In this session, we welcome submissions addressing the latest progress in new techniques applied to research on all aspects of atmospheric, climate, and environmental sciences, including but not limited to,
- The application of AI/ML and other techniques for:
• Advancing the understanding of the complex earth system, especially the underlying mechanisms of weather/climate system, atmospheric environmental system, and their interactions
• Facilitating faster and more accurate weather/climate/air quality modeling and forecasting, especially for extreme weather, climate change, and air pollution episodes
• Shedding new insights into the mechanisms of atmospheric chemistry and physics
• Achieving air pollution tracing and source attribution
• Assisting policymakers on decisions towards environmental sustainability (e.g., considering interactions between extreme weather, climate change, air quality, socio-economics, and public health
- The adaptation and development of AI/ML and other techniques by proposing:
• Explainable AI (XAI)
• Hybrid methods (e.g., hybrid ML, physics-integrated ML)
• Transfer learning
• New algorithms
• Advanced model frameworks
We believe that exchanges across research fields could help breaking down the limitations of thinking and enabling technological innovations. Therefore, contributions from fields other than atmospheric, climate, and environmental sciences are also encouraged.
Seasonal droughts and floods are among the world’s most severe natural disasters, threatening socioeconomic development and human life and property, making accurate seasonal prediction a key need for disaster risk reduction. Yet, the high complexity of precipitation anomalies driven by multi-scale variations, multiple factors, and atmospheric nonlinear chaos, which makes such forecasts a global challenge. Climate modes (e.g., ENSO), as major drivers of seasonal precipitation anomalies, are essential predictors. Systematically assessing their contributions to predictability and developing prediction methods based on their influence are therefore of great scientific and practical value for understanding and forecasting seasonal droughts and floods.
Addressing seasonal rainfall prediction challenges, the SMART, the acronym for Singular predictable climate Modes (SM) and Anomalous Relative Tendency (ART), climate prediction principle is proposed by Prof. Xiu-Qun Yang from Nanjing University. It contains 4 major steps: 1. Online temporal-scale separation (ART); 2. Extracting optimal singular climate modes (SM); 3. Constructing SMART model based on SM and ART; 4. Predicting with ART and Recent Background Anomalous (RBA). This study develops two ensemble prediction methods based on SMART principle, which combine the impacts of climate modes on China’s flood-season precipitation anomalous relative tendencies (ART) with multiple artificial intelligence (AI) models and multi-parameter perturbation scheme, including the SMART Optimal combined Multiple AI Method (OMAI) and SMART Ensemble AI Method (EAI). These two methods demonstrate significant predictive skill improvements over MME direct predictions. For example, using 160 stations historical precipitation data in China and historical circulation datasets, multiple key tropical and extratropical climate modes affecting to the ART of China’s flood-season (JJA) precipitation are extracted by SVD method. The SMART-OMAI method integrates these modes with multiple AI models, while SMART-EAI incorporates multi-parameter perturbations with LSTM model. Independent validation for flood-season precipitation anomalies in China during 1994-2016 via these two methods, yields PS scores of 76.5 and 76.4, respectively over 5% higher than dynamical-statistical models (73.2) and over 20% better than direct MME direct predictions (63.5). Anomaly correlation coefficients reach 0.16 and 0.14, marking qualitative improvements over MME direct predictions (0.01), with notable temporal correlation enhancements in North China and Northeast China. By merging the physical basis of the climate modes with non-linear predictive strengths of AI models, these two AI-based prediction methods offer a scientifically robust and practical solution, that is, a SMART solution for seasonal rainfall prediction in China.
How to cite: Wang, Y., Yang, X.-Q., and Sun, X.: SMART-AI: A High-Performance Prediction Method for Seasonal Rainfall in China Based on the Impacts of Climate Modes and the Artificial Intelligence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6177, https://doi.org/10.5194/egusphere-egu26-6177, 2026.
The extreme high temperature in western North America (WNA) exerts profound impacts on industrial and agricultural production, and trigger catastrophic wildfires. Exploring the underlying mechanisms influencing extreme hot days over WNA (WEHDs) and improving the seasonal prediction are of great scientific and social significance. This study reveals that two independent precursor signals, the persistent negative sea surface temperature (SST) anomalies in tropical eastern Pacific and the cooling tendency in tropical North Atlantic SST during springtime exhibit significant influence on WEHDs. A physics-based empirical model constructed using these two predictors exhibits robust independent prediction skills. Guided by the underlying physical mechanisms, we integrate SST tendency fields as critical input features into convolutional neural network (CNN) to further enhance the prediction accuracy. The physically informed CNN achieves significantly improved performance and successfully predicts the extreme WEHD events of 2021. The results emphasize the pivotal role of physical cognition in advancing deep learning-based climate prediction.
How to cite: Tan, H. and Zhu, Z.: Refine Extreme Hot Day Predictions with the Sea Surface Temperature Tendency, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8949, https://doi.org/10.5194/egusphere-egu26-8949, 2026.
AI and deep learning are rapidly becoming essential tools in weather and climate science. This presentation will cover our recent work using these techniques to enhance predictions across various timescales and phenomena. We have successfully applied architectures like convolutional neural networks, transformers, and generative models to forecast events like ENSO and the Indian Ocean Dipole, as well as to correct biases in traditional climate models and to downscale coarse-resolution outputs using diffusion framework. Looking ahead, I will also discuss our efforts in building AI large models for ensemble subseasonal-to-decadal forecasting and the exciting prospect of creating AI agents dedicated to climate research.
How to cite: Luo, J.-J. and Ling, F.: Developing deep-learning models for weather-climate forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-369, https://doi.org/10.5194/egusphere-egu26-369, 2026.
The rapid loss of Arctic sea ice is a striking consequence of anthropogenic global warming. Itsremote impacts on mid‐latitude weather and climate have attracted scientific and media attention. In this study,we use a hybrid (dynamical plus machine‐learning) atmospheric model—Google's NeuralGCM—to investigatethe mid‐latitude atmospheric circulation responses to Arctic sea‐ice loss for the first time. We conductexperiments in which NeuralGCM is forced with pre‐industrial and future sea‐ice concentrations following theprotocol of the Polar Amplification Model Intercomparisom Project. To assess the performance of NeuralGCM,we compare the results with those simulated by two physics‐based climate models. NeuralGCM produces acomparable response of near‐surface warming to sea‐ice loss and the subsequent weakened zonal wind in mid‐latitudes. However, there is a substantial discrepancy between the two models' stratospheric responses, wheredifferent temperature responses in these models are associated with different zonal wind and geopotential heightresponses. Further investigation of North Atlantic blocking shows that NeuralGCM produces stronger, morefrequent, and more realistic blocking events. Our results demonstrate the capability of NeuralGCM insimulating the tropospheric responses to Arctic sea‐ice loss, but improvements may be needed for thestratospheric representation.
How to cite: Liang, Y.-C., Lutsko, N., and Kwon, Y.-O.: Exploring the Atmospheric Responses to Arctic Sea-Ice Loss in Google's NeuralGCM, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15674, https://doi.org/10.5194/egusphere-egu26-15674, 2026.
Due to their fast computational performance and accuracy, neural networks are nowadays commonly used in the context of remote sensing. The issue of performance is especially important in the context of big data and near-real-time (NRT) operational processing. Classical retrieval algorithms typically use a radiative transfer model (RTM) as a forward model to solve the inverse problem of inferring the quantities of interest from the measured spectra. However, these RTMs are often computationally very expensive and therefore replacing them by a NN is desirable to increase performance. But the application of NNs is not straightforward and there are at least two main approaches:
1. NNs used as forward model, where a NN accurately approximates the radiative transfer model and can thus replace it in the inversion algorithm
2. NNs for solving the inverse problem, where a NN is trained to infer the atmospheric parameters from the measurement directly
The first approach is more straightforward to apply. However, the inversion algorithm still faces many challenges, as the spectral fitting problem is generally ill-posed. Therefore, local minima are possible and the results often depend on the selection of the a-priori values for the retrieval parameters.
For the second case, some of these issues can be avoided: no a-priori values are necessary, and as the training of the NN is performed globally, i.e. for many training samples at once, this approach is potentially less affected by local minima. However, due to the black-box nature of a NN, no indication about the quality of the results is available. In order to address this issue, novel methods like Bayesian neural networks (BNNs), invertible neural networks (INNs) or also variational auto-encoders (VAEs) should be considered as they allow the characterization of the retrieved values by an estimate of uncertainty describing a range of values that are probable to produce the observed measurement.
We apply and evaluate both approaches for the retrieval of cloud properties and consider their potential as operational algorithms for the Copernicus atmospheric composition missions Sentinel-4 and Sentinel-5P.
How to cite: Romahn, F., Loyola, D., Molina García, V., Doicu, A., Lutz, R., and Argyrouli, A.: Retrieval of Cloud Properties for the Copernicus Atmospheric Missions Sentinel-4 (S4) and TROPOMI / Sentinel-5 Precursor (S5P) using deep neural networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20381, https://doi.org/10.5194/egusphere-egu26-20381, 2026.
Generative models are increasingly used for quantitative remote-sensing retrievals, yet the physical interpretability and reliability of their ensemble-based uncertainty estimates remain insufficiently assessed. We introduce RTMDiff, a retrieval framework that couples a conditional diffusion model with radiative transfer model (RTM) simulations to retrieve cloud properties and associated uncertainties from multi-channel thermal infrared (TIR) observations of FY-4B AGRI, enabling consistent day–night retrievals. RTMDiff is evaluated against a Bayesian optimal-estimation (OE) baseline using the same forward RTM, showing that the diffusion-based ensembles yield stable uncertainty estimates while preserving physical consistency. Comparisons with independent MODIS and CALIPSO products further support the realism of the retrieved cloud fields, with particularly clear improvements for low-level, optically thick clouds where pixel-wise OE is constrained by limited spectral sensitivity in TIR.
How to cite: Li, W., Zhang, D., Zhang, F., Zhang, R., and Lu, F.: Cloud Property Retrieval and Uncertainty Estimation from FY-4B AGRI Using Conditional Diffusion, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3959, https://doi.org/10.5194/egusphere-egu26-3959, 2026.
Clouds play a crucial role in the global water cycle and the balance of the energy budget. The unique topographic and thermal conditions of the Tibetan Plateau (TP) have a profound impact on the formation of regional extreme weather and global climate change. However, existing official cloud products from geostationary satellites suffer from the spatiotemporal discontinuity over the TP.
Therefore, this study develops an all‑day retrieval algorithm of cloud physical properties (CPP) from geostationary satellite measurements using a deep learning model, achieving high-precision retrieval of cloud phase (CLP), cloud top height (CTH), cloud effective radius (CER), and cloud optical thickness (COT). This algorithm not only leverages the spatial structural information of clouds to compensate for the limitations of retrieving thick clouds from thermal infrared channels caused by their weak penetration ability, but also effectively combines the observed advantages of geostationary satellites with a wide coverage and polar-orbiting satellites with high precision.
Based on the retrieved CPP products with spatiotemporal continuity, we further adopted a Tracking Of Organized Convection Algorithm through a three-dimensional segmentatioN (TOOCAN-CPP) method to automatically identify and track the deep convection system (DCS) over the TP and its surrounding areas. The results show that, influenced by the South Asian Summer Monsoon and topographic conditions, DCSs are primarily concentrated in the Southern TP, the Southern Himalayas Front, and the Ganges Plain. The diurnal variation of DCS number follows a unimodal pattern, with a phase difference of approximately 2 hours between the two areas. Additionally, diurnal variation in cloud properties of DCSs and their internal regions is revealed for the first time. Quantitative analysis of the DCS properties with different sizes and lifetimes indicates that these two areas are dominated by small-sized DCS with initial DCS lifetimes under 6 hours. These discoveries provide valuable insights into understanding the development and evolution of DCSs and their climatic effects.
How to cite: Zhao, Z., Zhang, F., Li, W., Yang, B., He, Q., and Cai, M.: Retrieval of All-Day Cloud Physical Properties from Geostationary Satellite Measurements and Its Application to the Tibetan Plateau, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3669, https://doi.org/10.5194/egusphere-egu26-3669, 2026.
Sentinel-5P (S5P) plays a central role in global atmospheric and environmental monitoring, yet its coarse spatial resolution limits the analysis of localised emission sources and sharp concentration gradients. Super-resolution (SR) methods have been proposed to address this limitation, but most existing approaches rely on paired low and high-resolution data that are unavailable for S5P, restricting their applicability in real-world settings. In this work, we present a self-supervised hyperspectral SR framework specifically designed for S5P that enables training without high-resolution ground truth. The proposed framework integrates the S5P degradation operator and band-dependent noise characteristics derived from sensor signal-to-noise ratio metadata within a self-supervised learning strategy. Convolutional Neural Network (CNN) architectures tailored to S5P's spectral characteristics based on Depthwise Separable Convolutions (DSC) are introduced to efficiently enhance spatial detail while preserving spectral fidelity. The framework is evaluated across all S5P spectral bands under two settings: (i) reference experiments where supervised and self-supervised learning can be directly compared using synthetic ground truth, and (ii) fully self-supervised settings where high-resolution reference data are unavailable, and assessment relies on physics-based consistency metrics. Results show that the proposed self-supervised models achieve performance comparable to supervised counterparts and produce sharper spatial structures than standard bicubic interpolation. Additional validation using coincident EMIT hyperspectral observations demonstrates that the super-resolved outputs exhibit physically meaningful spatial enhancement, particularly along coastline regions. These findings highlight the potential of the proposed self-supervised framework to improve the effective spatial resolution of atmospheric satellite observations, enabling practical deployment in scenarios where high-resolution reference data are inherently unavailable.
How to cite: Omar Ali, H., Crosnier, A., Abraham, R., Combelles, B., Jegou, F., and Galerne, B.: Self-Supervised Super-Resolution for Sentinel-5P Hyperspectral Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21787, https://doi.org/10.5194/egusphere-egu26-21787, 2026.
How to cite: Xu, Q., Zhang, H., and Sui, Q.: Accelerating Investment Returns in Island Community Microgrids: An AI-Driven, Carbon-Aware Demand Response Framework with Techno-Economic and Environmental Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2511, https://doi.org/10.5194/egusphere-egu26-2511, 2026.
Orals: Tue, 5 May, 08:30–12:30 | Room E2
Midlatitude winter storms are a major cause of economic loss and infrastructure damage across Europe. Although reanalysis datasets, such as ERA5, offer reliable near-surface wind gust fields from 1940 onwards, the limited set of winter storm events remains inadequate for catastrophe models. The LArge Ensemble of Regional climaTe modEl Simulations for EUrope (LAERTES-EU) dataset addresses this limitation by providing over 12,000 years of synthetic climate data, yielding a substantially larger catalogue of possible winter storm events. However, a closer analysis revealed that its coarse spatial resolution (~27 km) systematically underestimates extreme wind gusts, which are critical for catastrophe models. High-resolution regional climate model (RCM) simulations using the Icosahedral Nonhydrostatic (ICON) model at 2.5 km grid spacing can accurately capture these extremes. As dynamical downscaling of the entire LAERTES-EU dataset is computationally extortionate, other solutions are required. This study presents a deep learning-based approach, commonly known as Super Resolution (SR), as a cost-effective alternative. Specifically, a probabilistic Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) was trained using pairs of coarse-resolution (ERA5, ~25 km) and high-resolution (ICON, ~2.5 km) data in order to downscale wind gust fields from approximately 300 historical winter storms. Our results show that the WGAN-GP model generates high-resolution wind gust fields that are statistically similar to the ICON simulations, but with much lower computational costs. The trained model is then employed to downscale wind gust fields of winter storm events from the LAERTES-EU ensemble, producing a large dataset of high-resolution synthetic storm events suitable for detailed risk assessment and climate impact studies.
How to cite: Rasheeda Satheesh, A., Sokol, L., Stadelmaier, K. H., Eisenstein, L., Ludwig, P., Ramos, A. M., Braun, L., Brocklehurst, A., Georgiadis, A., and Pinto, J. G.: An AI-based downscaling tool to generate a large ensemble of high-resolution wind storm footprints over Europe, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12848, https://doi.org/10.5194/egusphere-egu26-12848, 2026.
Urbanization modulates precipitation through thermodynamic, dynamical, and aerosol pathways, yet how the rapidly increasing three-dimensional urban height interacts with rising aerosol pollution to shape urban precipitation at the global scale remains unresolved. Here, we quantify the global urban imprint on precipitation using high-resolution satellite precipitation products and attribute its drivers with a glass-box explainable artificial intelligence (XAI) model, the Explainable Boosting Machine (EBM). After controlling for other factors, we find that higher aerosol burden and greater built-up height each tends to enhance urban precipitation when considered individually, but their interaction is antagonistic: at simultaneously high aerosol concentrations and urban heights, their combined effect suppresses precipitation. Simulations with the Weather Research and Forecasting model coupled with chemistry (WRF-Chem) for Delhi, Dakar, and Oklahoma City corroborate this pattern and indicate that the suppression arises primarily because the urban-height–aerosol interaction damps the circulation response associated with the aerosol direct effect. These results address a key gap in understanding how urban vertical growth and air pollution jointly regulate precipitation. They further suggest that, under continued vertical densification and anticipated emission reductions, the urban precipitation enhancement could intensify, with implications for future urban flood-risk management.
How to cite: Han, Y. and Gao, M.: Urban height-Aerosol synergy drives globally dampened urban rainfall, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1970, https://doi.org/10.5194/egusphere-egu26-1970, 2026.
In order to estimate pollutant plume dispersion at a local scale in accidental release scenarios, it is necessary to estimate the air flow behavior around the affected site. This flow can be computed using Computational Fluid Dynamics (CFD), but such an approach can be computationally intensive. As a promising alternative, deep learning surrogates learned on CFD-generated data usually require cheaper resources at inference time. However, local air flows depend strongly on urban geometry, which is challenging to take into account in deep learning surrogate models. Additionally, deep learning approaches tend to struggle to scale to large meshes required by real-case scenarios. Finally, flow behavior in the Atmospheric Boundary Layer is influenced by atmospheric stratification stability, which modifies the turbulence level in the flow and must be taken into account [2].
To tackle these challenges, we propose an Anchored Branched Steady-state WInd Flow Transformer (AB-SWIFT), a transformer-based model with an internal branched structure uniquely designed for atmospheric flow modeling. AB-SWIFT relies on the anchor attention mechanism [1], allowing scalability to hundreds of millions of mesh points. To the best of the authors’ knowledge, AB-SWIFT is among the first works to apply transformer-type neural networks to atmospheric modeling. It also explicitly accounts for variable atmospheric stratification stability, which is typically neglected in existing models.
We challenge our model on a specially designed database of atmospheric simulations around randomised urban geometries and with a mixture of unstable, neutral, and stable atmospheric stratification. Urban geometries are determined by randomly sampling buildings and positioning them in space. Additionally, for each simulation, the atmospheric stratification stability is varied by sampling values of the Monin-Obukhov length and of the ground roughness. Our model reaches the best accuracy on all predicted fields compared to the state-of-the-art transformers and graph-based models.
Figure 1: Horizontal slice at h = 2m above ground of an AB-SWIFT prediction on an unseen geometry under stable atmospheric conditions.
[1] B. Alkin, M. Bleeker, R. Kurle, T. Kronlachner, R. Sonnleitner, M. Dorfer, and J. Brandstetter. Ab-upt: Scaling neural cfd surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers. arXiv preprint arXiv:2502.09692, 2025.
[2] S. R. Hanna, G. A. Briggs, and R. P. Hosker Jr. Handbook on atmospheric diffusion. Technical report, National Oceanic and Atmospheric Administration, Oak Ridge, TN (USA . . . ,) 1982
How to cite: de Villeroché, A., Le Guen, V., Mouradi, R.-S., Massin, P., Bocquet, M., Farchi, A., Cheng, S., and Armand, P.: Anchored-Branched Steady-state WInd Flow Transformer (AB-SWIFT): a metamodel for 3D atmospheric flow in urban environments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12926, https://doi.org/10.5194/egusphere-egu26-12926, 2026.
This study addresses a critical challenge in AI-based weather forecasting by developing a physics-informed ensemble system (Orthogonal Conditional Nonlinear Optimal Perturbations, O-CNOPs) that bridges the gap between computational efficiency and physical consistency for tropical cyclone (TC) forecasting. Unlike conventional NWP ensembles constrained by computational costs or current AI ensembles limited by inadequate perturbation methods, O-CNOPs generates dynamically optimized perturbations that both capture fastest-growing errors of AI model and maintain physical plausibility. The key innovation lies in its ability to produce orthogonal perturbations that respect the nonlinear dynamics of AI model, yielding physically interpretable probability forecasting and structure of perturbations reflecting dominant dynamical controls. Demonstrating superior deterministic and probabilistic forecasting skills over operational Integrated Forecasting System ensemble prediction system, this work establishes a new paradigm for ensemble forecasting that combines AI's computational advantages with rigorous dynamical constraints. The success in TC track forecasting paves the way for reliable ensemble forecasting across other high-impact weather systems, marking a significant step toward operational AI-based ensemble forecasting system.
How to cite: Duan, W. and Li, Y.: A Synergistic Approach: Dynamics-AI Ensemble in Tropical Cyclone Forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3321, https://doi.org/10.5194/egusphere-egu26-3321, 2026.
The intensity of tropical cyclones (TCs) is highly associated with their structure. Geostationary satellite cloud products provide rich information about the TCs’ structure like storm morphology, and can be used for estimating TC intensity. This study utilizes the Swin-Unet architecture as a backbone model for an objective deep learning (DL)-based TC intensity estimation method over the Western North Pacific. This model incorporates several key components, including the self-attention mechanism, shift-window mechanism, and Unet structure. The most important point in this study is that the model introduces a rotation index and a dispersion index as part of the loss function to characterize storm morphology. These two indexes can be computed based on the comprehensive feature extraction from time-series geostationary satellites imagery. The input of this model includes five cloud products from the Fengyun series geostationary satellites: sectional image (SEC), cloud top temperature (CTT), temperature of the brightness black-body (TBB), precipitation estimation (PRE), and humidity profile derived from cloud analysis (HPF). Results show that the model obtains an exceptionally low mean absolute error (MAE) of 3.71 m/s and root mean square error (RMSE) of 5.05 m/s. Furthermore, the ablation study (component-impact analysis) was conducted to quantify the contribution of the rotation index and dispersion index which enhance the model’s estimation performance to some extent. Finally, through an analysis of feature importance across the five cloud products, HPF, CTT, and TBB received higher importance scores, indicating the model concentrates on the thermodynamic and dynamic features that are strongly associated with TC convective activities. This study is expected to provide hydrometeorological departments with technical support for real-time TC intensity estimation in coastal regions and contribute to disaster warning systems.
How to cite: Chen, S. and Tan, J.: Estimating tropical cyclone intensity based on deep learning and satellite imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16605, https://doi.org/10.5194/egusphere-egu26-16605, 2026.
Machine learning (ML), particularly deep learning (DL), is becoming increasingly central to geophysical data assimilation (DA), serving to enhance classical methods, complement them, or potentially replace parts of the DA cycle altogether. This talk reviews recent developments and outlines promising directions for integrating ML into DA, and ultimately improve forecasting in the geosciences. For instance, ML can be used to develop auto-differentiable emulators for dynamics, parametrisations, or model-error corrections, which can be seamlessly incorporated into variational DA frameworks. ML also enables adaptive and efficient exchanges of information among the state, observation, and latent spaces in which DA analysis computations occur. In ensemble DA, ML can improve forecast ensemble generation and facilitate the efficient tuning of hyperparameters through auto-differentiable DA implementations. Moreover, ML opens the possibility of learning and replacing the analysis step, or even the full DA and forecast cycle, in an end-to-end manner. I will illustrate these opportunities with two examples: one in which DL is used to discover new and efficient analysis operators, and another in which generative AI is embedded within classical DA schemes.
How to cite: Bocquet, M.: Machine learning–driven advances in geophysical data assimilation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3932, https://doi.org/10.5194/egusphere-egu26-3932, 2026.
Currently, weather forecasting still relies primarily on Numerical Weather Prediction (NWP) models. While recent advances in machine learning (ML) have demonstrated the potential of ML-based forecasting models to revolutionize NWP, these models often struggle to accurately estimate the initial atmospheric states from raw observations and generate precise weather forecasts. To address this challenge, FuXi Weather introduces an innovative machine learning-based paradigm that integrates multi-source global observations to generate high-resolution analysis fields and medium-range forecasts. Notably, its performance across the vast majority of forecast targets is comparable to the ECMWF High-Resolution (HRES) model. This breakthrough signifies that AI-driven meteorological systems have evolved from experimental prototypes into mature, real-world solutions capable of competing with the most sophisticated traditional NWP frameworks.
How to cite: xu, X. and sun, X.: FuXi Weather-2: A unified neural paradigm for accurate global weather assimilation and forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10666, https://doi.org/10.5194/egusphere-egu26-10666, 2026.
Hailstorms are among the costliest extreme weather events, causing major damage to agriculture and infrastructure, and leading to substantial losses in the insurance sector.
Hailstorm modeling is extremely challenging and requires computationally expensive high resolution physical modeling. Machine Learning approaches have recently emerged as a way to bypass some limitations of physical models, combining hail reports as target data with meteorological predictors. Limitations to this approach appear from the scarcity of consistent observational data in most regions.
In this study, we compare domain-shift adaptation methodologies to propose an optimal approach to produce hailstorm models in data scarce regions. Results show that models combining data from data-rich and data-scarce regions offer the best balance between regional skill and cross-domain generalization.
Furthermore, we introduce a probability calibration methodology to improve interpretability of the model inferences and demonstrate how these models can be used to construct hailstorm hazard maps, providing valuable tools for stakeholders.
In addition to historical climatology, we present results for hail climate projections.
This work has been partially funded by the EDF Project KOIOS GA 101103770
How to cite: Bueso, D., Sanchez-Marroquin, A., Lovo, A., Baladima, F., and Rodríguez, M.: Constructing statistical models for hailstorm occurrence in US and Europe. , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20845, https://doi.org/10.5194/egusphere-egu26-20845, 2026.
Ice aggregation is a fundamental driver of cloud evolution and precipitation formation. However, quantifying its rate in natural environments remains challenging due to the difficulty of tracking particle history in a Lagrangian frame. To address this issue, we use a unique dataset from 21 targeted glaciogenic seeding experiments (CLOUDLAB, Henneberger et al., 2023) conducted in supercooled stratiform clouds ranging from -7.8 °C to -4.7 °C. This experimental design establishes a controlled initial state and advection time (5–10 minutes). Central to our methodology is IceDetectNet (Zhang et al., 2024), a deep learning architecture that applies in situ holographic imagery to detect and classify individual monomers within complex aggregates. Quantifying the number of collisions per aggregate at the monomer level allows us to reconstruct the initial ice crystal number concentration (ICNCt0) directly from downwind observations.
To disentangle the microphysical and environmental drivers of aggregation, we implemented a comprehensive analytical workflow that integrated three distinct paradigms: data-driven causal inference, a theoretically derived physical equation, and machine learning regressors. These independent approaches converge on the conclusion that ICNCt0 parameter is governing aggregation, significantly outweighing the influence of temperature, turbulence, or aspect ratio. Our analysis reveals a significant departure from classical collection theory: the aggregation rate exhibits sub-quadratic power-law dependence on initial concentration (mean exponent 0.92; 95% confidence interval CI: 0.88–0.97), contrasting with the traditional quadratic scaling assumed in kinetic collection kernels. We hypothesize that this scaling involves aggregation among smaller crystals, where subsequent diffusional growth masks the boundaries between monomers, making early collisions difficult to detect. Furthermore, benchmarking eleven machine learning architectures against the physically derived formulation revealed a clear trade-off. While CatBoost's gradient boosting ensembles achieved higher statistical accuracy (R² = 0.87), the theoretical model showed greater robustness and generalizability in sensitivity testing. This multi-perspective framework uses a combination of experimental atmospheric physics and AI-driven interpretation to demonstrate how data-driven plasticity and physically-based stability complement each other. It provides a practical approach to understanding complex microphysical processes.
Reference:
Henneberger J, Ramelli F, Spirig R, et al. Seeding of supercooled low stratus clouds with a UAV to study microphysical ice processes: an introduction to the CLOUDLAB project[J]. Bulletin of the American Meteorological Society, 2023, 104(11): E1962-E1979. https://doi.org/10.1175/BAMS-D-22-0178.1
Zhang H, Li X, Ramelli F, et al. IceDetectNet: a rotated object detection algorithm for classifying components of aggregated ice crystals with a multi-label classification scheme[J]. Atmospheric Measurement Techniques, 2024, 17(24): 7109-7128. https://doi.org/10.5194/amt-17-7109-2024
How to cite: Zhang, H., Ramelli, F., Fuchs, C., Miller, A. J., Omanovic, N., Spirig, R., Wu, Z., Chu, Y., Li, X., Lohmann, U., and Henneberger, J.: Unveiling In-Situ Ice Aggregation: Deep Learning, Causal Discovery, and Physics, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14147, https://doi.org/10.5194/egusphere-egu26-14147, 2026.
Operational global air quality forecasting often faces a critical trade-off between computational efficiency and the high spatial resolution required for effective pollution governance. Conventional numerical models are computationally expensive when resolving sub-grid processes, while standard data-driven approaches often struggle to capture global long-range dependencies effectively. In this work, we present a purely data-driven yet geometry-aware framework designed to predict and downscale global atmospheric composition fields. The framework operates in two stages to balance global dynamics with local fidelity. The first stage employs a Spherical Fourier Neural Operator (SFNO), trained on two decades of reanalysis data, meteorological fields, and emission fields. This model learns to evolve global concentrations of seven key pollutants (including PM2.5, PM10, O3, CO, NO2, SO2, and NO) at a 0.75° resolution. To provide finer spatial detail in regions of interest, the coarse-resolution predictions are downscaled to 0.1° × 0.1° using a Schrödinger Bridge–based stochastic super-resolution approach, while maintaining statistical consistency between the original and refined fields. This two-stage framework allows efficient generation of high-resolution global and regional air quality fields, while reducing the computational demands compared to conventional chemical transport models. The resulting model provides a practical tool for investigating pollutant transport and for supporting the evaluation of emission control strategies across multiple spatial scales.
How to cite: Zhao, D., Tao, C., and Cheng, Z.: Data-driven global air quality model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9358, https://doi.org/10.5194/egusphere-egu26-9358, 2026.
FourCastNet 3 advances global weather modeling by implementing a scalable, geometric machine
learning (ML) approach to probabilistic ensemble forecasting. The approach is designed to respect
spherical geometry and to accurately model the spatially correlated probabilistic nature of the
problem, resulting in stable spectra and realistic dynamics across multiple scales. FourCastNet 3
delivers forecasting accuracy that surpasses leading conventional ensemble models and rivals the best
diffusion-based methods, while producing forecasts 8 to 60 times faster than these approaches. In
contrast to other ML approaches, FourCastNet 3 demonstrates excellent probabilistic calibration
and retains realistic spectra, even at extended lead times of up to 60 days. All of these advances
are realized using a purely convolutional neural network architecture tailored for spherical geometry.
Scalable and efficient large-scale training on 1024 GPUs and more is enabled by a novel training
paradigm for combined model- and data-parallelism, inspired by domain decomposition methods in
classical numerical models. Additionally, FourCastNet 3 enables rapid inference on a single GPU,
producing a 60-day global forecast at 0.25°, 6-hourly resolution in under 4 minutes. Its computational
efficiency, medium-range probabilistic skill, spectral fidelity, and rollout stability at subseasonal
timescales make it a strong candidate for improving meteorological forecasting and early warning
systems through large ensemble predictions.
How to cite: Bonev, B., Kurth, T., Mahesh, A., Bisson, M., Koch, M., Ertl, G., Foster, D., Carpentieri, A., Kossaifi, J., Kashinath, K., Anandkumar, A., Collins, W. D., Pritchard, M. S., and Keller, A.: FourCastNet 3: A geometric approach to probabilisticmachine-learning weather forecasting at scale, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11448, https://doi.org/10.5194/egusphere-egu26-11448, 2026.
Accurate and timely weather forecasting and air quality prediction are foundational to public safety, environmental policy, and sustainable urban planning. These forecasts are essential for mitigating the adverse impacts of extreme weather events, such as heatwaves, wildfires, or severe storms, and for managing chronic air pollution issues that affect human health, ecosystems, and climate. Moreover, they play a vital role in advancing our scientific understanding of aerosol–meteorology interactions, which influence cloud formation, precipitation patterns, and radiative forcing within weather and climate systems. Despite their value, traditional modeling approaches, most notably chemical transport models (CTMs), face significant limitations. CTMs simulate the transport, chemical transformation, and deposition of pollutants based on atmospheric dynamics and emissions data. While highly detailed, they are also extremely computationally demanding. Running CTMs at high spatial and temporal resolutions, especially over extended periods or across large regions, requires substantial computational infrastructure and time. These constraints limit their practicality for real-time forecasting and rapid policy evaluation, particularly in data-scarce or resource-limited settings. To overcome these challenges, we introduce DeepCTM4D, a novel deep learning–based modeling framework that emulates the functionality of CTMs while drastically enhancing computational efficiency. DeepCTM4D leverages modern neural network architectures to learn from historical CTM outputs, enabling it to replicate the dynamic behavior of atmospheric chemical concentrations across a four-dimensional domain (three spatial dimensions plus time). The model is trained on a rich set of input variables, including anthropogenic and natural precursor emissions, meteorological conditions (e.g., wind, temperature, humidity), and initial chemical states, allowing it to learn complex, nonlinear interactions that govern pollutant formation and dispersion. One of the key strengths of DeepCTM4D lies in its ability to retain interpretability and scientific relevance. The relationships it captures between emissions, meteorology, and pollutant concentrations are consistent with known atmospheric chemistry mechanisms, lending credibility to its predictions. Furthermore, the model enables sensitivity analyses to identify major pollution drivers under different scenarios making it a powerful tool for evaluating the impacts of emission control strategies, policy interventions, or changing meteorological conditions. Beyond accuracy and interpretability, DeepCTM4D offers a transformative reduction in computational cost. It can generate near-instantaneous forecasts once trained, making it well-suited for operational use in early-warning systems, daily air quality updates, and climate-health applications. This efficiency opens new opportunities for integrating high-resolution air quality simulations into coupled Earth system models, weather prediction platforms, and mobile or edge-based applications in real time. In summary, DeepCTM4D represents a significant advancement in atmospheric science and computational modeling. By blending domain knowledge with data-driven intelligence, it provides a scalable, adaptable, and scientifically robust alternative to traditional CTMs. As an AI-enhanced forecasting tool, DeepCTM4D holds great potential to support global environmental monitoring systems and equip decision-makers with timely, actionable insights for managing air quality and responding to weather-related risks.
How to cite: Fu, J. and Xing, J.: AI-powered models for air quality forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22203, https://doi.org/10.5194/egusphere-egu26-22203, 2026.
Chemical processes significantly impact air pollution and its effects on climate, human health, ecosystems, and food security. Therefore, accounting for atmospheric chemistry is essential for reliable air pollution assessments and effective mitigation strategies.
This is typically achieved through the use of chemistry-transport models, which involve solving large systems of ordinary differential equations (ODEs) derived from chemical kinetics. However, as more species and reactions are incorporated into the models, the chemical mechanisms considered become increasingly complex, and the computational burden of the ODE solvers limits atmospheric simulations. This calls for alternative approaches, with artificial intelligence (AI) emerging as one of the most promising.
The EACH (Emulating Atmospheric Chemistry) project, a collaboration between the Max Planck Institute for Chemistry, the Barcelona Supercomputing Center, and Freie Universität Berlin, investigates the potential of using artificial intelligence in atmospheric chemistry modelling. Key results of the project presented here include a comprehensive training and benchmark dataset for AI-driven chemistry models, which will be publicly available. We also address the integration of physical constraints into AI chemistry models, such as element conservation and the non-negativity of concentrations, which are crucial for realistic and stable simulations. While such constraints have been explored in simple chemical mechanisms, scaling their application to complex mechanisms presents new challenges.
How to cite: Klingmüller, K., Butler, T., Gromov, S., Jorba, O., Kuhn, L., Mas Magre, I., Melli, A., Mouchel-Vallon, C., Petetin, H., Sander, R., Schaap, M., Thürkow, M., Lelieveld, J., and Pozzer, A.: Modelling complex atmospheric chemistry with artificial intelligence: data, constraints, and scalability, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16748, https://doi.org/10.5194/egusphere-egu26-16748, 2026.
Air pollution is now the second highest risk factor globally, highlighting the importance of air quality simulations, policies, and pollution peaks mitigation. Chemical Transport Models (CTMs) such as MONARCH are essential tools for designing air pollution mitigation plans but are limited by their high computational cost.
In the frame of the AIRE Spanish national project, we are developing eMONARCH, an emission- and meteorology-sensitive deep-learning-based surrogate model of MONARCH. eMONARCH aims at providing a cost-effective tool for generating ensemble of atmospheric composition simulations, supporting needs of air quality planning and data assimilation. We chose to start using a U-Net type architecture as a baseline model for its simplicity. The training dataset is composed of an ensemble of multi-annual MONARCH simulations with pertirbued emissions. A high performance was obtained for one-hour predictions, and we are now engaged in investigating ways to reduce the error accumulation and instabilities in multi-days autoregressive predictions. The first implementation of the model focuses on surface NOₓ concentrations, while the following versions include PMs, and multiple pollutants across several layers of atmosphere. In parallel, we are also developing a new Graph Neural Network (GNN) architecture composed of an encoder-processor-decoder structure whose preliminary results will also be presented.
How to cite: Orieux, M., Mathas, D., Petetin, H., and Mas Magre, I.: eMONARCH: a Deep Learning emission sensitive Chemical-Transport Model (CTM) for Air Quality planning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7029, https://doi.org/10.5194/egusphere-egu26-7029, 2026.
While the health impacts of air pollution are established, how public perception evolves with air quality remains unclear. Here, we analyze 180 million geotagged Weibo posts from China (2013–2023) using a natural language processing model to quantify public satisfaction with air quality and examine its relationship with monitored pollutant levels. We find that air quality improved significantly (PM2.5 concentration decreased by 51%), but public satisfaction increased only marginally (8.2%). This reflects progressively stricter subjective standards over time. Regional disparities reveal that economically developed areas exhibit lower tolerance for pollution, driven by heightened public awareness and media exposure. Annually, air pollution triggered negative emotions in 80 million people, influencing governance priorities. The findings underscore the dynamic interplay between air quality, public perception, and socioeconomic factors, advocating for adaptive policies integrating behavioral metrics to align with evolving public expectations. This work highlights the need for perception-aware environmental governance globally.
How to cite: Cheng, Z.: Dynamic Public Perception of Air Quality in China: Implications for Adaptive Environmental Governance, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9091, https://doi.org/10.5194/egusphere-egu26-9091, 2026.
Air pollution is one of the main environmental and public health challenges in urban and rural areas, influenced by a wide range of factors, including traffic, biomass burning, and meteorology. In Brazil, about 326,478 deaths occurred between 2019 and 2021 due to exposure to air pollution. About 8,400 deaths per year are attributed to the Metropolitan Area of São Paulo (MASP), the largest metropolitan area of South America. Mitigating the effects of air pollution is only possible with a deep understanding of the spatial and temporal distributions of air pollutants at high resolution. We employed a machine learning framework based on Extreme Gradient Boosting (XGBoost) to spatialize particulate matter concentrations (PM2.5 and PM10) at MASP at 300 × 300 m². In addition, we developed a Ridge regression model to control multicollinearity and ensure stable estimates. We used this model to examine monthly hospitalizations associated with air pollution and heat exposure in MASP during 2023–2024, a period marked by severe biomass burning and heat waves. The study used integrated data from the Environmental Company of the State of São Paulo (CETESB), ERA5 reanalysis, land use and land cover (MapBiomas), emission inventories, terrain roughness and altitude, and hospitalizations (National Health Data Network, DATASUS) from 2022 to 2024. The XGBoost model has shown to be robust, with high R² values of 0.85 for PM2.5 and 0.88 for PM10, and RMSE of 3.3 µg/m³ and 5.2 µg/m³, respectively, for the test set (30% of the data). The analysis showed higher pollution levels in densely populated and industrialized areas, such as Guarulhos-Pimentas and Parque Don Pedro, while less urbanized regions, such as Pico do Jaraguá, had lower concentrations due to meteorological and topographical factors. The Ridge distributed-lag hospitalization model exhibited high explanatory power (R² = 0.88; RMSE = 214 hospitalizations per month). Chronic cumulative exposure over three months revealed that ozone and nitrogen dioxide were the dominant drivers of hospitalizations, associated with increases of approximately 65% and 57%, respectively, in monthly hospitalizations, while PM10 showed a moderate effect (~16%). Carbon monoxide did not present a significant association. These findings indicate that photochemical pollution combined with seasonal and thermal variability plays a critical role in respiratory morbidity in MASP, providing a robust quantitative basis for environmental health surveillance and urban air-quality management.
How to cite: Franco, M. A., Cruz, D. D., Santibañez, J. A. P., Fernandes, K., Nascimento, E. G. S., Kumar, P., and Andrade, M. D. F.: Machine Learning and Statistical Modeling of Air Pollution and Hospitalizations in South America’s Largest Metropolitan Area, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13242, https://doi.org/10.5194/egusphere-egu26-13242, 2026.
The chemical composition of fine particulate matter (PM2.5) critically shapes its impacts on climate, air quality and human health, yet its high-resolution spatiotemporal variability covering continental to global scales remains poorly constrained owing to sparse ground observations. Here we develop a 10-km-resolution global dataset of PM2.5 chemical composition for 2010-2020, including sulfate (SO42-), nitrate (NO3-), ammonium (NH4+), organic matter (OM) and black carbon (BC), using a physically constrained deep transfer learning framework. Model evaluation against surface observations yields high accuracies, with correlation coefficients ranging from 0.80 to 0.91 across compositions. We identify pronounced spatiotemporal heterogeneity in PM2.5 composition distribution and long-term evolution. During the study period, the global reduction in PM2.5 concentration was driven primarily by decreases in SO42-, with Europe and Asia contributing most prominently. The fractional contributions of BC and OM increased significantly and exhibited a sustained upward trend in North America (by 4.72% and 5.86%, respectively) and Africa (by 2.32% and 6.94%), whereas secondary inorganic aerosols declined in all the continents except Africa. Recent studies have reported substantial differences in toxicity among PM2.5 compositions. Composition-specific exposure data therefore enable more accurate assessments of PM2.5-related health risks and underscore the importance of sustained and comprehensive monitoring of PM2.5 composition.
How to cite: wang, Z., Xiao, Q., Geng, G., and Zhang, Q.: Global Shifts in PM2.5 Chemical Composition over decade (from 2010 to 2020), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17520, https://doi.org/10.5194/egusphere-egu26-17520, 2026.
Abstract: The accurate detection of low-visibility events such as fog, haze, and fog-haze is a persistent challenge for satellite remote sensing, hindered by poor spatial generalization in existing models and unreliable aerosol retrievals. This study introduces a unified deep learning framework that integrates geostationary satellite data, meteorological reanalysis, and fine particulate matter (PM₂.₅) observations to identify these events. The model is able to produce the intensity of different low-visibility events and can be linked to visibility reduction. By incorporating PM2.5, the polluted fog-haze can be discriminated from clean fog. The method can be extended to sea fog. To isolate the impact of PM₂.₅ on fog-haze formation, sensitivity experiments were conducted. The findings reveal that the high frequency of winter fog-haze is primarily driven by elevated pollution; reducing winter PM₂.₅ concentrations to summer-like levels (a 60% reduction) causes the simulated fog-haze distribution to align with summer observations. This response is linked to the microphysical role of aerosols, where the primary effect of reducing PM₂.₅ is to cause a transition of fog-haze to fog, rather than to suppress the formation of low-visibility events entirely. We are able to investigate the mitigation of PM2.5 in reducing the hazardous fog-haze. By reducing the overall concentration of PM2.5 by 40% can reduce 75% of fog-haze area. For the first time, this work dynamically attributes the seasonal characteristics of fog-haze to pollution levels, providing a quantitative framework for evaluating the visibility co-benefits of air quality policies.
Keywords: Deep Learning; Satellite Remote Sensing; Low-Visibility Events
How to cite: Liang, Y. and Liu, D.: Classification and Attribution of Low-Visibility Events Using Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-248, https://doi.org/10.5194/egusphere-egu26-248, 2026.
Fine particulate matter (PM2.5), ozone (O3) and nitrogen dioxide (NO2) each pose significant risks to public health and are among the World Health Organisation (WHO) criteria pollutants. Operational air quality forecasting relies on computationally expensive Chemical Transport Models (CTMs), and recent deep learning methods focus on station-based forecasts, limiting usability to areas with station networks. We present a deep learning framework for probabilistic, gridded ambient air pollution forecasting to address both limitations.
Our approach employs a latent dynamics architecture. A convolutional variational autoencoder (Conv-VAE) learns compressed latent representations of input channels. A temporal core block captures the dynamical evolution of ambient pollutants in the latent space, and a probabilistic decoder reconstructs forecasts with uncertainty intervals. Probabilistic forecasting allows for more trustworthy predictions, as stakeholders are also presented with relevant confidence. We systematically compare four latent cores: ConvLSTM, Mamba (state-space model), Transformer (attention) and Neural ODEs. This comparison will identify which inductive bias best represents the dynamics of ambient air pollution evolution.
Experiments utilise a dataset of CAMS European reanalysis and ERA5 reanalysis (by ECMWF), as well as EDGAR emissions inventories over the UK (2015-2022), targeting 24–72 hour forecast horizons. Multi-pollutant settings test the framework's capacity to represent species with distinct atmospheric and chemical interactions in a unified latent representation. We will evaluate forecast skill, uncertainty quantification and computational efficiency of all models. Ongoing work is exploring physics-informed constraints, stochastic latent dynamics, and self-supervised pre-training for improved generalisation.
How to cite: Gibbons, N., Kumar, P., Aurélio de Menezes Franco, M., Fernandes, K., and Giovani Sperandio Nascimento, E.: Probabilistic forecasting of multiple air pollutants via latent dynamics modelling with deep learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19495, https://doi.org/10.5194/egusphere-egu26-19495, 2026.
This study focuses on volatile organic compound (VOC) emissions at the western terminal of Taichung Port, where 15 advanced air quality sensors were deployed to establish an intelligent monitoring network. Site selection was based on historical pollution hotspots, prevailing wind directions, and the presence of high-emission industrial facilities. The deployment employed a “smart fence” strategy, featuring sensors equipped with wind speed and direction modules to identify pollutant sources and transport dynamics, thereby providing real-time data to support air quality management. To ensure data reliability, sensor calibration and validation were conducted for O₃, NO₂, CO, VOCs, temperature, and humidity. The temperature and humidity sensors demonstrated strong correlations (R² > 0.8) and were effectively corrected using linear regression. O₃ sensors showed high correlation (R² > 0.9) and were successfully adjusted (RMSE reduced to 4.63 ppb).
From September to December 2024, VOC concentrations were further monitored across 13 industrial parks in Taichung City. Results revealed considerable spatial and temporal variability, as well as short-term high concentration events. Parks such as Renhua, Wufeng, Central Taiwan Science Park, and Fengzhou exhibited notably high VOC levels and standard deviations, indicating the presence of occasional emission sources. Many monitoring sites displayed standard deviations 3–5 times greater than the mean, highlighting frequent transient pollution events. It is recommended that local authorities intensify source tracking and real-time control measures in identified hotspots.
Additionally, a Positive Matrix Factorization (PMF) analysis identified six major VOC sources: vehicle emissions, biomass burning, fuel evaporation, industrial emissions, background pollutants, and solvent use. The study’s finding of 28.3% for vehicle emissions in Taichung Port aligns with existing literature, indicating consistency in source profiles. An O₃ prediction model was also developed using data from Dali Traffic Station and the advanced sensors, applying both XGBoost and ANN algorithms. XGBoost demonstrated superior performance (R² up to 0.88). SHAP analysis identified relative humidity, temperature, and NOₓ as the most influential variables. This model supports real-time O₃ prediction and hotspot identification.
How to cite: Lin, G.-Y., Lee, Y.-M., and Hong, G.-H.: Smart Monitoring and Ozone Precursor Analysis in the Port Area of Central Taiwan, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-37, https://doi.org/10.5194/egusphere-egu26-37, 2026.
Exposure to fine particulate matter (PM2.5) poses a significant environmental health risk, particularly in regions with limited ground-based monitoring infrastructure like Pakistan. This study presents a machine learning framework that generates hourly PM2.5 concentration maps at a high spatial resolution. We integrate meterological features from ERA5-Land reanalysis hourly data published by the European Center for Medium-range Weather Forecasts (ECMWF) with ground-based observations from a citizen science network of low-cost sensors.
Our approach uses model ensembling techniques with multiple tree-based gradient boosted algorithms to improve predictive accuracy of the framework. The ensemble technique captures the complex, non-linear relationships between meteorological variables and surface PM2.5 concentrations, while improving generalizability and predictive variance.
Preliminary results from cross-validation on an independent test set indicate strong predictive performance, confirming the framework’s capability to reliably estimate pollution concentrations in areas lacking direct measurements. The framework produces spatially complete, high-resolution pollution maps, offering datasets for visualizing and analyzing particulate matter. This work provides a scalable foundation for enhanced exposure assessment, future epidemiological studies, and evidence-based policy-making to mitigate the health impacts of air pollution in data-sparse regions of Pakistan.
How to cite: Naveed, M., Ahmad, R., and Omar, A.: Spatiotemporal Estimation of PM2.5 Across Pakistan Using Machine Learning Methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-833, https://doi.org/10.5194/egusphere-egu26-833, 2026.
Air quality forecasting is crucial for assessing population exposure to pollutants and avoid health complications. The Air Quality Index (AQI) is a scale of air pollution that indicates how clean is the air and helps to evaluate these complications. However, predictive models for air quality in European urban environments remain limited.
This index is calculated from air quality monitored data related to different pollutants such as PM2.5, PM10, CO, NO2 and O3 and there are six categories for the AQI: Good, Moderate, Fair, Poor, Very Poor and Hazardous. This work focuses on the implementation and evaluation of AQI forecast models for Spanish metropolitan areas such as Granada, Madrid and Barcelona. This study provides one of the first applications of LSTM-based AQI forecasting with extended horizons in a southern European environment.
The input data used are pollutant concentrations (PM2.5, PM10, CO, NO2 and O3) from an urban background station and meteorological variables (T, RH, P, wind direction, wind velocity and precipitation) from the nearest available station. Missing values were imputed to address short-term gaps, and all input variables were scaled to ensure stable training. The dataset was then split into 70/15/15 for training, validation and testing, respectively. The model used is a Long Short-Term Memory (LSTM) neural network, implemented to forecast AQI levels for 1 to 3-day horizons. Two AQI formulations were tested: a continuous index and the discrete version defined in national guidelines. Both were used to evaluate the model, and the continuous AQI consistently outperformed the discrete one. Thus, the continuous AQI was selected for 1-, 2- and 3-day forecasts. In order to improve model performance, additional features were included to capture temporal patterns, and backtesting was applied to obtain a robust performance estimate.
Preliminary conducted for the city of Granada and hourly AQI of PM10 have shown an accuracy in the range of 0.86 to 0.73 for horizons of 1 to 3 days, decreasing with the forecast horizon. The model reproduces the main AQI variability and most index transitions, although its performance is limited by the lack of high AQI levels (<1 %). In sum, these results highlight the potential of LSTM models to support air-quality forecasting in Spanish urban environments, enabling synergistic work with local authorities for early warnings. Future work will focus on improving the 3-day forecast, extending it to other cities with transfer learning.
Acknowledgements:
This work is part of the project funded by the 2024 Leonardo Grant for Researchers and Cultural Creators from the BBVA Foundation and grant JDC2022-048231-I, funded by MICIU/AEI/10.13039/501100011033 and the EU NextGenerationEU/PRTR. The authors also acknowledge the Junta de Andalucía for providing the air quality data.
How to cite: Molina-Párraga, L., Canella-Ortiz, A., Castillo, S., Mirza-Montoro, F., Casquero-Vera, J. A., Alados-Arboledas, L., and del Águila, A.: LSTM model for multi-day forecasting of Air Quality Index in urban areas , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1154, https://doi.org/10.5194/egusphere-egu26-1154, 2026.
In recent years, sustainability has become a global priority, making the mitigation of air pollution—particularly hazardous Particulate Matter (PM)—a paramount societal task. Leveraging air quality data from the Taiwan EPA, this study employs a fine-tuned, pre-trained transformer model to capture the complex, non-linear relationships between various pollutants and PM concentrations. Our results demonstrate that this approach significantly outperforms traditional ANN benchmarks in one-day-ahead predictions. Furthermore, we validate the model’s practical applicability by evaluating its performance under conditions of spatial variability and extreme events. From a statistical and stochastic perspective, the proposed framework can be interpreted as a data-driven approximation of latent stochastic dynamical systems governing pollutant transport and dispersion. This enables probabilistic characterization of forecast uncertainty, tail risks, and rare extreme pollution events, which are critical for risk-sensitive urban environmental governance.
The study offers two main contributions. It applies a large-scale transformer model to capture complex temporal patterns and achieve markedly better PM forecasts than traditional baselines. It also demonstrates strong generalizability through evaluation across varied environmental contexts in Taiwan. The work supports UN SDGs 3, 11, and 13 by enabling more sustainable urban management, improving public health protection, and strengthening climate resilience, thereby linking advanced AI forecasting with sustainability policy.
How to cite: Kow, P.-Y. and Kow, P.-E.: Transformers for Air Quality: Enhancing PM2.5 Modelling with Deep Attention Mechanisms, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1459, https://doi.org/10.5194/egusphere-egu26-1459, 2026.
The representation of hydrometeor collision processes remains a significant source of inaccuracy in bulk microphysics schemes. This work reduces these uncertainties through a unified framework that incorporates theoretical refinements of collision kernels together with machine-learning (ML) parameterizations capable of emulating high-resolution kernel behavior with markedly lower computational expense. The theoretical component incorporates realistic liquid–ice collision efficiencies, terminal velocities, and coalescence or sticking efficiencies derived from laboratory studies, together with turbulence-induced enhancements to cloud-drop collision efficiency based on direct particle simulations. The resulting dataset includes the rate of change of the 0th, 2nd, 3rd, and 6th moments for gamma-type size distributions, along with predicted changes in the shape and density of ice particles. Using Latin Hypercube sampling, 100,000 samples were generated for each collision process and used to train XGBoost-based ML parameterizations.
The ML parameterizations were implemented in a two-moment bulk microphysics scheme within the WRF model and evaluated in an idealized squall-line simulation. Execution-time analyses demonstrate substantial performance gains, with runtime reductions of up to 40% relative to the baseline configuration, while maintaining or improving the physical fidelity of the simulated microphysical processes. These results indicate that the proposed ML-based parameterization framework enhances both physical realism and computational efficiency, offering a promising pathway for next-generation microphysics schemes.
How to cite: Chen, J.-P., Wang, L.-J., Tsai, P.-C., Liao, C.-S., Tsai, T.-C., and Hong, Y.-T.: Toward efficient and physically consistent collision parameterizations using ML methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1484, https://doi.org/10.5194/egusphere-egu26-1484, 2026.
Lead-containing fine particles (Pb-FPs) from industrial emissions pose significant health risks, but their source-specific characteristics and traceability remain significant knowledge gaps. This study constructed a nationwide Pb-FP multi-metal fingerprint dataset and developed a machine learning–based source apportionment approach for efficient and accurate source attribution of atmospheric Pb-containing particles. Specifically, we presented a comprehensive investigation of Pb-FPs derived from four major industrial sectors in China, i.e. coal-fired power (CFP), iron and steel smelting (ISS), waste incineration power (WIP), and biomass power generation (BP), through systematic analysis of 134 PM samples collected nationwide using single-particle inductively coupled plasma time-of-flight mass spectrometry (spICP-TOF-MS). Our results showed that WIP (5 ×107 particles/mg) and ISS (3.9 ×107 particles/mg) activities emitted significantly higher number concentrations of Pb-FPs compared to CFP and BP sources. Across all sources, Pb–multi-metal FPs accounted for 66.7–81.2 % of total Pb-FPs number concentrations, with the mass fraction of Pb was predominantly ≤ 10 %.
Hierarchical clustering resolved 36 elemental fingerprint clusters with distinct source signatures (e.g., Fe/Mn/Zn-enriched ISS particles versus Si/Al-dominated CFP particles). Building on these fingerprints, we evaluated five machine learning algorithms for source apportionment, with XGBoost emerging as the optimal classifier (F1 score = 0.76, accuracy = 0.77) after intra-fold parameter optimization and cross-validation strategies. Application of the model to PM2.5 samples from Beijing and Shanghai revealed persistent and substantial contributions from ISS-derived Pb-FPs (6.7–38.1 % in Beijing, 10.5–33.7 % in Shanghai), with additional average inputs from CFP (7.4 %), WIP (5.8 %), and BP (12.1 %). These results highlight the dominant role of ISS in atmospheric Pb pollution across industrialized regions of China and provide a basis for explainable source-attribution analysis and future transfer-learning applications.
How to cite: Zhao, X.: Source Apportionment of Lead-Containing Fine Particles from Typical Industrial Emissions: A Machine Learning Approach Based on Source-specific Fingerprints, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1718, https://doi.org/10.5194/egusphere-egu26-1718, 2026.
Sulfur dioxide (SO2) hydrolysis is a critical step in secondary sulfate formation, which significantly affects air quality and climate change. Since the 1980s, debate has persisted over whether this reaction occurs mainly at the air–water interface or in the bulk phase. In this study, we investigate SO2 hydrolysis in heterogeneous systems using molecular dynamics simulations that are driven by a deep neural network potential with ab initio accuracy. In previous studies, rapid interfacial reactions have been proposed to account for the unexpectedly high SO2 uptake coefficients. In contrast, our results reproduce the observed uptake coefficients but show that interfacial hydrolysis contributes only 1%. We find that hydrolysis is accelerated in the bulk phase, where the denser hydrogen-bond network enhances SO2 electrophilicity and lowers the reaction barrier. The theoretical simulations in this work help to improve the understanding of aqueous sulfate aerosol formation and microdroplet chemistry.
How to cite: Du, M., Yang, M., Wang, H., Song, Y., and Zhu, T.: Bulk Phase Dominates Sulfur Dioxide Hydrolysis over Interfacial Processes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2027, https://doi.org/10.5194/egusphere-egu26-2027, 2026.
Conventional satellite-based aerosol optical depth (AOD) products typically offer coarse spatial resolutions, suitable for large-scale atmospheric studies but inadequate for localized applications such as urban air quality assessments. To address this limitation, we developed a Physics-Informed Convolutional Neural Network (PI-CNN) that estimates AOD at 30m resolution using Top-of-Atmosphere (ToA) reflectance from Landsat imagery over the Delhi and Kanpur regions of the Indo-Gangetic Plain (IGP). The architecture incorporates the Radiative Transfer Model (RTM) equations into the CNN structure, ensuring physically consistent retrievals. The model was trained over Kanpur using physics-based AOD estimates as training targets, and fine-tuned to Delhi through transfer learning. Evaluation against AERONET observations yielded correlation coefficients of 0.81 and 0.78 for Kanpur and Delhi, respectively, with corresponding MAE/RMSE values of 0.046/0.21 and 0.066/0.25. Furthermore, PI-CNN was compared with the traditional SEMARA retrieval method, which captured extreme values more effectively. In contrast, PI-CNN provided smoother, more generalized outputs with higher spatial variability than SEMARA. PI-CNN effectively reproduced the spatial distribution of AOD across different land use land cover (LULC), showing strong consistency with SEMARA and demonstrating its reliability in capturing spatial variations. These findings highlight the potential of PI-CNN as a flexible and scalable framework for retrieving high-resolution, physics-consistent AOD datasets across local to global scales.
How to cite: Kumar Singh, R. and Satyanarayana, A. N. V.: Development of Physics-Informed Convolutional Neural Network (PI-CNN) Model for Retrieval of High-resolution AOD over Cities of Indo-Gangetic Plain, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2321, https://doi.org/10.5194/egusphere-egu26-2321, 2026.
Fine particulate matter (PM2.5) oxidative potential (OP) is an important indicator of health risk, and it varies substantially across different emission sources. Although concentration–response functions (CRFs) exist that relate PM2.5 mass to its OP, the absence of source-specific OP CRFs has limited accurate global risk assessment.
In this study, we developed global source-resolved OP CRFs by combining machine learning, statistical modeling, and extensive datasets on PM2.5 concentrations, source apportionment across 50 countries, and more than 10,000 OP measurements from 29 countries. Our results show clear differences in the intrinsic OP per unit mass for major emission sectors, with the following ranking: energy > transportation > industry > agriculture and residential combustion.
Using these CRFs with 2017 PM2.5 source data for 203 countries, we estimated a global average source-resolved OP of 0.78 nmol min⁻¹ m⁻³ (95% confidence interval: 0.39–1.2). The energy sector (33%) and the combined agriculture and residential combustion sector (31%) were the largest contributors at the global scale, though contributions vary widely among countries.
Poisson regression analysis shows that source-resolved OP is a substantially stronger predictor of mortality attributable to PM2.5 than either PM2.5 mass concentration or bulk PM2.5 OP. These findings demonstrate that source-resolved OP provides a more accurate and policy-relevant metric for evaluating mortality risks and guiding targeted air quality interventions.
A full version of this work has been published in the Journal of Environmental Management (2026).
How to cite: Esu, C. O. and Cho, K.: Explainable Machine Learning for Source-Resolved PM2.5 Oxidative Potential: Implications for Global Mortality Burden, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2686, https://doi.org/10.5194/egusphere-egu26-2686, 2026.
This study introduces ReSA-ConvLSTM, an artificial intelligence (AI) framework for systematic bias correction in numerical weather prediction (NWP). We propose three innovations by integrating dynamic climatological normalization, ConvLSTM with temporal causality constraints, and residual self-attention mechanisms. The model establishes a physics-aware nonlinear mapping between ECMWF forecasts and ERA5 reanalysis data. Using 41 years (1981–2021) of global atmospheric data, the framework reduces systematic biases in 2-m air temperature (T2m), 10-m winds (U10/V10), and sea-level pressure (SLP), achieving a maximum reduction in RMSE of up to 20% for the 7-day T2m forecasts compared to operational ECMWF outputs. The lightweight architecture (10.6M parameters) enables efficient generalization to multiple variables and downstream applications, reducing retraining time by 85% for cross-variable correction while improving ocean model skill through bias-corrected boundary conditions. The ablation experiments demonstrate that our innovations significantly improve the model's correction performance, suggesting that incorporating variable characteristics into the model helps enhance forecasting skills.
How to cite: Sun, Y., Zhou, X., and Huang, X.: A Spatiotemporal Deep Learning Framework for Correcting Bias in Global Atmospheric Core Variables, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2695, https://doi.org/10.5194/egusphere-egu26-2695, 2026.
Accurate visibility prediction faces challenges from extreme data imbalance and complex spatiotemporal dependencies. This study develops an enhanced deep learning framework based on Informer architecture for short-term visibility prediction, trained on station-based observations in 2019-2024. To address extreme sample imbalance in visibility data, we have optimized data preprocessing and implemented physical constraints to the Informer architecture, specifically targeting improved prediction of low-visibility events like fog that hold significant public safety implications. First, visibility values were confined to a threshold range of 0.01 to 15 km, followed by a logarithmic-reciprocal transformation to nonlinearly expand the value interval for low-visibility conditions and inherently enhance their weighting within the model. Correspondingly, the activate function at the final output layer was also constrained to this threshold range to ensure physically realistic predictions. In addition, we propose a differentiable Threat Score-based loss function (TSLoss) that complements the mean squared error (MSE) loss, strategically weighting errors in rare low-visibility events. This approach resolves the non-differentiability of regression-to-binary conversion through sigmoid-activated thresholds. For comparative analysis, three models were trained: LSTM, standard Informer, and our modified Informer_TS. Evaluated against two baseline models, the optimized Informer_TS achieves superior performance for rare low-visibility events (≤1 km TS = 0.3, peaking at 0.55 at t+0) especially for significant reduction in false alarms. It performs especially well at coastal fog-prone sites and effectively captures nocturnal low-visibility events with better stability. Interpretability analyses highlight visibility autocorrelation, diurnal cycles, and meridional wind as key features. The algorithm demonstrates significant operational value for maritime and aviation safety through nowcasting of rapid-onset fog.
How to cite: Xia, Y.: A Physics-Informed Deep Learning Framework for Enhancing Rare Low-Visibility Event Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2929, https://doi.org/10.5194/egusphere-egu26-2929, 2026.
This study presents a novel approach for conducting all-day retrieval of cloud macro-physical
properties (single-layer cloud phase, cloud top height, and cloud base height for optical thickness less than 10)
using the Advanced Geostationary Radiation Imager (AGRI) and the Geostationary Interferometric Infrared
Sounder (GIIRS) onboard the geostationary meteorological satellite Fengyun-4A based on machine learning
methods. Model accuracy was compared after integrating ECMWF Reanalysis v5 (ERA-5) data, atmospheric
temperature and moisture profiles, and GIIRS clear-column radiance. Results demonstrate that integrating
GIIRS clear-column radiances can enhance the precision of cloud phase classification and the retrieval of cloud
macro-physical properties. This effectively replaces the role of atmospheric temperature and humidity profiles,
which are typically required for thermal infrared remote sensing retrieval. Moreover, the issue of delayed
acquisition of ERA-5 atmospheric temperature and humidity profiles is mitigated, enabling near real-time and
all-day retrieval of cloud macro-physical properties.
How to cite: Guo, B., Zhang, F., Zhao, Z., Guo, J., and Li, W.: Retrieval of Cloud Macro-Physical Properties Using the FY-4A Advanced Geostationary Radiation Imager (AGRI)and the Geostationary Interferometric Infrared Sounder(GIIRS), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3451, https://doi.org/10.5194/egusphere-egu26-3451, 2026.
Urban air pollution is characterized by significant spatio-temporal heterogeneity resulting from the complex interplay between regional long-range transport and localized emission sources, posing a major source of uncertainty in exposure assessment and policy formulation. In particular, fine particulate matter (PM2.5) and nitrogen dioxide (NO2) are representative pollutants simultaneously influenced by regional background levels and urban traffic/industrial emissions, necessitating the generation of high-resolution concentration fields. While conventional chemical transport models (CTMs) effectively capture regional-scale distribution and transport processes, they are limited in resolving micro-scale variability driven by complex urban terrain, traffic networks, and localized emission characteristics. Conversely, local dispersion models can precisely depict concentration gradients at fine scales but struggle to consistently incorporate background concentrations transported from outside the domain. Thus, hybrid approach that integrates the strengths of both models is essential.
In this study, we propose a hybrid air quality modeling framework that couples a Graph Convolutional Network (GCN) with the CALPUFF dispersion model. Focusing on Seoul, South Korea, in November 2022, the GCN leverages CMAQ data assimilation outputs to estimate high-resolution (1km⨯1h) regional background fields for PM2.5 and NO2 across the metropolitan area. By integrating these background fields with CALPUFF simulation results, we simulated PM2.5 and NO2 variations at a 100-meter resolution, explicitly accounting for road traffic and localized emission characteristics.
The proposed GCN–CALPUFF hybrid approach overcomes the inherent limitations of single-model frameworks and provides a robust methodology for high-resolution air pollution prediction, with broad applications in urban air quality forecasting, high-resolution exposure and health impact assessments, and evidence-based policy monitoring.
Acknowledgments
"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., Jang, K., Lee, J.-B., Kang, J.-G., Yoon, H.-Y., and Choi, D.: Development of a GCN–CALPUFF Hybrid Model for High-Resolution Simulation of PM2.5 and NO2, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3688, https://doi.org/10.5194/egusphere-egu26-3688, 2026.
Transformer-based time series models are already widely used in PM2.5 prediction studies due to their ability to learn long-term dependencies. However, despite input sequence length being a key design factor governing prediction performance, systematic evaluations that independently control this factor and examine how its effects vary across forecast lead times and atmospheric conditions remain limited. In particular, quantitative evidence is lacking on whether extending the input sequence length consistently improves long-term forecasting skill, or whether, under certain conditions, excessive historical information can instead degrade forecast stability.
In this study, Seoul—characterized by frequent high-pollution episodes and pronounced seasonal variability—is selected as a case study region. Using hourly observational data from 2018 to 2024, we quantitatively analyze the effects of input sequence length (3, 7, and 15 days) on Transformer-based PM2.5 prediction performance. Vanilla Transformer, Informer, and Autoformer models are evaluated under identical data partitioning, preprocessing, input variable configuration, training strategies, and output structures, allowing the effects of input sequence length to be isolated from other modeling factors. Prediction performance is assessed for short-term (24 h) and long-term (72 h) forecast horizons using MAE and RMSE, enabling joint analysis of error reduction, error accumulation, and forecast stability as input sequence length increases.
The results show that extending the input sequence length from 3 to 7 days leads to reduced short-term prediction errors and improved stability in long-term forecasts across all models. However, further extension to 15 days yields diminishing returns and, in some cases, increased errors for long-term forecasts. In particular, differences in MAE associated with input sequence length reach up to approximately 10–15% for 72 h predictions, indicating that longer input sequences can introduce not only useful long-term dependencies but also redundant or irrelevant historical patterns. Seasonal analyses further reveal that sensitivity to input sequence length is amplified during wintertime conditions with frequent high-pollution events, suggesting that the utilization of accumulated historical information plays a more critical role under stagnant atmospheric regimes.
This study demonstrates that longer input sequences are not universally optimal across all forecast horizons and highlights the need to tailor input sequence length according to forecast lead time and environmental context, even within the same Transformer architecture. By reframing input sequence length as a purpose-driven design parameter rather than a fixed hyperparameter, this work provides empirical guidance for the development and application of Transformer-based PM2.5 forecasting models.
"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., Han, S.-H., Wang, K.-H., and Yun, H.-Y.: Sensitivity of Transformer-Based PM2.5 Forecasting to Input Sequence Length, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3690, https://doi.org/10.5194/egusphere-egu26-3690, 2026.
Particle number size distribution (PNSD) is a cornerstone property of atmospheric aerosols and is essential for quantifying aerosol–cloud interactions. Although PNSD over continents has been studied comprehensively in the past decades via an extensive in-situ observational network worldwide, estimating marine PNSD (where clouds are more susceptible to aerosol and exert larger climate forcing) remains highly uncertain because of sparse observations, and PNSD varies strongly in space and time during the transport of air parcels. Here, we introduce a framework that integrates air-parcel location history with co-located aerosol, cloud, meteorological, and gas-phase information into deep learning (DL) approaches to constrain aerosol size distributions better. We employ three DL models: two Long Short-Term Memory (LSTM) models and one Bidirectional Long Short-Term Memory (BiLSTM) model. Evaluated against measured PNSD at the Cape Verde Atmospheric Observatory (CVAO) in the central Atlantic over 10 years, all three models achieve a mean fractional error (MFE) below 0.17. We further transfer the well-trained models to Ascension Island (ASI) in the South Atlantic; the predicted PNSD agrees with measurements with an MFE below 0.14, demonstrating strong model transferability. These DL models can therefore be used to project PNSD in remote marine environments. We also assess feature importance across the three models using the SHapley Additive exPlanations (SHAP) method. The models yield inconsistent interpretations of input features, suggesting they do not capture the mechanisms of aerosol formation pathways during transport. We therefore caution that, when using deep learning for mechanistic interpretation, multiple models should be applied for cross-validation to ensure the stability and reproducibility of the results.
How to cite: Cheng, Y., Xu, X., Wang, L., Huang, Y., Chen, H., Wei, X., Rahaman, S., Cai, D., Qi, B., Chen, Y., Shen, C., Wang, M., and Gong, X.: Learning Aerosol Particle Size by Embedding Airmass Historical Pathways in Multi-Model Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3803, https://doi.org/10.5194/egusphere-egu26-3803, 2026.
Visible-band satellite observations provide critical information on cloud structure and organization but are fundamentally unavailable at night, creating a long-standing gap in all-day Earth system monitoring. This limitation restricts the use of visible-band information in tracking cloud evolution, characterizing diurnal variability, and assessing nighttime tropical cyclone (TC) intensity and structure. Here, we present RefDiff, a diffusion-based probabilistic generative framework that generates nighttime visible reflectance by learning the statistical mapping between thermal infrared brightness temperature (BT) and daytime visible reflectance from geostationary satellites. Trained exclusively on daytime data and applied to nighttime conditions without nighttime supervision, the proposed approach generates spatially coherent, daytime-consistent visible reflectance and enables uncertainty estimation. Quantitative evaluation shows that RefDiff achieves clear accuracy improvements relative to deterministic deep-learning baselines, with the most pronounced gains for cloud systems characterized by complex structures and high optical thickness. We further show that the generated visible reflectance (GVR) significantly improves the accuracy of TC intensity estimation during nighttime. These results establish a new paradigm for all-day visible satellite observations, enabling continuous monitoring of clouds and storms across the diurnal cycle.
How to cite: Zhou, T., Zhang, F., Fu, H., and Guo, B.: Generation of Nighttime Visible Reflectance and Its Applications, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3888, https://doi.org/10.5194/egusphere-egu26-3888, 2026.
Cloud parameterization introduces uncertainty in numerical weather prediction (NWP), partly arising from the “tunable parameters”. However, the selection of the parameter values, namely calibration, has long been criticized for its arbitrariness, its tendency to induce error compensation, and its high computational cost. The development of machine learning (ML) methods in geoscientific research offers new tools to improve traditional calibration approach. Here, we propose a new framework for the objective calibration of cloud parameterization in the state-of-art ICON-NWP model. A trained machine learning model based on Gaussian Process Regression (GPR) will serve as a surrogate model for the numerical model, which allows sufficiently large ensembles under limited computation resources. History matching is adopted to quantify a plausible range for parameter values. We expect this framework to reveal the spatiotemporal distribution and cloud-regime dependency of paramters, which will provide us with a new insight into cloud parameterization and the underlying physics. In the future, we will further analyse the calibration results, especially regarding its impacts on aerosol-cloud radiative forcing and cloud–climate feedback.
How to cite: Chen, Y. and Quaas, J.: Machine Learning Calibration of Cloud Parameterization in a Numerical Weather Prediction Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4034, https://doi.org/10.5194/egusphere-egu26-4034, 2026.
The growing integration of artificial intelligence (AI) and atmospheric observations is opening new opportunities to resolve fast, nonlinear processes in atmospheric chemistry. A key bottleneck is the limited temporal resolution of routine volatile organic compound (VOC) monitoring, which weakens observational constraints on rapid chemical evolution and can bias process-based simulations of secondary pollution. Current VOC measurements rely primarily on gas chromatography–mass spectrometry (GC–MS) and proton-transfer-reaction time-of-flight mass spectrometry (PTR-ToF-MS). GC–MS is favored for accurate compound identification but is limited by relatively low temporal resolution. Conversely, PTR-ToF-MS can achieve minute-scale resolution by directly ionizing samples, yet it struggles to detect compounds with low proton affinity. Here, based on five years of long-term online monitoring data, we propose an Adaptive Convolutional Tree Ensemble (ACTE) model to overcome the limitations of current instruments and reconstruct VOC concentrations at 5-minute resolution. Our results indicate that ACTE consistently achieves robust predictive accuracy across major chemical classes, with R2 values of 0.92 and 0.89 for alkanes and alkenes, respectively, many of which have relatively low proton affinity. Furthermore, using ozone photochemical simulations driven by VOC inputs at different temporal resolutions, we find that higher-resolution inputs more accurately capture rapidly evolving photochemical reactions, whereas hourly inputs tend to overlook short-term variability, potentially biasing mechanistic interpretation. Our findings demonstrate how machine learning (ML)-enabled temporal super-resolution can bridge routine monitoring and mechanism-based modeling, improving process-level diagnosis of atmospheric chemical evolution.
How to cite: Cheng, Y., Huang, X.-F., Peng, Y., and He, L.-Y.: AI-Augmented High-Frequency Reconstruction of Online VOC Observations and Implications for Atmospheric Chemistry Mechanism Modeling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4881, https://doi.org/10.5194/egusphere-egu26-4881, 2026.
Large-scale vegetation restoration has been performed to improve the fragile ecological environment of the Loess Plateau in the last decades. At present, the effects and mechanism of long-term vegetation restoration on soil properties in ecosystems require further exploration to provide a reference for rational ecological construction. Hence, we investigated the differences in vegetation attributes and soil properties between three typical vegetation types (Pinus tabulaeformis plantation forest, PTPF; Robinia pseudoacacia plantation forest, RPPF; natural secondary forest, NSF) after long-term (30–40 years) vegetation restoration in the western Loess Region, China. Our results showed that (1) the arborous synusia biomass of the plantation forests was twice that of NSF, whereas NSF had almost 50% higher near-surface synusia biomass than the plantation forests; (2) the soil nutrient contents of the plantation forests were lower (30%) than those of NSF; (3) the soil bulk density, organic matter, total nitrogen, and phosphorus were positively related to arborous and shrub synusia biomass; (4) the coupling effects of four biological synusiae (with the contribution of 47.67%) were the dominant factors affecting soil physicochemical properties. Natural forests have the better vegetation attributes and soil properties than plantation forests, indicating that the close-to-nature restoration should be considered in ecological restoration. These findings can provide scientific support and theoretical basis for reforestation and ecological restoration in the Loess Plateau region and similar areas in the future.
How to cite: Zhang, Y.: Effects of long-term vegetation restoration on soil physicochemical properties mainly achieved by the coupling contributions of biological synusiae to the Loess Plateau, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6482, https://doi.org/10.5194/egusphere-egu26-6482, 2026.
Air pollution poses a critical risk to both human health and the environment, particularly in the Arctic and Northern Europe, where pollutants are primarily transported from mid-latitudes by atmospheric circulation. Also, local sources further contribute to pollution levels in Arctic communities. Accurate short-term forecasts of atmospheric pollutant concentrations are vital for enabling adaptive measures and protecting public health during pollution episodes.
The Copernicus Atmospheric Monitoring Service (CAMS) provides 96-hour forecasts for key pollutants across Europe using 11 state-of-the-art models and an ensemble approach. However, these forecasts exhibit significant errors in Northern Europe and the Arctic. To address this, we investigate the applicability of deep learning (Transformer-based) models for 48-hour PM10 concentration forecasting at monitoring stations in Northern Europe. Our approach integrates in situ PM10 observations with CAMS model outputs and forecasted meteorological parameters as input features. We evaluated four time-series specialized models—Informer, Autoformer, FEDformer, and Crossformer—to identify the most effective architecture for this task. The Crossformer model demonstrated superior performance, outperforming CAMS by 30% in Mean Squared Error (MSE) and 23% in Mean Absolute Error (MAE). It also surpassed the newly introduced CAMS Model Output Statistics (MOS), reducing MSE by 12% and MAE by 14%.
With its low computational complexity, fast execution time, and minimal resource requirements, the Crossformer presents a viable alternative to traditional numerical models for local-scale predictions. Future work will extend the forecasting window to 72 hours and incorporate additional pollutants, such as PM2.5, NO2, and O3, to enhance predictive capabilities for Arctic and Northern European communities.
How to cite: Crotti, I., Cuzzucoli, A., Pasini, A., and Dobricic, S.: Boosting Arctic Air Quality Forecasts with Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6773, https://doi.org/10.5194/egusphere-egu26-6773, 2026.
Tropospheric Black Carbon (BC) aerosols are short lived positive radiative forcer with critical impacts on cardiovascular and pulmonary health. The MRI-ESM2(CMIP6) delivers the BC monthly global surface concentration from 1950-2100 at coarse special resolution of 1.875° × 1.875° which does not capture the city level BC hotspots at the global scale. The BC hotspots are essential for reducing the BC aerosols induced health burden, implementing air quality management policies and regional planning. The downscaling algorithm is an enhanced U-net with attention-based convolution neural network (Super Resolution Convolution Neural Network (SRCNN)). The SRCNN model executes downscaling of MRI-ESM2(CMIP6) BC monthly surface concentration with special resolution of 1.875° × 1.875° to NASA’s MERRA2 reanalysis BC monthly average surface concentration at a spatial resolution of 0.5° × 0.625°, thus achieving 3.6 times downscaling for identification city level BC-hotspots and cold spots at a global scale. The SRCNN model is trained on global monthly average BC surface concentration data from MRI-ESM2(CMIP6) and NASA’s MERRA2 reanalysis product. The model training is spread from 1980-2012(31 years) with validation from 2013-2016(4 years) and testing for 2017-2020(4 years). We have also examined the effects of Channel Based Attention Module (CBAM) with and without Residual Block (RB) and their effectiveness and efficacy in climate data downscaling with data-scarce condition. The training results showes that CBAM with RB based CNN outperforms then both (CNN without CBAM and without CBAM & RB) in the benchmarks such as stability, overfitting, validation losses etc. The training results for SRCNN (with CBAM and RB) shows a final validation losses of 0.0028, final R² value of 0.7162, final Pearson-r value of 0.8467 with Structural Similarity Index Measure (SSIM) at 0.9954. The SRCNN (with CBAM & RB) model testing reveals it performs exceptionally well in the identification of hotspots and cold-spots, with final testing RSME at 0.0015, final R2 at 0.88, final Pearson-r values at 0.94 and final SSIM at 0.99. Furthermore, testing outputs of SRCNN with attention module and residual blocks shows close fidelity with MERRA2 reanalysis vis-à-vis MRI-ESM2(CMIP6) at both seasonal and annual temporal resolution thus reducing systematic bias between ground truth and global climate models.
How to cite: Mishra, K., Kajino, M., Sekiyama, T. T., and Oshima, N.: Super-Resolution Surrogate Downscaling of MRI-ESM2(CMIP6) Black Carbon Surface Concentration Using Attention Based Convolutional Neural Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6806, https://doi.org/10.5194/egusphere-egu26-6806, 2026.
Reliable estimation of 10-hour dead fuel moisture content (10-h DFMC) is essential for forest fire risk forecasting, particularly in mountainous and fire-prone regions such as Gangwon Province, South Korea. This study presents a machine learning model trained on hourly observed 10-h DFMC data from 11 stations during 2024, using routine meteorological observations from Automatic Weather System (AWS) sensors as input. The goal was to estimate 10-h DFMC in real-time using standard operational inputs. The model inputs consisted of 1-h averages of 2 m air temperature, relative humidity, 10 m wind speed, and 1-h accumulated precipitation. As solar radiation data were unavailable, we included Julian day and hour of day (0–23) as proxy variables to partially account for diurnal and seasonal patterns in solar energy input. The observed 10-h DFMC data revealed distinct seasonal and spatial variation: spring and early winter showed persistently low moisture, consistent with peak fire seasons. High-elevation stations retained moisture longer due to snow cover, while coastal sites exhibited greater variability influenced by maritime air masses. The random forest model achieved high predictive accuracy (R² = 0.80; RMSE = 2.73%; MAE = 1.93%) on the test dataset. Station-level evaluation showed R² ranging from 0.76 to 0.86. Relative humidity was the most influential predictor, while precipitation had marginal impact, suggesting that 10-h DFMC is more sensitive to sustained atmospheric humidity than to short-term rainfall. Comparative experiments confirmed that the random forest approach outperformed linear regression and support vector regression and achieved similar performance to gradient boosting. Snow-affected high-altitude sites showed larger errors, indicating the need for future inclusion of snow-state and terrain-related covariates. This study offers a regionally calibrated, operationally feasible model for 10-h DFMC estimation based solely on widely available AWS data. Its structure is inherently transferable to other regions with localized training data, supporting scalable, real-time fire danger assessment systems under a changing climate. This abstract is based on findings from our peer-reviewed article published in December 2025 under the title: “Machine Learning–Based Analysis and Prediction of 10-h Dead Fuel Moisture Content Using Automated Weather Observations in Gangwon Province, South Korea.” This research was funded by the Korea Meteorological Administration Research and Development Program “Advanced Research on Bio- and Agricultural Meteorology” (Grant No. KMA2018-00626).
How to cite: Chae, S., Han, Y., and Kim, K. R.: Random Forest-Based Estimation of 10-h Dead Fuel Moisture Using Automatic Weather System Observations in Gangwon, Republic of Korea, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8546, https://doi.org/10.5194/egusphere-egu26-8546, 2026.
Summer precipitation in China, primarily driven by the East Asian summer monsoon, holds significant socioeconomic implications. Skillful prediction of summer precipitation requires effective multi-model integration of operational climate models. To improve the model integration, this study proposes a novel Bayesian deep learning (BDL) network that integrates convolutional neural networks (CNNs) with attention mechanisms. The BDL network is evaluated using four operational climate models: ECMWF_SEAS51, JMA_CPS3, NCC_CSM11, and NCEP_CFS2. Compared to conventional Bayesian Model Averaging (BMA), the BDL network more accurately captures the spatiotemporal patterns of summer precipitation, improving the anomaly correlation coefficient (ACC), the prediction score (PS), and the root-mean-square error (RMSE). These improvements are primarily attributed to the adaptive weighting of individual model over time. Further analysis identifies NCEP_CFS2 and ECMWF_SEAS51 as the primary contributors to the integrated prediction. This study presents a new perspective for model integration via deep learning, providing an effective approach to enhance summer precipitation prediction.
How to cite: Yang, Z., Gui, S., Gong, Z. G., Feng, G., Zi, P., Li, T., and Yang, R. Y.: Skillful summer precipitation prediction in China using an attention-based Bayesian deep learning network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8765, https://doi.org/10.5194/egusphere-egu26-8765, 2026.
Extreme wind and precipitation events result in signicant societal disruption in the South China coastal region, typically triggered by tropical cyclones (TCs) or mesoscale storms. The large-, meso-, and small-scale atmospheric circulation processes that can influence these high- impact weather events may be altered by climate change, potentially changing TC characteristics. However, quantifying the sensitivity of TCs and extreme precipitation to climate change is challenging, primarily due to the limited detail provided by global model simulations with coarse resolution. High-resolution simulations are essential to address such issues. We have developed a smart dynamical downscaling (SDD) model to downscale the climate simulations (100 km) to high-resolution simulations (15 km). The trained SDD model can be applied to ensemble climate simulations under the SSP585 scenario from 2020 to 2100 to explore the variations of the severe TC cases, regarding the spatial distribution, maximum surface wind, and precipitation, respond to global warming. We found the inland areas of China will be affected more by TC-induced extreme precipitation and the intense typhoons are increased in the future based on the ensemble downscale results. The high-resolution simulations are conducted for selected extreme precipitation events to further under the dynamical response to global warming.
How to cite: Wang, Y., Shi, X., and Fung, C.-J.: Employing Deep Learning to Quantify the Trends in Tropical Cyclones and Associated Extreme Precipitation Events in Southern China, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9209, https://doi.org/10.5194/egusphere-egu26-9209, 2026.
In 2024, the China Meteorological Administration (CMA), in collaboration with Tsinghua University, developed the “Fengqing” forecasting system following an innovative "AI-Physics" hybrid approach. Through designs such as a multi-scale latent space projection architecture and an energy-conservation loss function, the model has been equipped with global short-and medium-range weather forecasting capabilities and has been operationally implemented. This study comprehensively evaluated the forecasting ability of the Fengqing in China and its surrounding areas in 2024 from several metrics such as forecasting accuracy and bias distribution. It also focused on two kinds of typical synoptic processes, typhoons and rainstorms, to deeply explore the model's performance in forecasting of disastrous weather. The results show that the 500 hPa geopotential height forecasts maintain predictive skill beyond 10 days in Fenging. The Root Mean Square Error (RMSE) for the 2 m surface air temperature and the 850 hPa temperature in the upper air is significantly lower than that of the European Centre for Medium-Range Weather Forecasts (hereafter, ECMWF-IFS), with a maximum improvement of 37.66%. In terms of typical weather processes, the Fengqing model demonstrates marginally superior performance in typhoon track forecasting compared to ECMWF-IFS, though exhibits systematic underestimation in typhoon intensity prediction. In addition, the Fengqing model exhibits superior torrential rainfall forecasting capabilities, demonstrating precise prediction of typhoon-induced precipitation patterns and Mei-yu front rainfall belt positioning. The TS score for heavy rain forecasts in the medium-term (73-168h lead time) improvements reaching 43.53% compared to that of ECMWF-IFS forecasts. Overall, the Fengqing model demonstrates considerable potential in operational forecasting, although further improvements are needed in forecast activity and typhoon intensity prediction at medium- to long-range lead times.
How to cite: Li, N., Gong, Y., and Cao, Y.: Preliminary Evaluation the Operational Application Effect of "Fengqing", an AI-based Global Short and Medium Range Forecasting System , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9465, https://doi.org/10.5194/egusphere-egu26-9465, 2026.
The additional impact of emission-reduction measures in North China (NC) during autumn and winter on the air quality of downwind regions is an interesting but less addressed topic. The mass concentrations of routine air pollutants, the chemical compositions, and sources of fine particles (PM2.5) for January 2018, 2019, and 2020 at a megacity of Central China were identified, and meteorology-isolated by a machine-learning technique. Their variations were classified according to air mass direction. An unexpectedly sharp increase in emission-related PM2.5 by 22.7% (18.0 μg m−3) and 25.7% (19.4 μg m−3) for air masses from local and NC in 2019 was observed compared to those of 2018. Organic materials exhibited the highest increase in PM2.5 compositions by 6.90 μg m−3 and 6.23 μg m−3 for the air masses from local and NC. PM2.5 source contributions related to emission showed an upsurge from 1.39 μg m−3 (biomass burning) to 24.9 μg m−3 (secondary inorganic aerosol) in 2019 except for industrial processes, while all reduced in 2020. From 2018 to 2020, the emission-related contribution of coal combustion to PM2.5 increased from 10.0% to 19.0% for air masses from the local area. To support the priority natural gas quotas in northern China, additional coal in cities of southern China was consumed, raising related emissions from transportation activities and road dust in urban regions, as well as additional biofuel consumption in suburban or rural regions. All these activities could explain the increased primary PM2.5 and related precursor NO2. This study gave substantial evidence of air pollution control measures impacting the downwind regions and promote the necessity of air pollution joint control across the administration.
How to cite: Jiang, S., Kong, S., and Fu, P.: Winter-autumn air pollution control plan in North China modified the PM2.5 compositions and sources in Central China, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9795, https://doi.org/10.5194/egusphere-egu26-9795, 2026.
The volume and complexity of atmospheric data have expanded significantly, driven by the proliferation of low-cost sensor networks, high-fidelity research stations, and multi-platform remote sensing. However, the utility of these datasets is often hindered by inherent noise in low-cost hardware, the necessity for labor-intensive manual analysis, and limited spatial coverage. This work explores the integration of Artificial Intelligence (AI) and Machine Learning (ML) to automate data processing workflows, ensuring high-quality, scalable, and real-time atmospheric insights.
We present several case studies demonstrating the effectiveness of AI in bridging data gaps and enhancing analytical accuracy. First, we discuss the development of "virtual sensors" for Ozone (O3) monitoring, designed for deployment within micro-measurement stations where physical chemical sensors may be impractical. Second, we introduce a novel, robust fitting algorithm for Particle Number Size Distributions (PNSD) that operates in near real-time, offering superior reliability over traditional iterative methods. Third, we showcase a predictive model that fuses satellite remote sensing data with ground-level observations to estimate and spatially scale PM2.5 concentrations, providing high-resolution coverage in previously unmonitored areas.
Beyond traditional data processing, this work outlines the broader potential of emerging AI technologies to address remaining atmospheric challenges. We explore the implementation of EdgeAI for on-device sensor calibration and the use of Computer Vision to quantify traffic and human activity, thereby providing critical metadata for source apportionment. By integrating these automated technologies, we demonstrate a path toward a more responsive and comprehensive framework for air quality and atmospheric data analysis.
How to cite: Zaidan, M. A., Rahman, A., Sarwar, H., Chua, S., Rohal, D., Kangasluoma, J., Lehtipalo, K., Petäjä, T., and Tarkoma, S.: Leveraging Artificial Intelligence for the automated processing and analysis of real-time atmospheric data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9912, https://doi.org/10.5194/egusphere-egu26-9912, 2026.
Chemical transport models (CTMs) are essential tools for investigating the chemical processes at stake in the atmosphere and supporting needs on air quality assessment and planning. Yet, they typically require massive computational resources to solve the system of stiff ordinary differential equations governing atmospheric chemical kinetics (many reactions and species with highly variable abundances and kinetic time scales), which limits the resolution of simulations and/or the level of complexity of the chemistry representation.
In the frame of the EACH (Emulating Atmospheric Chemistry) project involving the Barcelona Supercomputing Center, the Max Planck Institute for Chemistry,and Freie Universität Berlin, we are investigating the potential of deep learning to emulate the chemistry, focusing first on gas phase chemistry, with the ultimate goal of being able to accelerate CTM simulations. More specifically, we assess the performance and generalization capability of dense deep feedforward neural networks based on the multilayer perceptron (MLP) architecture using two test mechanisms: POLLU[1], a simplified tropospheric ozone formation mechanism (20 species, 25 reactions), and CB05[2], a condensed mechanism of atmospheric oxidant chemistry (59 species, 156 reactions) used in many CTMs. Training datasets were generated from millions of 0D chemical box-model simulations, with initial conditions sampled from uniform multidimensional distributions. For each experimental setup, a systematic hyper-parameter search was conducted to identify the optimal configuration. We trained several MLP variants incorporating physical consistency through both hard (architectural) and soft (loss-function-based) physical constraints designed to preserve stoichiometric relationships and enforce non-negativity of concentrations, and we assessed mass conservation using tailored evaluation metrics. The sensitivity of the MLPs performances to the number of training time series and their length was explored to examine the impact of data design on model performance.
[1] Verwer, J. G. (1994). Gauss–Seidel iteration for stiff ODEs from chemical kinetics. SIAM Journal on Scientific Computing, 15(5), 1243-1250.
[2] Yarwood, Greg, et al. (2005). Final report to the US EPA, RT-0400675 8: 13
How to cite: Petetin, H., Melli, A., Mouchel-Vallon, C., Mas Magre, I., Jorba Casellas, O., Klingmüeller, K., Gromov, S., Kuhn, L., Sander, R., Butler, T., Thürkow, M., and Pozzer, A.: Emulating tropospheric chemistry mechanisms with deep neural networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11742, https://doi.org/10.5194/egusphere-egu26-11742, 2026.
In situ observations of electrically induced aggregation of cloud ice and frozen droplets have primarily been observed in mid- to upper-level clouds of summertime storms. These aggregates, distinguished by their elongated, quasi-linear structure, are specifically termed as chain aggregates. Cloud chamber experiments reveal that chain aggregation is temperature-dependent, and their formation is enhanced in an electric field exceeding approximately 60 kV m-1. However, various difficulties arise when connecting the laboratory experiments to in situ observations. While there is evidence that significant electric fields are required for chain aggregate formation, the precise locations and the mechanisms for chain aggregation within storms remain poorly understood. This knowledge gap hinders the accurate parameterization of chain aggregate formation processes in cloud models, impacting precipitation formation, radiative transfer, remote sensing retrievals, and precipitation forecasting.
During NASA’s Investigation of Microphysics and Precipitation for Atlantic Coast-Threatening Snowstorms (IMPACTS) field campaign, chain aggregates were observed in 30 of 34 research flights, across temperatures from –38.2 to 2.5 °C and altitudes from 1.5 to 9.7 km, including in weakly electrified winter storms. These frequent observations challenge prevailing assumptions and underscore the need for comprehensive analysis. Given that the Cloud Particle Imager (CPI) captured millions of particle images during IMPACTS, manual classification is infeasible. To address this, we developed a supervised convolutional neural network (CNN) classifier using transfer learning to distinguish chain aggregates from non-chains directly from CPI images. We benchmarked several common CNN backbones (ResNet18/34/50/101 and VGG16/19) and selected the final model using a precision-first criterion supported by PR-AUC/ROC-AUC and calibration metrics (log-loss/Brier). The resulting ResNet34 model provides reliable separation of chain aggregates vs. non-chains and achieves strong performance on unseen data (≈95% precision and ≈80% recall for the chain class), enabling confident campaign-scale mapping of chain aggregate occurrence and more robust comparisons with collocated ER-2 radar and lidar observations.
How to cite: Nairy, C., Delene, D., Wagner, S., Finlon, J., and Yorks, J.: Identifying Ice Crystal Chain Aggregates in Cold-Season Storms: Leveraging Machine Learning to Map Occurrence and Distribution, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13901, https://doi.org/10.5194/egusphere-egu26-13901, 2026.
Image analysis is integral in understanding the increasing abundance of atmospheric science imagery
data. Whereas time-consuming analysis of individual images or simplified image processing techniques
were previously necessary, machine learning can now quickly learn trends in and distinctions between
images. However, training deep learning models on large datasets can be computationally expensive.
Leveraging the architectures and weights of pre-trained neural networks can alleviate some expense. In
this work, we apply pre-trained networks to back trajectory images generated for the NASA Aerosol
Cloud meTeorology Interactions oVer the western north ATlantic Experiment (ACTIVATE) campaign.
We find this method outperforms the principal component analysis baseline and results in four
geographically distinct clusters. Pairing these images with the host of measurements taken during the
ACTIVATE campaign, we find the regions to also be distinct in their bulk characteristics of chemical and
microphysical variables.
How to cite: Butler, K. and Silva, S. J.: Convolutions, clusters, and characterizations: Using pretrained networks for back trajectory analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15004, https://doi.org/10.5194/egusphere-egu26-15004, 2026.
Mesoscale convective systems(MCSs) can generate severe disasters, including extreme precipitation, hail, flooding, thunderstorms, and strong winds, and are significantly influenced by weather circulation and local geographical conditions. For MCSs occurring in urban areas, urban forcing plays a crucial role in regulating their activity, though the associated impacts are highly complex. Utilizing 15 years of high-resolution radar network products from the China Meteorological Administration, this study identifies and tracks MCSs across three major urban agglomerations, analyzing their spatiotemporal distribution characteristics. It is observed that MCSs in the Pearl River Delta predominantly occur from April to June, in the Yangtze River Delta from May to July, and in the Beijing-Tianjin-Hebei region from June to August. In the Yangtze River Delta and Beijing-Tianjin-Hebei regions, MCSs are more frequently initiated at night (20:00–08:00 Beijing Time), whereas in the Pearl River Delta, they are more commonly initiated during the day (08:00–20:00 Beijing Time). Moreover, the spatial distribution patterns and movement directions of MCSs in these three major urban agglomerations exhibit distinct variations and differences between the cold half-year (October–March) and warm half-year (April–September), as well as between daytime and nighttime. To assess urban impacts, MCSs traversing urban areas were further selected to analyze the spatiotemporal distribution characteristics of their trajectory points and changes before and after crossing cities. It is evident that trajectory points within Pearl River Delta cities exhibit higher numbers, larger areas, and greater intensity during daytime. In contrast, trajectory points within Yangtze River Delta cities show higher numbers but smaller areas and lower intensity during daytime. Trajectory points within Beijing-Tianjin-Hebei cities demonstrate higher numbers, smaller areas, and lower intensity during nighttime. Additionally, significant changes in trajectory point characteristics were observed before and after urban crossing. For instance, trajectory points in the Pearl River Delta and Beijing-Tianjin-Hebei regions expanded in area upon entering cities, whereas those in the Yangtze River Delta contracted. Finally, correlation coefficients were used to identify relationships between various environmental and urbanization factors and the characteristics of MCSs traversing urban clusters.
How to cite: Jin, Q. and Zhao, K.: Spatial and Temporal Characteristics of Mesoscale Convective Systems in China's Three Major Urban Agglomerations and the Urban Impact on Them, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15637, https://doi.org/10.5194/egusphere-egu26-15637, 2026.
Traditional numerical weather models struggle with predicting the planetary boundary layer (PBL) in urban areas and during the morning transition. Accurately predicting this part of the atmosphere is crucial because of its downstream impacts, such as accurate air quality forecasts and the representation of convection-based processes. The rapid growth of information technology has increased the capability of machine learning, but several limitations and sensitivities arise when using it to predict the PBL. In this work, aircraft observations from the Aircraft Meteorological Data Relay Program are compiled into half-hourly temperature profiles of the PBL. Dallas-Fort Worth, TX, USA is chosen because it is far from topographical and coastal influences, and there are two large airports near the center of the city. Profiles are compiled into daily bins of five half-hourly profiles prior to and including sunrise as the inputs and eight half-hourly profiles after sunrise as the outputs. This provides the opportunity to test the performance of machine learning models under a variety of stability classifications and PBL heights. To determine the sensitivity of the model configuration, five machine learning model types are tested, learning rates from 0.01 to 0.000001, various training epochs, the order of the layers, the number of neurons in each layer, and eight optimizers. At the start, mean square error (MSE) is used as the loss function to find the optimal model configuration. However, standard summary statistics may not produce larger errors when the physically more important parts of the PBL are astray, such as near the surface and the inversion at the top of the PBL. To test the sensitivity of the loss function, MSE and correlation coefficient are used to gauge the performance of using loss functions of MSE, mean absolute error, Huber loss, and the logarithm of the hyperbolic cosine, in addition to several custom weighted profiles that place higher weights at different parts of the PBL. The optimal model configuration found using MSE as the loss function is a long-short term memory network layer with 2,000 nodes followed by two dense layers with 1,000 nodes, a learning rate of 0.0001, 100 epochs, and an AdamW optimizer, which had an overall MSE of 0.882°C. The MSE was larger for predictions further after sunrise, and the model generally underestimated (overestimated) the onset of mixing and near-surface temperature in the summer (winter). Mean absolute error was the most accurate loss function with an overall MSE of 0.538°C and a correlation coefficient of 0.958. The work shown here highlights the importance in methodically testing various machine learning configurations to back out the sensitivity of the model, which can influence the confidence of the conclusions. It also shows the potential of using a simple machine learning model to produce rapidly updated short-term weather forecasts that can be used in conjunction with traditional numerical weather models.
How to cite: Nielsen, K. and Rahn, D.: The Sensitivity of Machine Learning Configuration for Predicting Temperature Profiles in the Planetary Boundary Layer, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16030, https://doi.org/10.5194/egusphere-egu26-16030, 2026.
We present a deep learning model that regionally downscales relatively coarse (~25 km) ERA5 reanalysis data to a 1-km grid. The model is trained on hourly fields from GeoSphere Austria’s high-resolution INCA model, a regional data assimilation and nowcasting system. Once trained, it can generate hourly high-resolution climate fields using only coarse ERA5 data and a digital elevation model as input. Early results show the deep learning model outperforms simple interpolation of ERA5 data. By comparing our model with baseline models that apply only constant bias correction and lapse rate based elevation adjustment we can quantify how much skill comes from basic statistical corrections versus the additional skill provided by deep-learning downscaling. This comparison allows us to determine whether the deep learning model is capturing nonlinear terrain/flow effects beyond what bias and elevation corrections can provide.
How to cite: Ennis, K. and Scher, S.: A Deep Learning Model Framework for High-Resolution Downscaling of ERA5 in the Austrian Alps, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16275, https://doi.org/10.5194/egusphere-egu26-16275, 2026.
Biases in boundary-layer clouds and cloud–radiation interactions remain a leading source of uncertainty in Earth’s energy budget and global-model performance. Using multiple satellite datasets (MODIS, PATMOS-x, CLARA-AVHRR, and CERES-EBAF), we diagnose biases in cloud fraction and cloud–radiation interaction in the Korea Meteorological Administration’s global Korean Integrated Model (KIM; 8-km horizontal resolution). We further propose an observation-constrained approach to improve a cloud fraction parameterization by combining observations with machine-learning-based diagnostics. Evaluation of forecasts for July 2022 and January 2023 shows that KIM overestimates the global-mean low-level cloud fraction by about 30%, while underestimating cloud fraction over major marine stratocumulus decks. In the tropics, KIM simulates excessive high-level cloud fraction, consistent with overly vigorous deep convection. Neural-network-based permutation importance and sensitivity analyses indicate that temperature, relative humidity, and lower-tropospheric stability (e.g., inversion strength) are key controls on cloud fraction. However, KIM fails to capture the dependence of cloud fraction on these controls in stratus and stratocumulus regimes. To address this limitation, we retune the parameters of the previously developed symbolic regression based cloud-fraction diagnostic parameterization for the KIM grid scale. We retune it using a CloudSat–ERA5 matched dataset. Specifically, we sample ERA5 along CloudSat tracks, upscale the matched dataset to 8 km (horizontal) and 20 hPa (vertical), and optimize the diagnostic parameters using differential evolution. The retuned diagnostic formulation reduces low- and high-cloud biases across the low to mid-latitudes and correspondingly reduces biases in surface shortwave radiation and outgoing longwave radiation (OLR). Notably, cloud fraction in the cumulus regime within the stratocumulus-to-cumulus transition region decreases substantially and becomes much closer to observations. These improvements are accompanied by a more realistic thermodynamic structure near the planetary boundary-layer top.
How to cite: Suhan, K. and Jihoon, S.: Improving Cloud–Radiation Interaction Simulations with an Observation-Constrained Symbolic-Regression Cloud-Fraction Diagnostic, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16584, https://doi.org/10.5194/egusphere-egu26-16584, 2026.
The background error covariance matrix (B) plays a central role in variational data assimilation by controlling how observational and background information are combined and is therefore essential for producing an accurate analysis. However, explicitly constructing B and computing its inverse are severely constrained in practice due to the extremely high dimensionality and associated computational cost.
To mitigate this limitation, this study proposes an unsupervised learning approach that directly estimates the inverse square root of the background error covariance (B⁻¹ᐟ²) from the forecast error patterns. A feedforward neural network that learns a linear matrix corresponding to B⁻¹ᐟ² is trained under whitening constraints, which are intrinsic properties of B⁻¹ᐟ². The learned operator satisfies symmetry and positive definiteness, and the inverse background error covariance (B⁻¹) is obtained in a numerically stable manner by squaring the learned B⁻¹ᐟ².
The performance of the learned B⁻¹ᐟ² is evaluated through verification of its whitening properties and comparison with a reference B⁻¹ᐟ² constructed by the pseudo inversion of B using singular value decomposition, demonstrating that it reproduces the dominant structural characteristics and leading modes of the reference. The learned B⁻¹ is further implemented within a three-dimensional variational data assimilation (3D-Var) framework, where it stably controls the spatial structure of analysis increments without numerical instability. These results indicate that the proposed unsupervised approach provides a practical and effective alternative for estimating and applying the B⁻¹ in variational data assimilation.
How to cite: Lee, H.-J. and Ham, Y.-G.: Unsupervised learning of background error covariance matrix for variational data assimilation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17084, https://doi.org/10.5194/egusphere-egu26-17084, 2026.
Exceptional and extreme events like wildfires, pollution episodes, or volcanic eruptions require near-real-time (NRT) detection to enable effective mitigation and impact reduction. While satellite geophysical products provide valuable information, their NRT availability is limited to targeted atmospheric species. In contrast, radiances (raw satellite data) provide the full spectral information, within of which a wide variety of atmospheric compounds and geophysical parameters simultaneously exist. The IASI atmospheric sounders aboard the Metop satellites provide an extensive archive of such spectra, covering the spectral signature of stable greenhouse gases and highly variable trace gases relevant to extreme events.
This work builds on these observations to develop an AI-based automated detection system. By validating our approach on IASI’s long historical record, we aim to establish a robust framework capable of fully exploiting the higher spectral resolution and enhanced trace-gas sensitivity of the next-generation IASI-NG launched aboard the Metop-SG satellite in summer 2025.
The methodology is organized in two phases. First, the long-term IASI dataset (since 2007) is used to develop AI models for extreme event detection. An event "atlas" is built associating spectral signatures with documented events, and used to train supervised models, including neural networks, directly on radiance data. Unsupervised techniques are also applied to identify unlabeled anomalies and potential unknown atmospheric species.
In the second phase, these models will be adapted to IASI-NG, accounting for instrumental differences and ensuring consistency over the operational overlap period. An operational processing system will then be deployed to provide continuous and reliable monitoring of extreme events.
At this stage, we will focus on presenting the development and results of the fire event atlas produced in the first phase.
How to cite: Luque Lázaro, Á., Boynard, A., Safieddine, S., Hadji-Lazaro, J., and Prunet, P.: Exceptional events detection using remote sensing and artificial intelligence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18854, https://doi.org/10.5194/egusphere-egu26-18854, 2026.
Fine particulate matter (PM2.5) is one of the most harmful air pollutants, posing severe risks to human health and contributing significantly to premature mortality worldwide. Accurate high-resolution monitoring and forecasting of PM2.5 are therefore essential for air quality management, public health assessment, and the design of effective mitigation policies. However, operational atmospheric composition models such as the Copernicus Atmosphere Monitoring Service (CAMS) provide global fields at relatively coarse spatial resolution (~40 km), limiting their ability to represent local-scale pollution patterns driven by complex interactions between emissions, meteorology, and topography. Higher-resolution regional CAMS products (~10 km) partly address this limitation but are computationally expensive and are restricted to specific geographical domains, mainly Europe. As a result, high-resolution information remains difficult to obtain consistently at the global scale.
In this work, we present a deep learning–based super-resolution approach to downscale PM2.5 concentration fields from 40 km to 10 km resolution, bridging the gap between global model outputs and regional-scale applications. The proposed approach is based on a SwinFIR architecture, a hierarchical Vision Transformer that leverages shifted window self-attention to efficiently capture multiscale spatial dependencies. The model ingests multiple low-resolution dynamic variables from CAMS, including PM2.5, 2-meter temperature (T2M), 10-meter wind speed components (U10, V10), dewpoint (D2M) and boundary layer height (BLH), providing both chemical and meteorological context. In addition, high-resolution static data, such as orography and population, are introduced through a secondary branch, enabling the model to condition the super-resolutionprocess on fine-scale geographical features that strongly influence pollutant distributions. The output consists of high-resolution (10 km) PM2.5 fields. Model performance is evaluated using both a held-out test period and independent ground-based PM2.5 observations from the European Environment Agency.
Results show that the model effectively reconstructs fine-scale PM2.5 structures and reduces biases present in the global forecasts. Verification against ground-based observations indicates that the model achieves performance comparable to high-resolution CAMS Europe regional forecasts. The proposed SwinFIR model consistently outperforms a carefully optimized state-of-the-art U-Net baseline across multiple evaluation criteria, including error metrics, spatial correlation, and structural consistency. These improvements reflect the ability of self-attention mechanisms to capture long-range spatial interactions that are difficult to model using purely convolutional approaches.
Beyond predictive performance, we also focus on interpretability. Feature importance analyses quantify the relative contribution of each input variable, demonstrating that static inputs used play a key role in the downscaling process. Attention maps further reveal that the model focuses on physically meaningful events, including high-concentration peaks and regions of strong wind, indicating physically consistent behavior.
Finally, transferability was assessed by applying the model to North America, a region unseen during training. Evaluation against AirNow observations shows reasonable generalization performance, while highlighting the need for further research to improve robustness and extrapolation to unseen regions.
Overall, this study demonstrates the potential of Transformer-based architectures for data-driven downscaling of atmospheric composition fields, providing both improved accuracy and enhanced physical interpretability. The proposed framework offers a promising tool for high-resolution air
quality applications based on global model outputs.
How to cite: Martínez-Roig, M., Granell-Haro, F., Monsalvez-Pozo, K., Plaza-Martin, N. P., Galván Fraile, V., Martin Otto, P. R., Bieser, J., Flemming, J., Harder, P., Razinger, M., and Azorin-Molina, C.: Interpretable Swin Transformer–Based Downscaling of PM2.5 Air Pollution Field, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19262, https://doi.org/10.5194/egusphere-egu26-19262, 2026.
Posters virtual: Wed, 6 May, 14:00–18:00 | vPoster spot 5
EGU26-2796 | ECS | Posters virtual | VPS4
Investigating the formation mechanisms of hydroxyl dicarboxylic acids based on machine learningWed, 06 May, 14:15–14:18 (CEST) vPoster spot 5
Secondary organic aerosol (SOA) has been shown to significantly impact climate, air quality, and human health. Hydroxyl dicarboxylic acids (OHDCA) are generally of secondary origin and ubiquitous in the atmosphere, with high concentrations in South China. This study explored the formation of representative OHDCA species based on time-resolved measurements and explainable machine learning. Malic acid, the most commonly studied OHDCA, had higher concentrations in the noncontinental air (63.7 ± 33.3 ng m–3) than in the continental air (7.5 ± 1.4 ng m–3). Machine learning quantitatively revealed the high relative importance of aromatics and monoterpenes SOA, as well as aqueous processes, in the noncontinental air, due to either shared precursors or similar formation pathways. Isoprene SOA, particle surface area, and ozone corrected for titration loss (Ox) also elevated the concentrations of malic acid in the continental air. Aqueous photochemical formation of malic acid was confirmed given the synergy between LWC, temperature, and Ox. Moreover, the OHDCA-like SOA might have facilitated a relatively rare particle growth from early afternoon to midnight in the case with the highest malic acid concentrations. This study enhances our understanding of the formation of OHDCA and its climate impacts.
How to cite: Li, H. and Lyu, X.: Investigating the formation mechanisms of hydroxyl dicarboxylic acids based on machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2796, https://doi.org/10.5194/egusphere-egu26-2796, 2026.
EGU26-9696 | Posters virtual | VPS4
Fog Risk Monitoring and Assessment for India Using Bayesian Networks and ECMWF IFS Ensemble Prediction SystemWed, 06 May, 14:18–14:21 (CEST) vPoster spot 5
Dense fog events across India severely disrupt aviation, surface transportation, and daily activities during winter months, with northern districts experiencing extended periods of visibility issues. Building upon the WRF-based ensemble fog forecasting over the Indo-Gangetic Plain and BOFFIN-Melbourne's Bayesian Decision Network framework, this study proposes a continuous risk monitoring and decision support system at district-level (admin-2).
The operational system will conduct daily continuous risk assessment, leveraging satellite observations from MODIS/VIIRS/INSAT-3D and the ECMWF IFS ensemble forecasts (51 members, 0.25° resolution) including probabilistic meteorological predictions of temperature, dewpoint, wind speed, boundary layer height, relative humidity profiles, and cloud cover, to characterize antecedent fog conditions and to establish baseline occurrence patterns.
A Bayesian Network will integrate these layers to provide real-time short-term forecasts using pre-defined conditional probability tables which encode relationships between stable boundary layer conditions, radiative cooling, and regional fog formation mechanisms. The operational output of the algorithms will be in the form of traffic light decision matrix for each district: Green (Minimal/Low risk - Monitor), Yellow (Moderate risk - Be Aware), Orange (High risk - Be Prepared), Red (Extreme risk - Take Action).
This paper will present the validation results from pilot districts and the development framework for scaling to nationwide continuous risk assessment, demonstrating the system's potential for proactive decision-making in transportation management, aviation operations, and public safety advisories.
References
- Parde, Avinash N., et al. "Operational probabilistic fog prediction based on ensemble forecast system: A decision support system for fog." Atmosphere 13.10 (2022): 1608.
- Boneh, Tal, et al. "Fog forecasting for Melbourne Airport using a Bayesian decision network." Weather and Forecasting 30.5 (2015): 1218-1233.
How to cite: Guttikunda, S. K., Kalladath, N., Tucci, R. R., Ouma, J., Amdihun, A., and Dammalapati, S. K.: Fog Risk Monitoring and Assessment for India Using Bayesian Networks and ECMWF IFS Ensemble Prediction System, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9696, https://doi.org/10.5194/egusphere-egu26-9696, 2026.
EGU26-10657 | ECS | Posters virtual | VPS4
Machine learning analysis of global LAI trends and their relationship with climate variability (1982–2022)Wed, 06 May, 14:21–14:24 (CEST) vPoster spot 5
Understanding vegetation responses to climate variability is essential for assessing long-term ecosystem dynamics. Leaf Area Index (LAI) is a widely used variable to characterise vegetation state and productivity. However, attributing observed global LAI trends to specific climatic drivers remains challenging due to non-linear interactions, strong spatial heterogeneity, and scale-dependent processes.
This study is conducted within the framework of the PROFECIA project, which aims to improve the monitoring and interpretation of vegetation responses to climate change by combining remote sensing observations and artificial intelligence techniques. We analyse global LAI trends over the period 1982–2022 using the GEOV2-AVHRR long-term satellite record and examine their relationship with trends in key climatic variables obtained from the ERA5 reanalysis, including temperature, precipitation, radiation, and several indicators of water availability and drought conditions. All trends are computed consistently over the 1982-2022 temporal record to ensure a homogeneous assessment of long-term vegetation–climate relationships at the global scale.
The vegetation–climate relationships are modelled using a suite of machine learning approaches, including tree-based methods and neural networks, designed to capture non-linear responses across diverse climatic and ecological conditions. Particular emphasis is placed on the role of the training strategy: different spatio-temporal sampling schemes are evaluated to assess their impact on model performance, robustness, and generalisation capability when analysing long-term trends at the global scale.
To move beyond purely predictive modelling, the study systematically applies explainable artificial intelligence (XAI) techniques to interpret the trained models. Methods such as SHAP-based attribution and partial dependence analyses are used to quantify the relative contribution of individual climatic drivers to observed LAI trends and to examine how these contributions vary across regions and time periods.
Overall, this work highlights the importance of combining robust machine learning training strategies with interpretability tools to improve the attribution of long-term vegetation trends to climatic drivers, providing new insights into global vegetation–climate interactions over the last four decades.
How to cite: García-Diaz, D., Aguilar, F., Schauman, S., and Verger, A.: Machine learning analysis of global LAI trends and their relationship with climate variability (1982–2022), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10657, https://doi.org/10.5194/egusphere-egu26-10657, 2026.
EGU26-10823 | ECS | Posters virtual | VPS4
A Hybrid Neural Network and Cellular Automata Model for spatiotemporal Forecasting of PM10 and PM2.5 in Lima, PeruWed, 06 May, 14:24–14:27 (CEST) vPoster spot 5
Particulate matter (PM) pollution represents a significant public health concern, in Lima, Peru. This issue is further compound by the lack of accurate forecasting tools due to limited monitoring networks. This study addresses this gap by developing and validating a hybrid model combining a Multilayer Perceptron (MLP) neural network and a type of rule-based Cellular Automata (CA) simulation. This model simulates and forecasts the spatiotemporal dispersion of PM10 and PM2.5. Using a decade of historical PM data (2015-2024) from seven monitoring stations and NASA's meteorological data, an optimized MLP was trained to learn the complex, non-linear transition rules from 47 engineered features. The model demonstrated remarkable performance in historical validation (R2 > 0.90), outperforming standard baseline models. When fed with weather forecast data, the model can operate as an Early Warning System (EWS), providing a reliable prediction horizon to anticipate the exceedance of Air Quality Standards. The resulting hotspot maps accurately identify high-risk areas, confirming the potential of this hybrid model as a robust, proactive, and quantitative tool for air quality management and public health protection in complex urban environments.
How to cite: Maita, B., Condezo, P., Llanto, J., Huaman, S., and Sanabria, J.: A Hybrid Neural Network and Cellular Automata Model for spatiotemporal Forecasting of PM10 and PM2.5 in Lima, Peru, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10823, https://doi.org/10.5194/egusphere-egu26-10823, 2026.
EGU26-18107 | Posters virtual | VPS4
Satellite-Based PM2.5 Estimation in Data-Sparse Urban Environments: Comparing Machine Learning and Geostatistical Approaches in Kolkata, IndiaWed, 06 May, 14:27–14:30 (CEST) vPoster spot 5
Fine particulate matter (PM2.5) is among the foremost environmental determinants of human health, contributing to cardiovascular disease, respiratory illness, and premature mortality. In rapidly urbanizing regions of the Global South, accurate spatial characterization of PM2.5 exposure requires spatially continuous concentration surfaces that also provide reliable uncertainty estimates, yet ground-based monitoring networks remain severely sparse. Kolkata, India’s third-largest metropolitan area (population 14.9 million), exemplifies this challenge: only seven regulatory monitoring stations cover the entire city, leaving large areas unobserved.
This study evaluates how different PM2.5 surface generation strategies—satellite-based machine learning (ML) and spatial interpolation—differ not only in predictive accuracy but also in their ability to provide decision-relevant uncertainty under sparse monitoring conditions. Using six years of daily observations (2019–2024), we compare two complementary approaches. The first employs satellite-based ML, integrating Sentinel-5P trace gases, MODIS aerosol optical depth, ERA5 meteorological reanalysis, and static urban features (VIIRS nightlights, population density) to predict PM2.5. The second evaluates spatial interpolation methods—ordinary kriging, inverse distance weighting (IDW), and simple averaging—using station observations alone.
For satellite-based ML (Random Forest), the station-level model achieved R2 = 0.79 under leave-one-station-out (LOSO) validation, while grid-based model trained on kriging-interpolated targets reached R2 = 0.70 under temporal out-of-sample validation (train: 2019–2022, test: 2023–2024). Feature importance analysis consistently identified dewpoint temperature, air temperature, and surface albedo as dominant predictors, indicating that atmospheric conditions exert stronger control on PM2.5 variability than emission proxies or land-use variables.
For spatial interpolation evaluated under daily LOSO, all methods achieved comparable point prediction accuracy (R2 ≈ 0.85). However, uncertainty calibration diverged sharply. Ordinary kriging achieved 88% empirical coverage for nominal 95% prediction intervals (90% when including observation noise)—approaching theoretical calibration—whereas IDW and simple averaging exhibited severe under-coverage (45–52%), substantially underestimating true prediction error.
These findings yield three key insights: (1) satellite-derived predictors enable spatially complete PM2.5 estimation beyond monitoring locations, though with moderate accuracy; (2) when temporally aligned station data are available, interpolation achieves higher point accuracy than satellite-based ML; and (3) regardless of estimation strategy, only geostatistical approaches provide uncertainty estimates suitable for health-protective decision-making. We conclude that hybrid frameworks combining satellite-based spatial prediction with kriging-derived uncertainty characterization offer a principled pathway for generating spatially complete and risk-aware PM2.5 maps in data-sparse urban environments.
How to cite: Raj, A., Dasgupta, T., Sinha, M., and Mitra, A.: Satellite-Based PM2.5 Estimation in Data-Sparse Urban Environments: Comparing Machine Learning and Geostatistical Approaches in Kolkata, India, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18107, https://doi.org/10.5194/egusphere-egu26-18107, 2026.
EGU26-6386 | ECS | Posters virtual | VPS4 | Highlight
Automated Analysis of City Level Climate Action Plans using Natural Language Processing TechniqueWed, 06 May, 14:42–14:45 (CEST) vPoster spot 5
The growing urgency of climate action at the city level has led to an exponential rise in documents that describe a city’s policy, action plan, or progress towards climate action. The increased number of documents has made it increasingly difficult for governments to track commitments and compare approaches across jurisdictions. These documents are essential for informed decision-making, but extracting useful information from unstructured PDF reports remains a largely manual, resource-intensive, and inconsistent process. Recent advances in AI and large language model (LLM) based document understanding offer strong potential, but their application in urban climate governance workflows is still limited. Integrating AI-driven document analysis into this workflow offers opportunity for building scalable, standardized, and transparent climate policy assessment.
This study presents an AI-assisted natural language processing (NLP) pipeline that automatically extracts, segments, and classifies climate actions from diverse policy documents. The workflow integrates layout-aware text extraction with an action-segmentation mechanism to identify action statements across heterogeneous formats. A fine-tuned, two-stage ClimateBERT classifier then categorizes actions: Stage 1 differentiates mitigation and adaptation measures (F1 = 93%), while Stage 2 assigns domain-specific sub-categories, achieving 92% F1 for mitigation and 91% for adaptation. An equity-detection module further identifies references to vulnerable groups, inclusivity, and justice-oriented themes.
The pipeline significantly reduces manual review effort and enhances consistency in understanding climate action. By enabling standardized comparisons, the approach directly supports mayors, policymakers, and urban practitioners in evaluating progress and designing more effective and equitable interventions.
As AI capabilities advance, such automated tools will strengthen climate governance by improving the accessibility, reliability, and strategic value of climate policy data.
How to cite: Sahu, S. and Shanker, S. and the MU NLP team: Automated Analysis of City Level Climate Action Plans using Natural Language Processing Technique, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6386, https://doi.org/10.5194/egusphere-egu26-6386, 2026.
