They transport water mass, heat, nutrition, and carbon and have been identified as hot spots of biological activity. Monitoring eddies is therefore of interest among others to marine biologists and fishery. 
Recent advances in satellite-based observation for oceanography such as sea surface height (SSH) and sea surface temperature (SST) result in a large supply of different data products in which eddies are visible. In radar altimetry observations are acquired with repeat cycles between 10 and 35 days and cross-track spacing of a few 10 km to a few 100 km. Therefore, ocean eddies are clearly visible but typically covered by only one ground track. In addition, due to their motion, eddies are difficult to reconstruct, which makes creating detailed maps of the ocean with a high temporal resolution a challenge. In general, they are considered a perturbation, and their influence on altimetry data is difficult to determine, which is especially limiting for the determination of an accurate time-averaged dynamic topography of the ocean.
Due to their spatio-temporal dynamic behavior the identification and tracking are challenging. There is a number of methods that have been developed to identify and track eddies in gridded maps of sea surface height derived from multi-mission data sets. However, these procedures have shortcomings since the gridding process removes information that is valuable in achieving more accurate results.
Therefore, in the project EDDY carried out at the University of Bonn we intend to use ground track data from satellite altimetry and - as a long-term goal - additional remote sensing data such as SST, optical imagery, as well as statistical information from model outputs. The combination of the data will serve as a basis for a multi-modal deep learning algorithm. In detail, we will utilize transformers, a deep neural network architecture, that originates from the field of Natural Language Processing (NLP) and became popular in recent years in the field of computer vision. This method shows promising results in terms of understanding temporal and spatial information, which is essential in detecting and tracking highly dynamic eddies.
In this presentation, we introduce the deep neural network used in the EDDY project and show the results based on gridded data sets for the Gulf stream area for the period 2017 and first results of single-track eddy identification in the region.

How to cite: Bolmer, E., Abulaitijiang, A., Kusche, J., Fenoglio-Marc, L., Stolzenberger, S., and Roscher, R.: Machine learning-based identification and classification of ocean eddies, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9281,, 2022.

EGU22-9461 | Presentations | ITS2.7/AS5.2

Data Driven Approaches for Climate Predictability 

Balasubramanya Nadiga

Reduced-order dynamical models play a central role in developing our understanding of predictability of climate. In this context, the Linear Inverse Modeling (LIM) approach (closely related to Dynamic Mode Decomposition DMD), by helping capture a few essential interactions between dynamical components of the full system, has proven valuable in being able to give insights into the dynamical behavior of the full system. While nonlinear extensions of the LIM approach have been attempted none have gained widespread acceptance. We demonstrate that Reservoir Computing (RC), a form of machine learning suited for learning in the context of chaotic dynamics, by exploiting the phenomenon of generalized synchronization, provides an alternative nonlinear approach that comprehensively outperforms the LIM approach.  Additionally, the potential of the RC approach to capture the structure of the climatological attractor and to continue the evolution of the system on the attractor in a realistic fashion long after the ensemble average has stopped tracking the reference trajectory is highlighted. Finally, other dynamical systems based methods and probabilistic deep learning methods are considered and a broader perspective on the use of data-driven methods in understanding climate predictability is offered

How to cite: Nadiga, B.: Data Driven Approaches for Climate Predictability, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9461,, 2022.

EGU22-9877 | Presentations | ITS2.7/AS5.2

A Conditional Generative Adversarial Network for Rainfall Downscaling 

Marcello Iotti, Paolo Davini, Jost von Hardenberg, and Giuseppe Zappa

Predicting extreme precipitation events is one of the main challenges of climate science in this decade. Despite the continuously increasing computing availability, Global Climate Models’ (GCMs) spatial resolution is still too coarse to correctly represent and predict small-scale phenomena as convection, so that precipitation prediction is still imprecise. Indeed, precipitation shows variability on both spatial and temporal scales (much) smaller than the current state-of-the-art GCMs resolution. Therefore, downscaling techniques play a crucial role, both for the understanding of the phenomenon itself and for applications like e.g. hydrologic studies, risk prediction and emergency management. Seen in the context of image processing, a downscaling procedure has many similarities with super-resolution tasks, i.e. the improvement of the resolution of an image. This scope has taken advantage from the application of Machine Learning techniques, and in particular from the introduction of Convolutional Neural Networks (CNNs).

In our work we exploit a conditional Generative Adversarial Network (cGAN) to train a generator model to perform precipitation downscaling. This generator, a deep CNN, takes as input the precipitation field at the scale resolved by GCMs, adds random noise, and outputs a possible realization of the precipitation field at higher resolution, preserving its statistical properties with respect to the coarse-scale field. The GAN is being trained and tested in a “perfect model” setup, in which we try to reproduce the ERA5 precipitation field starting from an upscaled version of it.

Compared to other downscaling techniques, our model has the advantage of being computationally inexpensive at run time, since the computational load is mostly concentrated in the training phase. We are examining the Greater Alpine Region, upon which numerical models performances are limited by the complex orography. Nevertheless the approach, being independent of physical, statistical and empirical assumptions, can be easily extended to different domains.

How to cite: Iotti, M., Davini, P., von Hardenberg, J., and Zappa, G.: A Conditional Generative Adversarial Network for Rainfall Downscaling, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-9877,, 2022.

EGU22-10120 | Presentations | ITS2.7/AS5.2

A Convolutional Neural Network approach for downscaling climate model data in Trentino-South Tyrol (Eastern Italian Alps) 

Alice Crespi, Daniel Frisinghelli, Tatiana Klisho, Marcello Petitta, Alexander Jacob, and Massimiliano Pittore

Statistical downscaling is a very popular technique to increase the spatial resolution of existing global and regional climate model simulations and to provide reliable climate data at local scale. The availability of tailored information is particularly crucial for conducting local climate assessments, climate change studies and for running impact models, especially in complex terrain. A crucial requirement is the ability to reliably downscale the mean, variability and extremes of climate data, while preserving their spatial and temporal correlations.

Several machine learning-based approaches have been proposed so far to perform such task by extracting non-linear relationships between local-scale variables and large-scale atmospheric predictors and they could outperform more traditional statistical methods. In recent years, deep learning has gained particular interest in geoscientific studies and climate science as a promising tool to improve climate downscaling thanks to its greater ability to extract high-level features from large datasets using complex hierarchical architectures. However, the proper network architecture is highly dependent on the target variable, time and spatial resolution, as well as application purposes and target domain.

This contribution presents a Deep Convolutional Encoder-Decoder Network (DCEDN) architecture which was implemented and evaluated for the first time over Trentino-South Tyrol in the Eastern Italian Alps to derive 1-km climate fields of daily temperature and precipitation from ERA-5 reanalysis. We will show that in-depth optimization of hyper-parameters, loss function choice and sensitivity analyses are essential preliminary steps to derive an effective architecture and enhance the interpretability of results and of their variability. The validation of downscaled fields of both temperature and precipitation confirmed the improved representation of local features for both mean and extreme values, even though lower performances were obtained for precipitation in reproducing small-scale spatial features. In all cases, DCEDN was found to outperform classical schemes based on linear regression and the bias adjustment procedures used as benchmarks. We will discuss in detail the advantages and recommendations for the integration of DCEDN as an efficient post-processing block in climate data simulations supporting local-scale studies. The model constraints in feature extraction, especially for precipitation, over the limited extent of the study domain will also be explained along with potential future developments of such type of networks for improved climate science applications.

How to cite: Crespi, A., Frisinghelli, D., Klisho, T., Petitta, M., Jacob, A., and Pittore, M.: A Convolutional Neural Network approach for downscaling climate model data in Trentino-South Tyrol (Eastern Italian Alps), EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10120,, 2022.

EGU22-10773 | Presentations | ITS2.7/AS5.2 | Highlight

Choose your own weather adventure: deep weather generation for “what-if” climate scenarios 

Campbell Watson, Jorge Guevara, Daniela Szwarcman, Dario Oliveira, Leonardo Tizzei, Maria Garcia, Priscilla Avegliano, and Bianca Zadrozny

Climate change is making extreme weather more extreme. Given the inherent uncertainty of long-term climate projections, there is growing need for rapid, plausible “what-if” climate scenarios to help users understand climate exposure and examine resilience and mitigation strategies. Since the 1980s, such “what-if” scenarios have been created using stochastic weather generators. However, it is very challenging for traditional weather generation algorithms to create realistic extreme climate scenarios because the weather data being modeled is highly imbalanced, contains spatiotemporal dependencies and has extreme weather events exacerbated by a changing climate.

There are few works comparing and evaluating stochastic multisite (i.e., gridded) weather generators, and no existing work that compares promising deep learning approaches for weather generation with classical stochastic weather generators. We will present the culmination of a multi-year effort to perform a systematic evaluation of stochastic weather generators and deep generative models for multisite precipitation synthesis. Among other things, we show that variational auto-encoders (VAE) offer an encouraging pathway for efficient and controllable climate scenario synthesis – especially for extreme events. Our proposed VAE schema selects events with different characteristics in the normalized latent space (from rare to common) and generates high-quality scenarios using the trained decoder. Improvements are provided via latent space clustering and bringing histogram-awareness to the VAE loss.

This research will serve as a guide for improving the design of deep learning architectures and algorithms for application in Earth science, including feature representation and uncertainty quantification of Earth system data and the characterization of so-called “grey swan” events.

How to cite: Watson, C., Guevara, J., Szwarcman, D., Oliveira, D., Tizzei, L., Garcia, M., Avegliano, P., and Zadrozny, B.: Choose your own weather adventure: deep weather generation for “what-if” climate scenarios, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10773,, 2022.

EGU22-10888 | Presentations | ITS2.7/AS5.2

How to utilize deep learning to understand climate dynamics? : An ENSO example. 

Na-Yeon Shin, Yoo-Geun Ham, Jeong-Hwan Kim, Minsu Cho, and Jong-Seong Kug

Many deep learning technologies have been applied to the Earth sciences, including weather forecast, climate prediction, parameterization, resolution improvements, etc. Nonetheless, the difficulty in interpreting deep learning results still prevents their applications to studies on climate dynamics. Here, we applied a convolutional neural network to understand El Niño–Southern Oscillation (ENSO) dynamics from long-term climate model simulations. The deep learning algorithm successfully predicted ENSO events with a high correlation skill of 0.82 for a 9-month lead. For interpreting deep learning results beyond the prediction skill, we first developed a “contribution map,” which estimates how much each grid point and variable contribute to a final output variable. Furthermore, we introduced a “sensitivity,” which estimates how much the output variable is sensitively changed to the small perturbation of the input variables by showing the differences in the output variables. The contribution map clearly shows the most important precursors for El Niño and La Niña developments. In addition, the sensitivity clearly reveals nonlinear relations between the precursors and the ENSO index, which helps us understand the respective role of each precursor. Our results suggest that the contribution map and sensitivity would be beneficial for understanding other climate phenomena.

How to cite: Shin, N.-Y., Ham, Y.-G., Kim, J.-H., Cho, M., and Kug, J.-S.: How to utilize deep learning to understand climate dynamics? : An ENSO example., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10888,, 2022.

EGU22-11111 | Presentations | ITS2.7/AS5.2

Machine learning based estimation of regional Net Ecosystem Exchange (NEE) constrained by atmospheric inversions and ecosystem observations 

Samuel Upton, Ana Bastos, Fabian Gans, Basil Kraft, Wouter Peters, Jacob Nelson, Sophia Walther, Martin Jung, and Markus Reichstein

Accurate estimates and predictions of the global carbon fluxes are critical for our understanding of the global carbon cycle and climate change. Reducing the uncertainty of the terrestrial carbon sink and closing the budget imbalance between sources and sinks would improve our ability to accurately project future climate change. Net Ecosystem Exchange (NEE), the net flux of biogenic carbon from the land surface to the atmosphere, is only directly measured at a sparse set of globally distributed eddy-covariance measurement sites. To estimate the terrestrial carbon flux at the regional and global scale, a global gridded estimate of NEE must be accurately upscaled from a model trained at the ecosystem level. In this study, the Fluxcom system* is used to train a site-level model on remotely-sensed and meteorological variables derived from site measurements, MODIS and ECMWF ERA5 atmospheric reanalysis data. The non-representative distribution of these site-level data along with missing disturbance histories impart known biases to current upscaling efforts. Observations of atmospheric carbon may provide important additional information, improving the accuracy of the upscaled flux estimate. 

This study adds an atmospheric observational operator to the model training process that connects the ecosystem-level flux model to top-down observations of atmospheric carbon by adding an additional term to the objective function. The target data are regionally integrated fluxes from an ensemble of atmospheric inversions corrected for fossil-fuel emissions and lateral fluxes.  Calculating the regionally integrated flux estimate at each training step is computationally infeasible. Our hypothesis is that the regional flux can be modeled with a limited set of points and that this sparse model preserves sufficient information about the phenomena to act as a constraint for the underlying ecosystem-level model, improving regional and global upscaled products.  Experimental results show improvements in the machine learning based regional estimates of NEE while preserving features such as the seasonal variability in the estimated flux.


*Jung, Martin, Christopher Schwalm, Mirco Migliavacca, Sophia Walther, Gustau Camps-Valls, Sujan Koirala, Peter Anthoni, et al. 2020. “Scaling Carbon Fluxes from Eddy Covariance Sites to Globe: Synthesis and Evaluation of the FLUXCOM Approach.” Biogeosciences 17 (5): 1343–65. 


How to cite: Upton, S., Bastos, A., Gans, F., Kraft, B., Peters, W., Nelson, J., Walther, S., Jung, M., and Reichstein, M.: Machine learning based estimation of regional Net Ecosystem Exchange (NEE) constrained by atmospheric inversions and ecosystem observations, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11111,, 2022.

EGU22-11216 | Presentations | ITS2.7/AS5.2

Unsupervised clustering of Lagrangian trajectories in the Labrador Current 

Noémie Planat and Mathilde Jutras

Lagrangian studies are a widely-used and powerful way to analyse and interpret phenomenons in oceanography and atmospheric sciences. Such studies can be based on dataset either consisting of real trajectories (e.g. oceanic drifters or floats) or of virtual trajectories computed from velocity outputs from model or observation-derived velocities. Such data can help investigate pathways of water masses, pollutants or storms, or identify important convection areas to name a few. As many of these analyses are based on large volumes of data that can be challenging to examine, machine learning can provide an efficient and automated way to classify information or detect patterns.

Here, we present an application of unsupervised clustering to the identification of the main pathways of the shelf-break branch of the Labrador Current, a critical component of the North Atlantic circulation. The current flows southward along the Labrador Shelf and splits in the region of the Grand Banks, either retroflecting north-eastward and feeding the subpolar basin of the North Atlantic Ocean (SPNA) or continuing westward along the shelf-break, feeding the Slope Sea and the east coast of North America. The proportion feeding each area impacts their salinity and convection, as well as their biogeochemistry, with consequences on marine life.

Our dataset is composed of millions of virtual particle trajectories computed from the water velocities of the GLORYS12 ocean reanalysis. We implement an unsupervised Machine Learning clustering algorithm on the shape of the trajectories. The algorithm is a kernalized k-means++ algorithm with a minimal number of hyperparameters, coupled to a kernalized Principal Component Analysis (PCA) features reduction. We will present the pre-processing of the data, as well as canonical and physics-based methods for choosing the hyperparameters. 

The algorithm identifies six main pathways of the Labrador Current. Applying the resulting classification method to 25 years of ocean reanalysis, we quantify the relative importance of the six pathways in time and construct a retroflection index that is used to study the drivers of the retroflection variability. This study highlights the potential of such a simple clustering method for Lagrangian trajectory analysis in oceanography or in other climate applications.

How to cite: Planat, N. and Jutras, M.: Unsupervised clustering of Lagrangian trajectories in the Labrador Current, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11216,, 2022.

EGU22-11388 | Presentations | ITS2.7/AS5.2 | Highlight

Learning ENSO-related Principal Modes of Vegetation via a Granger-Causal Variational Autoencoder 

Gherardo Varando, Miguel-Ángel Fernández-Torres, and Gustau Camps-Valls

Tackling climate change needs to understand the complex phenomena occurring on the Planet. Discovering  teleconnection patterns is an essential part of the endeavor. Events like El Niño Southern Oscillation (ENSO) impact essential climate variables at large distances, and influence the underlying Earth system dynamics. However, their automatic identification from the wealth of observational data is still unresolved. Nonlinearities, nonstationarities and the (ab)use of correlation analyses hamper the discovery of true causal patterns.  Classical approaches proceed by first, extracting principal modes of variability and second, by performing lag-correlations or Granger causal analysis to identify possible teleconnections. While the principal modes are an effective representation of the data, they could be causally not meaningful. 
To address this, we here introduce a deep learning methodology that extracts nonlinear latent representations from spatio-temporal Earth data that are Granger causal with the index altogether. The proposed algorithm consists of a variational autoencoder trained with an additional causal penalization that enforces the latent representation to be (partially) Granger-causally related to the considered signal. The causal loss term is obtained by training two additional autoregressive models to forecast some of the latent signals, one of them including the target signal as predictor. The causal penalization is finally computed by comparing the log variances of the two autoregressive models, similarly to the standard Granger causality approach. 

The major drawback of deep autoencoders with respect to the classical linear principal component approaches is the lack of a straightforward interpretability of the representations learned. 
To address this point we perform synthetic interventions in the latent space and analyse the differences in the recovered NDVI signal.
We illustrate the feasibility of the approach described to study the impact of ENSO on vegetation, which allows for a more rigorous study of impacts on ecosystems globally. The output maps show NDVI patterns which are consistent with the known phenomena induced by El Niño event. 

How to cite: Varando, G., Fernández-Torres, M.-Á., and Camps-Valls, G.: Learning ENSO-related Principal Modes of Vegetation via a Granger-Causal Variational Autoencoder, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11388,, 2022.

EGU22-11451 | Presentations | ITS2.7/AS5.2

Time evolution of temperature profiles retrieved from 13 years of IASI data using an artificial neural network 

Marie Bouillon, Sarah Safieddine, Simon Whitburn, Lieven Clarisse, Filipe Aires, Victor Pellet, Olivier Lezeaux, Noëlle A. Scott, Marie Doutriaux-Boucher, and Cathy Clerbaux

The IASI remote sensor measures Earth’s thermal infrared radiation over 8461 channels between 645 and 2760 cm-1. Atmospheric temperatures at different altitudes can be retrieved from the radiances measured in the CO2 absorption bands (645-800 cm-1 and 2250-2400 cm-1) by selecting the channels that are the most sensitive to the temperature profile. The three IASI instruments on board of the Metop suite of satellites launched in 2006, 2012 and 2018, will provide a long time series for temperature, adequate for studying the long term evolution of atmospheric temperature. However, over the past 14 years, EUMETSAT, who processes radiances and computes atmospheric temperatures, has carried out several updates on the processing algorithms for both radiances and temperatures, leading to non-homogeneous time series and thus large difficulties in the computation of trends for temperature and atmospheric composition.


In 2018, EUMETSAT has reprocessed the radiances with the most recent version of the algorithm and there is now a homogeneous radiance dataset available. In this study, we retrieve a new temperature record from the homogeneous IASI radiances using an artificial neural network (ANN). We train the ANN with IASI radiances as input and the European Centre for Medium-Range Weather Forecasts reanalysis ERA5 temperatures as output. We validate the results using ERA5 and in situ radiosonde temperatures from the ARSA database. Between 750 and 7 hPa, where IASI has most of its sensitivity, a very good agreement is observed between the 3 datasets. This work suggests that ANN can be a simple yet powerful tool to retrieve IASI temperatures at different altitudes in the upper troposphere and in the stratosphere, allowing us to construct a homogeneous and consistent temperature data record.


We use this new dataset to study extreme events such as sudden stratospheric warmings, and to compute trends over the IASI coverage period [2008-2020]. We find that in the past thirteen years, there is a general warming trend of the troposphere, that is more important at the poles and at mid latitudes (0.5 K/decade at mid latitudes, 1 K/decade at the North Pole). The stratosphere is globally cooling on average, except at the South Pole as a result of the ozone layer recovery and a sudden stratospheric warming in 2019. The cooling is most pronounced in the equatorial upper stratosphere (-1 K/decade).

How to cite: Bouillon, M., Safieddine, S., Whitburn, S., Clarisse, L., Aires, F., Pellet, V., Lezeaux, O., Scott, N. A., Doutriaux-Boucher, M., and Clerbaux, C.: Time evolution of temperature profiles retrieved from 13 years of IASI data using an artificial neural network, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11451,, 2022.

Existing databases for extreme weather events such as floods, heavy rainfall events, or droughts are heavily reliant on authorities and weather services manually entering details about the occurrence of an event. This reliance has led to a massive geographical imbalance in the likelihood of extreme weather events being recorded, with a vast number of events especially in the developing world remaining unrecorded. With continuing climate change, a lack of systematic extreme weather accounting in developing countries can lead to a substantial misallocation of funds for adaptation measures. To address this imbalance, in this pilot study we combine socio-economic data with climate and geographic data and use several machine-learning algorithms as well as traditional (spatial) econometric tools to predict the occurrence of extreme weather events and their impacts in the absence of information from manual records. Our preliminary results indicate that machine-learning approaches for the detection of the impacts of extreme weather could be a crucial tool in establishing a coherent global disaster record system. Such systems could also play a role in discussions around future Loss and Damages.

How to cite: Schwarz, M. and Pretis, F.: Filling in the Gaps: Consistently detecting previously unidentified extreme weather event impacts, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12165,, 2022.

EGU22-12720 | Presentations | ITS2.7/AS5.2 | Highlight

Interpretable Deep Learning for Probabilistic MJO Prediction 

Hannah Christensen and Antoine Delaunay

The Madden–Julian Oscillation (MJO) is the dominant source of sub-seasonal variability in the tropics. It consists of an Eastward moving region of enhanced convection coupled to changes in zonal winds. It is not possible to predict the precise evolution of the MJO, so subseasonal forecasts are generally probabilistic. Ideally the spread of the forecast probability distribution would vary from day to day depending on the instantaneous predictability of the MJO. Operational subseasonal forecasting models do not have this property. We present a deep convolutional neural network that produces skilful state-dependent probabilistic MJO forecasts. This statistical model accounts for intrinsic chaotic uncertainty by predicting the standard deviation about the mean, and model uncertainty using a Monte-Carlo dropout approach. Interpretation of the mean forecasts from the neural network highlights known MJO mechanisms, providing confidence in the model, while interpretation of the predicted uncertainty indicates new physical mechanisms governing MJO predictability.

How to cite: Christensen, H. and Delaunay, A.: Interpretable Deep Learning for Probabilistic MJO Prediction, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12720,, 2022.

EGU22-12822 | Presentations | ITS2.7/AS5.2

Assessing model dependency in CMIP5 and CMIP6 based on their spatial dependency structure with probabilistic network models 

Catharina Elisabeth Graafland and Jose Manuel Gutiérrez Gutiérrez

Probabilistic network models (PNMs) are well established data-driven modeling and machine learning prediction techniques used in many disciplines, including climate analysis. These techniques can efficiently learn the underlying (spatial) dependency structure and a consistent probabilistic model from data (e.g. gridded reanalysis or GCM outputs for particular variables; near surface temperature in this work), thus constituting a truly probabilistic backbone of the system underlying the data. The complex structure of the dataset is encoded using both pairwise and conditional dependencies and can be explored and characterized using network and probabilistic metrics. When applied to climate data, it is shown that Bayesian networks faithfully reveal the various long‐range teleconnections relevant in the dataset, in particular those emerging in el niño periods (Graafland, 2020).


In this work we apply probabilistic Gaussian networks to extract and characterize most essential spatial dependencies of the simulations generated by the different GCMs contributing to CMIP5 and 6 (Eyring 2016). In particular we analyze the problem of model interdependency (Boe, 2018) which poses practical problems for the application of these multi-model simulations in practical applications (it is often not clear what exactly makes one model different from or similar to another model).  We show that probabilistic Gaussian networks provide a promising tool to characterize the spatial structure of GCMs using simple metrics which can be used to analyze how and where differences in dependency structures are manifested. The probabilistic distance measure allows to chart CMIP5 and CMIP6 models on their closeness to reanalysis datasets that rely on observations. The measures also identifies significant atmospheric model changes that underwent CMIP5 GCMs in their transition to CMIP6. 




Boé, J. Interdependency in Multimodel Climate Projections: Component Replication and Result Similarity. Geophys. Res. Lett. 45, 2771–2779, DOI: 10.1002/2017GL076829 (2018).


Eyring, V. et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model. Dev. 9, 1937–1958, DOI: 10.5194/gmd-9-1937-2016  (2016).


Graafland, C.E., Gutiérrez, J.M., López, J.M. et al. The probabilistic backbone of data-driven complex networks: an example in climate. Sci Rep 10, 11484 (2020). DOI: 10.1038/s41598-020-67970-y



The authors would like to acknowledge project ATLAS (PID2019-111481RB-I00) funded by MCIN/AEI (doi:10.13039/501100011033). We also acknowledge support from Universidad de Cantabria and Consejería de Universidades, Igualdad, Cultura y Deporte del Gobierno de Cantabria via the “instrumentación y ciencia de datos para sondear la naturaleza del universo” project for funding this work. L.G. acknowledges support from the Spanish Agencia Estatal de Investigación through the Unidad de Excelencia María de Maeztu with reference MDM-2017-0765.

How to cite: Graafland, C. E. and Gutiérrez, J. M. G.: Assessing model dependency in CMIP5 and CMIP6 based on their spatial dependency structure with probabilistic network models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12822,, 2022.

EGU22-12858 | Presentations | ITS2.7/AS5.2

Identifying drivers of extreme reductions in carbon uptake of forests with interpretable machine learning 

Mohit Anand, Gustau Camps-Valls, and Jakob Zscheischler

Forests form one of the major components of the carbon cycle and take up large amounts of carbon dioxide from the atmosphere, thereby slowing down the rate of climate change. Carbon uptake by forests is a highly complex process strongly controlled by meteorological forcing, mainly because of two reasons. First, forests have a large storage capacity acting as a buffer to short-duration changes in meteorological drivers. The response can thus be very complex and extend over a long time. Secondly, the responses are often triggered by combinations of multiple compounding drivers including precipitation, temperature and solar radiation. Effects may compound between variables and across time. Therefore, a large amount of data is required to identify the complex drivers of adverse forest response to climate forcing. Recent advances in machine learning offer a suite of promising tools to analyse large amounts of data and address the challenge of identifying complex drivers of impacts. Here we analyse the potential of machine learning to identify the compounding drivers of reduced carbon uptake/forest mortality. To this end, we generate 200,000 years of gross and net carbon uptake from the physically-based forest model FORMIND simulating a beech forest in Germany. The climate data is generated through a weather generator (AWEGEN-1D) from bias-corrected ERA5 reanalysis data.  Classical machine learning models like random forest, support vector machines and deep neural networks are trained to estimate gross primary product. Deep learning models involving convolutional layers are found to perform better than the other classical machine learning models. Initial results show that at least three years of weather data are required to predict annual carbon uptake with high accuracy, highlighting the complex lagged effects that characterize forests. We assess the performance of the different models and discuss their interpretability regarding the identification of impact drivers.

How to cite: Anand, M., Camps-Valls, G., and Zscheischler, J.: Identifying drivers of extreme reductions in carbon uptake of forests with interpretable machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12858,, 2022.

EGU22-13345 | Presentations | ITS2.7/AS5.2

A novel approach to systematically analyze the error structure of precipitation datasets using decision trees 

Xinxin Sui, Zhi Li, Guoqiang Tang, Zong-Liang Yang, and Dev Niyogi
Multiple environmental factors influence the error structure of precipitation datasets. The conventional precipitation evaluation method over-simply analyzes how the statistical indicators vary with one or two factors via dimensionality reduction. As a result, the compound influences of multiple factors are superposed rather than disassembled. To overcome this deficiency, this study presents a novel approach to systematically and objectively analyze the error structure within precipitation products using decision trees. This data-driven method can analyze multiple factors simultaneously and extract the compound effects of various influencers. By interpreting the decision tree structures, the error characteristics of precipitation products are investigated. Three types of precipitation products (two satellite-based: ‘top-down’ IMERG and ‘bottom-up’ SM2RAIN-ASCAT, and one reanalysis: ERA5-Land) are evaluated across CONUS. The study period is from 2010 to 2019, and the ground-based Stage IV precipitation dataset is used as the ground truth. By data mining 60 binary decision trees, the spatiotemporal pattern of errors and the land surface influences are analyzed.
Results indicate that IMERG and ERA5-Land perform better than SM2RAIN-ASCAT with higher accuracy and more stable interannual patterns for the ten years of data analyzed. The conventional bias evaluation finds that ERA5-Land and SM2RAIN-ASCAT underestimate in summer and winter, respectively. The decision tree method cross-assesses three spatiotemporal factors and finds that underestimation of ERA5-Land occurs in the eastern part of the rocky mountains, and SM2RAIN-ASCAT underestimates precipitation over high latitudes, especially in winter. Additionally, the decision tree method ascribes system errors to nine physical variables, of which the distance to the coast, soil type, and DEM are the three dominant features. On the other hand, the land cover classification and the topography position index are two relatively weak factors.

How to cite: Sui, X., Li, Z., Tang, G., Yang, Z.-L., and Niyogi, D.: A novel approach to systematically analyze the error structure of precipitation datasets using decision trees, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-13345,, 2022.

ITS3 – A new era of Earth and planetary observation: instrumentation for what and to whom?

EGU22-2024 | Presentations | ITS3.1/SSS1.2 | Highlight

Understanding natural hazards in a changing landscape: A citizen science approach in Kigezi highlands, southwestern Uganda 

Violet Kanyiginya, Ronald Twongyirwe, Grace Kagoro, David Mubiru, Matthieu Kervyn, and Olivier Dewitte

The Kigezi highlands, southwestern Uganda, is a mountainous tropical region with a high population density, intense rainfall, alternating wet and dry seasons and high weathering rates. As a result, the region is regularly affected by multiple natural hazards such as landslides, floods, heavy storms, and earthquakes. In addition, deforestation and land use changes are assumed to have an influence on the patterns of natural hazards and their impacts in the region. Landscape characteristics and dynamics controlling the occurrence and the spatio-temporal distribution of natural hazards in the region remain poorly understood. In this study, citizen science has been employed to document and understand the spatial and temporal occurrence of natural hazards that affect the Kigezi highlands in relation to the multi-decadal landscape change of the region. We present the methodological research framework involving three categories of participatory citizen scientists. First, a network of 15 geo-observers (i.e., citizens of local communities distributed across representative landscapes of the study area) was established in December 2019. The geo-observers were trained at using smartphones to collect information (processes and impacts) on eight different natural hazards occurring across their parishes. In a second phase, eight river watchers were selected at watershed level to monitor the stream flow characteristics. These watchers record stream water levels once daily and make flood observations. In both categories, validation and quality checks are done on the collected data for further analysis. Combining with high resolution rainfall monitoring using rain gauges installed in the watersheds, the data are expected to characterize catchment response to flash floods. Lastly, to reconstruct the historical landscape change and natural hazards occurrences in the region, 96 elderly citizens (>70 years of age) were engaged through interviews and focus group discussions to give an account of the evolution of their landscape over the past 60 years. We constructed a historical timeline for the region to complement the participatory mapping and in-depth interviews with the elderly citizens. During the first 24 months of the project, 240 natural hazard events with accurate timing information have been reported by the geo-observers. Conversion from natural tree species to exotic species, increased cultivation of hillslopes, road construction and abandonment of terraces and fallowing practices have accelerated natural hazards especially flash floods and landslides in the region. Complementing with the region’s historical photos of 1954 and satellite images, major landscape dynamics have been detected. The ongoing data collection involving detailed ground-based observations with citizens shows a promising trend in the generation of new knowledge about natural hazards in the region.

How to cite: Kanyiginya, V., Twongyirwe, R., Kagoro, G., Mubiru, D., Kervyn, M., and Dewitte, O.: Understanding natural hazards in a changing landscape: A citizen science approach in Kigezi highlands, southwestern Uganda, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2024,, 2022.

EGU22-2929 | Presentations | ITS3.1/SSS1.2

Possible Contributions of Citizen Science in the Development of the Next Generation of City Climate Services 

Peter Dietrich, Uta Ködel, Sophia Schütze, Felix Schmidt, Fabian Schütze, Aletta Bonn, Thora Herrmann, and Claudia Schütze

Human life in cities is already affected by climate change. The effects will become even more pronounced in the coming years and decades. Next-generation of city climate services is necessary for adapting infrastructures and the management of services of cities to climate change. These services are based on advanced weather forecast models and the access to diverse data. It is essential to keep in mind that each citizen is a unique individual with their own peculiarities, preferences, and behaviors. The base for our approach is the individual specific exposure, which considers that people perceive the same conditions differently in terms of their well-being. Individual specific exposure can be defined as the sum of all environmental conditions that affect humans during a given period of time, in a specific location, and in a specific context. Thereby, measurable abiotic parameters such as temperature, humidity, wind speed, pollution and noise are used to characterize the environmental conditions. Additional information regarding green spaces, trees, parks, kinds of streets and buildings, as well as available infrastructures are included in the context. The recording and forecasting of environmental parameters while taking into account the context, as well as the presentation of this information in easy-to-understand and easy-to-use maps, are critical for influencing human behavior and implementing appropriate climate change adaptation measures.

We will adopt this approach within the frame of the recently started, EU-funded CityCLIM project. We aim to develop and implement approaches which will explore the potential of citizen science in terms of current and historical data collecting, data quality assessment and evaluation of data products.  In addition, our approach will also provide strategies for individual climate data use, and the derivation and evaluation of climate change adaptation actions in cities.

In a first step we need to define and to characterize the different potential stakeholder groups involved in citizen science data collection. Citizen science offers approaches that consider citizens as both  organized target groups (e.g., engaged companies, schools) and individual persons (e.g. hobby scientists). An important point to be investigated is the motivation of citizen science stakehoder groups to sustainably collect data and make it available to science and reward them accordingly. For that purpose, strategic tools, such as value proposition canvas analysis, will be applied to taylor the science-to-business and the science-to-customer communications and offers in terms of the individual needs.

How to cite: Dietrich, P., Ködel, U., Schütze, S., Schmidt, F., Schütze, F., Bonn, A., Herrmann, T., and Schütze, C.: Possible Contributions of Citizen Science in the Development of the Next Generation of City Climate Services, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2929,, 2022.

EGU22-4168 | Presentations | ITS3.1/SSS1.2

Extending Rapid Image Classification with the Picture Pile Platform for Citizen Science 

Tobias Sturn, Linda See, Steffen Fritz, Santosh Karanam, and Ian McCallum

Picture Pile is a flexible web-based and mobile application for ingesting imagery from satellites, orthophotos, unmanned aerial vehicles and/or geotagged photographs for rapid classification by volunteers. Since 2014, there have been 16 different crowdsourcing campaigns run with Picture Pile, which has involved more than 4000 volunteers who have classified around 11.5 million images. Picture Pile is based on a simple mechanic in which users view an image and then answer a question, e.g., do you see oil palm, with a simple yes, no or maybe answer by swiping the image to the right, left or downwards, respectively. More recently, Picture Pile has been modified to classify data into categories (e.g., crop types) as well as continuous variables (e.g., degree of wealth) so that additional types of data can be collected.

The Picture Pile campaigns have covered a range of domains from classification of deforestation to building damage to different types of land cover, with crop type identification as the latest ongoing campaign through the Earth Challenge network. Hence, Picture Pile can be used for many different types of applications that need image classifications, e.g., as reference data for training remote sensing algorithms, validation of remotely sensed products or training data of computer vision algorithms. Picture Pile also has potential for monitoring some of the indicators of the United Nations Sustainable Development Goals (SDGs). The Picture Pile Platform is the next generation of the Picture Pile application, which will allow any user to create their own ‘piles’ of imagery and run their own campaigns using the system. In addition to providing an overview of Picture Pile, including some examples of relevance to SDG monitoring, this presentation will provide an overview of the current status of the Picture Pile Platform along with the data sharing model, the machine learning component and the vision for how the platform will function operationally to aid environmental monitoring.

How to cite: Sturn, T., See, L., Fritz, S., Karanam, S., and McCallum, I.: Extending Rapid Image Classification with the Picture Pile Platform for Citizen Science, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4168,, 2022.

EGU22-5094 | Presentations | ITS3.1/SSS1.2

Life in undies – Preliminary results of a citizen science data collection targeting soil health assessement in Hungary 

Mátyás Árvai, Péter László, Tünde Takáts, Zsófia Adrienn Kovács, Kata Takács, János Mészaros, and László Pásztor

Last year, the Institute for Soil Sciences, Centre for Agricultural Research launched Hungary's first citizen science project with the aim to obtain information on the biological activity of soils using a simple estimation procedure. With the help of social media, the reactions on the call for applications were received from nearly 2000 locations. 

In the Hungarian version of the international Soil your Undies programme, standardized cotton underwear was posted to the participants with a step-by-step tutorial, who buried their underwear for about 60 days, from mid of May until July in 2021, at a depth of about 20-25 cm. After the excavation, the participants took one digital image of the underwear and recorded the geographical coordinates, which were  uploaded to a GoogleForms interface together with several basic information related to the location and the user (type of cultivation, demographic data etc.).

By analysing digital photos of the excavated undies made by volunteers, we obtained information on the level to which cotton material had decomposed in certain areas and under different types of cultivation. Around 40% of the participants buried the underwear in garden, 21% in grassland, 15% in orchard, 12% in arable land, 5% in vineyard and 4% in forest (for 3% no landuse data was provided).

The images were first processed using Fococlipping and Photoroom softwares for background removing and then percentage of cotton material remaining was estimated based on the pixels by using R Studio ‘raster package’.

The countrywide collected biological activity data from nearly 1200 sites were statistically evaluated by spatially aggregating the data both for physiographical and administrative units. The results have been published on various platforms (Facebook, Instagram, specific web site etc.), and a feedback is also given directly to the volunteers.

According to the experiments the first citizen science programme proved to be successful. 


Acknowledgment: Our research was supported by the Hungarian National Research, Development and Innovation Office (NKFIH; K-131820)

Keywords: citizen science; soil life; soil health; biological activity; soil properties

How to cite: Árvai, M., László, P., Tak