Understanding long-term total water storage (TWS) changes in the Yangtze River Basin (YRB) is essential for optimizing water resource management and mitigating hydrological extremes. While the Gravity Recovery and Climate Experiment (GRACE) and its follow-on (GRACE-FO) mission have provided valuable observations for investigating global or regional TWS changes, the approximately one-year data gap between these missions and their relatively short 20-year data record limits our ability to study the continuous and long-term variability of YRB's TWS. In this study, two deep learning models are employed to bridge the data gap and reconstruct the historical TWS changes within YRB, respectively. For the data gap filling task, a noise-augmented u-shaped network (NA-UNet) is presented to address UNet's overfitting issues associated with training on limited GRACE observations. Results show that NA-UNet can accurately bridge the data gap, exhibiting favourable and stable performance at both the basin and grid scales. Subsequently, we introduce another deep learning model named RecNet, specifically designed to reconstruct the climate-driven TWS changes in YRB from 1923 to 2022. RecNet is trained on precipitation, temperature, and GRACE observations using a weighted mean square error (WMSE) loss function. We show that RecNet can successfully reconstruct the historical TWS changes, achieving strong correlations with GRACE, water budget estimates, hydrological models, drought indices, and existing reconstruction datasets. We also observe superior performance in RecNet when trained with WMSE compared to its non-weighted counterpart. In addition, the reconstructed datasets reveal a recurring occurrence of diverse hydrological extremes over the past century within YRB, influenced by major climate patterns. Together, NA-UNet and RecNet provide valuable observations for studying long-term climate variability and projecting future hydrological extremes in YRB, which can inform effective water resource management and contribute to the development of adaptive strategies for climate change.

How to cite: Wang, J., Shen, Y., Awange, J., Yang, L., and Chen, Q.: Reconstructing total water storage changes in the Yangtze River Basin based on deep learning models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3272, https://doi.org/10.5194/egusphere-egu24-3272, 2024.

This study introduces an approach for land heatwave forecasting, using spatiotemporal machine learning models trained with ERA5 reanalysis data. We focused on key environmental variables like soil moisture, vegetation, and meteorological factors for modelling. The study utilized linear regression as a base model, augmented by more complex algorithms such as Random Forest (RF), XGBoost, and Graph Neural Networks (GNN). We defined heatwaves using temperature data from 1970-2000, and the training phase involved data from 2000 to 2020, focusing on predictive accuracy for 2021-2023. This methodology enabled a detailed exploration of heatwave trends and dynamics over an extended period. Finally, we used explainable AI methods to further deepen our understanding of the complex interplay between environmental variables and heatwave occurrences.

How to cite: Mazumder, S., Hahn, S., and Wagner, W.: Monitoring The Development Of Land Heatwaves Using Spatiotemporal Models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1101, https://doi.org/10.5194/egusphere-egu24-1101, 2024.

The biogeoscience community has increasingly embraced the application of machine learning models across various domains from fire prediction to vegetation forecasting. Yet, as these models become more widely used, there is sometimes a gap in understanding between what we assume the model learns and what the model actually learns. For example, Long-short Term Memory (LSTM) models are applied to long time series, hoping they benefit from access to more information, despite their tendency to rapidly forget information. This can lead to erroneous conclusions, misinterpretation of results, and an overestimation of the models, ultimately eroding trust in their reliability. 

To address this issue, we employ an explainable artificial intelligence (XAI) post hoc perturbation technique that is task-agnostic and model-agnostic. We aim to examine the extent to which the model leverages information for its predictions, both in terms of time and space. In other words, we want to observe the actual receptive field utilized by the model. We introduce a methodology designed to quantify both the spatial impact of neighboring pixels on predicting a specific pixel and the temporal periods contributing to predictions in time series models. The experiments take place after training the model, during inference. In the spatial domain, we define ground-truth pixels to predict, then examine the increase in prediction error, caused by shuffling their neighboring pixels at various distances from the selection. In the temporal domain, we investigate how shuffling a sequence of frames within the context period at different intervals relative to the target period affects the increase in prediction loss. This method can be applied across a broad spectrum of spatio-temporal tasks. Importantly, the method is easy-to-implement, as it only relies on the inference of predictions at test time and the shuffling of the perturbation area. 

For our experiments, we focus on the vegetation forecasting task, i.e., forecasting the evolution of the Vegetation Index (VI) based on Sentinel-2 imagery using previous Sentinel-2 sequences and weather information to guide the prediction. This task involves both spatial non-linear dependencies arising from the spatial context (e.g., the surrounding area, such as a river or a slope, directly influencing the VI) and non-linear temporal dependencies such as the gradual onset of drought conditions and the rapid influence of precipitation events. We compare several models for spatio-temporal tasks, including ConvLSTM and transformer-based architectures on their usage of neighboring pixels in space, and context period in time. We demonstrate that the ConvLSTM relies on a  restricted spatial area in its predictions, indicating a limited utilization of the spatial context up to 50m (5 pixels). Furthermore, it utilizes the global order of the time series sequence to capture the seasonal cycle but loses sensitivity to the local order after 15 days (3 frames). The introduced XAI method allows us to quantify spatial and temporal behavior exhibited by machine learning methods.

How to cite: Robin, C., Benson, V., Requena-Mesa, C., Alonso, L., Poehls, J., Russwurm, M., Carvalhais, N., and Reichstein, M.: Analyzing Spatio-Temporal Machine Learning Models through Input Perturbation, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17389, https://doi.org/10.5194/egusphere-egu24-17389, 2024.

In recent years, the intersection of machine learning (ML) and climate science has yielded profound insights into understanding and predicting extreme climate events, particularly heatwaves and droughts. Various approaches have been suggested to define and model extreme events, including extreme value theory (Sura, 2011), random forests (e.g., (Weirich-Benet et al., 2023) and, more recently, deep learning (e.g., (Jacques-Dumas et al., 2022)). Within this context, quantile regression (QR) is valuable for modelling the relationship between variables by estimating the conditional quantiles of the response variable. This provides insights into the entire distribution rather than just the mean but also aids in unravelling the complex relationships among climate variables (Barbosa et al., 2011; Franzke, 2015). QR has been extended in many ways to address critical issues such as nonlinear relations, nonstationary processes, compound events, and the complexities of handling spatio-temporal data. 

This study presents a novel approach for predicting and better understanding heatwaves. We introduce an interpretable, nonlinear, non-parametric, and structured Spatio-Temporal Quantile Regression (STQR) method that incorporates the QR check function, commonly known as pinball loss, into machine learning models. We focus on analysing how the importance of predictors changes as the quantile being modelled increases. This allows us to circumvent arbitrary definitions of what constitutes a heatwave and instead observe if a natural definition of a heatwave emerges in predictor space. By analysing European heatwaves over recent decades using reanalysis and weather data, we demonstrate the advantages of our methodology over traditional extreme event modelling methods.

References

Barbosa, S.M., Scotto, M.G., Alonso, A.M., 2011. Summarising changes in air temperature over Central Europe by quantile regression and clustering. Nat. Hazards Earth Syst. Sci. 11, 3227–3233. https://doi.org/10.5194/nhess-11-3227-2011

Franzke, C.L.E., 2015. Local trend disparities of European minimum and maximum temperature extremes. Geophys. Res. Lett. 42, 6479–6484. https://doi.org/10.1002/2015GL065011

Jacques-Dumas, V., Ragone, F., Borgnat, P., Abry, P., Bouchet, F., 2022. Deep Learning-based Extreme Heatwave Forecast. Front. Clim. 4, 789641. https://doi.org/10.3389/fclim.2022.789641

Sura, P., 2011. A general perspective of extreme events in weather and climate. Atmospheric Res. 101, 1–21. https://doi.org/10.1016/j.atmosres.2011.01.012

Weirich-Benet, E., Pyrina, M., Jiménez-Esteve, B., Fraenkel, E., Cohen, J., Domeisen, D.I.V., 2023. Subseasonal Prediction of Central European Summer Heatwaves with Linear and Random Forest Machine Learning Models. Artif. Intell. Earth Syst. 2. https://doi.org/10.1175/AIES-D-22-0038.1

How to cite: Bassotto, D., Diaz, E., and Camps-Valls, G.: Spatio-temporal Nonlinear Quantile Regression for Heatwave Prediction and Understanding, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17694, https://doi.org/10.5194/egusphere-egu24-17694, 2024.

Earth observation (EO) repositories comprise Petabytes of data. Due to their widespread use, these repositories experience extremely large volumes of data transfers. For example, users of the Sentinel Data Access System downloaded 78.6 PiB of data in 2022 alone. The transfer of such data volumes between data producers and consumers causes substantial latency and requires significant amounts of energy and vast storage capacities. This work introduces Neural Embedding Compression (NEC), a method that transmits compressed embeddings to users instead of raw data, greatly reducing transfer and storage costs. The approach uses general purpose embeddings from Foundation Models (FM), which can serve multiple downstream tasks and neural compression, which balances between compression rate and the utility of the embeddings. We implemented the method by updating a minor portion of the FM’s parameters (approximately 10%) for a short training period of about 1% of the original pre-training iterations. NEC’s effectiveness is assessed through two EO tasks: scene classification and semantic segmentation. When compared to traditional compression methods applied to raw data, NEC maintains similar accuracy levels while reducing data by 75% to 90%. Notably, even with a compression rate of 99.7%, there’s only a 5% decrease in accuracy for scene classification. In summary, NEC offers a resource-efficient yet effective solution for multi-task EO modeling with minimal transfer of data volumes.

How to cite: Gomes, C. and Brunschwiler, T.: Earth Observation Applications through Neural Embedding Compression from Foundation Models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19460, https://doi.org/10.5194/egusphere-egu24-19460, 2024.

The application of automatic differentiation and deep learning approaches to tackle current challenges is now a widespread practice. The biogeosciences community is no stranger to this trend; however, quite often, previously known physical model abstractions are discarded.

In this study, we model the ecosystem dynamics of vegetation, water, and carbon cycles adopting a hybrid approach. This methodology involves preserving the physical model representations for simulating the targeted processes while utilizing neural networks to learn the spatial variability of their parameters. These models have historically posed challenges due to their complex process representations, varied spatial scales, and parametrizations.

We show that a hybrid approach effectively predicts model parameters with a single neural network, compared with the site-level optimized set of parameters. This approach demonstrates its capability to generate predictions consistent with in-situ parameter calibrations across various spatial locations, showcasing its versatility and reliability in modelling coupled systems.
Here, the physics-based process models undergo evaluation across several FLUXNET sites. Various observations—such as gross primary productivity, net ecosystem exchange, evapotranspiration, transpiration, the normalized difference vegetation index, above-ground biomass, and ecosystem respiration—are utilized as targets to assess the model's performance. Simultaneously, a neural network (NN) is trained to predict the model parameters, using input features(to the NN) such as plant functional types, climate types, bioclimatic variables, atmospheric nitrogen and phosphorus deposition, and soil properties. The model simulation is executed within our internal framework Sindbad.jl (to be open-sourced), designed to ensure compatibility with gradient-based optimization methods.

This work serves as a stepping stone, demonstrating that incorporating neural networks into a broad collection of physics-based models holds significant promise and has the potential to leverage the abundance of current Earth observations, enabling the application of these methods on a larger scale.

How to cite: Alonso, L., Koirala, S., Carvalhais, N., Gans, F., Ahrens, B., Cremer, F., Wutzler, T., Ayoub Chettouh, M., and Reichstein, M.: Hybrid Modelling: Bridging Neural Networks and Physics-Based Approaches in Terrestrial Biogeochemical Ecosystems, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16513, https://doi.org/10.5194/egusphere-egu24-16513, 2024.

The exchange of water and carbon between the land-surface and the atmosphere is regulated by meteorological conditions as well as plant physiological processes. Accurate modeling of the coupled system is not only crucial for understanding local feedback loops but also for global-scale carbon and water cycle interactions. Traditional mechanistic modeling approaches, e.g., the Earth system model ICON-ESM with the land component JSBACH4, have long been used to study the land-atmosphere coupling. However, these models are hampered by relatively rigid functional representations of terrestrial biospheric processes, e.g., semi-empirical parametrizations for stomatal conductance.

Here, we develop data-driven, flexible parametrizations controlling terrestrial carbon-water coupling based on eddy-covariance flux measurements using machine learning (ML). Specifically, we introduce a hybrid modeling approach (integration of data-driven and mechanistic modeling), that aims to replace specific empirical parametrizations of the coupled photosynthesis (GPP ) and transpiration (E_tr ) modules with ML models pre-trained on observations. First, as a proof-of-concept, we train parametrizations based on original JSBACH4 output to showcase that our approach succeeds in reconstructing the original parametrizations, namely latent dynamic features for stomatal (g_s) and aerodynamic (g_a) conductance, the carboxylation rate of RuBisCO (V_cmax), and the photosynthetic electron transport rate for RuBisCO regeneration (J_max). Second, we replace JSBACH4’s original parametrizations by dynamically calling the emulator parameterizations trained on the original JSBACH4 output using a Python-FORTRAN bridge. This allows us to assess the impact of data-driven parametrizations on the output in the coupled land-surface model. In the last step, we adopt the approach to infer these parametrizations from FLUXNET observations to construct an observation-informed model of water and carbon fluxes in JSBACH4.

Preliminary results in emulating JSBACH4 parametrizations reveal R² ranging between 0.91-0.99 and 0.92-0.97 for GPP, E_tr, and the sensible heat flux Q_H  at half-hourly scale for forest and grassland sites, respectively. JSBACH4 with the plugged-in ML-emulator parametrizations provides very similar, but not identical predictions as the original JSBACH4. For example, R² for E_tr (g_s) amounts to 0.91 (0.84) and 0.93 (0.86) at grassland and forest sites, respectively. These differences in the transpiration flux between original predictions and JSBACH4 with emulating parametrizations only result in minor changes in the system, e.g., the soil-water budget in the two models is almost the same (R² of ~0.99). Based on these promising results of our proof-of-concept, we are now preparing the hybrid JSBACH4 model with parametrizations trained on FLUXNET observations.

This modeling framework will then serve as the foundation for coupled land-atmosphere simulations using ICON-ESM, where key biospheric processes are represented by our hybrid observation-informed land-surface model.

How to cite: ElGhawi, R., Reimers, C., Schnur, R., Reichstein, M., Körner, M., Carvalhais, N., and Winkler, A. J.: Hybrid-Modeling of Land-Atmosphere Fluxes Using Integrated Machine Learning in the ICON-ESM Modeling Framework, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2819, https://doi.org/10.5194/egusphere-egu24-2819, 2024.

Heatwaves over western Europe are increasing faster than elsewhere, which recent studies have attributed at least partly to changes in atmospheric dynamics. To increase our understanding of the dynamical drivers of western European heatwaves, we developed a heatwave classification method taking into account the spatio-temporal atmospheric dynamics. Our deep learning approach consists of several steps: 1) heatwave detection using the Generalized Density-based Spatial Clustering of Applications with Noise (GDBSCAN) algorithm; 2) dimensionality reduction of the spatio-temporal heatwave samples using a 3D Variational Autoencoder (VAE); and 3) a clustering of heatwaves using K-means, a Gaussian Mixture Model, and opt-SNE. We show that a VAE can extract meaningful features from high-dimensional climate data. Furthermore, we find four physically distinct clusters of heatwaves that are interpretable with known circulation patterns, i.e. UK High, Scandinavian High, Atlantic High, and Atlantic Low. Our results indicate that the heatwave phase space, as found with opt-SNE, is continuous with soft boundaries between these circulation regimes, indicating that heatwaves are best categorized in a probabilistic way.

How to cite: Happe, T., Wijnands, J., Fernández-Torres, M. Á., Scussolini, P., Muntjewerf, L., and Coumou, D.: Detecting spatio-temporal dynamics of western European heatwaves using deep learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6655, https://doi.org/10.5194/egusphere-egu24-6655, 2024.

Large deep neural network emulators are poised to revolutionize numerical weather prediction (NWP). Recent models like GraphCast or NeuralGCM can now compete and sometimes outperform traditional NWP systems, all at much lower computational cost. Yet to be explored is the applicability of large deep neural network emulators to other dense prediction tasks such as the modeling of 3D atmospheric composition. For instance the inverse modeling of carbon fluxes essential for estimating carbon budgets relies on fast CO₂ transport models.

Here, we present a novel approach to atmospheric transport modeling of CO₂ and other inert trace gases. Existing Eulerian transport modeling approaches rely on numerical solvers applied to the continuity equation, which are expensive: short time steps are required for numerical stability at the poles, and the loading of driving meteorological fields is IO-intensive. We learn high-fidelity transport in latent space by training graph neural networks, analogous to approaches used in weather forecasting, including an approach that conserves the CO₂ mass. For this, we prepare the CarbonBench dataset, a deep learning ready dataset based on Jena Carboscope CO₂ inversion data and NCEP NCAR meteorological reanalysis data together with ObsPack station observations for model evaluation.

Qualitative and quantitative experiments demonstrate the superior performance of our approach over a baseline U-Net for short-term (<40 days) atmospheric transport modeling of carbon dioxide. While the original GraphCast architecture achieves a similar speed to the TM3 transport model used to generate the training data, we show how various architectural changes introduced by us contribute to a reduced IO load (>4x) of our model, thereby speeding up forward runs. This is especially useful when applied multiple times with the same driving wind fields, e.g. in an inverse modeling framework. Thus, we pave the way towards integrating not only atmospheric observations (as is done in current CO₂ inversions), but also ecosystem surface fluxes (not yet done) into carbon cycle inversions. The latter requires backpropagating through a transport operator to optimize a flux model with many more parameters (e.g. a deep neural network) than those currently used in CO₂ inversions – which becomes feasible if the transport operator is fast enough. To the best of our knowledge, this work presents the first emulator of global Eulerian atmospheric transport, thereby providing an initial step towards next-gen inverse modeling of the carbon cycle with deep learning.

How to cite: Benson, V., Bastos, A., Reimers, C., Winkler, A. J., Yang, F., and Reichstein, M.: Graph Neural Networks for Atmospheric Transport Modeling of CO2 , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10922, https://doi.org/10.5194/egusphere-egu24-10922, 2024.

Studying low-likelihood high-impact extreme weather and climate events in a warming world requires massive
ensembles to capture long tails of multi-variate distributions. In combination, it is simply impossible to generate
massive ensembles, of say 10,000 members, using traditional numerical simulations of climate models at high
resolution. We describe how to bring the power of machine learning (ML) to replace traditional numerical
simulations for short week-long hindcasts of massive ensembles, where ML has proven to be successful in terms of
accuracy and fidelity, at five orders-of-magnitude lower computational cost than numerical methods. Because
the ensembles are reproducible to machine precision, ML also provides a data compression mechanism to
avoid storing the data produced from massive ensembles. The machine learning algorithm FourCastNet (FCN) is
based on Fourier Neural Operators and Transformers, proven to be efficient and powerful in modeling a wide
range of chaotic dynamical systems, including turbulent flows and atmospheric dynamics. FCN has already been
proven to be highly scalable on GPU-based HPC systems. 

We discuss our progress using statistics metrics for extremes adopted from operational NWP centers to show
that FCN is sufficiently accurate as an emulator of these phenomena. We also show how to construct huge
ensembles through a combination of perturbed-parameter techniques and a variant of bred vectors to generate a
large suite of initial conditions that maximize growth rates of ensemble spread. We demonstrate that these
ensembles exhibit a ratio of ensemble spread relative to RMSE that is nearly identical to one, a key metric of
successful near-term NWP systems. We conclude by applying FCN to severe heat waves in the recent climate
record.

How to cite: Collins, W., Pritchard, M., Brenowitz, N., Cohen, Y., Harrington, P., Kashinath, K., Mahesh, A., and Subramanian, S.: Huge Ensembles of Weather Extremes using the Fourier Forecasting Neural Network, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4460, https://doi.org/10.5194/egusphere-egu24-4460, 2024.

In recent years, deep learning models have rapidly emerged as a standalone alternative to physics-based numerical models for medium-range weather forecasting. Several independent research groups claim to have developed deep learning weather forecasts which outperform those from state-of-the-art physics-basics models, and operational implementation of data-driven forecasts appears to be drawing near. Yet, questions remain about the capabilities of deep learning models to provide robust forecasts of extreme weather.

Our current work aims to provide an overview of recent developments in the field of deep learning weather forecasting, and highlight the challenges that extreme weather events pose to leading deep learning models. Specifically, we problematise the fact that predictions generated by many deep learning models appear to be oversmooth, tending to underestimate the magnitude of wind and temperature extremes. To address these challenges, we argue for the need to tailor data-driven models to forecast extreme events, and develop models aiming to maximise the skill in the tails rather than in the mean of the distribution. Lastly, we propose a foundational workflow to develop robust models for extreme weather, which may function as a blueprint for future research on the topic.

How to cite: Olivetti, L. and Messori, G.: Advances and Prospects of Deep Learning for Medium-Range Extreme Weather Forecasting, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5611, https://doi.org/10.5194/egusphere-egu24-5611, 2024.

Subseasonal to seasonal (S2S) weather forecasts play an important role as a decision making tool in several sectors of modern society. However, the time scale on which these forecasts are skillful is strongly dependent on atmospheric and oceanic background conditions. While deep learning-based weather prediction models have shown impressive results in the short- to medium range, S2S forecasts from such models are currently limited, partly due to fewer available training data and larger fluctuations in predictability. In order to develop more reliable S2S predictions we leverage Masked Autoencoders, a state-of-the-art deep learning framework, to extract large-scale representations of tropical precipitation and sea-surface temperature data.  We show that the learned representations are highly predictive for the El Niño Southern Oscillation and the Madden-Julian Oscillation, and can thus serve as a foundation for identifying windows of opportunity and generating skillful S2S forecasts.

How to cite: Thümmel, J., Schlör, J., Strnad, F., and Goswami, B.: Subseasonal to seasonal forecasts using Masked Autoencoders, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10129, https://doi.org/10.5194/egusphere-egu24-10129, 2024.

In order to improve climate projections, machine learning (ML)-based parameterizations have been developed for Earth System Models (ESMs) with the goal to better represent subgrid-scale processes or to accelerate computations by emulating existent parameterizations. These data-driven models have shown success in approximating subgrid-scale processes based on high-resolution storm-resolving simulations. However, most studies have used a particular machine learning method such as simple Multilayer Perceptrons (MLPs) or Random Forest (RFs) to parameterize the subgrid tendencies or fluxes originating from the compound effect of various small-scale processes (e.g., turbulence, radiation, convection, gravity waves). Here, we use a filtering technique to explicitly separate convection from these processes in data produced by the Icosahedral Non-hydrostatic modelling framework (ICON) in a realistic setting. We use a method improved by incorporating density fluctuations for computing the subgrid fluxes and compare a variety of different machine learning algorithms on their ability to predict the subgrid fluxes. We further examine the predictions of the best performing non-deep learning model (Gradient Boosted Tree regression) and the U-Net. We discover that the U-Net can learn non-causal relations between convective precipitation and convective subgrid fluxes and develop an ablated model excluding precipitating tracer species. We connect the learned relations of the U-Net to physical processes in contrast to non-deep learning-based algorithms. Our results suggest that architectures such as a U-Net are particularly well suited to parameterize multiscale problems like convection, paying attention to the plausibility of the learned relations, thus providing a significant advance upon existing ML subgrid representation in ESMs.

How to cite: Heuer, H., Schwabe, M., Gentine, P., Giorgetta, M. A., and Eyring, V.: Interpretable multiscale Machine Learning-Based Parameterizations of Convection for ICON, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10325, https://doi.org/10.5194/egusphere-egu24-10325, 2024.

Combining process-based models in Earth system science with data-driven machine learning methods holds tremendous promise. Can we harness the best of both approaches? In our study, we integrate components of atmospheric models into artificial neural networks (ANN). The resulting hybrid atmospheric model can learn atmospheric dynamics from short trajectories while ensuring robust generalization and stability. We achieve this using the neural differential equations framework, combining ANNs with a differentiable, GPU-enabled version of the well-studied Marshall Molteni Quasigeostrophic Model (QG3). Similar to the approach of many atmospheric models, part of the model is computed in the spherical harmonics domain, and other parts in the grid domain. In our model, ANNs are used as parametrizations in both domains, and form together with the components of the QG3 model the right-hand side of our hybrid model. We showcase the capabilities of our model by demonstrating how it generalizes from the QG3 model to the significantly more complex primitive equation model of SpeedyWeather.jl. 

How to cite: Gelbrecht, M. and Boers, N.: Hybrid neural differential equation models for atmospheric dynamics, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12495, https://doi.org/10.5194/egusphere-egu24-12495, 2024.

We explore the task of learning the dynamics of a system from observed data without prior knowledge of the laws governing the system. Our extensive simulation study focuses on ordinary differential equation (ODE) problems that are specifically designed to reflect key aspects of various machine learning tasks for dynamical systems - namely, chaos, complexity, measurement uncertainty, and variability in measurement intervals. The study evaluates a variety of methods, including neural ODEs, transformers, Gaussian processes, echo state networks, and spline-based estimators. Our results show that the relative performance of the methods tested varies widely depending on the specific task, highlighting that no single method is universally superior. Although our research is predominantly in low-dimensional settings, in contrast to the high-dimensional nature of many climate science challenges, it provides insightful comparisons and understanding of how different approaches perform in learning the dynamics of complex systems.

How to cite: Schötz, C., White, A., and Boers, N.: Comparing Machine Learning Methods for Dynamical Systems, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12826, https://doi.org/10.5194/egusphere-egu24-12826, 2024.

Inverse modelling systems relying on Lagrangian Particle Dispersion Models (LPDMs) are a popular way to quantify greenhouse gas emissions using atmospheric observations, providing independent evaluation of countries' self-reported emissions. For each GHG measurement, the LPDM performs backward-running simulations of particle transport in the atmosphere, calculating source-receptor relationships (“footprints”). These reflect the upwind areas where emissions would contribute to the measurement. However, the increased volume of satellite measurements from high-resolution instruments like TROPOMI cause computational bottlenecks, limiting the amount of data that can be processed for inference. Previous approaches to speed up footprint generation revolve around interpolation, therefore still requiring expensive new runs. In this work, we present the first machine learning-driven LPDM emulator that once trained, can approximate satellite footprints using only meteorology and topography. The emulator uses Graph Neural Networks in an Encode-Process-Decode structure, similar to Google’s Graphcast [1], representing latitude-longitude coordinates as nodes in a graph. We apply the model for GOSAT measurements over Brazil to emulate footprints produced by the UK Met Office’s NAME LPDM, training on data for 2014 and 2015 on a domain of size approximately 1600x1200km at a resolution of 0.352x0.234 degrees. Once trained, the emulator can produce footprints for a domain of up to approximately 6500x5000km, leveraging the flexibility of GNNs. We evaluate the emulator for footprints produced across 2016 on the 6500x5000km domain size, achieving intersection-over-union scores of over 40% and normalised mean absolute errors of under 30% for simulated CH4 concentrations. As well as demonstrating the emulator as a standalone AI application, we show how to integrate it with the full GHG emissions pipeline to quantify Brazil’s emissions. This method demonstrates the potential of GNNs for atmospheric dispersion applications and paves the way for large-scale near-real time emissions emulation.

 [1] Remi Lam et al.,Learning skillful medium-range global weather forecasting. Science 382,1416-1421 (2023). DOI:10.1126/science.adi2336

How to cite: Fillola, E., Santos-Rodriguez, R., and Rigby, M.: A Graph Neural Network emulator for greenhouse gas emissions inference, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15144, https://doi.org/10.5194/egusphere-egu24-15144, 2024.

Leveraging Machine Learning (ML) models, particularly Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) like Long-Short Term Memory (LSTM), and Artificial Neural Networks (ANN), has become pivotal in addressing the escalating frequency and severity of extreme events such as heatwaves, hurricanes, floods, and droughts. In climate modeling, ML proves invaluable for analyzing diverse datasets, including climate data and satellite imagery, outperforming traditional methods by adeptly handling vast information and identifying intricate patterns. Focusing on the study's emphasis on extreme precipitation events, the urgency arises from climate change, demanding more accurate and timely methods to predict and manage the impacts of these events.

In this study, we completed two main experiments to understand if ML algorithms can detect the extreme events. In both experiment the predictors that have been used are eastern and northern wind (u,v), geopotential height (z), specific humidity (q) and temperature (t) at four pressure levels, which are 1000hpa, 850hpa, 700hpa, and 500hpa. The frequency for the predictors is 3 hours, while the predictand being the precipitation accumulated over 3 hours. The data used in this study are the Re-Analysis -5th generation- (ERA5) produced by European Center for Medium-Range Weather Forecast (ECMWF), which provides global hourly estimates of large number of atmospheric, land and oceanic climate variables with a resolution of 25 km at different pressure levels and for the surface (precipitation in our case).

In this study, two main architectures have been applied. The first emulator, ERA-Emulator, contains 14 layers, divided in 4 blocks (input, convolutional, dense, output). In the convolutional block we have 6 convolutional layers, one layer of type ConvLSTM2D, that combines a 2D Convolutional layer and an LSTM layer, and five simple 2D convolutional layers, with two of them followed by a MaxPooling layer. In the Dense block there are three fully connected Dense layers followed by one Flatten layer and one Dropout layer. Then, we have the output layer, also a Dense layer. We used the same architecture for the second emulator, GRIPHO-Emulator, with one extra MaxPooling in the convolutional block, for a total of 15 layers. The first emulator uses variables from ERA5 both as input and output at 25 km resolution, while the second one uses variables from ERA5 as input, and the Gridded Italian Precipitation Hourly Observations dataset (GRIPHO) as output at 3 km resolution.

The ERA-Emulator is designed to approximate the downscaling function by utilizing low-resolution simulations to generate equivalent low resolution precipitation fields. ERA-Emulator resulted in a viable approach to address this challenge. The emulator demonstrates the capability to derive precipitation fields that align with ERA5 low-resolution simulations.  GRIPHO-emulator aims to downscale high resolution precipitation from low-resolution large-scale predictors. The emulator aims to estimate the downscaling function. GRIPHO-Emulator is able to create realistic high-resolution precipitation fields that well represent the observed precipitation distribution from the high resolution GRIPHO dataset.

How to cite: Abed, W. and Coppola, E.: Detection of High Convective Precipitation Events Using Machine Learning Methods, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-21760, https://doi.org/10.5194/egusphere-egu24-21760, 2024.

Decision making and efficient early warning systems for extreme weather rely on subseasonal-to-seasonal (S2S) forecasts. However, the chaotic nature of the atmosphere impedes predictions by dynamical forecast systems on the S2S time scale. Improved predictability may arise due to remote drivers and corresponding teleconnections in so-called windows of opportunities, but using knowledge of such drivers to boost S2S forecast skill is challenging. Here, we present a spatio-temporal deep neural network (DNN), predicting a time series of weekly North Atlantic European (NAE) weather regimes on lead-times of one to six weeks during boreal winter. The spatio-temporal architecture combines a convolutional Long-short-term-memory (convLSTM) encoder with an Long-short-term-memory (LSTM) decoder and was built to consider both short and medium-range variability as information. As predictors it uses 2D (image) time series input data of expected drivers of European winter weather, including the stratospheric polar vortex  and tropical sea surface temperatures, alongside the 1D time series of NAE regimes. Our results indicate that additional information provided in the image time series yield a skill score improvement for longer lead times. In addition, by analysing periods of enhanced or decreased predictability of the DNN, we can infer further information regarding prevalent teleconnections.

How to cite: Bommer, P. L., Kretschmer, M., Boehnke, P., and Hoehne née Vidovic, M. M.-C.: Using spatio-temporal neural networks to investigating teleconnections and enhance S2S forecasts of european extreme weather , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15174, https://doi.org/10.5194/egusphere-egu24-15174, 2024.

Analyzing climate scenarios is crucial for quantifying uncertainties, identifying trends, and validating models. Objective statistical methods provide decision support for policymakers, optimize resource allocation, and enhance our understanding of complex climate dynamics. These tools offer a systematic and quantitative framework for effective decision-making and policy formulation amid climate change, including accurate projections of extreme events—a fundamental requirement for Earth system modeling and actionable future predictions. 

This study applies dynamic mode decomposition with control (DMDc) to assess temperature and precipitation variability in climate model projections under various future shared socioeconomic pathways (SSPs). We leverage global greenhouse gas emissions and local aerosol emissions as control parameters to unveil nuanced insights into climate dynamics.Our approach involves fitting distinct DMDc models over a high-ambition/low-forcing scenario (SSP126), a medium-forcing scenario (SSP245) and a high-forcing scenario (SSP585). By scrutinizing the eigenvalues and dynamic modes of each DMDc model, we uncover crucial patterns and trends that extend beyond traditional climate analysis methods. Preliminary findings reveal that temporal modes effectively highlight variations in global warming trends under different emissions scenarios. Moreover, the spatial modes generated by DMDc offer a refined understanding of temperature disparities across latitudes, effectively capturing large-scale oscillations such as the El Niño Southern Oscillation. 

The proposed data-driven analytical framework not only enriches our comprehension of climate dynamics but also enhances our ability to anticipate and adapt to the multifaceted impacts of climate change. Integrating DMDc into climate scenario analysis may help formulate more effective strategies for mitigation and adaptation.

References

Allen, Myles R., et al. "Warming caused by cumulative carbon emissions towards the trillionth tonne." Nature 458.7242 (2009): 1163-1166.

Zelinka, Mark D., et al. "Causes of higher climate sensitivity in CMIP6 models." Geophysical Research Letters 47.1 (2020): e2019GL085782.

Proctor, Joshua L., Steven L. Brunton, and J. Nathan Kutz. "Dynamic mode decomposition with control." SIAM Journal on Applied Dynamical Systems 15.1 (2016): 142-161.

How to cite: Mankovich, N., Bouabid, S., and Camps-Valls, G.: Analyzing Climate Scenarios Using Dynamic Mode Decomposition With Control, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11831, https://doi.org/10.5194/egusphere-egu24-11831, 2024.

To account for uncertainties in future projections associated with the level of greenhouse gas emissions, most climate models are run using different forcing scenarios, like the Shared Socioeconomic Pathways (SSPs). Although it is possible to compare real-world greenhouse gas concentrations with these hypothetical scenarios, it is less clear how to determine whether observed patterns of weather and climate anomalies align with individual scenarios, especially at the interannual timescale. As a result, this study designs a data-driven approach utilizing artificial neural networks (ANNs) that learn to classify global maps of annual-mean temperature or precipitation with a matching emission scenario using a high-resolution, single model initial-condition large ensemble. Here we construct our ANN framework to consider whether a climate map is from SSP1-1.9, SSP2-4.5, SSP5-8.5, a historical forcing scenario, or a natural forcing scenario using the Seamless System for Prediction and EArth System Research (SPEAR) by the NOAA Geophysical Fluid Dynamics Laboratory. A local attribution technique from explainable AI is then applied to identify the most relevant temperature and precipitation patterns used for each ANN prediction. The explainability results reveal that some of the most important geographic regions for distinguishing each climate scenario include anomalies over the subpolar North Atlantic, Central Africa, and East Asia. Lastly, we evaluate data from two overshoot simulations that begin in either 2031 or 2040, which are a set of future simulations that were excluded from the ANN training process. For the rapid mitigation experiment that starts a decade earlier, we find that the ANN links its climate maps to the lowest emission scenario by the end of the 21^st century (SSP1-1.9) in comparison to the more moderate scenario (SSP2-4.5) that is selected for the later mitigation experiment. Overall, this framework suggests that explainable machine learning could provide one possible strategy for assessing observations with future climate change pathways.

How to cite: Labe, Z., Delworth, T., Johnson, N., and Cooke, W.: Explainable AI for distinguishing future climate change scenarios, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9110, https://doi.org/10.5194/egusphere-egu24-9110, 2024.

Accurately assessing precipitation impacts due to anthropogenic global warming relies on numerical Earth system model (ESM) simulations. However, the discretized formulation of ESMs, where unresolved small-scale processes are included as semi-empirical parameterizations, can introduce systematic errors in the simulations. These can, for example, lead to an underestimation of spatial intermittency and extreme events.
 Generative deep learning has recently been shown to skillfully bias-correct and downscale precipitation fields from numerical simulations [1,2]. Using spatial context, these methods can jointly correct spatial patterns and summary statistics, outperforming established statistical approaches.
However, these approaches require separate training for each Earth system model individually, making corrections of large ESM ensembles computationally costly. Moreover, they only allow for limited control over the spatial scale at which biases are corrected and may suffer from training instabilities.
Here, we follow a novel diffusion-based generative approach [3, 4] by training an unconditional foundation model on the high-resolution target ERA5 dataset only. Using fully coupled ESM simulations of precipitation, we investigate the controllability of the generative process during inference to preserve spatial patterns of a given ESM field on different spatial scales.

[1] Hess, P., Drüke, M., Petri, S., Strnad, F. M., & Boers, N. (2022). Physically constrained generative adversarial networks for improving precipitation fields from Earth system models. Nature Machine Intelligence, 4(10), 828-839.

[2] Harris, L., McRae, A. T., Chantry, M., Dueben, P. D., & Palmer, T. N. (2022).A generative deep learning approach to stochastic downscaling of precipitation forecasts. Journal of Advances in Modeling Earth Systems, 14(10), e2022MS003120.

[3] Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J. Y., & Ermon, S. (2021).  Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073.

[4] Bischoff, T., & Deck, K. (2023). Unpaired Downscaling of Fluid Flows with Diffusion Bridges. arXiv preprint arXiv:2305.01822.

How to cite: Hess, P., Gelbrecht, M., Aich, M., Pan, B., Bathiany, S., and Boers, N.: Downscaling precipitation simulations from Earth system models with generative deep learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10298, https://doi.org/10.5194/egusphere-egu24-10298, 2024.

How to cite: Lütjens, B., Selin, N., Souza, A., Geogdzhayev, G., Newman, D., Giani, P., Tebaldi, C., Watson-Parris, D., and Ferrari, R.: Is linear regression all you need? Clarifying use-cases for deep learning in climate emulation, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10759, https://doi.org/10.5194/egusphere-egu24-10759, 2024.

Underrepresentation of cloud formation is a known failing in current climate simulations. The coarse grid resolution required by the computational constraint of integrating over long time scales does not permit the inclusion of underlying cloud generating physical processes. This work employs a multi-output Gaussian Process (MOGP) trained on high resolution Unified Model (UM) simulation data to predict the variability of temperature and specific humidity fields within the climate model. A proof-of-concept study has been carried out where a trained MOGP model is coupled in-situ with a simplified Atmospheric General Circulation Model (AGCM) named SPEEDY. The temperature and specific humidity profiles of the SPEEDY model outputs are perturbed at each timestep according to the predicted high resolution informed variability. 10-year forecasts are generated for both default SPEEDY and ML-hybrid SPEEDY models and output fields are compared ensuring hybrid model predictions remain representative of Earth's atmosphere. Some changes in the precipitation, outgoing longwave and shortwave radiation patterns are observed indicating modelling improvements in the complex region surrounding India and the Indian sea.

How to cite: Briant, J., Giles, D., Morcrette, C., and Guillas, S.: A Hybrid Machine Learning Climate Simulation Using High Resolution Convection Modelling, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3499, https://doi.org/10.5194/egusphere-egu24-3499, 2024.

Historical records of climate fields are often sparse due to missing measurements, especially before the introduction of large-scale satellite missions. Several statistical and model-based methods have been introduced to fill gaps and reconstruct historical records. Here, we employ a recently introduced deep-learning approach based on Fourier convolutions, trained on numerical climate model output, to reconstruct historical climate fields. Using this approach we are able to realistically reconstruct large and irregular areas of missing data, as well as reconstruct known historical events such as strong El Niño and La Niña with very little given information. Our method outperforms the widely used statistical kriging method as well as other recent machine learning approaches. The model generalizes to higher resolutions than the ones it was trained on and can be used on a variety of climate fields. Moreover, it allows inpainting of masks never seen before during the model training.

How to cite: Bochow, N., Poltronieri, A., Rypdal, M., and Boers, N.: Reconstructing Historical Climate Fields With Deep Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5103, https://doi.org/10.5194/egusphere-egu24-5103, 2024.

Understanding atmospheric variability is essential for adapting to future climate extremes. Key ways to do this are through analysing climate field reconstructions and reanalyses. However, producing such reconstructions can be limited by high production costs, unrealistic linearity assumptions, or uneven distribution of local climate records. 

Here, we present a machine learning-based non-linear climate variability reconstruction method using a Recurrent Neural Network that is able to learn from existing model outputs and reanalysis data. As a proof-of-concept, we reconstructed more than 400 years of global, monthly temperature anomalies based on sparse, realistically distributed pseudo-station data.

Our reconstructions show realistic temperature patterns and magnitude reproduction costing about 1 hour on a middle-class laptop. We highlight the method’s capability in terms of mean statistics compared to more established methods and find that it is also suited to reconstruct specific climate events. This approach can easily be adapted for a wide range of regions, periods and variables. As additional work-in-progress we show output of this approach for reconstructing European weather in 1807, including the extreme summer heatwave of that year.

How to cite: Wegmann, M. and Jaume-Santero, F.: From climate to weather reconstruction with inexpensive neural networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3614, https://doi.org/10.5194/egusphere-egu24-3614, 2024.

This study focuses on the application of machine learning techniques to better characterize predictability of the spatiotemporal variability of sea surface temperature (SST) on the basin scale. Both, sub-seasonal variability including extreme events (cf. marine heatwaves) and interannual variability are considered. 

We rely on dimensionality reduction techniques---linear principal component analysis (PCA)  and nonlinear autoencoders and their variants---to then perform the actual prediction tasks in the corresponding latent space using disparate methodologies ranging from linear inverse modeling (LIM) to reservoir computing (RC), and attention-based transformers. 

After comparing performance, we examine various issues including the role of generalized synchronization in RC and implicit memory of RC vs. explicit long-term memory of transformers with the broad aim of shedding light on the effectiveness of these techniques in the context of data-driven climate prediction.

How to cite: Nadiga, B. and Srinivasan, K.: Climate Prediction in Reduced Dimensions: A Comparative Analysis of Linear Inverse Modeling, Reservoir Computing and Attention-based Transformers, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12141, https://doi.org/10.5194/egusphere-egu24-12141, 2024.

Here we develop a machine learning emulator based on the Norwegian Earth System Model (NorESM) to predict regional climate responses to aerosol emissions and use it to study the sensitivity of surface temperature to anthropogenic emission changes in key policy regions. Aerosol emissions have both an immediate local effect on air quality, and regional effects on climate in terms of changes to temperature and precipitation distributions via direct radiative impacts and indirect cloud-aerosol interactions. Regional climate change depends on a balance between aerosol and greenhouse gas forcing, and in particular extreme events are very sensitive to changes in aerosol emissions. Our goal is to provide a tool which can be used to test the impacts of policy-driven emission changes efficiently and accurately, while retaining the spatio-temporal complexity of the larger physics-based Earth System Model.

How to cite: Dewey, M., Ekman, A., Watson-Parris, D., and Hansson, H. C.: Machine learning aerosol impacts on regional climate change., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17601, https://doi.org/10.5194/egusphere-egu24-17601, 2024.

High resolution projections are useful for planning climate change adaptation [1] but are expensive to produce using physical simulations. We make use of a state-of-the-art generative machine learning (ML) method, a diffusion model [2], to predict variables from a km-scale model over England and Wales. This is trained to emulate daily mean output from the Met Office 2.2km UK convection-permitting model (CPM) [3], averaged to 8.8km scale for initial testing, given coarse-scale (60km) weather states from the Met Office HadGEM3 general circulation model. This achieves downscaling at much lower computational cost than is required to run the CPM and when trained to predict precipitation the emulator produces samples with realistic spatial structure [4, 5]. We show the emulator learns to represent climate change over the 21^st century. We present some diagnostics indicating that there is skill for extreme events with ~100 year return periods, as is necessary to inform decision-making. This is made possible by training the model on ~500 years of CPM data (48 years from each of 12 ensemble members). We also show the method can be useful in scenarios with limited high-resolution data. The method is stochastic and we find that it produces a well-calibrated spread of high resolution precipitation samples for given large-scale conditions, which is highly important for correctly representing extreme events.

Furthermore, we are extending this method to generate coherent multivariate samples including other impact-relevant variables (e.g. 2m temperature, 2m humidity and 10m wind). We will show the model’s performance at producing samples with coherent structure across all the different variables and its ability to represent extremes in multivariate climate impact indices.

References

[1] Kendon, E. J. et al. (2021). Update to the UKCP Local (2.2km) projections. Science report, Met Office Hadley Centre, Exeter, UK. [Online]. Available: https://www.metoffice.gov.uk/pub/data/weather/uk/ukcp18/science-reports/ukcp18_local_update_report_2021.pdf

[2] Song, Y. et al. (2021). Score-Based Generative Modeling through Stochastic Differential Equations. ICLR.

[3] Kendon EJ, E Fischer, CJ Short (2023) Variability conceals emerging trend in 100yr projections of UK local hourly rainfall extremes, Nature Comms, doi: 10.1038/s41467-023-36499-9

[4] Addison, Henry, Elizabeth Kendon, Suman Ravuri, Laurence Aitchison, and Peter AG Watson. (2022). Machine learning emulation of a local-scale UK climate model. arXiv preprint arXiv:2211.16116.

[5] Addison, H., Kendon, E., Ravuri, S., Aitchison, L., and Watson, P. (2023). Downscaling with a machine learning-based emulator of a local-scale UK climate model, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-14253, https://doi.org/10.5194/egusphere-egu23-14253

How to cite: Addison, H., Kendon, E., Ravuri, S., Aitchison, L., and Watson, P.: Machine learning-based emulation of a km-scale UK climate model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10262, https://doi.org/10.5194/egusphere-egu24-10262, 2024.