Accurate classification of precipitation type (convective vs. stratiform) from passive microwave satellite observations is fundamental for understanding global precipitation patterns and improving weather prediction models. While previous studies have demonstrated the effectiveness of deep learning approaches with accuracies above 90% on smaller temporal windows, questions remain about model generalization across extended time periods and different surface types. This study presents a comprehensive analysis using a nine-year Global Precipitation Measurement (GPM) mission dataset, followed by a detailed investigation of surface-type specialization.
Our primary analysis leverages over 400 million samples from 2014-2022, using 32x32 pixel patches and a ResNet architecture. This large-scale model achieved an accuracy of 85% on a holdout test-year, with balanced performance across both precipitation types (F1-score of 0.85 for both convective and stratiform classes). While matching the general performance range of previous approaches, these results demonstrate robust generalization capabilities across a much longer temporal span and diverse global conditions, using a significantly larger training dataset.
To further investigate model generalization, we conducted a specialized analysis to examine performance across different surface types creating distinct datasets for land and ocean. ResNet-50 architectures were trained for three comparative models: a baseline model using combined data, a land-only, and an ocean-only model. Analysis revealed that the baseline model achieved robust performance across both surface types (82 % accuracy over land, 86 % over ocean). Surprisingly, surface-specific models showed minimal improvement, with the land-specific model achieving 81% and the ocean-specific model reaching 85% accuracy on their respective domains. This suggests that larger, diverse datasets enable models to learn more robust features that generalize well across data subsets.
These findings demonstrate that deep learning models can effectively maintain consistent performance when scaling to multi-year global datasets, and suggest that investing in larger, more diverse training datasets provides robust generalization across both temporal and spatial dimensions without requiring specific subsetting. The approach could lead to more efficient operational systems while maintaining reliable classification accuracy across different surface types and extended time periods.
How to cite: Orescanin, M., Duvio, D., and Petkovic, V.: Large-Scale Deep Learning for Global Precipitation Type Classification, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14109, https://doi.org/10.5194/egusphere-egu25-14109, 2025.
We use machine learning (ML) to map column-integrated water vapor (IWV) to characterize atmospheric rivers (ARs) and to map planetary boundary layer (PBL) height given Global Navigation Satellite Systems (GNSS) radio occultation (RO) data. GNSS RO of the Earth’s atmosphere obtains vertical profiles of microwave refractivity with vertical resolution approaching 100 meters. RO is effectively a water vapor sounder in the lower troposphere and is especially sensitive to vertical gradients associated with the top of the PBL. RO soundings undersample synoptic variability of the atmosphere with severe nonuniformity in spatial and solar angle distribution, making the creation of atmospheric model-agnostic level 3 climatologies a complicated task. ML methods have already shown great promise for mapping RO retrieved quantities in the horizontal and time. We present two applications of ML to RO data, the first to characterize ARs, and the second to map PBL height.
In a mission architecture trade study, we use an ML approach to determine what type of small-satellite constellation would be appropriate to map ARs with detail sufficient for atmospheric process studies and for the prediction of the severe weather on the U.S. Pacific coast that results from ARs. Because ARs are high-volume flows of water vapor in filaments within the PBL and RO sounds water vapor, RO data are ideally suited as input to ML algorithms for the study of ARs. How many low-Earth orbiting RO sounders would be needed to gain desired information on ARs remains an open question, as does our ability to map ARs with existing program-of-record RO data. We answer these questions by formulating ML algorithms to map ARs in the North Pacific Ocean from simulated and real RO data. Simulated RO sounding geolocations are defined by various sizes and types of Walker constellations, realistic GNSS orbits, and the interpolation of refractivity profiles from the ECMWF operational forecast system. We develop two neural networks, one to convert refractivity soundings between 0 and 10 km to IWV and another to map IWV in the horizontal in 1-hr time windows. We find that optimal performance is obtained with Walker constellations of 36 or more RO satellites in near-polar orbits, appropriate orbits for temporal uniformity of the RO’s sampling density. The advent of GNSS RO satellites in micro-satellite and nano-satellite form factors makes such constellations feasible and affordable in the very near future.
We then use ML to map PBL height. The PBL is the part of the atmosphere closest to the Earth’s surface. In the PBL, turbulent processes often affect the vertical redistribution of heat and moisture and their exchange influences cloud evolution and large- to meso-scale circulation. The PBL height can be used to describe climatological processes in a specific region, including cloud characterization. The high vertical resolution of RO observations is suitable to model PBL height in individual profiles. Initially, we use the changes in refractivity to model the PBL height at RO locations and times. Then, we apply ML to produce global PBL heights with a high temporal resolution.
How to cite: Shehaj, E., Leroy, S., and Cahoy, K.: A machine learning framework to map Atmospheric Rivers and Planetary Boundary Layer height from GNSS radio occultation observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13623, https://doi.org/10.5194/egusphere-egu25-13623, 2025.
Cloud fraction significantly affects the short- and long-wave radiation. However, its realistic representation in models has been difficult due to inadequate understanding of the sub-grid scale cloud processes. Recently, we have developed a neural network-based scale-adaptive (NSA) cloud-fraction scheme using the CloudSat data and found that the new scheme could greatly improve the simulation of cloud spatial distribution and vertical structure. In this study, we present two applications of the NSA scheme in the WRF model. The first is the simulation of the regional winter climate of the Tibet Plateau, where the NSA scheme was shown to significantly reduce the longstanding bias of too-cold surface temperature. The second is a tropical cyclone simulation, showing that the NSA scheme better simulated the track of In-Fa (2021). The underlying mechanisms will be presented.
How to cite: Chen, G.: The application of a neural network-based scale-adaptive cloud-fraction scheme in the WRF model , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7683, https://doi.org/10.5194/egusphere-egu25-7683, 2025.
A deep ANN-based dimension reduction (DR) method, called DIRESA (distance-regularized Siamese twin autoencoder), has been developed to capture nonlinearities while preserving distance (ordering) and producing statistically independent latent components. The architecture is based on a Siamese twin autoencoder, with three loss functions: reconstruction, covariance, and distance loss. An annealing method is used to automate the otherwise time-consuming process of tuning the different weights of the loss function. DIRESA has been compared with PCA and state-of-the-art DR methods for two conceptual models, Lorenz ’63 and MAOOAM (Modular Arbitrary-Order Ocean-Atmosphere Model), and significantly outperforms them in terms of distance (ordering) preservation KPIs and reconstruction fidelity. The latent components have a physical meaning as the dominant modes of variability in the system. DIRESA correctly identifies the major coupled modes associated with the low-frequency variability of the coupled ocean-atmosphere system. Next to the conceptual model results, the first DIRESA results for reanalysis data will be presented.
DIRESA is provided as an open-source Python package, based on Tensorflow. With one line of code convolutional and/or dense layers DIRESA models can be build. On top of that, the package allows the use of custom encoder and decoder submodels to build a DIRESA model. The DIRESA package acts as a meta-model, which can use submodels with various kinds of layers, such as attention layers, and more complicated designs, such as graph neural networks. Thanks to its extensible design, the DIRESA framework can handle more complex data types, such as three-dimensional, graph, or unstructured data. Its flexibility and robust performance make DIRESA an promising new tool in weather and climate science to distil meaningful low-dimensional representations from the ever-increasing volumes of high-resolution climate data, for applications ranging from analog retrieval to attribution studies.
Tutorial: https://diresa-learn.readthedocs.io/
Preprint: https://arxiv.org/abs/2404.18314
How to cite: De Paepe, G. and De Cruz, L.: DIRESA – A Deep Learning-based, nonlinear "PCA", EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6004, https://doi.org/10.5194/egusphere-egu25-6004, 2025.
Weather and climate simulations produce petabytes of high-resolution data that are later analyzed by researchers to understand climate change or severe weather. We propose a new error-bounded compression method targeting for weather and climate data. It contains a JPEG2000 compression layer to capture the bulk part of the data, and a sparse wavelet layer to record the sparse signal that excess the given error bound. The sparse wavelet layer encodes the wavelet coefficients using the SPIHT algorithm. We test our method with established compression methods on a suite of benchmarks including basic statistics, case study of the hurricane data, derived variable computation, and Lagrangian trajectory simulation. Our method is favourable in most benchmark cases at given range relative error targets from 0.1% to 10% achieving compression ratios from 10x to more than 800x. It can reconstruct the derivatives of the compressed data with a best fidelity and does not add high frequency artifacts found in other compression methods. In Lagrangian trajectory simulations, our method can produce less distortion in trajectories and distribution of particles compared with SZ3. We are able to produce a 16x compressed wind data achieving less error metric than adding 5% random noise to the data, making it ready for practical use.
How to cite: Huang, L. and Hoefler, T.: Error bounded compression for weather and climate applications, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10282, https://doi.org/10.5194/egusphere-egu25-10282, 2025.
Data-driven weather forecasting models, such as FuXi and Pangu-Weather, have made significant advancements in global forecasting accuracy and computational efficiency. However, these models lack physical constraints, a limitation that traditional numerical weather prediction (NWP) models address through the dynamical core and physical parameterization schemes. Recent efforts, like NeuralGCM and PINNs, have successfully integrated the dynamical core or Navier-Stokes equations with machine learning models. Yet, effective integration of physical parameterization schemes remains uncharted in this field, primarily due to the greater uncertainty and complexity of physical processes compared to the dynamical core. To bridge this gap, we integrated the shortwave radiative transfer scheme with FuXi, by modeling the Rapid Radiative Transfer Model for General Circulation Models Applications (RRTMG) as a neural network. This represents the first successful integration of a physical parameterization scheme with large-scale weather forecasting models. This integration yielded substantial improvements in forecasting performance and physical consistency, reducing root mean square error (RMSE) by approximately 15% for radiatively related variables, such as albedo and cloud water mixing ratio, especially for longer lead times. Moreover, the optimized model demonstrated significantly enhanced atmospheric moisture energy conservation. This work provides a promising pathway for integrating physical processes into machine learning based weather forecasting models, paving the way for more accurate and physically consistent weather forecasts.
How to cite: huang, Q., zhong, X., fan, X., and li, H.: Physics-Constrained Machine Learning Model Enhances Weather Forecast Accuracy and Physical Consistency, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14355, https://doi.org/10.5194/egusphere-egu25-14355, 2025.
Satellite-based research on global sea surface wind is essential for understanding and monitoring the air-sea interface dynamical processes, driving the need for accurate and efficient assessment. Spaceborne scatterometers, which are among the most relevant sensors for sea surface wind observation, play a vital role in obtaining global ocean surface wind information. However, a significant challenge associated with polar-orbiting satellites lies in data gaps caused by their orbital paths, resulting in missing observations between swaths. While several satellite-derived sea surface wind products have been developed using data assimilation (DA) techniques, these existing methods are time-consuming and require large amounts of diverse data, rendering them computationally expensive. Additionally, the iterative steps of variational algorithms can only perform linear or weakly nonlinear adjustments to the governing equations, which may pose challenges given the highly non-linear nature of these equations.
In this study, we leveraged physics-informed neural networks (PINNs) techniques to reconstruct large-scale sea surface wind fields with sparse scatterometer observations from different satellites, integrating observations and filling gaps. By incorporating physical constraints into the loss function, specifically the Navier-Stokes equations, we efficiently fill the data gaps and reconstruct wind fields that not only match observational data but also adhere to physical principles. Another objective of this work is to introduce the wind speed gradient and direction parallel consistent constraints into the loss function in order to enhance the detail of the reconstructed wind field and increase the accuracy of both wind speed and direction. Structurally, the PINN resembles a fully connected neural network (FCNN), offering the advantage of automatic feature extraction. Our model not only extracts valuable information from existing data but also uncovers complex patterns and correlations in data that are difficult for traditional algorithms to capture. This approach provides a novel perspective and an alternative methodology for wind field reconstruction.
The results show that PINNs can reconstruct wind fields that closely resemble realistic wind patterns, capturing large-scale structures while preserving fine-scale details, thanks to the introduction of wind speed gradient and direction parallel consistent constraints. The training time for this model is about 3 hours, using only a single GPU core. This efficiency is partly due to the fact that PINNs do not rely on ensemble methods or large datasets to produce results. Unlike traditional DA methods, PINN does not depend on an initial best-guess field for assimilating observations. While we use only a small amount of scatterometer observation data, no initial field is required to complete the reconstruction. Additionally, since the PINN represents a continuous and differentiable function, it can produce outputs at any spatial or temporal point within the training domain.
Recognizing their potential for forecast models and data integration, PINNs offer a promising approach for accurate sea surface wind field reconstruction and could serve as an effective alternative to current methods.
How to cite: Bo, R., Zhou, Z., Du, H., Yang, P., Zhao, X., Li, Q., and Zang, Z.: An Alternative Large-scale Sea Surface Wind Field Reconstruction Method Using Sparse Scatterometer Data Based on Physics-informed Neutral Network, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7915, https://doi.org/10.5194/egusphere-egu25-7915, 2025.
Tropical cyclones (TCs) pose significant risks, particularly in coastal regions, making accurate prediction of their track and intensity is crucial for effective disaster preparedness and response. Traditional numerical models struggle with balancing accuracy and computational efficiency, although TC track prediction has achieved substantial progress, challenges remain in forecasting TC intensity, especially rapid intensification (RI). This study aims to (1) develop a Transformer-based deep learning (DL) model to predict TC track, intensity, and 24-hour future intensity change simultaneously, and (2) investigate the relative importance of input variables to the contribution of improving model forecast ability. Based on 2001–2021 best track data and ERA5 reanalysis data over the western North Pacific (WNP), we develop an optimal model called OWZP-Transformer, which leverages the multi-head self-attention mechanism and incorporates the 13 input parameters categorized into four factors: basic, environmental, gradient, and structural. Specifically, our study incorporates structural parameters, which is represented by Okubo-Weiss-Zeta Parameter (OWZP). Our OWZP-Transformer model achieves competitive results for track prediction and shows excellent performance in intensity forecasting for the next 6 hour, with an overall root mean square error (RMSE) of 0.91 m/s. This result represents an improvement of 61.9% to 68.4% compared to existing DL models, which generally have RMSE values above 2 m/s. In addition, our model demonstrates superior performance in predicting 24-hour future intensity change, achieving an overall lower mean absolute error (MAE) of 1.57 m/s, which is 63.5% to 74.6% lower than existing DL models. Furthermore, our model successfully identifies all 11 RI events out of 30 TCs samples from WNP test dataset during 2020-2021. We further evaluate the contributions of each parameter for the first time using two explainable feature importance methods: DeepLIFT and DeepLiftShap. The results indicate that self-contributions and the basic factors play a dominant role in short-term forecasts, while the OWZ parameter plays a significant role following them. This study is the first attempt to comprehensively predict a broad range of TC forecasting tasks using a single DL model, highlighting the potential the OWZP-Transformer model as a reliable tool for enhancing both the accuracy and efficiency of TC predictions.
How to cite: Lin, Z. and Chu, J. E.: Enhancing tropical cyclone track and intensity predictions with the OWZP-Transformer model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-80, https://doi.org/10.5194/egusphere-egu25-80, 2025.
Understanding and predicting wind flow structures within the atmospheric boundary layer (ABL) across varying stability conditions remains a key challenge in atmospheric science and environmental modeling. Although large-eddy simulation (LES) provides high-fidelity insights, it is computationally prohibitive for near-real-time or large-scale applications. To address this, we propose a deep learning framework that predicts the two-dimensional wind flow fields in the u, v, and w components at a target height z, using corresponding flow fields at three levels above z. By integrating vertical flow correlations and continuity principles, our approach captures essential turbulent features while reducing input dimensionality and eliminating the need for full 3D simulations.
A modified convolutional neural network (CNN) forms the core of this framework, capturing complex spatial and temporal patterns from high-resolution LES datasets. Mass conservation is embedded in the training process to ensure physically consistent results. Preliminary results indicate that the model preserves large-scale turbulence features and captures the influence of higher-elevation dynamics, although smaller high-frequency turbulent features require further refinement. To address this, our ongoing work includes adopting a scale-specific approach to explicitly handle the diverse turbulent length scales observed in the ABL. We are also incorporating multitemporal dynamics and attention mechanisms into the architecture of our model to better account for long-range dependencies over time, thereby enabling the model to adapt to different stability regimes and transitions among them. To enhance interpretability, we will employ explainable AI (XAI) tools such as SHAP and GradCAM, revealing how specific regions in the input influence the emergence of particular turbulent footprints in the predicted flow. These insights guide improvements in both model design and understanding of atmospheric processes governing ABL flow development.
This research underscores the transformative potential of deep learning in boundary layer meteorology. By significantly reducing computational demands while retaining essential flow dynamics, our model enables real-time, high-resolution predictions of ABL flows. This scalable and efficient framework opens new possibilities for diverse applications, including weather forecasting, wind energy optimization, and environmental analysis.
How to cite: Difuntorum, J. K., Katurji, M., Zhang, J., and Zawar-Reza, P.: Multilevel Deep Learning for Non-Local Prediction of ABL Flow Fields Across Varying Stability Regimes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14604, https://doi.org/10.5194/egusphere-egu25-14604, 2025.
The Arctic Ocean plays a critical role in global climate dynamics, yet direct observations of its surface energy budget are sparse and largely constrained to individual field campaigns. Current estimates often rely on reanalysis data, notably ERA5, which demonstrates systematic biases in boundary-layer properties and surface fluxes over Arctic sea-ice. In this study, we train an artificial neural network (ANN) to predict surface fluxes observed during MOSAiC, SHEBA, Arctic Ocean 2018 and ARTofMELT expeditions using ERA5 data as input. Data from shorter field campaigns, such as N-ICE, are used for testing our model against unseen data. Our results indicate that ERA5 Arctic surface fluxes exhibit very low correlations with observations and are characterized by large RMSE values. Our predictions demonstrate significant error reductions across key variables: sensible heat flux (~39%), 2m temperature (~39%), net shortwave radiation (~37%), downward longwave radiation (~21%) and net longwave radiation (~17%). Furthermore, we find a higher correlation (~0.58) with observations compared to ERA5 (~0.24) and approximately 50% reductions in RMSE of the hourly total surface energy budget, i.e. the sum of the individual fluxes. We produce a bias-corrected estimate of surface energy fluxes over Arctic sea-ice. We use our bias correction to revise previous estimates of the climatological surface energy budget over Arctic sea ice. We will make the trained weights available to allow for the custom derivation of bias-corrected fluxes in individual case studies and for climate model evaluation.
How to cite: Hossain, A., Keil, P., Grover, H., and Pithan, F.: Leveraging observations to bias-correct the Arctic surface energy budget in reanalysis through machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18459, https://doi.org/10.5194/egusphere-egu25-18459, 2025.
Large scale machine learning is currently revolutionizing Earth system modeling and the next generation of models will likely be machine learning-based or contain substantial machine learning components. The lack of a complete equation-based description of the Earth system as well as the availability of a plethora of high-quality data make this a tantalizing possibility to obtain models with unprecedented capabilities. In the first part of the talk, we will discuss grand challenges for building machine learning-based Earth system models and what milestones have already been achieved in the last years. We will also examine cases where machine learning models already surpass state-of-the-art equation-based models, e.g. for medium range weather forecasting. In the second half of the talk, we will introduce the WeatherGenerator project that aims to build a next generation, machine learning-based Earth system model. Led by leading European modeling centers but open-source from day-1, the project will train on a wide range of datasets to build a seamless prediction model that can faithfully represent Earth system dynamics from sub-km scale, short term processes to multi-decadal projects. The project will also consider selected applications to ensure the real-world applicability of the developed model.
How to cite: Lessig, C.: Towards machine learning-based Earth system models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17016, https://doi.org/10.5194/egusphere-egu25-17016, 2025.
Fine particulate matter (PM2.5) and ozone pollution cause a significant number of premature deaths in China each year. Evaluating the effects of emission mitigation and climate change on air quality with climate-chemistry models is computationally expensive. In this study, we develop two machine learning models — a deep learning framework based on U-Net image segmentation and long short-term memory (LSTM) neural networks, and an extreme value model — to predict China’s future air quality trends. These models are trained using model simulation results from 2014 to 2022 under high to low anthropogenic emission scenarios. The deep learning model yields promising results for PM2.5, with an R2 of 0.79 and root mean squared error (RMSE) of 12.58 ug/m3. The extreme value model for ozone episode exceedance rates achieves an R2 of 0.97 and RMSE of 1.43%. This work demonstrates the potential of deep learning and extreme value models in efficiently modeling air quality, offering robust tools for future air quality assessments.
How to cite: Wan, F., Shen, L., and Jiang, Z.: Efficient machine learning frameworks for prediciting China’s future air quality trends, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14723, https://doi.org/10.5194/egusphere-egu25-14723, 2025.
Dust aerosol forecasting is of significant scientific and societal importance. Currently, the most accurate forecasting systems rely on numerical weather prediction methods, which solve differential equations to simulate the physical and chemical processes of dust aerosols and predict dust concentrations. However, errors introduced by initial and boundary conditions, along with the complex nonlinear interactions between aerosol physical-chemical processes and atmospheric dynamics, result in uncertainties and high computational costs in numerical prediction methods. In recent years, artificial intelligence (AI) methods have demonstrated significant potential in the field of weather forecasting. However, AI-based approaches for addressing such challenging extreme weather events remain in their infancy. Here, we introduce DustWatcher, an AI-based forecasting method for global dust aerosols. DustWatcher integrates spatiotemporal Transformers with conditional generative networks to develop a neural network framework that optimizes forecasting errors in an end-to-end manner. Compared to current state-of-the-art global and regional aerosol forecasting systems, DustWatcher, trained on 41 years of global reanalysis aerosol data, delivers more accurate deterministic forecasts for most aerosol variables including surface dust concentration and aerosol optical depth. DustWatcher provides skillful forecasts every three hours for the next seven days at a resolution of 0.5°×0.625°. Our results demonstrate the potential of AI in improving the dust forecasts accuracy and advancing its application in dust aerosol forecasting field.
How to cite: Du, S. and Chen, S.: Deep Learning for Accurate Global Dust Aerosol forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3633, https://doi.org/10.5194/egusphere-egu25-3633, 2025.
Diffusion models have been shown highly capable for image generation tasks and more recently have been adapted for weather forecasting, allowing sharper predictions than forecasts generated by previous methods, and straight-forward generation of ensemble predictions. Their value for image-based predictions is evident. Dust storms are frequent high-impact weather phenomena in West Africa that directly impact human life, e.g., by disrupting land and air traffic, posing health threats, and affecting energy delivery from solar-energy systems. Timely and precise prediction of these phenomena is crucial to mitigate adverse impacts.
State-of-the-art machine learning-based weather prediction (MLWP) models do not predict dust since they are limited by computational constraints and by the need of high-quality aerosol reanalyses. Moreover, the current operational numerical weather prediction (NWP) models for Africa still need improvement for resolving the short-scale dynamics and surface properties which leads to the formation of convective dust storms, and also often the convection itself. This is where observation-based short term forecasts (“nowcasts”) become particularly valuable. Nowcasts can provide greater skill than NWP on short time-scales, can be frequently updated, and have the potential to predict phenomena currently operational NWP and MLWP models do not reproduce. However, despite routine high frequency and high resolution observations from satellites, as of January 2025, no nowcast of dust storms is available.
In this study, we present an image-based approach for nowcasting dust storms: we apply a diffusion model to predict next frames of the SEVIRI desert dust RGB composite, a product of false-colour satellite images highlighting both dust and deep convection. We create nowcasts of this RGB composite for a large domain over West Africa up to 6 hours ahead and show that our nowcasts can predict both convective storms and convectively generated dust storms which currently operational NWP may not reliably reproduce. Furthermore, we create ensemble predictions, allowing a probabilistic forecast assessment.
Our approach provides a valuable tool that could be used in operational forecasting to improve the prediction of convective storms, dust storms, and indeed other weather events. Due to the technical similarity of RGB composite imagery from geostationary satellites, this approach could also be adapted to nowcast other RGB composites, such as those for ash, or convective storms. In the wider context, such nowcasts of brightness temperatures and brightness temperature differences, which the RGB composites are based on, could be used for predicting other products which use these satellite retrievals.
How to cite: Hermes, K., Marsham, J., Klose, M., Marenco, F., Brooks, M., and Bollasina, M.: Diffusion models for image-based nowcasting of desert dust for West Africa, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13224, https://doi.org/10.5194/egusphere-egu25-13224, 2025.
Machine Learning (ML) has been widely explored for its potential in modelling air quality in numerous studies in the past. However, these approaches approximated function that maps the finite-dimensional input and output vectors. This restricts their extrapolation to unseen data and different discretization. Neural operators, a class of neural networks approximate the operator between infinite dimensional input and output functions. These models learn the underlying operator between input functions and their time-evolved state directly from data. In this work, we introduce our contribution to the field of neural operators termed Complex Neural Operator (CoNO) to learn the evolution of PM2.5 and CO concentrations over India. We trained our models using WRF-Chem simulated data over India for the years 2016-2018 and evaluated it for the year 2019. We assess our models for forecasting high pollution events, long-term forecasting (up to 72 hours) and city-level forecasts for six cities targeting two key pollutants.
How to cite: Bedi, S., Krishnan, N. M. A., and Kota, S. H.: Developing Neural Operators for Modeling PM2.5 and CO over India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10865, https://doi.org/10.5194/egusphere-egu25-10865, 2025.
Urban and industrial areas are vulnerable to accidental releases of pollutants. To accurately determine the pollutant's plume position and affected areas, it is essential to estimate the atmospheric flow around the affected site. This flow can be precisely computed using numerical methods of Computational Fluid Dynamics (CFD). However, CFD computation is expensive and slow, making it unsuitable for emergency response. As reduced order approximations, machine learning surrogates offer a promising alternative as they are usually much faster; but they must first be trained on CFD-generated data. In this study, we propose a database of atmospheric simulations with varying meshes and atmospheric stability conditions. Meshes are built by randomly sampling buildings and positioning them in space. For each mesh, values of the Monin-Obukhov length and of the ground roughness are sampled, leading to different turbulent regimes and overall atmoshperic flow behaviour. We then train a MeshGraphNet on this database, i.e. a graph neural network built on the mesh structure. The performance of the trained neural network on unseen scenarios with different initial conditions has been evaluated and will be presented.
How to cite: de Villeroché, A., Le Guen, V., Mouradi, R.-S., Massin, P., Bocquet, M., Farchi, A., Cheng, S., and Armand, P.: MeshGraphNets for 3D atmospheric flow in Urban Environment for Atmospheric Dispersion, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10297, https://doi.org/10.5194/egusphere-egu25-10297, 2025.
Here we present the joint German-Brazilian project QUALARIA (Artificial Intelligence based system for sub-urban scale air quality prediction/Sistema baseado em Inteligência Artificial para previsão de qualidade do ar em escala sub-urbana).
Through joint effort between research and business partners in Brazil and Germany, QUALARIA proposes to develop an operational artificial intelligence-based system for monitoring, simulating and predicting air quality in urban environments with unprecedented spatial resolution availability (https://meteoia.com/qualaria/). New downscaling approaches based on artificial intelligence have recently shown promising performance to simulate sub-grid atmospheric processes. This approach is designed to monitor and predict atmospheric pollutant concentrations and air quality indexes with high spatial resolution, through the development of the QUALARIA system. Advanced global and regional chemical-meteorological models, such as reanalysis data from ERA5 and EAC4, and WRF-Chem simulations are applied to derive the climatological state of air composition, specifically the average levels of air pollutant based on existing emission inventories. Measurements of PM10, PM2.5 NO2, and O3 concentrations from Air Quality Automatic Stations of the Environmental Company of São Paulo State (CETESB) are used to train the downscaling AI algorithm to capture the sub-grid spatial variations of the pollutant concentrations. Low-cost sensors are deployed to increase and complement the spatial coverage of the CETESB network. Artificial intelligence will transform air quality maps at a horizontal resolution of 10 km to street-level maps with an increased resolution of 100 m. From its simulated and predicted downscaled pollutant concentration fields, QUALARIA will provide its users with relevant air quality indicators, informing about the impacts of air pollution in human health and activities via an online dashboard. To achieve the optimal dashboard design, public and private sector stakeholders are being engaged and consulted for the co-development of the dashboard design and features.
How to cite: Li, C. W. Y., Peli, V. M. L., Calderón, M. E. G., Perez, G. M. P., Martin, T. C. M., de Lucena, A. V., Barbosa, E. L. S. Y., Schindler, M., Laimer, F., Andrade, M. D. F., Dias de Freitas, E., and Brasseur, G.: An AI System to Predict Street-Level Air Quality: Introduction to the QUALARIA Project, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4569, https://doi.org/10.5194/egusphere-egu25-4569, 2025.
The applications of deep learning (DL) technique in atmospheric environment research are expanding rapidly. Here we developed a DL framework to quantify the responses of surface ozone (O3) to anthropogenic and meteorological changes in China. The DL-based analysis suggests volatile organic compound (VOC)-limited regimes in urban areas over northern inland China, and thus, reductions of nitrogen oxide (NOx) emissions have resulted in increases in surface O3 concentrations. In contrast, changes in meteorological conditions led to a dramatic decrease in surface O3 concentrations in 2019-2021, particularly, in the North China Plain, whereas the decline in surface O3 concentrations driven by beneficial meteorological conditions in 2019-2021 has been completely reversed due to the occurrence of long-lasting heatwave in 2022, particularly in central China. The DL framework, developed in this work, provides a novel data-driven pathway to assess the causes of surface O3 changes, and is helpful for a comprehensive understanding of the driving factors of surface O3 evolution in China.
How to cite: Jiang, Z., Chen, X., Wang, M., and He, T.-L.: Deep learning-based surface O3 responses to anthropogenic and meteorological changes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2638, https://doi.org/10.5194/egusphere-egu25-2638, 2025.
The rapid advancement of Information Technology is transforming research in atmospheric and environmental sciences, with Artificial Intelligence and Machine Learning (AI/ML) offering novel tools to explore complex environmental systems. AI/ML techniques have demonstrated significant potential in atmospheric research and pollutant dynamics [6]. Machine learning’s capability to capture non-linear relationships among environmental variables has been validated in prior studies [5].
This study leverages a feed-forward neural network (FFNN) to investigate nitrogen dioxide (NO2) transport from a coastal urban environment in Central Italy to an inland rural area, leading to increased ozone (O3) production downwind. Such transport phenomena underscore the need to address both direct and transported emissions, as observed in urban-rural gradients worldwide [4,6].
By integrating observational data and meteorological parameters, including wind speed and direction alongside NOx and O3 levels, the FFNN model effectively predicted O3 concentrations at the inland site. Results showed consistently higher O3 levels at the rural site compared to the urban area, reflecting significant O3 production during transport. The model exhibited a high correlation (R = 0.82) between observed and predicted O3 concentrations, underscoring AI’s value in enhancing air pollution dynamics understanding. These findings align with broader research demonstrating AI’s role in refining air quality predictions and improving source attribution [1,2].
This study highlights the effectiveness of AI techniques in environmental research, particularly in elucidating interactions between transportation emissions and secondary pollutants like O3. The results stress the importance of regional air quality modeling and advanced computational approaches in supporting environmental policy and decision-making. AI-driven insights can inform more effective mitigation strategies, enhance air quality forecasting, and assist policymakers in addressing public health concerns related to air pollution [1,2]. Recent reviews emphasize the necessity of integrating AI into air quality management frameworks [7].
Additionally, this research underscores the potential of hybrid AI methods and physics-informed machine learning to further improve atmospheric models and source attribution accuracy. Such innovations are critical for advancing air quality modeling and developing targeted strategies to mitigate environmental and public health impacts [7].
[1] World Health Organization. (2021). Air pollution and health.
[2] Lelieveld, J. et al. (2019). The contribution of outdoor air pollution sources to premature mortality on a global scale. Nature, 525(7569), 367-371.
[3] Heal, M. R., et al. (2013). Particles, air quality, policy, and health. Chemical Society Reviews, 41(19), 6606-6630.
[4] Crutzen, P. J. (2006). The role of NO and NO2 in the chemistry of the troposphere and stratosphere. Annual Review of Earth and Planetary Sciences, 7, 443-472.
[5] Jacob, D. J. (1999). Introduction to Atmospheric Chemistry. Princeton University Press.
[6] Monks, P. S., et al. (2015). Tropospheric ozone and its precursors from the urban to the global scale. Atmospheric Chemistry and Physics, 15(15), 8889-8973.
[7] Kumar, P., et al. (2018). Ambient volatile organic compounds in urban environments: Techniques for sampling, analysis, and implications for air quality. Progress in Environmental Science and Technology, 2(1), 3-13.
How to cite: Chiacchiaretta, P., Aruffo, E., Mascitelli, A., Colangeli, C., Palermi, S., Bianco, S., and Di Carlo, P.: Inland Ozone Production Due to Nitrogen Dioxide Transport Downwind of a Coastal Urban Area: A Neural Network Assessment , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1413, https://doi.org/10.5194/egusphere-egu25-1413, 2025.
Nitrogen oxides (NOx = NO + NO₂) are significant air pollutants, mainly emitted from anthropogenic sources. Bottom-up methods for the estimation of anthropogenic NOx emissions are based on energy consumption data, which, if outdated, result in a delayed response in the produced emission inventories. The TROPOspheric Monitoring Instrument (TROPOMI) provides high-resolution NO₂ column densities, offering valuable data for estimating NOx emissions. Given the short atmospheric lifetime of NOx, horizontal transport influences over distances within a few tens to a few hundred kilometers must be taken into account. To address this, we developed a convolutional neural network (CNN) which incorporates the NO₂ divergence and horizontal transport features to estimate the anthropogenic NOx emissions. Our model operates on a monthly timescale with a spatial resolution of 0.1°, utilizing TROPOMI NO₂ column densities and ERA5 wind field data as inputs, and the EDGARv8.1 0.1° gridded NOx inventories as targets. The training set comprised data from 2019 and 2020, of which 70 % were used for training, and the remaining 30 % for testing of the model. The model achieved an R² of 0.922 and an RMSE of 11.214 Mg/month on the test set when estimating NOx emissions in Europe and the USA. Additionally, the model demonstrated temporal generalization capabilities, achieving an average R² of 0.853 (±0.066) and an average RMSE of 16.545 (±3.804) Mg/month in monthly estimations for Europe and the USA during 2021-2022. The proposed method integrates satellite observations with emission inventories, employing CNNs to facilitate rapid updates of anthropogenic NOx emissions.
How to cite: Zhang, Y., Beirle, S., Kuhn, L., Wagner, T., and Liu, L.: Estimating Anthropogenic NOx Emissions Using Convolutional Neural Networks with Horizontal Transport Considerations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11971, https://doi.org/10.5194/egusphere-egu25-11971, 2025.
Predicting PM2.5 concentrations using air quality data is often hindered by the presence of missing values, which can compromise accuracy and reliability. Traditional prediction models frequently suffer significant data loss during the handling of missing values, necessitating new approaches that address data quality issues while improving prediction performance.
This study proposes an efficient prediction methodology leveraging transfer learning to minimize the impact of missing values. A pre-trained model was constructed using integrated data from all monitoring stations in Seoul, followed by fine-tuning for specific monitoring stations to develop a PM2.5 prediction model. Transfer learning is a machine learning technique that utilizes knowledge from previously trained models to enhance learning efficiency and performance in new tasks or domains. Unlike traditional approaches that require training from scratch, transfer learning reuses the weights and structure of pre-trained models, reducing training time and improving performance.
In this study, a deep neural network (DNN) pre-trained model was built using data from all monitoring stations in Seoul, and fine-tuning was applied using specific station data. The model was trained on six-hour average PM2.5 data to predict the next six hours, effectively addressing missing values.
Preliminary results indicate that the transfer learning-based model effectively handles missing values and demonstrates improved prediction accuracy compared to independently modeled traditional approaches. By leveraging domain-wide information, the model compensates for the limitations of individual monitoring station data, achieving higher accuracy and reliability.
This study provides a scalable solution for addressing data gaps in air quality prediction and contributes to research on the health impacts of air pollution and urban air quality management. Future research will explore the application of this methodology to other pollutants and regions, further enhancing its generalizability and effectiveness.
Acknowledgement
"This research was supported by Particulate Matter Management Specialized Graduate Program through the Korea Environmental Industry & Technology Institute(KEITI) funded by the Ministry of Environment(MOE)"
How to cite: Han, S.-H., Lee, J.-Y., Jang, K., Wang, K.-H., and Yun, H.: A Transfer Learning-Based Model for PM2.5 Prediction in Seoul, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2683, https://doi.org/10.5194/egusphere-egu25-2683, 2025.
The importance of predicting fine particulate matter (PM2.5) concentrations has grown significantly due to deteriorating urban air quality. Seoul presents unique modeling challenges due to its intensive industrial activities, high population density, significant data variability between monitoring stations, and numerous missing values. This study analyzes and compares the predictive performance of LSTM (Long Short-Term Memory) and Bi-LSTM (Bidirectional LSTM) models using hourly PM2.5 concentration data collected from 25 monitoring stations in Seoul from 2018 to 2022.
In the data preprocessing phase, we employed the MICE (Multiple Imputation by Chained Equations) method to handle missing values, which effectively preserved the data's structural characteristics by considering inter-variable relationships. The predictive performance of both models was evaluated using metrics such as RMSE (Root Mean Square Error) and MAE (Mean Absolute Error). While LSTM focuses on forward learning, Bi-LSTM can capture complex time series patterns by utilizing both forward and backward information.
The results demonstrated that Bi-LSTM effectively captured complex time series patterns through its bidirectional learning structure, and optimal learning rates and dropout ratios were determined through various parameter tuning experiments. These findings present the characteristics and potential applications of both models in PM2.5 concentration prediction, expected to contribute to improved air quality forecasting systems and policy decision support.
Future research will focus on enhancing model accuracy through the implementation of additional algorithms and exploring applications for PM2.5 prediction in other cities. This aims to increase the spatial resolution of air pollution prediction and contribute to broader air quality improvements.
Acknowledgement
"This research was supported by Particulate Matter Management Specialized Graduate Program through the Korea Environmental Industry & Technology Institute(KEITI) funded by the Ministry of Environment(MOE)"
How to cite: Jang, K., Lee, J.-Y., Han, S.-H., Wang, K.-H., and Yun, H.-Y.: Comparison of LSTM and Bi-LSTM Models for Predicting PM2.5 Concentration in Seoul, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2682, https://doi.org/10.5194/egusphere-egu25-2682, 2025.
Aerosol composition information is crucial for determining the detrimental effects of aerosols on climate, air quality, and human health, given the differential effects of varying aerosol types. Conversely, developing south Asian countries lack systematic data on aerosol composition, with the available composition information limited to a few sites and a limited time. Moreover, the large spatio-temporal coverage of satellite observations is relatively unexplored for aerosol characterization.
In this study, we utilize satellite sensor Multi-angle Imaging Spectro-Radiometer (MISR) Level 2 version 23 aerosol products (spatial resolution = 4.4 km x 4.4 km) to calculate fractional aerosol optical depths (fAODs) for 2015-2016. These fAODs represent the proportion of total AOD attributable to the eight aerosol models assumed in MISR's aerosol retrieval algorithm, categorized based on size, shape, and refractive indices. The fractional AOD of aerosol model i is represented by fAODi. In this study, we use these eight fAODs, EDGAR emission data, together with land use and meteorological variables, as predictors in a machine learning (ML) model. The model is trained on a quarter-degree grid covering south Asia, using the chemical model simulated aerosol species mass fraction as the target variable.
We train models to predict six aerosol species-sulfate, nitrate, ammonium, black carbon (BC), organic carbon (OC) and dust. We employ two models: Random Forest with out-of-bag bootstrap sampling for cross-validation and Support Vector Regression (SVR) with a 5-fold cross-validation, utilizing 80% of the data for training and 20% for testing. The SVR model shows a mean cross-validation R² of 0.79, 0.83, 0.72, 0.81, 0.73 and 0.81 for sulfate, nitrate, ammonium, dust, OC and BC, respectively, with corresponding RMSE values of 0.02, 0.03, 0.01, 0.05, 0.04 and 0.01 on the test data. The Random Forest model performs better, with R² values of 0.87, 0.92, 0.85, 0.89, 0.90 and 0.88 for the same aerosol species and RMSE values of 0.02, 0.03, 0.01, 0.001, 0.03 and 0.01 for the test data. Permutation feature importance analysis shows that MISR-derived fAODs significantly influence the model’s predictions. The model anticipates aerosol composition to strengthen climate and health effect assessments of aerosols by focussing on their differential effects in low-income south Asian countries.
How to cite: Srivastava, S. and Dey, S.: Estimation of surface-based aerosol composition from satellite data-driven machine learning model over south Asia, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-873, https://doi.org/10.5194/egusphere-egu25-873, 2025.
Understanding the interplay between weather conditions and fine particulate matter (PM2.5) is crucial for improving air quality management and public health. This study investigates the diurnal, seasonal, and spatial variability in the influence of weather factors on PM2.5 concentrations across Poland during 2015–2024. Hourly PM2.5 data from a nationwide network of monitoring stations were analyzed alongside meteorological parameters derived from the Weather Research and Forecasting (WRF) model. Key weather variables included wind speed and direction, precipitation, temperature, atmospheric pressure, relative humidity, solar radiation, and planetary boundary layer (PBL) height.
The machine learning model, XGBoost, was employed to predict PM2.5 concentrations, and the SHAP (SHapley Additive exPlanations) method, optimized using TreeSHAP, was used to assess variable importance and uncover patterns in the data. The results reveal complex, nonlinear relationships between meteorological factors and PM2.5, with notable variability across time and space. Seasonal and diurnal analyses highlight that weather influences are context-dependent: for instance, precipitation has negative impact on PM2.5 concentrations in winter, but its effect diminishes during summer months. Additionally, the role of PBL height was found to vary diurnally, reflecting their influence on pollutant dispersion. Spatial differences in weather-PM2.5 relationships were also observed, emphasizing the role of local topography, urbanization, and emission sources.
The comprehensive analysis provides a decade-long perspective on how weather factors influenced PM2.5 concentrations in Poland, offering valuable insights for regional air quality modeling and policy interventions aimed at mitigating air pollution under varying climatic conditions.
How to cite: Vovk, T. and Kryza, M.: Diurnal, seasonal, and spatial patterns of weather influences on PM2.5 concentrations in Poland: An explainable machine learning approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10255, https://doi.org/10.5194/egusphere-egu25-10255, 2025.
This study introduces a novel artificial neural network (ANN)-based methodology for predicting the Incremental Lifetime Cancer Risk (ILCR) in urban environments, leveraging weather parameters and PM2.5 concentrations. The innovative approach addresses the limitations of conventional ANN models, enabling superior performance and applicability across diverse geographical locations. The model incorporates a unique method to preprocess wind direction and speed into a singular representative factor, enhancing its adaptability and generalizability to different sites.
To validate the proposed methodology, one year of weather data and polycyclic aromatic hydrocarbons (PAHs) data were collected from two distinct sites in India. PAHs were analyzed using gas chromatography-mass spectrometry (GC-MS) to calculate ILCR. These data served as inputs for the ANN models. The conventional ANN model yielded a coefficient of determination (R²) of 0.73 and a mean squared error (MSE) of 0.0100. In contrast, the proposed method achieved significantly improved performance, with an R² of 0.93 and an MSE of 0.0031.
This improvement highlights the efficacy of the novel preprocessing technique, which optimally integrates meteorological parameters, particularly wind-related factors, into the modelling framework. Moreover, the proposed model’s ability to generalize across multiple sites allows it to be trained on larger datasets, thereby enhancing its robustness and reliability for predicting ILCR in various urban areas.
The study's findings emphasize the importance of refining input parameter representation in ANN-based environmental risk models to achieve superior accuracy and broader applicability. This work not only demonstrates the feasibility of using advanced AI techniques to assess public health risks but also offers a scalable solution for multi-site applications, paving the way for better-informed environmental and public health policies.
How to cite: Singh, S., Vishwakarma, P., and Gupta, T.: Development of a Novel ANN-Based Predictive Model for Multi-Site ILCR Estimation Using Weather Parameters and PM2.5, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-628, https://doi.org/10.5194/egusphere-egu25-628, 2025.
Ground-based weather radar provides crucial information for severe weather monitoring and forecasting, but it faces coverage limitations in regions with complex terrain, especially for oceanic and mountainous regions. To address the limitation, this study proposes "Echo Reconstruction UNet (ER-UNet)", a novel deep learning approach that reconstructs radar composite reflectivity (CREF) using Fengyun-4A geostationary satellite observations with broad coverage. The proposed ER-UNet enhances the UNet architecture by integrating wavelet transforms and multi-scale feature extraction mechanisms, significantly improving the network's capacity to capture detailed radar echo characteristics. Experimental results demonstrate that ER-UNet achieves superior performance compared to UNet, with improvements in statistical metrics that include root mean square error (RMSE), mean absolute error (MAE), and structural similarity index measure (SSIM), as well as categorical verification scores including probability of detection (POD), false alarm rate (FAR), critical success index (CSI), and Heidke skill score (HSS). Case studies further reveal ER-UNet's enhanced capability in reconstructing strong echo features, particularly in terms of intensity distribution and spatial structure. The proposed method shows potential for providing reliable radar reflectivity estimates in areas with limited radar coverage, offering valuable support for severe weather monitoring and early warning services.
How to cite: Zhao, J., Chen, S., Tan, J., Huang, Q., Gao, L., Li, Y., and Wei, C.: Reconstruction of Radar Composite Reflectivity Based on Satellite Observations and Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8958, https://doi.org/10.5194/egusphere-egu25-8958, 2025.
Clouds of any kind play a substantial role in a wide variety of atmospheric processes. They are directly linked to the formation of precipitation, and significantly affect the atmospheric energy budget via radiative effects and latent heat. Hence, knowledge of currently occurring cloud types allows the observer to draw conclusions about the short-term evolution of the state of the atmosphere and hence also the weather. However, the number of operational cloud observations is rather decreasing than increasing due to high monetary and personnel expenses.
To show that automatized methods can be used to close this emerging gap, we trained an ensemble of 10 identically initialized residual neural network architectures from scratch to classify clouds from ground-based RGB pictures into 30 different classes. 4 pictures are used as input at each instance, so that the whole visible sky is covered. Operational manual cloud classification reports at the nearby station Vienna HoheWarte are used as ground truth, where for each instance up to 3 out of 30 categories are reported according to the state-of-the-art WMO cloud classification scheme for operational synoptic observations, making this a multi-label classification task. To the best of our knowledge we are the first to automatically classify clouds based on this elaborate classification scheme. Weutilize class specific resampling to reduce prediction biases because of highly imbalanced observation frequencies among categories. Results show that precision and recall scores are high in all classes, although in initially small classes overfitting is supposed to be the reason for exceptionally high accuracy. Still, every member of our ensemble outperforms both random and climatological predictions in each class. A substantial ratio of wrongly assigned pictures is made up by false negative predictions, where the model recognized the correct class in the input but the assigned probability was too small. For further improvement of current results, we aim to include also satellite images and measurement data, e.g. cloud base height, into our classifier. Though additional data is not supposed to solve overfitting issues, we expect to reduce the number of false negative and false positive predictions substantially.
Autonomy and output consistency are the main advantages of such a trained classifier, hence we consider operational cloud monitoring as main application. Either for consistent cloud class observations or to observe the current state of the weather and its short time evolution with high temporal resolution, e.g. in proximity of solar power plants. There, upcoming clouds can substantially change the possible energy output, which leads to the necessity of taking precautions.
How to cite: Rosenberger, M., Dorninger, M., and Weissmann, M.: Utilizing a residual neural network ensemble for ground-based cloud classifications, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10253, https://doi.org/10.5194/egusphere-egu25-10253, 2025.
The organization of mesoscale cloud fields is crucial for understanding atmospheric dynamics and their modeling.
The role played by these cloud structures on their direct environment and, more generally, on the climate remains challenging to incorporate in climate models.
In this contribution, we propose a methodology for automatically identifying and studying these cloud organizations, combining two innovative approaches:
(1) an Ising-like model (called BICIM) capable of reproducing cloud fields with specific organization patterns, and
(2) graph theory, applied to the outputs of this model and satellite observations, which allows us to derive distributions of surfaces and perimeters of clouds as well as inter-cloud distance distributions.
For the first point, sensitivity tests on input data and fluxes revealed that BICIM consistently responds to changes, produces realistic results, and highlights humidity and wind as key factors in the formation of cloud organizations.
From the second point, we propose a new quantity, denoted as M, as a Metric for Assessing Similarity between Cloud Organization Layouts (MASCOL).
As its name suggests, M quantifies the similarity between two cloud fields.
We apply this methodology to satellite data, identifying cloud structures, with MASCOL aiding in classifying cloud fields relative to reference organizations (from BICIM).
This approach therefore enables a faster and more objective identification of each structure compared to visual methods.
The reason is that it is based on graph theory, which is an efficient mathematical tool.
Furthermore, the methodology's reliance on graph theory and robust pattern recognition metrics makes it particularly well-suited for integration with machine learning and artificial intelligence techniques, opening avenues for automated analysis and large-scale applications.
How to cite: Mascaut, F., Pujol, O., and Forkman, P.: Objective characterization of mesoscale cloud patterns from graph theory and an Ising-like model (BICIM)., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10486, https://doi.org/10.5194/egusphere-egu25-10486, 2025.
Climate change is a pressing reality, with increasing extreme weather events such as droughts, floods, and heatwaves causing significant damages and casualties [1]. To mitigate these impacts, climate neutrality has become a global priority. This requires a transition to renewable energy, necessitating accurate weather and climate models [2]. High-resolution Regional Climate Models (RCMs) offer detailed projections but are computationally expensive. Statistical downscaling techniques provide a more efficient alternative but have limitations, such as the inability to capture relevant climate change signals and underestimating extreme events. To address these issues, we propose a physics-informed artificial intelligence (AI) model bridging the gap between data-driven and model-driven approaches, by incorporating known physical principles and domain knowledge into the learning and prediction process [3,4].
In this research, we focus on developing a physics-informed AI model for efficient downscaling of climate and weather data, enabling high-resolution projections that enhance renewable energy predictions. More specifically, we aim for improving downscaling techniques, reducing uncertainties, and accurately representing extreme weather events. Key research questions include identifying suitable datasets for downscaling, evaluating errors, and improving multivariate downscaling from coarse (100 km for GCM, 10 km for RCM) to high resolutions (5 km to 1 km). Our developed method is compared against dynamical downscaling techniques across different temporal and spatial resolutions. This research aims to advance climate and weather predictions for impact sectors in need of very high spatial resolutions through providing an efficient and fast AI-based downscaling method, particularly for renewable energy applications, aiming at supporting decision-making and adaptation strategies in the face of climate change.
References
[1] IPCC, 2021 IPCC, 2021: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Masson-Delmotte, V., P. Zhai, A. Pirani, S.L. Connors, C. Péan, S. Berger, N. Caud, Y. Chen, L. Goldfarb, M.I. Gomis, M. Huang, K. Leitzell, E. Lonnoy, J.B.R. Matthews, T.K. Maycock, T. Waterfield, O. Yelekçi, R. Yu, and B. Zhou (eds.)]. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, In press, doi:10.1017/9781009157896.
[2] Schaeffer, Roberto, Alexandre Salem Szklo, André Frossard Pereira De Lucena, Bruno Soares Moreira Cesar Borba, Larissa Pinheiro Pupo Nogueira, Fernanda Pereira Fleming, Alberto Troccoli, Mike Harrison, and Mohammed Sadeck Boulahya. 2012. ‘Energy Sector Vulnerability to Climate Change: A Review’. Energy 38 (1): 1–12. https://doi.org/10.1016/j.energy.2011.11.056.
[3] Harder, Paula, Alex Hernandez-Garcia, Venkatesh Ramesh, Qidong Yang, Prasanna Sattigeri, Daniela Szwarcman, Campbell Watson, and David Rolnick. 2022. ‘Hard-Constrained Deep Learning for Climate Downscaling’. arXiv. https://doi.org/10.48550/ARXIV.2208.05424.
[4] Karniadakis, George Em, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. 2021. ‘Physics-Informed Machine Learning’. Nature Reviews Physics 3 (6): 422–40. https://doi.org/10.1038/s42254-021-00314-5.
How to cite: Thiele, P. L., Lampert, J., Bügelmayer-Blaschek, M., Baier, K., Hasel, K., Schellander-Gorgas, T., and Schicker, I.: Physics-Informed AI for Enhanced Climate Downscaling and Extreme Event Prediction in the Energy Sector, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13106, https://doi.org/10.5194/egusphere-egu25-13106, 2025.
With the advent of the AI era, deep learning has been actively applied to global weather prediction, achieving remarkable progress. Furthermore, these deep learning-based global prediction models are being utilized for seasonal forecasting, with efforts underway to extend the forecast lead time. Leveraging the memory effect of the ocean is essential for such advancements. In this study, we developed a deep learning-based global three-dimensional ocean model, incorporating three key innovations: (1) expanding the receptive field and reducing the number of parameters using a visual attention network, (2) eliminating ocean/land boundary effects through the application of partial convolution, and (3) aligning prediction value distributions with observations using adversarial loss. Compared to persistence forecasts and NMME models, our model demonstrated global three-dimensional ocean simulation capabilities comparable to state-of-the-art coupled general circulation models, achieving significant improvements, particularly in predicting horizontal ocean currents. Furthermore, the model realistically simulated the ocean’s response to surface boundary forcing. These results highlight the potential for developing a deep learning-based ocean-atmosphere coupled general circulation model.
How to cite: Kim, J.-H., Kang, D., Yang, Y.-M., and Park, J.-H.: Deep learning for global three-dimensional ocean modeling with physical response consistency, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15264, https://doi.org/10.5194/egusphere-egu25-15264, 2025.
Deep convective systems (DCSs) play a crucial role in the tropical hydrological cycle and radiative budget (Stephens et al., 2023; Roca et al., 2014). In particular, the largest and longest-lived of those cloud systems contribute to a high fraction of the extreme precipitation in the Tropics (Roca and Fiolleau, 2020). Therefore understanding what drives these types of systems is crucial. To that end, Abramian (2023) developed a new method to predict the maximum area of DCSs using the DYAMOND-Summer simulation with the cloud-resolving global model SAM, and the TOOCAN algorithm to track cloud systems. The method uses simple machine learning models, trained on information on the early stage of the systems and their surrounding environment, including dynamical and thermodynamical variables, morphological features of the systems and the characteristics of their neighbors.
We improve this method by incorporating an integrated gradients (IG) approach, which provides a more precise quantification of the importance of each input variable directly from the neural network model. Furthermore, we embedded the neural network outputs into a causal discovery framework by identifying the variables that explain the most variance, using the IG method. These key variables were then subjected to a causal discovery analysis, enabling the identification of causal drivers that influence the maximum extent of the systems at various stages of their lifecycle.
This approach improves both interpretability and includes causal inference to avoid non-causal relations. Preliminary results suggest that during the early stages (0.5 hours after the onset of the DCS), the strength of vertical velocity and upper tropospheric saturation explain most of the variance in the system’s maximum area. Interestingly, the presence of neighboring systems also plays a significant role, likely because a smaller number of neighbors allows more moisture and energy to be available for the DCS to grow. In contrast, during the later stages (around 3.5 hours after the DCS onset), when the area reaches its maximum, neighboring systems no longer contribute significantly to the variance. At this stage, thermodynamic factors, particularly
moisture and temperature, emerge as the primary drivers, with the 2-meter temperature playing a particularly important role, suggesting a potential role of cold pools in determining the maximum area of the system.
How to cite: Casallas, A., Polesello, A., Muller, C., and Abramian, S.: From Prediction to Causation: Understanding theDrivers of Maximum Deep Convective Systems Area, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15746, https://doi.org/10.5194/egusphere-egu25-15746, 2025.
Accurate numerical weather prediction is a prerequisite for disaster prevention and mitigation. Based on the numerical ocean-atmosphere-wave coupling forecasting model developed by our team, artificial intelligence technology is introduced to correct prediction bias, addressing the systematic forecast error inherent in numerical models, and further enhancing the accuracy and reliability of our forecast products. This study aims to establish a bias correction model for the numerical forecast products and integrate multi-source geographic information, such as elevation, land cover, and soil type, to improve the forecast results. Convolutional Neural Networks (CNNs) effectively extract spatial features through the computational mechanisms of their convolutional modules, making them well-suited for tasks like meteorological forecast correction, where spatial correlations have a significant impact. We employ a Residual Convolutional Neural Network to efficiently extract the spatial features from numerical model forecast results, leading to improved correction performance compared to traditional correction methods. A spatiotemporal feature extraction module is utilized to adaptively detect terrain and surface features at different scales, addressing the extraction and fusion of multi-source heterogeneous features. We further evaluate and optimize the impact of terrain and land surface features at different resolutions on model correction effectiveness, maximizing prediction accuracy. This research will provide strong technical support for enhancing the precision of ocean-atmosphere-wave coupling forecasting products and offer reliable data assurance for disaster prevention and mitigation efforts.
How to cite: Zhang, S., Chen, R., Huang, Y., and Zhang, X.: Deep Learning based Bias Correction of Forecast Products with Fusion of Geographical Information, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16620, https://doi.org/10.5194/egusphere-egu25-16620, 2025.
The increasing frequencies of extreme weather events like heavy precipitation, drought, heatwaves, etc, have been associated with climate change in recent years. The reliability of air temperature forecasts at 2 meters above the surface is vital when trying to prepare for potential weather-related disasters, such as heat waves. In recent years, there has been a lot of emphasis placed on the prediction of heatwave conditions over India by using deterministic Numerical Weather Prediction (NWP) models. Despite improvements in model physics and resolution, deterministic NWP models have difficulties predicting extreme events at longer lead times. As the model integrates over time, errors grow due to the uncertainty associated with the initial conditions. This uncertainty is taken into account using ensemble prediction systems (EPSs). Heatwaves are now being predicted in India using EPSs due to their better performance in predicting events with longer lead times. The intensity of extreme events is typically underestimated by these models because EPSs typically have a low resolution and are also affected by the systematic biases present in the parent deterministic models. So to make the forecast more reliable, bias correction of the maximum temperature forecasts from EPSs is required.
This study focuses on the comparative assessment of various machine learning techniques for bias correction of maximum temperature from ensemble forecasts of maximum temperature for Hyderabad station. Three machine learning techniques were used in this study, namely Random Forest, Gradient Boost, and Support Vector Machine. The temperature forecasts used in this study were obtained from the National Centre for Medium Range Weather Forecasting (NCMRWF) global ensemble prediction system (EPS—called the NEPS). The NEPS configuration is based on the UK Met Office Global and Regional Ensemble Prediction System (MOGREPS). The climatology used in this study is obtained from the MOGREPS data available on TIGGE and the observations for the maximum temperature are from the Indian Meteorological Department (IMD) station data (1985-2021), this data is used as the training set. The objective of this research study is to improve the accuracy of temperature forecasts by utilizing machine learning techniques for the bias correction of maximum temperature in order to improve model performance, primarily based on metrics such as Root Mean Square Error (RMSE). The initial raw RMSE values for Day 3, Day 5, and Day 7 are recorded as 2.1461, 2.4741, and 2.811, respectively. By examining the refined RMSE values for these specific forecast days, model corrections are revealed using Support Vector Machines (SVM), Gradient Boosting (GB), and Random Forest (RF) . After correction, the SVM model achieves improvements of 18.42%, 26.29%, and 27.42% in RMSE, demonstrating its increased predictive accuracy for Days 3, 5, and 7. Similarly, the RMSE reductions for GB on Day 3, Day 5, and Day 7 are observed at 18.77%, 26.23%, and 28.16%, while RF exhibits reductions of 39.21%, 28.24%, and 22.5% for the corresponding forecast days. The percentage reductions indicate the improved accuracy attained by bias correction employing various machine learning methods.
Keywords: Heat Waves, Ensemble Prediction Systems, Support Vector Machine, Random Forest, Gradient boost.
How to cite: Sharma, S., Chakraborty, A., Dube, A., Singh, H., and Ashrit, R.: Bias Correction of Maximum Temperature Forecasts for Ensemble-Based model using Various Machine Learning Techniques for Hyderabad Station, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18422, https://doi.org/10.5194/egusphere-egu25-18422, 2025.
Posters virtual: Tue, 29 Apr, 14:00–15:45 | vPoster spot 5
EGU25-14960 | Posters virtual | VPS2
Air Quality Assessment In The University Of The Philippines Diliman Campus Through The Integration Of Small Sensors, Satellite Data, And Kriging Interpolation TechniquesTue, 29 Apr, 14:00–15:45 (CEST) | vP5.16
Air quality monitoring is an essential procedure to ensure that pollutant levels remain within safe limits and do not pose a threat to public health, particularly for vulnerable populations. The deployment and maintenance of stationary air quality monitoring stations can be expensive, especially when a large number is required to create a comprehensive network. As a result, there has been growing interest in utilizing small, low-cost sensors that are easier to deploy and provide a more flexible and cost-effective alternative. In addition to these sensors, satellite systems have become valuable tools for air quality monitoring, offering high temporal resolution data that facilitates the assessment of air pollution over larger areas. This study looks into data fusion techniques to combine data from both stationary and mobile low-cost sensors with satellite data to analyze the air quality at the University of the Philippines, Diliman campus. Seven small sensors were deployed across the university, a mixed-use area with both vegetation and buildings, to measure pollutant concentrations, such as particulate matter. Satellite data from MODIS, Sentinel-5P, and ERA5 reanalysis were used to monitor aerosol optical depth (AOD), sulfur dioxide (SO2), nitrogen dioxide (NO2), and meteorological conditions. The time-series analysis focused on a three-day period during which mobile air quality data from an e-trike were collected around the university. The data from these mobile sensors, along with the stationary sensor measurements, were used to estimate PM2.5 concentrations across the campus. Kriging interpolation, a geostatistical method that estimates unknown values based on the spatial correlation of known data points, was employed to generate smooth surfaces of PM2.5 concentration across the university. Kriging interpolation was used on the stationary sensor dataset to predict the PM2.5 levels at the location of the mobile sensors at a given timeframe. Moreover, cokriging was also applied by incorporating multiple correlated variables, improving predictions by utilizing relationships between the primary variable (PM2.5) and secondary variables, such as aerosol optical depth or SO2 and NO2 concentrations. The results obtained from both Kriging and Cokriging methods were compared with data collected from mobile sensors to assess the air quality at the University of the Philippines, Diliman. The interpolated PM2.5 values were compared with the data from the mobile sensors (SEN55 and PMS7003) as ground truth, and a mean absolute percentage error (MAPE) of 43.00% to 57.23% was obtained. Initial results of cokriging with NO2 showed MAPE of 36.67% to 52.55%. Further work is expanding the dataset and refining the interpolation models to enhance the accuracy and reliability of air quality assessments across the university. By integrating more data and conducting additional tests, this approach can provide more comprehensive air quality monitoring at reduced costs and address data gaps.
How to cite: Hizon, J. R., Torres, R. A., Togonon, A. C. E., Recto, B. A., Apostol, F. A., Magpantay, P., Eslit, J. J., Ganhinhin, J., Rosales, M., Austria, I., de Guzman, J., de Leon, M. T., Cajote, R., Co, P. J., and Ramos, R.: Air Quality Assessment In The University Of The Philippines Diliman Campus Through The Integration Of Small Sensors, Satellite Data, And Kriging Interpolation Techniques, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14960, https://doi.org/10.5194/egusphere-egu25-14960, 2025.
EGU25-1542 | ECS | Posters virtual | VPS2
Rainfall Prediction using Hybrid CNN-LSTM approach: A case study in the Boudh district, Odisha, IndiaTue, 29 Apr, 14:00–15:45 (CEST) | vP5.17
The forecast of monthly rainfall is a significant topic for water resource management and hydrological disaster prevention. A critical need for precise hydrological forecasts in water resource management is addressed in this study by analyzing machine learning (ML) models for precipitation forecasting in the Boudh district of Odisha, India. Although machine learning (ML) models have demonstrated significant promise in rainfall forecasting due to their high performance, often surpassing that of certain physical models, the intricate physical processes involved in rainfall creation mean that a single ML model is typically insufficient to provide reliable rainfall projections. A thorough set of meteorological parameters, including precipitation wind speed, temperature, and humidity, are utilized to create four distinct models: Support Vector Regression (SVR), long and short memory neural networks (LSTM), Bi-LSTM and Convolutional neural network with LSTM (CNN-LSTM). The performance of these models is thoroughly assessed utilizing a range of evaluation metrics. In this work, the correlations between precipitation and climate factors are assessed using the cross-correlation function (XCF). With maxima consistently reported during months across all four sites, the XCF analysis shows a number of significant trends, including a strong correlation amid precipitation and maximum temperature. Moreover, precipitation is significantly correlated with wind speed and relative humidity. The results demonstrate the effectiveness of hybridized ML techniques in raising the precision of precipitation forecasts. The CNN-LSTM models, which have R2 values between 0.93 and 0.97, generally perform better. Their remarkable accuracy highlights their efficacy in precipitation forecasting, outperforming rival models during both training and testing. These findings have important ramifications for hydrological processes, particularly in Odisha's Boudh region, where sustainable water resources management depends on precise precipitation forecasting.
How to cite: Samantaray, S., Sahoo, A., and Satapathy, D. P.: Rainfall Prediction using Hybrid CNN-LSTM approach: A case study in the Boudh district, Odisha, India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1542, https://doi.org/10.5194/egusphere-egu25-1542, 2025.
EGU25-13882 | ECS | Posters virtual | VPS2
Leveraging Large Language Models for Enhancing and Reasoning Adverse Weather Hazard ClassificationTue, 29 Apr, 14:00–15:45 (CEST) | vP5.18
Severe weather events often develop rapidly and cause extensive damage, resulting in billions of dollars in losses annually. This paper explores Large Language Models (LLMs) to effectively reason about the adversity of weather hazards. To tackle this issue, we gathered National Weather Service (NWS) flood reports covering the period from June 2005 to September 2024. Two pre-trained LLMs including Bidirectional and Auto-Regressive Transformer (BART) models (large) and Bidirectional Encoder Representations from Transformers (BERT) were employed to classify flood reports according to predefined labels. These models encompass a range of sizes with parameter counts of 406 million, and 110 million parameters, respectively. We employed the Low-Rank Adaptation (LoRA) fine-tuning technique to enhance performance and memory efficiency. The fine-tuning and few-shot learning capabilities of these models were evaluated to adapt pre-trained language models for specific tasks or domains. The methodology was applied in Charleston County, South Carolina, USA— a vulnerable region to compound flooding. Extreme events reported during the training periods were unevenly distributed across training period, resulting in imbalanced datasets. The “Cyclonic” category represents significantly fewer instances in the report text data, while the “Flood” and “Thunderstorm” categories appeared more frequent. The findings revealed that while few-shot learning significantly reduced computational costs, fine-tuned models resulted in more stable and reliable performance. Among multiple LLMs applied in this research, the BART model achieved higher F1 scores in the “Flood,” “Thunderstorm,” and “Cyclonic” categories—requiring fewer training epochs to reach optimized performance levels. Furthermore, the BERT model demonstrated a shorter overall training time (12 hours 17 minutes) compared to other LLMs, demonstrating efficient cost of computing. This comprehensive evaluation of LLMs across diverse NWS flood reports enhanced our understanding of their capabilities in text classification and offered valuable insights into leveraging these advanced techniques for weather disaster assessment.
How to cite: Neupane, A., Zafarmomen, N., and Samadi, V.: Leveraging Large Language Models for Enhancing and Reasoning Adverse Weather Hazard Classification, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13882, https://doi.org/10.5194/egusphere-egu25-13882, 2025.
EGU25-7965 | Posters virtual | VPS2
An Explainable AI-Driven Feature Reduction Framework for Enhanced Agricultural Yield PredictionTue, 29 Apr, 14:00–15:45 (CEST) | vP5.19
Agricultural yield prediction plays a crucial role in food security and economic planning, yet existing models often struggle with the complexity and high dimensionality of agricultural data. This study presents a framework that combines explainable artificial intelligence (XAI) with feature reduction methodology to enhance the accuracy and efficiency of rice yield prediction. Our approach addresses the dual challenges of model interpretability and computational efficiency while maintaining high prediction accuracy.
The framework begins with a systematic development of prediction models utilizing advanced machine learning (ML) and deep learning (DL) techniques. We implemented comprehensive pre-processing steps, including data normalization, feature engineering, and missing value handling, to ensure data quality. Our evaluation encompassed various models, including Random Forest, Gradient Boosting Machines, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks with attention mechanisms. To optimize model performance, we employed hyperparameter tuning through grid search, effectively mitigating issues of overfitting and underfitting.
A notable innovation of our framework is the incorporation of SHapley Additive exPlanations (SHAP), enabling transparent insights into the model's decision-making process. Leveraging this XAI approach, we introduced a novel feature reduction methodology that systematically identifies and removes negatively contributing features while maintaining model accuracy. Our analysis of a multivariate dataset which is a public dataset from rice fields in the an Giang province of the Mekong Delta, Vietnam, required the integration of diverse satellite datasets, including optical data from Landsat and radar data from Sentinel-1. This revealed distinct patterns of feature influence on yield prediction, facilitating the optimization of the feature set for maximum effectiveness. Key radar polarization bands, VV (Vertical-Vertical) and VH (Vertical-Horizontal), provided crucial surface backscatter data, capturing information on crop structure, growth stages, and post-harvest soil conditions. Notably, the feature min_vh consistently emerged as the most significant predictor.
The implementation of our feature reduction strategy resulted in significant improvements in both model performance and computational efficiency. By removing 15-20 number of identified negatively contributing features, we achieved approximately 3-5% improvement in prediction accuracy while substantially reducing the computational overhead and model training time. This enhancement in efficiency did not compromise the model's interpretability, demonstrating the robust nature of our framework.
Our methodology represents a significant advancement in agricultural modeling by successfully addressing the challenges of high-dimensional data while maintaining model interpretability. The framework's ability to identify and eliminate non-contributing features while improving prediction accuracy demonstrates its potential for wide-scale application in agricultural yield prediction. Furthermore, the reduced computational requirements make it a practical solution for real-world applications where computational resources may be limited.
These results validate the effectiveness of our integrated approach in handling complex agricultural data while providing actionable insights for yield prediction. The framework offers a scalable, interpretable, and computationally efficient solution that can be adapted for various agricultural prediction tasks, potentially transforming how we approach agricultural yield forecasting.
How to cite: Dey, A., Saha, A., Sarkar, S., Mondal, A., and Mitra, P.: An Explainable AI-Driven Feature Reduction Framework for Enhanced Agricultural Yield Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7965, https://doi.org/10.5194/egusphere-egu25-7965, 2025.
EGU25-243 | ECS | Posters virtual | VPS2
Envisioning the Role of Physics-Informed Neural Networks in Atmospheric Science: Advancements, Challenges, and Future ProspectsTue, 29 Apr, 14:00–15:45 (CEST) | vP5.20
Since the inception of physics-informed neural networks (PINNs) by Raissi et al. in 2019, it has been seen as a promising approach to outperform conventional algorithms in terms of computational efficiency, reduced costs, and improved prediction accuracy, especially in small data regimes.PINNs incorporate known physical governing equations in the form of partial differential equations (PDEs) or ordinary differential equations (ODEs) into neural networks, and occasionally the governing equations are derived from observational or simulated data, allowing PINNs to address specific atmospheric systems.Moreover, depending on the problem being solved, most work adds the physical constraints directly into the loss or cost function, while others enhance performance using modified architectures or preprocessing techniques.In the realm of atmospheric sciences, challenges remain, including a heavy reliance on simulated data and limited use of observational datasets, which does not show the real-world applicability of PINNs. A detailed review of available results shows critical gaps in scalability, hybrid data integration, and standardization in atmospheric science.We identified a hybrid methodology by combining simulated and observational data, which includes optimizing hybrid loss functions to balance physics-based and observational accuracy, applying adaptive training techniques, and standardizing preprocessing schemes to handle multi-scale atmospheric phenomena.Results demonstrate the ability of PINNs to deliver faster computation, enhanced prediction accuracy, and robustness in sparse data environments. This highlights the transformative advantages of PINNs over traditional methods and suggests future directions for leveraging their capabilities in atmospheric science applications.
How to cite: Ekue, J. A., Hammond, D., and Agyei-Yeboah, E.: Envisioning the Role of Physics-Informed Neural Networks in Atmospheric Science: Advancements, Challenges, and Future Prospects, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-243, https://doi.org/10.5194/egusphere-egu25-243, 2025.
