This session aims to highlight and bring together recent efforts in hydrological forecasting, using machine learning based techniques and/or hybrid approaches. Contributions are welcome showcasing examples of model developments (ranging from implementations to operational setups), studies ranging from local to global scales and across different time horizons (short-, medium- and long-term), as well as studies showcasing the efforts data-driven/hybrid approaches to tackle challenges in hydrological forecasting. We particularly welcome talks that reach beyond the description of machine learning architectures to uncover physical and human-induced processes, account for uncertainties, generate novel insights about hydrological forecasting, or support efforts in reducing common forecasting difficulties.
Other topics related to the subdivision of Hydrological Forecasting and the corresponding sessions can be found here: https://www.egu.eu/hs/about/subdivisions/hydrological-forecasting/
Orals: Wed, 6 May, 10:45–15:45 | Room B
Reliable global streamflow forecasting is essential for flood preparedness, yet data-driven models often suffer from a performance gap when transitioning from historical reanalysis to operational forecast products. This study introduces AIFL (Artificial Intelligence for Floods), a deterministic LSTM-based model designed for global daily streamflow forecasting. Trained on 18,588 basins curated from the CARAVAN dataset, AIFL utilizes a novel two-stage training strategy to bridge the reanalysis-to-forecast domain shift. The model is first pretrained on 40 years of ERA5 reanalysis to capture robust hydrological processes, then fine-tuned on operational Integrated Forecasting System (IFS) control forecasts to adapt to the specific error structures and biases of operational numerical weather prediction. To our knowledge, this is the first global model trained end-to-end within the CARAVAN and MultiMet ecosystem. Evaluated on an independent temporal test set (2021–2024), AIFL achieves a median KGE’ of 0.66 and a median NSE of 0.53. Benchmarking results show that AIFL is highly competitive with current state-of-the-art global systems, maintaining a transparent and reproducible forcing pipeline while demonstrating exceptional reliability in extreme event detection. The resulting model provides a streamlined and operationally robust baseline for the global hydrological community.
How to cite: Taccari, M. L., Tazi, K., M. Morrison, O., Grafberger, A., Colonese, J., Carton de Wiart, C., Prudhomme, C., Mazzetti, C., Chantry, M., and Pappenberger, F.: AIFL: A New Global Flood Forecasting Model Trained on CARAVAN, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13650, https://doi.org/10.5194/egusphere-egu26-13650, 2026.
Compound floods arise from interacting drivers such as rainfall, river discharge, and coastal surge, posing challenges for risk assessment due to their stochastic nature. Traditional hydrodynamic models, while accurate, are computationally expensive for ensemble forecasting and probabilistic analysis. Surrogate models offer a promising alternative, but most existing approaches focus on single drivers or static flood conditions, limiting their applicability to compound events. The hybrid SFINCS–LSG surrogate model addresses these gaps by integrating low-resolution SFINCS simulations with Empirical Orthogonal Function (EOF) decomposition and Sparse Gaussian Process learning to emulate high-resolution flood dynamics. Two case studies, Charleston, USA, and Brisbane, Australia, were selected to evaluate model performance under diverse flood conditions. Training datasets were generated by scaling historical events decomposed to individual flood drivers to ensure coverage of diverse flood conditions. Model skill was assessed against high-fidelity SFINCS simulations using Critical Success Index (CSI) for flood extent and Root Mean Square Error (RMSE) for flood depth. Our results showed that SFINCS–LSG achieved speed-ups of 50–150× compared to high-fidelity SFINCS simulations with good accuracy. The median RMSE for flood depth was 0.06 m for the Charleston and a CSI of 0.96 and 0.91 respectively. However, performance varied by flood type due to large variability in extent between coastal and compound or pluvial-fluvial events. The compression of spatial information through EOF analysis introduced noise, which constrained the model’s ability to reproduce dominant flood driver zones. Despite these limitations, the approach demonstrates potential for real-time forecasting and probabilistic risk analysis where many simulations are required. This research advances state-of-the art surrogate models by capturing dynamic spatiotemporal flood evolution under multi-driver conditions rather than static peak inundation. Overall, the SFINCS–LSG framework offers a scalable solution for accelerating compound flood modelling at very limited loss of accuracy.
How to cite: Eilander, D., de Goede, R., Leijnse, T., and Fræhr, N.: Hybrid surrogate modeling of compound flood events using SFINCS-LSG , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11062, https://doi.org/10.5194/egusphere-egu26-11062, 2026.
Predicting high-resolution inundation at large spatial and temporal scales is important in understanding future water availability and flood risk. Classically, hydrologic models have been used to model inundation, but the computational expense associated with applying hydrologic models at large scales has motivated the use of other methodologies, such as machine learning. We propose a hybrid physics-based and machine learning modeling approach to produce high-resolution inundation maps at a much lower computational cost than high-resolution physics-based modeling while still maintaining high accuracy. This methodology is then tested in a representative watershed in Colorado, USA.
The proposed hybrid physics-based and machine learning modeling approach consists of a coarse spatial resolution hydrologic model and a random forest downscaling postprocessing step. First, a 1km resolution integrated surface-subsurface hydrologic model, ParFlow-CLM, is ran for the spatial and temporal domain of interest. Then, the resultant modeled inundation as well as additional climate and geographical parameters are fed into a random forest model which predicts inundation at a higher spatial resolution. We tested this methodology in a 1800km2 watershed in Colorado, USA during the 2019 water year to predict modeled inundation produced by a 100m resolution hydrologic model. In our test case, this hybrid methodology predicted whether each fine resolution cell is inundated at each hourly timestep correctly >97% of the time and maintained high accuracy in unseen timesteps as well as in unseen locations within the same region. We will also discuss next steps to predict real-world inundation by training the random forest model on 30m resolution satellite data. This study shows the potential for this methodology to be applied at the continental scale to predict high-resolution inundation accurately and efficiently.
How to cite: Peeples, A., Leonarduzzi, E., Condon, L. E., and Maxwell, R. M.: Accurately and Efficiently Predicting High-Resolution Inundation using a Hybrid Machine Learning and Physics-Based Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14903, https://doi.org/10.5194/egusphere-egu26-14903, 2026.
Long Short-Term Memory (LSTM) networks have demonstrated state-of-the-art performance for operational flood forecasting at the daily scale (Nearing et al., 2024). Recent advances have extended LSTM-based models to higher temporal resolutions through multi-frequency LSTM architectures (Acuña Espinoza et al., 2025) and introduced robust strategies for handling missing data, such as masked-mean embeddings (Gauch et al., 2025).
Building on this work, we introduce an LSTM-based approach that allows producing hourly flood forecasts in an operational setting, while being robust to missing data. Moreover, using masked-mean embeddings plus teacher-forcing (Williams et al., 1989) and noise injection strategies during training, allows the model to integrate observed stream flow observations when available, for enhanced prediction accuracy, while keeping the flexibility to operate without this signal.
To evaluate model performance, we benchmarked the new approach against LARSIM, the current operational model in many federal states in Germany. Our results show that the LSTM-based architecture outperforms the LARSIM model in predictive accuracy, while additionally offering robustness to missing inputs and faster inference times.
These findings highlight the potential of deep learning–based models for operational flood forecasting at an hourly resolution, while introducing strategies to increase robustness and add valuable information, when available.
Reference:
Acuña Espinoza, E., Kratzert, F., Klotz, D., Gauch, M., Álvarez Chaves, M., Loritz, R., & Ehret, U. (2025). Technical note: An approach for handling multiple temporal frequencies with different input dimensions using a single LSTM cell. Hydrology and Earth System Sciences, 29(6), 1749–1758. https://doi.org/10.5194/hess-29-1749-2025
Gauch, M., Kratzert, F., Klotz, D., Nearing, G., Cohen, D., & Gilon, O. (2025). How to deal with missing input data. Hydrology and Earth System Sciences, 29(21), 6221–6235. https://doi.org/10.5194/hess-29-6221-2025
Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T. Y., Weitzner, D., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559–563. https://doi.org/10.1038/s41586-024-07145-1
Williams, R. J., & Zipser, D. (1989). A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2), 270-280.
How to cite: Acuna Espinoza, E., Kratzert, F., Gauch, M., Álvarez Chaves, M., Loritz, R., and Ehret, U.: Robust hourly flood forecasting using LSTM: Handling missing inputs and integrating discharge observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3180, https://doi.org/10.5194/egusphere-egu26-3180, 2026.
Flash flooding is one of the deadliest natural hazards worldwide, causing severe loss of life and infrastructure damage. Predicting extreme flood events remains highly challenging because they often fall outside the range of historical observations, involve small-scale storm processes that are poorly resolved by existing forecasting systems, and include nonlinear flood-generation mechanisms that are inadequately represented in current models. Although pure AI models, such as LSTMs, generally outperform traditional hydrologic models in simulation accuracy, they often fail to predict extreme streamflow beyond a certain threshold and tend to underestimate extremes due to structural limitations, such as bounded activation functions. Differentiable models (DMs), which jointly train neural networks with process-based models, can overcome these limitations through interpretable physical modules and physically consistent representations, thereby achieving improved accuracy in extreme-event prediction compared with LSTMs. Here, we will demonstrate (1) how DMs improve extreme-event predictions and how dynamic parameters contribute to this improvement; (2) the feasibility of high-resolution, hourly differentiable models for operational extreme flood forecasting by resolving short-lived, small-scale storms; (3) the importance of incorporating different nonlinear flood-generation mechanisms; and (4) the robustness of DMs for long-term climate change impact assessment.
How to cite: Song, Y., Yang, W., Shen, C., Ji, H., Lonzarich, L., Bindas, T., Sawadekar, K., and Liu, J.: Improving Operational Extreme Flood Forecasting and Climate Change Impact Assessment with Physics-Embedded Differentiable Modeling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15795, https://doi.org/10.5194/egusphere-egu26-15795, 2026.
Catchment hydrology relies on models with complementary strengths and limitations in terms of physical interpretability, data requirements, structural flexibility, and predictive performance. Conceptual process-based models provide parsimony and physical meaning, but often sacrifice predictive performance. Data-driven approaches offer greater flexibility and predictive skill, yet often sacrifice interpretability and physical consistency. Hybrid strategies seek to combine these strengths, but many existing implementations rely on loosely constrained components that obscure process understanding.
In this study, we explore a structured reservoir network architecture for rainfall–runoff modelling as an intermediate approach between conceptual hydrological models and fully data-driven approaches. Rather than framing the method as a neural network, we focus on assembling physically interpretable reservoir elements into a network structure that can be calibrated using gradient-based optimization while preserving hydrological meaning.
The proposed model represents runoff generation through multiple parallel reservoir chains, each governed by a conceptual soil moisture balance with physically interpretable parameters. Excess rainfall from each chain is routed using convolution with a gamma transfer function to represent delayed runoff response. The routed contributions are combined through convex weighting, enabling a transparent and controlled aggregation of parallel runoff pathways. The overall architecture remains mass-conservative and avoids black-box recurrent components.
The approach is evaluated across multiple catchments within the Moselle basin (27,100 km²), which exhibits substantial heterogeneity in elevation and land use. The model is driven by daily precipitation and potential evapotranspiration and evaluated against observed discharge. Performance is assessed using the Nash–Sutcliffe Efficiency and hydrological signatures, and results are compared to the GR4J conceptual model under identical calibration conditions.
Results indicate that the proposed reservoir network achieves performance comparable to or slightly better than GR4J, with a mean validation NSE of 0.70 compared to 0.67. Improvements are particularly evident for low-flow metrics and flow-duration curve characteristics. Beyond predictive performance, the model enables interpretation of the relative contributions and temporal dynamics of parallel runoff generation pathways.
Overall, this work demonstrates the potential of reservoir network architectures as a transparent and flexible modelling framework for rainfall–runoff simulation and process exploration. Future work will focus on incorporating additional catchment information, such as permeability and physiographic descriptors, and on extending the approach toward regional-scale applications across heterogeneous catchments.
How to cite: Farfán-Durán, J. F., V. M. do Nascimento, T., Kavetski, D., and Fenicia, F.: Exploring reservoir network structures for rainfall–runoff modelling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11200, https://doi.org/10.5194/egusphere-egu26-11200, 2026.
This study addresses the limitations of traditional deep learning models in hydrological forecasting, which often lack physical interpretability and struggle with extreme events. We propose a hybrid framework that combines data-driven deep learning with physical constraints to enhance model accuracy and robustness. Traditional deep learning models (e.g., LSTM) have shown promise in rainfall-runoff prediction but are criticized for their "black-box" nature and poor performance under extreme conditions, while physical-based models, though interpretable, are computationally expensive and rely on detailed parameterization. To bridge this gap, we integrate physical constraints (e.g., water balance, monotonicity) into LSTM networks through three key approaches: extreme event constraints that add penalties for violating physical laws, monotonicity constraints ensuring runoff increases with rainfall intensity via ReLU-based loss functions, and hard constraints projecting outputs to strictly adhere to hydrological laws. Applied to 2683 basins globally, our physics-guided LSTM (PHY-LSTM) improved the Nash-Sutcliffe Efficiency (NSE) by 0.10 and reduced the Root Mean Square Error (RMSE) by 15% compared to standard LSTM, with a 20% enhancement in flood peak prediction using synthetic extreme samples. Additionally, we identified time-varying parameters across climate zones, revealing trends in water storage capacity (1.7mm/decade in wet regions, -0.6mm/decade in arid regions). This framework bridges data-driven efficiency and physical interpretability, enabling reliable predictions under extreme conditions and providing insights into hydrological processes, validated globally and applicable to water resource management and climate change impact assessments.
How to cite: Xie, K., Shu, Z., ning, Z., Ruan, Y., Zheng, Y., Liu, P., Wang, G., Jin, J., and Zhang, J.: Physics-Guided Deep Learning for Rainfall-Runoff Modeling: Integrating Physical Constraints and Data-Driven Approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1781, https://doi.org/10.5194/egusphere-egu26-1781, 2026.
Hydrological modeling has reached a point where deep learning models, especially those based on the LSTM architecture, are being used operationally by national agencies, the private sector, as well as across thousands of academic publications. However, by far the most common strategy is to use these models in a lumped setup, no matter the scale of the application. For example, Nearing et al. (2024) apply LSTM based models in a lumped setup at a global scale. Unfortunately, this withholds information that is relevant to precisely predict floods, such as the location of a precipitation event relative to the prediction point. Similarly, the spatial averaging over the entire upstream area dampens the precipitation signal that is provided to the model.
Classical hydrologic models use distributed or semi-distributed setups to solve this problem: they divide the basin into pixels or subpolygons and route streamflow along the river graph. There are first attempts to translate this semi-distributed modeling paradigm to end-to-end deep learning models, but so far they are typically trained only on individual river networks or select geographical regions (e.g., Kratzert et al., 2021, Kraft et al., 2025), lag behind the performance of lumped models (e.g., Kirschstein et al., 2021), cannot generalize to unseen river networks (e.g., Vischer et al., 2025), or are global and applicable ungauged basins but not trained end-to-end (e.g., Yang et al., 2025).
With the learnings and experience from operating lumped LSTM models at a global scale for multiple years, we revisit semi-distributed modeling with deep learning at a global scale with a focus on end-to-end training. In this submission, we present our version of a global end-to-end semi-distributed hydrologic model. We detail the model setup, its training procedure, and compare this model to the lumped setup. Our evaluation shows that the semi-distributed model has superior performance compared to the lumped model, especially for large, ungauged rivers. Finally, we highlight how this modeling approach is a step towards a broader multi-output, multi-modal system that propagates more information than just streamflow or physical quantities in general.
References:
- Kirschstein, N., et al. "The Merit of River Network Topology for Neural Flood Forecasting." Forty-first International Conference on Machine Learning. 2024.
- Kraft, B., et al. DROP: A scalable deep learning approach for runoff simulation and river routing. Authorea. November 25, 2025.
- Kratzert, F., et al. "Large-scale river network modeling using graph neural networks." EGU General Assembly Conference Abstracts. 2021.
- Nearing, G., et al. "Global prediction of extreme floods in ungauged watersheds." Nature 627.8004 (2024): 559-563.
- Vischer, M., et al. "Spatially Resolved Rainfall Streamflow Modeling in Central Europe." EGUsphere 2025 (2025): 1-26.
- Yang, Y., et al. (2025). Global daily discharge estimation based on grid long short-term memory (LSTM) model and river routing. Water Resources Research, 61.
How to cite: Kratzert, F., Gauch, M., Metzger, A., Klotz, D., Fronman, S., and Cohen, D.: Semi-Distributed Deep Learning Models at Global Scale, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19528, https://doi.org/10.5194/egusphere-egu26-19528, 2026.
Today water and resource managers face a significant challenge managing systems that are rapidly evolving in a warming climate, where historical observations are no longer a reliable guide. Capturing interactions from bedrock to treetops is important to understand water stresses and is a critical gap in our current models. Simulations with integrated hydrology models (that solve the 3D Richards' equation and 2D shallow water equations in a globally-implicit manner) provide robust results out to continental scales, yet are computationally expensive. Groundwater-surface water are tightly coupled and can have a large impact on watershed dynamics, yet are challenging for all models to accurately resolve.
We have developed a hybrid physics-based, machine learning digital twin over the entire continental US (CONUS). This proof-of-concept forecast system runs operationally, providing all hydrologic states and fluxes from bedrock to the top of the canopy at hourly timesteps and greater than 1km resolution. Automated comparison to observations is enabled through the HydroData platform, supporting continuous evaluation and model improvement. This talk will highlight the technical challenges of combining integrated hydrologic modeling with machine learning in a national forecast system, including physics-based approaches that improve solver performance by more than an order of magnitude for continental-scale simulations. Machine learning emulators embedded within integrated hydrology models can also drastically reduce computational burden and provide 30m spatial resolution for groundwater and surface water. We advance a vision that deploys these models and openly available forcing and parameter datasets to understand future water challenges from local to continental scales.
How to cite: Maxwell, R., Sandoval, L., Ma, Y., Peeples, A., Sawma, M. J., Defnet, A., Artavanis, G., Bennett, A., and Condon, L.: Advancing integrated continental-scale hydrologic forecasting through democratized data and ML-accelerated modeling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15738, https://doi.org/10.5194/egusphere-egu26-15738, 2026.
Artificial Intelligence (AI) approaches are nowadays well-established tools to make hydro-meteorological forecasts. While several AI-based models are available to provide probabilistic meteorological (e.g. Lang et al., 2024) or hydrological (Nevo et al., 2022) short-range predictions at global scale, seasonal hydrological probabilistic forecasts at large scale are still lagging behind. Here, we present the updated results of our AI-based seasonal hydrological forecasts across the European hydro-climatic gradient (Bertini et al., 2025). We use an Encoder-Decoder model trained at the pan-European scale with a combination of in-situ hydrological observations, reanalysis data from the process-based E-HYPE hydrological model, and bias-adjusted seasonal meteorological forecasts from the ECMWF SEAS5 prediction system. The model is trained over 500 catchments across Europe, grouped in 11 clusters based on their hydrological regime (Pechlivanidis et al., 2020), and the predictions are compared against both climatology and the E-HYPE streamflow forecasts. Compared to our previous results, the updated Encoder-Decoder model provides improved deterministic and probabilistic performances, proving once again the potential of AI approaches for operational hydrological forecasting.
Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, F., Raoult, B., ... & Rabier, F. (2024). AIFS--ECMWF's data-driven forecasting system. arXiv preprint arXiv:2406.01465.
Nevo, S., Morin, E., Gerzi Rosenthal, A., Metzger, A., Barshai, C., Weitzner, D., ... & Matias, Y. (2022). Flood forecasting with machine learning models in an operational framework. Hydrology and Earth System Sciences, 26(15), 4013-4032.
Bertini, C., Du, Y., van Andel, S. J., and Pechlivanidis, I.: AI-based seasonal streamflow forecasts across Europe’s hydro-climatic gradient, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10567, https://doi.org/10.5194/egusphere-egu25-10567, 2025.
Pechlivanidis, I.G., Crochemore, L., Rosberg, J., & Bosshard, T. (2020). What are the key drivers controlling the quality of seasonal streamflow forecasts? Water Resources Research, 56, e2019WR026987. https://doi.org/10.1029/2019WR026987
How to cite: Bertini, C., Du, Y., van Andel, S. J., and Pechlivanidis, I.: AI-based seasonal probabilistic hydrological forecasts across Europe, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11878, https://doi.org/10.5194/egusphere-egu26-11878, 2026.
Structural uncertainty in physics-based models (PBMs) and the limited generalisability of purely data-driven techniques limit the ability of predicting daily streamflow in catchments in data-scare region. In order to enhance prediction accuracy and geographic transferability, this study proposes a coupled physics-based-machine learning (PBM-ML) framework that combines knowledge of hydrological processes with data-driven learning. The framework was tested in multiple catchments with different hydroclimatic conditions, encompassing basins in Japan and New Zealand. PBM-derived states and fluxes were fed into machine-learning models after a PBM (SWAT) was first calibrated to simulate daily streamflow. The Nash-Sutcliffe efficiency (NSE) and coefficient of determination (R2) were used to evaluate the performance of the model. Coupled PBM-ML models consistently performed better than standalone SWAT in all basins. Testing NSE improved from 0.69-0.76 for SWAT to 0.80-0.89 for coupled models in New Zealand and from 0.67-0.68 to 0.74-0.86 in Japan. SWAT-LSTM had the best prediction ability among the hybrid methods. Regionalization approaches were used to investigate the transferability of the model. The coupled models retained robust performance under partially gauged and fully ungauged conditions. These findings demonstrate that PBM-ML coupling could enhance streamflow prediction and transferability in data-scarce regions.
How to cite: Anand, V., Oki, T., and Singh, S. K.: A coupled physics-based and machine-learning approach for enhancing daily streamflow simulations in data-scarce regions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15401, https://doi.org/10.5194/egusphere-egu26-15401, 2026.
Traditionally streamflow forecasting is accomplished using process-based hydrological models. These models could range from simple lumped type to more detailed distributed models. Lumped type models are easy to setup while distributed models require considerable skill and experience for setup. Due to the growing availability of large amounts of spatial and temporal data from various sources, such as remote sensing and re-analyses, and recent advances in the computational power, machine learning models are gaining momentum for solving applied engineering problems, triggering conceptual shifts perhaps led by rapid progress in data science and availability of ready-to-be-deployed software tools. These models have the ability to extract complex dynamical nonlinearities without explicitly defining the involved physical processes and their governing mathematical formulations, as is followed in the case of hydrological models. It is believed that new trends and conceptual shifts are essential for generating new knowledge, challenging or validating prevailing assumptions, and enhancing operational applications, which may include several water management-related functions, hydropower generation operations, and flood risk management across a range of temporal and spatial scales.
In this study, two deep learning variants of machine learning models, i.e., (1) the attention-based encoder-decoder bidirectional long short-term memory (AB-ED-BiLSTM) network and (2) the attention-based encoder-decoder bidirectional gated recurrent units (AB-ED-BiGRU) network, were tested on multiple watersheds selected from the Ottawa River Basin, Canada. After developing and successfully evaluating watershed-specific models, regional versions of both models were developed and tested based on the leave-one-watershed-out strategy to emulate an ungauged scenario. Both models were driven mainly by soil moisture states of watersheds and meteorological data in order to evaluate their usefulness for streamflow forecasting at ungauged locations. Although not as ideal as one would desire, these models demonstrated reasonable skill in forecasting streamflow with one to seven days lead time when assessed in terms of coefficient of determination, Nash-Sutcliff Efficiency, and Kling-Gupta Efficiency performance metrics. However, considerable discrepancies were noticed in simulating peak flow values for certain watersheds. Overall results of the study suggest that soil moisture driven machine learning models can potentially be used to develop streamflow forecasting tools for ungauged locations, with AB-ED-BiGRU being computationally an inexpensive option compared to the AB-ED-BiLSTM model. Additional investigations will be required to improve their performance further, e.g., by employing multiple soil moisture products, available through remote sensing and re-analyses sources, and ensemble modelling techniques. Based on continuous scientific progress, emerging machine learning frameworks and architectures, and better understanding of the origins and limitations of existing models, improved hydrological forecasting at ungagged locations can be made possible. In essence, this study contributes towards enhancing our understanding of the role of soil moisture in developing machine learning based streamflow modelling and forecasting tools to support operational applications at ungauged locations, which are often neglected when developing real-time streamflow forecasting systems.
How to cite: Khaliq, M., Dina, V., Sushama, L., and Elshorbagy, A.: Streamflow Forecasting at Ungauged Locations Using Deep Learning Networks, Driven by Soil Moisture States and Meteorological Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15989, https://doi.org/10.5194/egusphere-egu26-15989, 2026.
Recent advances in machine learning (ML) for hydrology demonstrate strong potential for improving short to sub-seasonal streamflow forecasting under increasing frequent extreme events. These models leverage large collections of meteorological and hydrological time series dataset that often combined with spatial and network-based information to learn transferable forecasting relationships across diverse hydroclimatic regimes. Here we discuss our recent progress and remaining challenges in developing robust ML-based streamflow forecasting systems that operate beyond traditional short lead times. We developed Future Time Series Transformer (FutureTST), a deep learning architecture designed to explicitly integrate past hydrometeorological conditions with future weather information for streamflow forecast. Unlike conventional autoregressive or process-based approaches, FutureTST independently encodes historical streamflow and meteorological forcings while conditioning forecasts on future atmospheric drivers which helps to capture complex temporal dependencies that govern streamflow at extended lead times. Evaluating across multiple basins, we demonstrate three key advances: (1) Forecast skill improvement across lead times: FutureTST achieves strong performance from short to sub-seasonal period with a mean Nash-Sutcliffe Efficiency (NSE) value of 0.82 at 1-day lead time to 0.67 at 30-day lead time which substantially outperform calibrated process-based hydrological models beyond 4 days, (2) Data filling and network-informed forecasting: Reconstructing lost streamflow information highlights the importance of data filling and spatial connection in a river network for improving forecast for partially gauged or data-sparse basins, and (3) Implications for compound flood prediction: By jointly conditioning on antecedent hydrologic states and meteorological extremes, the ML framework provides an interpretable variable importance for identifying compound flood drivers. Finally, we outline key challenges and future directions from ensemble weather forecasts, uncertainty quantification, and compound event aware training strategies to further improve streamflow forecasting.
How to cite: Krishnan Kutty Ambika, A., Tayal, K., Feng, D., Mishra, V., Lu, D., and Hoffman, F. M.: Improving short to sub-seasonal streamflow forecast using machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22393, https://doi.org/10.5194/egusphere-egu26-22393, 2026.
Accurate groundwater level (GWL) forecasting is crucial for effective water resource management, particularly under changing climatic conditions. In this study, we investigate the potential of the Kolmogorov–Arnold Network (KAN), an emerging neural architecture, for time series forecasting of GWL across the Normandy region in France. The performance of the KAN model was compared to classical recurrent neural network (RNN) architectures, including the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU).
Using ERA5 precipitation and temperature as predictors, all models were trained to simulate groundwater level variations at multiple monitoring stations. Results indicate that, although the standalone one-layer KAN model underperformed relative to LSTM and GRU in terms of predictive accuracy, it provided valuable interpretability by effectively capturing input importance and nonlinear dependencies between meteorological drivers and groundwater dynamics. Moreover, integrating a KAN layer within LSTM and GRU architectures improved performance at several stations, suggesting that hybrid KAN–RNN frameworks can combine the interpretability of KAN with the sequential learning capability of recurrent models. Based on our findings, we recommend a two-step approach: employing KAN alone for input relevance analysis, followed by applying hybrid KAN–LSTM architectures to enhance predictive accuracy.
As the KAN-based model architectures continue to evolve with frequent updates and new variants, future research should further explore and benchmark these improved versions for hydrological and, particularly, GWL forecasting tasks. These results highlight the potential of KAN-based hybrid models for interpretable and adaptive groundwater forecasting, opening promising perspectives for data-driven understanding of subsurface processes in data-scarce regions.
How to cite: Janbain, I., Massei, N., Jardani, A., and Fournier, M.: Evaluating Kolmogorov–Arnold Networks (KAN) for Time Series Forecasting: Influence on Interpretability and Accuracy in Groundwater Level Prediction in Normandy, France , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21634, https://doi.org/10.5194/egusphere-egu26-21634, 2026.
Groundwater is an essential part of the hydrological system and is increasingly affected by climate variability and human pressures. Spatial and temporal variation of groundwater depth (GWD), defined as the depth to the saturated zone below ground surface, is a key variable for assessing groundwater–surface interactions and for evaluating risks to infrastructure, land use, droughts and flooding. In-situ measurements of GWD are often too sparse to capture its variability both in time and space and modelling becomes a necessity for consistent assessment.
In this study, we evaluated hybrid machine learning (ML) models that combine the strengths of existing hydrological simulations from Physically Based Models (PBM) with the predictive power of ML methods for improved GWD estimation in Denmark. The hybrid model reduced mean bias of PBM GWD estimates from 1.65 m to 0.21 m and decreased the Root Mean Square Error (RMSE) by about 1.5 m at national scale. Furthermore, we demonstrated that increasing the availability of GWD observations over time and space enhances model performance.
We also illustrated the flexibility and effectiveness of the hybrid approach for GWD estimation across scales, and results showed that the hybrid model developed with coarse spatial resolutions can be effectively used for high-resolution GWD estimation while maintaining a satisfactory level of accuracy. Specifically, the mean RMSE is reduced from 2.66 m for the model trained and applied at 500 m to only 2.30 m for the model trained at 500 m but applied at 10 m. Similarly, for the model trained at 100 m but applied at 10 m for prediction the RMSE is reduced from 2.35 m to 2.28 m.
This study highlights the potential of hybrid modeling as a practical solution for improving groundwater quantification accuracy and shows avenues for higher-resolution estimates.
How to cite: Liu, J., Schneider, R., Troldborg, L., ma, Y., Maxwell, R., and Koch, J.: A National Scale Hybrid Model for Enhanced Groundwater Depth Estimation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6669, https://doi.org/10.5194/egusphere-egu26-6669, 2026.
Posters on site: Wed, 6 May, 08:30–10:15 | Hall A
We evaluate hybrid deep learning architectures and ensemble strategies for monthly precipitation prediction over Boyacá, Colombia (3,965 CHIRPS grid cells, 145–5,490 m elevation, horizons 1–12 months). Three spatial encoding paradigms are compared: convolutional (ConvLSTM), spectral (Fourier Neural Operator hybrids), and graph-based (Graph Neural Network with Temporal Attention, GNN-TAT). GNN-TAT matches ConvLSTM accuracy (R2: 0.628 vs 0.642) with 95% fewer parameters and lower variance, leveraging elevation-weighted edges for interpretable spatial reasoning. Beyond individual models, late fusion via Ridge regression (R2=0.668) improves over all single architectures by exploiting complementary grid-based and graph-based error structures. Conversely, early fusion stacking collapses to R2=0.212, showing that combining predictions preserves inductive biases while merging intermediate representations destroys them. We also report the first evaluation of State Space Models (Mamba) for regional precipitation, which fail to transfer from sequence modeling (R2=0.200). Three operational guidelines emerge: graph-based encoders are efficient alternatives in complex terrain, ensemble gains depend on late-stage combination, and documenting architectural failures narrows the search space for future practitioners. All experiments use standardized CHIRPS/SRTM inputs and fixed random seeds for reproducibility.
Keywords: monthly precipitation prediction, deep learning, hybrid architectures, Graph Neural Networks, ConvLSTM, ensemble learning, mountainous terrain, Colombian Andes, State Space Models
Related Publications:
1. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. Spatiotemporal Prediction of Monthly Precipitation: A Systematic Review of Hybrid Models. Hydrology Research (IWA Publishing), under review — Revision 3.
2. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. Hybrid Deep Learning Architectures for Multi-Horizon Precipitation Forecasting in Mountainous Regions: Systematic Comparison of Component-Combination Models in the Colombian Andes. Hydrology (MDPI), accepted.
3. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. A Data-Driven Deep Learning Framework for Monthly Precipitation Prediction in Complex Mountainous Terrain: Systematic Evaluation of Hybrid Architectures, Ensemble Strategies, and Emerging Paradigms. Hydrology (MDPI), ready for submission.
How to cite: Pérez Reyes, M. R., Suárez Barón, M. J., and García Cabrejo, Ó. J.: Deep Learning for Monthly Precipitation Prediction in Mountainous Terrain: From Individual Architectures to EnsembleStrategies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-39, https://doi.org/10.5194/egusphere-egu26-39, 2026.
In recent decades, climate change has led to a significant increase in the frequency and intensity of extreme weather events, such as heavy rainfall and flash floods, resulting in higher hydrogeological risk and increased vulnerability of both ecosystems and urban infrastructures. These phenomena, characterized by strong spatial and temporal variability, are particularly impactful in urban areas, where impervious surfaces, high population density, and the presence of critical infrastructure amplify consequences of flooding and inundation.
The hydraulic system of Milan represents a critical case study: natural watercourses and artificial canals are closely intertwined with the urban fabric. In particular, floods of the River Seveso recurrently cause inundation in the northern part of the city, producing widespread damage to people, infrastructure, and mobility.
In this context, the ability to accurately forecast meteorological and hydrological variables at very short lead times is crucial for risk management and the development of timely early-warning systems. This study proposes the use of machine learning models, such as LDCast and GPTCast, developed by MeteoSwiss and the Bruno Kessler Foundation in Trento, respectively, for radar-based nowcasting. The estimates produced by these models are subsequently coupled both as input for physically based hydrological models and within artificial intelligence algorithms developed by the Politecnico di Milano.
The objective of the study is to evaluate the overall performance of this forecasting system and to demonstrate how it might represent a significant advancement in the implementation of very short-term early-warning systems.
How to cite: Gambini, E., Mazza, M., Franch, G., Wanjari, R., and Ceppi, A.: Integrating Radar Nowcasting and Machine Learning in an Advanced Early-Warning System for Milan’s Hydraulic Node, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5701, https://doi.org/10.5194/egusphere-egu26-5701, 2026.
Brightness temperature (TB) data acquired from microwave satellite systems constitute a fundamental component of global environmental monitoring and earth system analysis. This data serves as critical variables for understanding our Earth systems and predicting carbon, water and energy fluxes. While these satellite systems offer global-scale observational data comparable to physics-based land surface models, they remain subject to fundamental limitations inherent to Low Earth Orbit satellite missions. In particular, their spatial and temporal sampling is constrained by orbital geometry and revisit cycles, resulting in observational gaps and reduced capability to resolve rapidly evolving hydrometeorological processes. Moreover, the continuity and availability of satellite-derived products are strongly dependent on mission lifetimes and launch schedules, leading to potential discontinuities across different satellite generations.
This study proposes a new deep learning-based framework to emulate TB data from microwave satellite systems. Recently, foundation models based on the Transformer architecture have been successful in specific downstream tasks. Foundation models provide superior zero-shot or out-of-distribution performance due to their broad pre-training. This has led to an increasing number of studies applying foundation models to various hydrological challenges.
Using available TB data from various microwave satellite systems as the target, the proposed model is trained to learn nonlinear relationships between latent vectors and global-scale TB dynamics. Based on these learned relationships, the model subsequently infers a suite of hydrological variables, including soil moisture and vegetation water content, thereby enabling consistent reconstruction of land surface states across space and time.
How to cite: Lee, D., Kim, S., Kim, S., and Kim, H.: Bridging Observational Gaps in Microwave Satellite Signals Using a Meteorological Foundation Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17329, https://doi.org/10.5194/egusphere-egu26-17329, 2026.
In recent years, data-driven models have demonstrated remarkable performance in capturing complex land-atmosphere interactions. In particular, the emergence of weather foundation models, which are pre-trained on vast unlabeled meteorological datasets through self-supervision and can be applied to diverse downstream tasks, has introduced robust backbones capable of representing global atmospheric dynamics. However, fine-tuning these massive models to specific downstream hydrological tasks presents significant challenges. Full fine-tuning is computationally prohibitive, and even parameter-efficient fine-tuning methods, such as Low-Rank Adaptation, also have an amount of computational overhead over the large embedding dimensions of foundation models. Furthermore, modifying the backbone's weights can be a risk of catastrophic forgetting or destabilize the learned representations, which are essential for maintaining their physical consistency during iterative long-term forecasts.
To address these challenges, this study investigates a transfer learning approach that utilizes a weather foundation model backbone with lightweight decoders. This strategy allows the model to handle the robust feature space of the pre-trained backbone while maintaining computational efficiency and architectural stability. We design and evaluate two representative classes of lightweight decoder architectures that differ in their structural complexity and information integration strategy. The first decoder adopts a minimalistic mapping scheme that directly transforms the latent representations of the foundation model into hydrological estimates, allowing us to assess whether the backbone features alone contain sufficient information for soil moisture inference. The second decoder employs a more expressive architecture capable of capturing multi-scale spatial dependencies and structural coherence in the output fields. A key architectural distinction between the two decoders lies in their input configuration: the simpler decoder relies exclusively on backbone representations, whereas the more advanced decoder additionally incorporates prior hydrological state information to reinforce physical consistency and temporal continuity.
Our results indicate that both lightweight decoders successfully reconstruct patterns of hydrological variables (e.g., soil moisture), demonstrating that the weather foundation models' backbone contains sufficient information to infer hydrological variables effectively. This study highlights the immense potential of weather foundation models as a new paradigm for hydrological research, providing a stable and efficient pathway to achieve high-fidelity results without the need for exhaustive fine-tuning.
How to cite: Kim, S., Lee, D., Kim, S., and Kim, H.: Leveraging Weather Foundation Models for Hydrological Applications: Enhancing Hydrological Prediction through Sophisticated Decoder Design, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17199, https://doi.org/10.5194/egusphere-egu26-17199, 2026.
Reliable high-resolution precipitation is crucial for monitoring hydrologic extremes and guiding climate-risk decisions, especially for compound events such as dry-to-wet “whiplash” . However, satellite precipitation products are often too coarse (5–25 km) and show strong region- and intensity-dependent biases, limiting their value for local hazard assessment. We develop a physics-aware geospatial machine-learning downscaling and fusion framework (PIGMLD) to generate 1-km daily precipitation over China for 2000–2020 by combining 10-km GPM IMERG with sparse gauges, ERA5-Land precipitation, and physically interpretable covariates linked to moisture, clouds, and land–atmosphere coupling. Validation against independent gauges across China and nine major basins shows broad skill gains (84.1% of stations with KGE > 0.60), improved event detection, and reduced bias; improvements are smaller in terrain-complex, gauge-scarce regions but remain useful. Performance gains are strongest for extremes, with large RMSE reductions for heavy and torrential rainfall and substantial bias corrections for both dry and wet percentile-defined extremes. Overall, PIGMLD provides more reliable 1-km precipitation to better characterize hydroclimate extremes and support water hazard related risk assessment.
How to cite: He, K., Zhao, D., Zhao, W., Brocca, L., and Chen, X.: PIGMLD: A Physics-Aware Geospatial Machine Learning for High-Resolution Extreme Precipitation Reconstruction , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20412, https://doi.org/10.5194/egusphere-egu26-20412, 2026.
In regions where the hydrological cycle is strongly influenced by seasonal snow dynamics, accurate estimation and prediction of Snow Water Equivalent (SWE) are essential for water resource management, hydropower planning, and flood forecasting. While traditional methods like in-situ observations, numerical models, and remote sensing provide robust and reliable approaches for monitoring SWE, challenges remain with respect to ungauged regions and precdictions. These include the difficulty of installing measurement stations in certain regions and resulting observation scarcity, high computational costs, complex parametrization, and low spatial resolution or limited temporal data availability.
Data-driven methods enable the creation and transfer of surrogate models capable of learning complex spatiotemporal relationships between meteorological forcings and SWE dynamics directly from high-fidelity simulations. This study develops a surrogate model using a Convolutional Long Short-Term Memory (ConvLSTM) architecture to provide high-resolution daily SWE estimates and forecasts for Norway. Specifically, the ConvLSTM is trained to emulate the operational SeNorge snow model, creating a portable and computationally efficient tool that can generate accurate SWE fields from diverse meteorological inputs.
The proposed ConvLSTM framework integrates spatial and temporal dependencies by processing sequences of gridded meteorological forcings (precipitation and temperature), static topographic features, and cyclical temporal indicators. To enable robust multi-day forecasting, the model employs an autoregressive training scheme with scheduled sampling. This approach gradually shifts the model from using true SWE values to its own previous predictions as inputs during training, effectively reducing error accumulation within a 7-day prediction horizon.
To evaluate the potential for areal transfer of the surrogate model for pan-European applications, we additionally forced the trained architecture with bias-corrected meteorological data from the ERA5-Land reanalysis. The results demonstrate that the ConvLSTM surrogate model accurately captures the spatiotemporal evolution of SWE across Norway's complex terrain, which suggests that the model indeed learned general physical relationships between input feature and target. Therefore, when driven by SeNorge data, the model achieves good fidelity with a median KGE of 0.8, effectively replicating seasonal accumulation, peak SWE magnitudes, and melt dynamics. Notably, when forced with the global ERA5 reanalysis dataset, the model maintains robust performance (KGE ~ 0.60), indicating its ability to generate reliable SWE estimates and potential transferability to other regions worldwide. .
This work is funded by the European Union’s HORIZON Research and Innovation Actions Program under Grant Agreement No. 101059372 (STARS4Water project) and the BMBF BioökonomieREVIER funding scheme with its BioRevierPlus project (funding reference 031B1137D/031B1137DX).
How to cite: Avila, L., Engeland, K., Jahr, T., and Kollet, S.: A ConvLSTM surrogate model to predict high-resolution daily snow water equivalent in Norway, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8502, https://doi.org/10.5194/egusphere-egu26-8502, 2026.
Recently, advances in deep learning (DL) have enabled the development of various surrogate modeling approaches to emulate traditional land surface models (LSMs), which typically require expensive computational resources. These surrogate models provide computationally efficient alternatives to conventional LSMs. In this study, we develop a DL-based autoregressive surrogate model to predict surface soil moisture (SSM) using meteorological forcing variables and the previous SSM state as inputs.
The developed surrogate model is further employed as a forecast model within a land data assimilation (LDA) framework, replacing the traditional LSMs. Since the true SSM state is unknown in the real-world applications, the fraternal twin experiments are conducted using a synthetic ground truth SSM, which is generated from an LSM nature run. In addition, a synthetic imperfect LSM SSM is generated by applying spatially correlated noise to the synthetic ground truth. Then, the surrogate model is trained to emulate this imperfect LSM simulation.
Synthetic satellite observations are generated from the synthetic ground truth by introducing controlled observational uncertainties derived from prior studies. These experiments systematically evaluate the sensitivity of LDA performance to satellite observation errors under a wide range of realistic observational scenarios. Therefore, the proposed framework is expected to serve as a computationally efficient and scientifically rigorous testbed for exploring LDA strategies, with potential applications in future satellite mission design and water resource management.
How to cite: Kim, S. and Kim, H.: Fraternal Twin Experiments for Satellite-Constrained Land Data Assimilation Using Deep Learning Surrogate Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17278, https://doi.org/10.5194/egusphere-egu26-17278, 2026.
Reliable projections of hydrological drought are essential for climate-resilient water management; however, many basins lack calibrated, process-based models. Here, we develop and test a purely data‑driven framework to forecast the Streamflow Drought Index (SDI) for the Sava River basin, using only widely available meteorological drought indices, and apply it to project future drought conditions under different climate scenarios. We assembled a monthly dataset for 1961–2020 comprising the Standardized Precipitation Index (SPI), a standardised temperature index (STI), the Standardised Precipitation–Evapotranspiration Index (SPEI), and SDI derived from observed streamflow. All indices are approximately standardised and show frequent negative excursions, indicating recurrent meteorological and hydrological droughts. Correlation analysis revealed that short-term precipitation anomalies significantly influence the linear control of SDI. Specifically, SPI at a one-month lag exhibits the strongest association (r ≈ 0.50), followed by contemporaneous SPI (r ≈ 0.44) and a two-month lag (r ≈ 0.24). By contrast, STI and SPEI lags exhibit negligible correlations, indicating that temperature-driven evaporative demand plays a secondary role in the initial onset of drought in this temperate, precipitation-dominated basin. We evaluated several machine-learning models for one-month-ahead SDI prediction, including Random Forest (RF), XGBoost, Elastic Net, Support Vector Regression (SVR), and a Multilayer Perceptron. Models were trained on the first 80% of the record and evaluated on the remaining 20% using a strictly chronological split. For RF, key hyperparameters (number of trees, maximum depth, leaf size and feature subsampling) were tuned using Randomized Search with Time Series Split cross‑validation. A linear‑scaling bias correction was applied to align the mean and variance of predicted SDI with observations in the training period. Random Forest clearly outperformed the alternative models. In the independent test period (2009–2020), the bias‑corrected RF achieved MAE ≈ 0.62, RMSE ≈ 0.83, and NSE ≈ 0.49, explaining almost half the variance in SDI. KGE ≈ 0.65 indicates good joint reproduction of correlation, variability and bias. The model accurately captured the timing and sign of most wet and dry episodes, while moderately underestimating the most extreme peaks. Other algorithms exhibited similar or larger errors and substantially lower KGE, confirming RF as the most suitable SDI forecasting approach in this index‑only setting. Finally, we drove the optimised RF with SPI/STI/SPEI projections from RCP2.6, RCP4.5 and RCP8.5 to generate monthly SDI projections for 2021–2050. Hydrostripes and distributions show clear scenario‑dependent changes: RCP2.6 maintains mainly mild, short‑lived droughts; RCP4.5 produces more persistent and clustered deficits; and RCP8.5 yields the most frequent and severe hydrological droughts. The framework demonstrates that a carefully tuned Random Forest, using only standardised meteorological indices, can provide skilful and interpretable SDI projections to support climate‑informed drought risk management.
Acknowledgment: This research was supported by the “Streamflow Drought Through Time” project funded by the EU NextGenerationEU through the Recovery and Resilience Plan of the Slovak Republic within the framework of project no. 09I03-03-V04-00186.
How to cite: Leščešen, I., Josić, M., Gnjato, S., Petrović, A. M., Bajtek, Z., and Pekárová, P.: Random Forest–Based Projection of Streamflow Drought Index from Meteorological Drought Indices under RCP Climate Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5096, https://doi.org/10.5194/egusphere-egu26-5096, 2026.
Accurate forecasting of reservoir inflow is crucial for effective water management, especially in regions with limited water resources and high demand from various sectors, including irrigation, domestic, and industrial uses. For the effective planning and management of reservoir operations, flood control, hydroelectric power generation, and drought mitigation, predicting reservoir inflow plays a crucial role. With the rapid increase in population and industrialization, uncertainty in reservoir storage has increased, leading to a risk of water stress and compromised water security. Therefore, predicting reservoir inflow is crucial for reservoir operation and efficient water management. The inflow prediction is challenging due to the complex and dynamic nature of the rainfall-runoff process in a river basin. Hydrological models provide a simplified representation of real hydrological systems; despite this, due to the complexities and uncertainties in hydrological processes, it is challenging to achieve accurate predictions. In recent years, machine learning (ML) techniques have been widely used for simulating the streamflow due to their accuracy in capturing complex and non-stationary relationships between rainfall and streamflow. However, these ML models do not account for the physical characteristics of the watershed. Therefore, to increase the accuracy of prediction by gaining a better understanding of the hydrological patterns, physics-based, hybrid machine learning models have been developed in this study and applied in a river basin of Maharashtra, India. A physics-based HEC-HMC model was combined with ML models, such as long short-term memory (LSTM) and extreme gradient boosting (XGBoost), to develop a hybrid ML model using 2001 to 2021 hydro-meteorological data. The hybrid ML model was found to be capable of predicting the inflow (QIF) into the reservoir. The daily values of hydro-meteorological variables, viz., rainfall, temperature, relative humidity, wind speed, and reservoir inflow, were used to simulate the HEC-HMC model. The HEC-HMS simulated reservoir inflow (Qh), along with its lagged values (Qh-1, Qh-2), reservoir storage, rainfall, evaporation loss, and other factors, were used as inputs to the machine learning models. The preliminary results indicated that Qh, Qh-1 and lag-1 rainfall variables are essential inputs to machine learning models for the accurate prediction of the reservoir inflow.
How to cite: Dhivar, Y. N. and Jha, M. K.: Reservoir Inflow Prediction using Machine Learning Techniques, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11106, https://doi.org/10.5194/egusphere-egu26-11106, 2026.
The amount of dissolved organic matter (DOM) in freshwaters impacts many processes in aquatic ecology and, therefore, on derived ecosystem services such as water supply. Surface drinking water sources (e.g., lakes and reservoirs), in particular, are increasingly subjected to unforeseen increases in both the concentration and variability of DOM. This makes it more difficult to deal with such water along the drinking water cycle (abstraction, treatment, storage and network distribution), which can affect tap quality and users through the unintentional formation of toxic disinfection by-products (DBPs). DBPs such as trihalomethanes (THMs) and haloacetic acids (HAAs) are known to affect human health under long-term exposure, increasing risks for different cancers and congenital malformations. Anticipating seasonal changes in DOM in source waters is therefore important for both improved drinking water source protection measures and a reduction of DBPs in supplies.
Ecological forecasting provides a way to support water quality management by generating predictions of future environmental conditions within decision-relevant timeframes. In lakes, DOM dynamics often vary strongly at intra-annual and seasonal scales, suggesting that seasonal forecasts could help managers anticipate periods of increased treatment risk and plan mitigation measures in advance. However, forecasting DOM remains challenging due to the complex interactions between in-lake processes and catchment-scale drivers. Recent applications of machine learning have shown skill in simulating historical DOM dynamics in lakes, offering opportunities to extend these approaches to forecasting.
In this study, we developed a seasonal forecasting framework to predict monthly average concentrations of surface fluorescent DOM (fDOM) one to seven months ahead. The framework consists of a machine-learning workflow that simulates daily fDOM using random forest regression, and is applied to two contrasting study sites: a lake in Ireland and a reservoir in Spain. Forecasting is driven by a set of predictors selected based on their relative importance in historical simulations and their availability in open-access seasonal forecast datasets.
The workflow integrates meteorological variables, soil conditions, hydrological outputs, lake model variables and a seasonal indicator. Forecast skill and uncertainty were evaluated over multiple periods (1993–2023, 1993–2016 and 2016–2023) to reflect changes in forecast input characteristics, and results were compared against a climatological baseline. The analysis highlights how seasonal forecasts of DOM can support drinking water management by providing information on expected conditions in source waters in advance. The framework is designed to be transferable to and tested in other lake and reservoir systems where similar data are available.
How to cite: Paíz, R., Mercado-Bettín, D., Marcé, R., Jennings, E., and McCarthy, V.: Machine learning-based seasonal forecasting of dissolved organic matter to support drinking water management at the source , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15667, https://doi.org/10.5194/egusphere-egu26-15667, 2026.
Large-sample hydrologic studies often require calibrating multiple model structures across numerous catchments, which can be computationally intensive with traditional optimization algorithms. Alternatively, recent advances in Machine Learning (ML) have enabled computationally frugal calibration strategies that rely on model emulators. Such approaches leverage information across sites, enabling improved calibration efficiency and parameter transferability to unseen catchments. However, exploring the parameter space using emulators is challenging because of emulator error and the need to explore high-dimensional parameter spaces. In this work, we investigate ML-based emulation and optimization strategies designed to improve parameter-space exploration, with the broader goal of supporting reproducible and computationally feasible large-sample hydrologic simulation. To this end, we use the Framework for Understanding Structural Errors (FUSE), which systematically represents alternative process formulations through multiple model configurations. Our framework is calibrated for 1,070 catchments across North America, spanning a wide range of hydroclimatic conditions. We develop Random Forest (RF) and Quantile Random Forest (QRF) emulators to approximate the relationship between model parameters, catchment attributes, and the Kling–Gupta Efficiency (KGE). While RF provides point estimates, QRF captures predictive uncertainty through conditional quantiles. These emulators are integrated into two calibration strategies: (1) a standard Genetic Algorithm (GA) that efficiently searches for high-performing parameter sets, and (2) a two-step hybrid optimizer that first performs a broad global search using Markov chain Monte Carlo sampling and then refines promising solutions using local gradient-based optimization. By more fully evaluating the parameter space and avoiding premature convergence, the two-step strategy captures a more diverse ensemble of near-optimal parameter solutions. This diversity is particularly valuable for emulator-based calibration, as it allows the emulator to be retrained iteratively on a broader range of the parameter space, improving robustness and reducing reliance on narrowly sampled regions. These improvements are expected to support more stable parameter estimates and improved hydrologic simulations across a large sample of catchments. Overall, this hybrid framework enables reproducible and computationally efficient calibration across multiple model structures and hundreds of catchments, providing a scalable pathway for integrating ML emulators into large-sample hydrologic modeling workflows.
How to cite: Hatami, S., Vásquez, N., Thébault, C., Knoben, W., Eythorsson, D., Papalexiou, S. M., and Clark, M.: Machine Learning Emulator for Large-Sample Hydrologic Model Calibration across Multiple FUSE Structures, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13767, https://doi.org/10.5194/egusphere-egu26-13767, 2026.
Hydrologic model calibration can be challenging even with sufficient observations to constrain local model parameters, and far more difficult when estimating parameters across large domains – a process known as parameter regionalization. In recent years, the use of machine learning in differentiable hydrologic modeling has shown potential to address this regionalization problem. Here, a neural network learns to predict model parameters from meteorological forcings and geophysical catchment attributes by updating its weights using gradient-based optimization to minimize a loss function that quantifies the discrepancy between the conceptual model’s simulations and the observations. Such a model trained over a large set of basins at once will learn regional hydrological behaviors and can be used for parameter regionalization. We investigate whether this approach can be used to determine static parameters for NOAA’s Next Generation Water Resources Modeling Framework (NextGen), specifically for the National Water Model Conceptual Functional Equivalent (CFE) model by embedding a differentiable version (dCFE) into the NeuralHydrology (NH) platform for training and extracting the trained neural network to use in CFE parameter regionalization across CONUS. We introduce two ways of extracting static parameters from the neural network, and compare these to dynamic parameters obtained using the same workflow. This presentation describes this effort, including the validation of NH-dCFE to dCFE and CFE, successes in three modes of training, and the challenges encountered. We also offer recommendations on strategies to advance this parameter estimation approach in the future.
How to cite: Li, Z., Wood, A., McKenzie, D., and Frame, J. M.: Training a differentiable conceptual functional equivalent of the US national water model to estimate parameters for use in NextGen, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14714, https://doi.org/10.5194/egusphere-egu26-14714, 2026.
Hydrologic deep learning models have made their way from research to applications. More and more national hydrometeorological agencies, hydro power operators, and consulting companies are building Long Short-Term Memory (LSTM) models for operational use cases. However, all of these efforts are confronted with similar sets of challenges—issues that are different from those in controlled scientific studies. One common issue is the question: how to deal with missing input data? Operational systems depend on the real-time availability of various data products—most notably, meteorological forcings. Additional forcings generally improve the model performance, but at the same time, every new dependency increases the likelihood of an outage in one of the input data products.
In a recent study, we evaluated different solutions to generate predictions even when some of the meteorological input data do not arrive in time, or not arrive at all (Gauch et al., 2025). In this presentation, we will introduce these methods and discuss how they can help (1) operational forecasters to run reliable real-time flood forecasting systems, and (2) researchers and modelers to build accurate models that leverage as much data as possible.
Gauch, M., Kratzert, F., Klotz, D., Nearing, G., Cohen, D., and Gilon, O.: How to deal w___ missing input data, Hydrol. Earth Syst. Sci., 29, 6221–6235, 2025.
How to cite: Gauch, M., Kratzert, F., Klotz, D., Nearing, G., Cohen, D., and Gilon, O.: How to deal w___ missing input data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16581, https://doi.org/10.5194/egusphere-egu26-16581, 2026.
Recent advances in integrated hydrologic modeling have advanced our ability to simulate coupled surface–subsurface processes, offering critical insights into water dynamics at large scales. Specifically, simulations of the Contiguous United States (CONUS) have been recently conducted with ParFlow-CLM, a highly parallelizable integrated hydrologic modeling platform. These large-scale simulations have supported investigations of stream–aquifer connectivity and hydrologic sensitivity to climate forcing, yet their implementation remains technically demanding.
Assembling such simulations involves numerous challenges, including the configuration of complex HPC environments, management of evolving and voluminous climate forcings, large-scale input and output data handling, and heavy postprocessing workflows. These barriers limit the broader adoption and operational use of integrated models.
Here we present a semi-automated workflow for running ParFlow-CLM simulations over the CONUS domain in quasi–real-time. The workflow, designed to operate in weekly cycles, integrates Python and Shell scripting with the hf_hydrodata Python package to automate data preparation, model execution, and output management. We demonstrate its application on three widely used HPC platforms, highlighting its scalability and adaptability.
This contribution directly supports the community’s effort to make integrated modeling more accessible, reproducible, and operationally feasible at continental scales. By reducing technical overhead, the workflow promotes broader participation in high-resolution hydrologic modeling and facilitates timely water resources decision-making.
How to cite: Sandoval, L., Defnet, A., Artavanis, G., Maxwell, R., and Condon, L.: Streamlining Integrated Hydrologic Modeling at Continental Scale: A Workflow for Quasi–Real-Time ParFlow-CLM Simulations over CONUS, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21593, https://doi.org/10.5194/egusphere-egu26-21593, 2026.
Reliable multi-step streamflow prediction is essential for effective water resources management. In recent years, deep learning (DL) approaches have been increasingly adopted for streamflow forecasting as alternatives to process-based hydrological models. These approaches have partially reduced the reliance on high-quality and comprehensive hydrological observations required for robust parameterization of the process-based models. Nonetheless, the predictive performance of DL-based hydrological models often deteriorates as the forecast horizon extends, posing critical challenges to their reliability and practical applicability. Moreover, due to the scarcity of storage-related observations, most existing DL-based hydrological models are primarily driven by flux variables (e.g., precipitation and streamflow), while watershed memory effects related to storage regulation remain largely underrepresented. To address these limitations, this study proposes a multi-step streamflow forecasting model that incorporates a proxy representation of watershed memory through a DL approach, namely the Embedding Multi-Layer Perceptron (E-MLP). The proposed model was developed using only precipitation and streamflow time-series, without relying on explicit storage-related variables. Two widely used DL models, i.e., Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) models, were employed as benchmark approaches. Each model was evaluated in a flood-prone watershed, the Upper Wapsipinicon River watershed near Anamosa gauging station (USGS-05421740) in Iowa, United States. Comparative analyses across the three models demonstrated that incorporating a proxy representation of watershed memory yielded more stable predictive skill at longer forecast horizons, effectively mitigating performance degradation with increasing lead time. These findings highlight the critical role of watershed memory in DL-based streamflow forecasting and point to a viable pathway toward more robust multi-step forecasting frameworks.
Acknowledgements
This work was supported by the Creative-Pioneering Researchers Program through Seoul National University and by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2025-00523350). Additional support was provided by the National Institute of International Education of Korea (NIIED-230724-0041) and the China Scholarship Council (CSC No. 202208230007).
How to cite: Li, Y. and Yang, S.: Advancing Multi-step Streamflow Forecasting with an Embedding Multi-Layer Perceptron, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2903, https://doi.org/10.5194/egusphere-egu26-2903, 2026.
In hydrology, neural networks (NN) are often used to replace hydrological models. While they have been proven to perform well for forecasting and simulating streamflow, they may not always be adequate when transparency and process understanding are required. However, NN can also be used as complements to process-based hydrological models (e.g., physics-based or conceptual) as part of a forecasting chain. For instance, in Boucher et al. (2020), an ensemble of multilayer perceptrons was used to perform data assimilation of streamflow in the GR4J conceptual model, which yielded promising results.
Building on the approach introduced in Boucher et al. (2020), the research presented here aims to improve the NN-based data assimilation method and to alleviate its limitations. To achieve this, multilayer perceptrons are replaced by long short-term memory (LSTM) networks, with an additional attention component. Both streamflow and snow are assimilated, the main focus being on the latter. For this reason, this new methodology has been tested on watersheds located in Canada, Norway and Sweden, including the Mistassibi watershed, which was also used in Boucher et al. (2020). Each watershed has been modelled using two hydrological model structures (GR4J and HMETS) within the Raven modelling framework.
Results show a successful assimilation of both streamflow and snow, which translates into improved daily streamflow simulations compared to the open-loop (according to the CRPS and reliability diagrams), for all catchments and for both models. In particular, results for Mistassibi show an improvement of the post-assimilation simulations compared to Boucher et al. (2020). This presentation will explain those results in detail and also describe the next steps to further expand and generalize the proposed data assimilation method.
How to cite: Lapointe, F., Zhang, L., Hamelin, J., Radenkov, S., Kaddache, Y., Boucher, M.-A., Quilty, J., and Craig, J. R.: Deep learning for the assimilation of process-based hydrological models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5780, https://doi.org/10.5194/egusphere-egu26-5780, 2026.
Advancing hydrological modeling requires simultaneous improvements in predictive skill and process understanding. While conceptual rainfall–runoff models remain widely used for their physical interpretability and reasonable robustness, their empirical flux formulations limits flexibility and generalizability across contrasting hydro-climatic conditions. Recent hybrid modeling studies [1,2] have shown that integrating neural networks or universal differential equations into conceptual models for flux correction can improve performance while preserving physical constraints. Following this perspective, we introduce a new modeling paradigm termed « neural reservoir », in which traditional empirical reservoir flux laws are replaced by physics–neural operators [3]. These operators are constructed by combining neural operators with shape functions derived from functional analysis of the original flux equations, ensuring mass balance and physically admissible bounds while remaining fully flexible and trainable. This framework enables the learning of internal water fluxes governing reservoir dynamics directly from data, while retaining the interpretability and structural consistency of reservoir-based models. Preliminary results show that the neural reservoir consistently outperforms both classical conceptual and purely data-driven LSTM benchmark models, while exhibiting physically meaningful behaviors and enhanced responsiveness to hydro-climatic variability. Ongoing work focuses on extending the evaluation to large-sample and national-scale settings, as well as on integrating additional data sources to further refine process representation.
[1] Huynh, N. N. T., Garambois, P.-A., Renard, B., et al. (2025). A distributed hybrid physics–AI framework for learning corrections of internal hydrological fluxes and enhancing high-resolution regionalized flood modeling. Hydrol. Earth Syst. Sci., https://doi.org/10.5194/hess-29-3589-2025.
[2] Huynh, N. N. T., Garambois, P.-A., Colleoni, F., et al. ( 2026). A hybrid physics–AI approach using universal differential equations with state-dependent neural networks for learnable, regionalizable, spatially distributed hydrological modeling. Geosci. Model Dev., https://doi.org/10.5194/gmd-19-1055-2026.
[3] Huynh, N. N. T., Garambois, P.-A., Ettalbi, M., et al. (2026). Physics-Constrained Neural Reservoirs: A Powerful Neural Replacement of Conceptual Hydrological Laws for Learning Spatially Distributed Flow Dynamics. ESS Open Archive, https://doi.org/10.22541/essoar.177100580.07190684/v1.
How to cite: Huynh, N. N. T., Garambois, P.-A., Ettalbi, M., Colleoni, F., Nguyen, N. B., and Renard, B.: On the potential of neural reservoirs for learning flow dynamics from data to enhance rainfall–runoff modeling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8196, https://doi.org/10.5194/egusphere-egu26-8196, 2026.
Despite recent advances in operational streamflow forecasting systems, anticipating and forecasting droughts and associated low-flow conditions remain major challenges in hydrology, with substantial impacts on water-dependent sectors such as agriculture and industry. Enhancing sub-seasonal to seasonal streamflow forecasting is therefore critical for improving water resources management. This study investigates the performance of a Handoff forecast Long Short-Term Memory (LSTM) architecture (Nearing et al., 2024) for probabilistic streamflow forecasting at lead times extending up to six months, with a particular emphasis on low-flow conditions.
The Handoff forecast LSTM is trained regionally on a subset of 292 basins from the Catchment Attributes and MEteorology for Large-sample Studies - FRance dataset (CAMELS-FR) (Delaigue et al., 2025), after excluding basins affected by unreliable low-flow measurements. Model training relies on basin-averaged hydro-meteorological reanalysis data provided by CAMELS-FR. Evaluation of the model is conducted using ensemble streamflow forecasts generated from historical scenarios and meteorological ensemble predictions from the SEAS5 model from the European Center for Medium-Range Weather Forecasts (ECMWF) (Johnson et al., 2019)
The generated ensemble streamflow forecasts are evaluated using a set of probabilistic metrics such as the Continuous Ranked Probability Score (CRPS), the Brier Score, the Area under the ROC curve, and the Talagrand diagram, and using the natural streamflow climatology as a reference. In addition, a sensitivity analysis of static catchment attributes is performed to assess their relative contribution to model performance and to better understand the drivers of predictability across basins.
Delaigue, O., Guimarães, G. M., Brigode, P., Génot, B., Perrin, C., Soubeyroux, J.-M., Janet, B., Addor, N., & Andréassian, V. (2025). CAMELS-FR dataset: a large-sample hydroclimatic dataset for France to explore hydrological diversity and support model benchmarking. Earth System Science Data, 17(4), 1461–1479. https://doi.org/10.5194/essd-17-1461-2025
Johnson, S. J., Stockdale, T. N., Ferranti, L., Balmaseda, M. A., Molteni, F., Magnusson, L., Tietsche, S., Decremer, D., Weisheimer, A., Balsamo, G., Keeley, S. P. E., Mogensen, K., Zuo, H., & Monge-Sanz, B. M. (2019). SEAS5: The new ECMWF seasonal forecast system. Geoscientific Model Development, 12(3), 1087–1117. https://doi.org/10.5194/gmd-12-1087-2019
Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T. Y., Weitzner, D., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559–563. https://doi.org/10.1038/s41586-024-07145-1
How to cite: Jack, Z., Surmont, F., Saint Fleur, B. E., Cooper, O., and Gaume, E.: Sub-seasonal to seasonal ensemble streamflow forecasting using a Handoff forecast LSTM, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10806, https://doi.org/10.5194/egusphere-egu26-10806, 2026.
Distributed hydrological models are the mainstream paradigm for watershed hydrological simulation, integrating heterogeneous factors through spatial discretization and demonstrating significant advantages in revealing the spatial differentiation patterns of hydrological processes. However, the high-dimensional parameter space leads to high computational costs and low robustness in parameter calibration in traditional distributed hydrological models. Taking Weihe River Basin and Water Allocation and Cycle Model (WACM), a traditional distributed hydrological model, as the study area and base model, this study proposed a surrogate Deep Neural Operator (DeepONet) model to enhance the calibration efficiency. The surrogate DeepONet uses a branch network to project the high-dimensional model parameters into a compact latent space and a trunk network to encode the spatiotemporal coordinates of runoff outputs, jointly learning a nonlinear mapping from parameters to runoff that replaces direct calibration in the original parameter space and thus greatly reduces both the effective parameter dimensionality and the computational cost of calibration. The results show that the median Kling–Gupta Efficiency (KGE) coefficient across all gauging stations exceeds 0.85, whereas the parameter calibration time is reduced to less than 10% of that required by traditional genetic algorithms. In addition, the surrogate model achieved high accuracy in runoff prediction, with KGE values above 0.80 at these ungauged stations. This study demonstrates that the deep integration of physical mechanisms and data-driven approaches can effectively enhance the trade-off in the efficiency-accuracy dilemma in hydrological simulations and presents a sound solution for high-dimensional parameter calibration in distributed hydrological models.
How to cite: Wang, T.: DeepONet Surrogate for Accelerating Distributed Hydrological Model Simulations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10995, https://doi.org/10.5194/egusphere-egu26-10995, 2026.
Accurate streamflow prediction is fundamental for water resource management and disaster response. However, predicting streamflow with station-based meteorological observations faces challenges due to low spatial density. In contrast, gridded meteorological data provide spatially continuous information, leading to improved streamflow prediction accuracy. Deep learning (DL) models have been widely adopted in water management and mostly use precipitation as an input. Therefore, this study tests whether gridded precipitation improves the predictive accuracy of DL models for streamflow in the Miho River Watershed, South Korea. Modified Korean Parameter-elevation Regression on Independent Slopes Model (MK-PRISM) is used as gridded precipitation. The MK-PRISM data with 1 km spatial resolution consider elevation, topographic facet, and coastal proximity. This study utilizes six meteorological variables: precipitation, average temperature, maximum temperature, minimum temperature, wind speed, and relative humidity. Three DL models, Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Convolutional Neural Networks-LSTM (CNN-LSTM), are used in this study. Five experimental cases are developed for this study. Cases 1 through 4 utilize LSTM and Bi-LSTM, while Case 5 implements a CNN-LSTM. Case 1 uses station-averaged data across the watershed. Case 2 employs the average of MK-PRISM at the watershed level. Case 3 uses meteorological data from individual stations. Case 4 utilizes the average of MK-PRISM at the sub-basin level. Finally, Case 5 employs a CNN-LSTM to use the original format of MK-PRISM as input data. The results of this study will demonstrate the advantages of gridded precipitation to predict streamflow with DL models and propose a suitable format of gridded precipitation.
How to cite: Kim, M., Lee, Y., Lee, Y., and Lee, S.: Comparative analysis of gridded and station-based meteorological data for deep learning-based streamflow prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15490, https://doi.org/10.5194/egusphere-egu26-15490, 2026.
Flood forecasts are a crucial component of flood hazard mitigation strategies and forecasted flood inundation maps are essential for transitioning from forecast information to decision-making to reduce flood risk. A national medium-range streamflow forecasting system has been developed with ILDAS as its physical modeling core and integrated with an AI-based postprocessor. The forecasting system consists of integrated Noah-MP and mizuRoute that produce daily streamflow forecasts with 1-5 days lead time at more than half a million streams across the country. While the system is computationally very efficient, extending it to generate inundation forecasts remains a drawback, as vector-based routing models do not produce inundation maps. To bridge this gap, we integrate TRITON, a GPU-accelerated 2D hydrodynamic model with improved terrain representation, into the system to simulate flood inundation. A GPU-based hydrodynamic model has computational superiority over a CPU-based model and hence reduces forecast generation time, which is a major bottleneck in inundation forecasting systems. In this framework, streamflow forecasts from the existing system are taken as input to the GPU-based model, which generates inundation forecasts with lead times of 1-5 days across India, along with streamflow and water level forecasts. Initial results show reasonable forecast skills when compared with observed water level data and SAR-based flood maps. The system demonstrates its potential to support operational flood preparedness and disaster risk management, and is also an important step towards building an impact-based flood forecasting system for India.
How to cite: Deka, P., Prakash, V., Kondapalli, N., and Saharia, M.: Developing a Flood Inundation Forecasting System for India. , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1094, https://doi.org/10.5194/egusphere-egu26-1094, 2026.
Urban flooding poses persistent challenges in rapidly urbanizing regions, where the lack of ground-based observations limits both flood characterization and predictive modeling. In many flood-prone cities, physics-based hydrologic and hydrodynamic models are constrained by the availability of high-resolution drainage data, boundary conditions, and event-specific calibration, limiting their applicability for rapid or scalable urban flood assessment. To address this gap, we present a scalable, physics-informed machine learning framework for predicting event-scale urban flood intensity at 10 m resolution and at the sub-seasonal time scales without reliance on ground flood calibration.
The proposed approach integrates multi-sensor information, including SAR-derived flood intensity, with high-resolution hydro-meteorological and geospatial predictors that encode rainfall forcing, terrain controls, urban morphology, and surface imperviousness. A key component of the framework is the use of physically interpretable static predictors, such as height above nearest drainage (HAND), to represent drainage proximity and inundation potential, thereby introducing hydrologically meaningful constraints into the learning process. Flooding is modeled as a continuous spatial variable rather than a binary state, enabling a more realistic representation of flood severity and spatial heterogeneity across urban landscapes.
The framework is applied to Mumbai, India, as a primary testbed and evaluated across multiple rainfall-driven flood events. Model performance is assessed through cross-event consistency and spatial generalization, internal agreement with independently derived flood intensity patterns, and coherence with known flood-prone zones shaped by drainage networks and urban form. Results demonstrate stable and physically plausible flood intensity predictions across events without local recalibration, highlighting the framework’s capacity to generalize in the absence of in situ flood measurements.
By combining observation-driven learning with physically informed predictors, this work advances a transferable pathway for high-resolution urban flood intensity prediction in data-scarce environments. The proposed framework is intended to support scalable flood risk assessment and early-stage decision-making in rapidly urbanizing regions facing increasing flood hazards.
How to cite: Haider, A., Mondal, A., Khanbilvardi, R., and Devineni, N.: A physics-informed machine learning framework for event-scale urban flood intensity prediction at the sub-seasonal time scales in data-scarce cities, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4502, https://doi.org/10.5194/egusphere-egu26-4502, 2026.
Urban flooding is intensifying under climate change and rapid urbanization, particularly in densely built metropolitan basins with limited drainage capacity. Taipei City, located within a low-lying alluvial basin and frequently affected by typhoons and short-duration extreme rainfall, experiences recurrent flash flood hazards. Conventional urban flood forecasting systems primarily rely on static rain gauge or satellite precipitation products, which suffer from coarse temporal resolution, sparse spatial coverage, and high forecast latency, constraining effective real-time early warning.
This study develops an IoT-enabled real-time urban flood forecasting framework for Taipei City by assimilating high-frequency rainfall observations into an operational hydrologic-hydraulic modeling chain. The Keelung River basin is selected as a representative urban catchment. Rainfall observations at a 10-minute temporal resolution are retrieved from Taiwan’s Civil IoT SensorThings API and dynamically injected into a HEC-HMS rainfall-runoff model within the HEC-RTS forecasting environment. The hydrologic model employs the SCS Curve Number method for loss estimation, SCS Unit Hydrograph for runoff transformation, linear reservoir baseflow representation, and Muskingum channel routing. Model calibration and validation are conducted using observed discharge data from historical typhoon events.
Model performance is evaluated using Kling-Gupta Efficiency(KGE), Nash-Sutcliffe Efficiency(NSE), RMSE, and percent bias. The system targets a KGE ≥ 0.75 while achieving a minimum 15-minute reduction in warning latency compared to traditional hourly gauge-driven simulations. The simulated discharge hydrographs are designed for coupling with a 2D HEC-RAS hydraulic model to generate urban flood inundation maps, with spatial performance assessed using an IoU threshold of ≥ 0.65.This study demonstrates that assimilating high-frequency IoT rainfall observations into an operational urban flood forecasting framework can significantly reduce warning latency without degrading hydrologic or hydraulic predictive skill.
How to cite: Safar, S., Wen, T.-H., and Hua, Y.-M.: Real-time Flood Forecasting Model for Taipei City Using IoT Sensor Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8720, https://doi.org/10.5194/egusphere-egu26-8720, 2026.
High resolution gravity field data sets provide valuable information about the wetness state of a catchment, which is a useful indicator in flood early warning systems.
The HiGrav Project aims to increase the temporal and spatial resolution of GRACE/GRACE-Follow-On (FO) data and derived terrestrial water storage anomalies (TWSAs) and wetness index data. Further, the project will assess the utility of this enhanced information basis for flood forecasting in terms of forecast accuracy and flood early warning reliability.
We will explore if downscaled GRACE/GRACE-FO data sets are useful to improve flood forecasting and warning at a regional scale (areas of around 10.000 km²). For this purpose, the semi-distributed, process-based hydrological model PANTA-RHEI will be used. The model is used operationally by the Flood Forecasting and Warning Centre in Lower Saxony (NLWKN - HWVZ). An extended PANTA RHEI model will be developed to dynamically assimilate high-resolution GRACE/GRACE-FO TWS data for regional flood forecasting by associating these data with model parameters and state variables that represent catchment water storage in terms of e.g. soil moisture and snow using machine learning. The performance of the extended model will be assessed in the Aller-Leine-Oker, Ilmenau, Hase, Wümme, Hunte, Vechte, Große Aue river basins in Lower Saxony with catchment areas between 3.000 and 15.000 km² using hindcast simulations in the period from 2017 to 2025 which includes significant flood events.
With this contribution, we aim to present our scientific motivation, discuss the methodological framework, ideas and anticipated outcomes.
How to cite: Özgür, S. and Schröter, K.: Enhancing regional Flood Early Warning Systems using High-Resolution GRACE/GRACE-FO total water storage (TWS) and wetness index data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9818, https://doi.org/10.5194/egusphere-egu26-9818, 2026.
Accurate streamflow forecasts are essential for flood risk management and impact mitigation. In recent years, the coupling of hydrological models with machine learning techniques has gained increasing attention to improve forecast skill, with post-processing emerging among the various existing approaches to correct systematic model errors. However, the interaction between machine learning post-processing, data assimilation and calibration strategies remains insufficiently explored. In this study, we assess the contribution of machine learning-based post-processing to hourly streamflow forecasts across 687 catchments in metropolitan France, covering a wide range of hydroclimatic conditions. Streamflow forecasts are generated using the GR5H-RI hydrological model under three forecasting approaches that differ in calibration strategy and use of data assimilation. Two machine learning models, Random Forest and Multilayer Perceptron, are applied to post-process raw forecasts at lead times of 3, 6, 12 and 24 hours. Forecast performance is evaluated using both continuous skill metrics relative to persistence and threshold-based metrics for flood event detection. Results show that post-processing consistently improves forecast skill at short lead times, especially for catchments with slower hydrological responses. The largest relative gains are observed for open-loop forecasts (i.e., without data assimilation), indicating that post-processing can mitigate the absence of state updating, although it does not fully replace it. Neural network-based post-processing slightly outperforms tree-based models in continuous metrics, while differences are more limited for event detection. Overall, results highlight the complementary roles of data assimilation and machine learning post-processing and demonstrate the potential of such frameworks for operational flood forecasting.
How to cite: Gabbardo dos Reis, G., Astagneau, P. C., Bourgin, F., Andréassian, V., and Perrin, C.: The combined impact of data assimilation and machine learning post-processing in improving flood forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10432, https://doi.org/10.5194/egusphere-egu26-10432, 2026.
Flood forecasting in data-scarce catchments remains a major challenge due to limited observations and heterogeneity among basins. In this study, a regional long short-term memory model (R-LSTM) is proposed, in which runoff data are scalarised using catchment attributes, to reduce local influences and generate unified geomorphological-runoff factors for regional modeling. The proposed model is evaluated in the Jiaodong Peninsula, China, and compared with local LSTMs and regional LSTMs that incorporate catchment attributes in different ways. Results indicate that the R-LSTM consistently outperforms the benchmark models, especially in flood peak simulation. These findings demonstrate the effectiveness of the proposed regionalization strategy, providing a reference for flood forecasting in data-scarce regions.
How to cite: Ye, K., Liang, Z., Hu, Y., wang, J., and Li, B.: A Regionalization-Guided LSTM Model for Flood Forecasting in Data-Scarce Catchments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10649, https://doi.org/10.5194/egusphere-egu26-10649, 2026.
Timely and reliable urban flood forecasting is essential for mitigating damage and supporting emergency decision-making. Ensemble data assimilation can improve forecast reliability by updating model states with observations, but real-time use with high-resolution hydrodynamic models is often constrained by computational cost. We propose an integrated forecasting framework that couples a physics-guided AI emulator with data assimilation to enable efficient, high-resolution spatiotemporal inundation prediction. The emulator is trained on high-fidelity hydrodynamic simulations and reproduces key flood dynamics with substantially lower runtime than conventional solvers, allowing large ensembles generated by perturbing initial conditions and meteorological forcings to quantify uncertainty. Real-time inundation-depth observations are assimilated to update evolving flood states, using both synthetic data for controlled testing and ground-based depth information derived from surveillance-camera imagery for real-event conditions. The framework is applied for an urban drainage basin in Seoul, South Korea. The presentation will discuss key challenges for real-time urban flood assimilation, including observation uncertainty and representativeness, intermittent availability and latency, and the balance between ensemble size and update frequency. We also examine how emulator design affects physical consistency during assimilation and outline remaining limitations for operational deployment.
How to cite: Woo, H., Kim, B., Choi, H., Kim, M., Shin, E. T., Song, C. G., and Noh, S. J.: Accelerating Urban Flood Data Assimilation: Coupling Physics Guided AI Emulators with Real Time Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16407, https://doi.org/10.5194/egusphere-egu26-16407, 2026.
Karst aquifers provide vulnerable water resources accounting for 25 % of the world groundwater resources. Croatian karst aquifers are also well known as highly karstified aquifers presenting very valuable and sensitive water resources. State of the art of the karst flow and transport modelling indicates that only hybrid distributed models, also known as hydrological integrated flow and transport models, can potentially resolve complex karst Multiphysics, especially interrelation between matrix and conduit exchange flow dynamics. However, there are many limitations of distributed hydrological karst models to successful application in practice, especially at the scale of whole watershed. The main problem is requirement for so many parameters and measurements to completely describe complex karst processes. Despite progress of computational resources, hydraulic and specially geophysics equipment and measurement technologies, lot of information usually remain unresolved. The most missing information are usually matrix heterogeneity distribution (i.e. hydraulic conductivity, sorption, porosity), unsaturated (i.e. Van Genunchten) parameters and conduit network structure (depth, spatial location of conduits and/or its diameters and dimensions). Computationally expensive numerical distributed hydrological karst models could be replaced by surrogate models such as deep learning neural networks (for instance see review of Herrmann and Kollmansberger, 2024). Therefore, we discuss here Multiphysics modelling of flow and transport karst process from the classical analytical and numerical approaches to the novel machine learning approaches such as Physical Informed Neural Network - PINN. Particularly, novel advantages of inverse PINN modelling are discussed, especially due to parameter deduction, uncertainty quantification and modelling efficiency.
How to cite: Gotovac, H.: Multiiphysics karst flow and transport modelling: From the classical numerical to the novel machine learning approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16554, https://doi.org/10.5194/egusphere-egu26-16554, 2026.
Posters virtual: Fri, 8 May, 14:00–18:00 | vPoster spot A
EGU26-16119 | ECS | Posters virtual | VPS11
Differentiable, Learnable MILC: Balancing Predictive Skill and Physical InterpretabilityFri, 08 May, 14:12–14:15 (CEST) vPoster spot A
Deep learning models, particularly LSTMs, have transformed large-sample hydrology by achieving high streamflow predictive performance, yet they remain largely black-box approaches with limited physical interpretability and no explicit representation of multiphysical hydrological processes. Differentiable, learnable process-based models (or δ-models) overcome these limitations by embedding neural networks within differentiable physics frameworks. While existing benchmarks like HBV-δ have proven this concept across 671 US basins, they rely on conceptual foundations (e.g., empirical beta-functions) that approximate, rather than resolve, underlying soil physics. This study introduces MILC-δ (Modular Differentiable Physic-Informed Learning), designed to bridge this gap. The MILC model utilizes continuous soil water retention curves and physically derived drainage laws, which can aid in more accurate hydrological flux simulation. Thus, we developed a MILC-δ - a hydrologic model embedded with neural networks and trained in a differentiable programming framework. Consequently, MILC-δ is anticipated to match or exceed HBV-δ by leveraging neural networks to map static catchment attributes directly to physically measurable properties (e.g., pore size distribution, hydraulic conductivity) rather than abstract calibration parameters. Initial testing of the developed model shows that the model performs at par in some basins and better than HBV-δ in other basins. This approach gives LSTM-level accuracy and generalizability as well as the clear physical story stakeholders actually need to explain the decline in baseflow, threats to the groundwater recharge, etc.
How to cite: Sharma, V., Barbhuiya, S., and Gupta, V.: Differentiable, Learnable MILC: Balancing Predictive Skill and Physical Interpretability, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16119, https://doi.org/10.5194/egusphere-egu26-16119, 2026.
