Runoff forecasting is a long-standing challenge in hydrology, particularly in unmeasured catchments and rapid flash flood prediction. For unmeasured catchment forecasting, we introduce the encoder-decoder-based dual-layer long short-term memory (ED-DLSTM) model[1]. This model fuses static spatial granularity attributes with temporal dynamic variables to achieve streamflow forecasting at a global scale. ED-DLSTM reaches an average Nash efficiency coefficient (NSE) of 0.75 across more than 2000 catchments from historical datasets in the United States, Canada, Central Europe, and the United Kingdom. Additionally, ED-DLSTM is applied to 150 fully ungauged catchments in Chile, achieving a high NSE of 0.65. The interpretability of the transfer capacities of ED-DLSTM is effectively tracked through the cell state induced by adding a spatial attribute encoding module, which can spontaneously form hydrological regionalization effects after performing spatial coding for different catchments.
Moreover, rapid flood prediction with daily resolution is challenged to capture changes in runoff over short periods. To address this, we also propose a benchmark evaluation for runoff and flood forecasting based on deep learning (RF-Bench) at an hourly scale. We introduce the Mamba model to hydrology for the first time. The benchmark also includes Dlinear, LSTM, Transformer, and its improved versions (Informer, Autoformer, Patch Transformer). Results indicate that the Patch Transformer exhibits optimal predictive capability across multiple lead times, while the traditional LSTM model demonstrates stable performance, and the Mamba model strikes a good balance between performance and stability. We reveal the attention patterns of Transformer models in hydrological modeling, finding that attention is time-sensitive and that the attention scores for dynamic variables are higher than those for static attributes.
Our work [2,3] provides the hydrological community with an open-source, scalable platform, contributing to the advancement of deep learning in the field of hydrology.
[1] Zhang, B., Ouyang, C., Cui, P., Xu, Q., Wang, D., Zhang, F., Li, Z., Fan, L., Lovati, M., Liu, Y., Zhang, Q., 2024. Deep learning for cross-region streamflow and flood forecasting at a global scale. The Innovation 5, 100617. https://doi.org/10.1016/j.xinn.2024.100617
[2] Zhang, B., Ouyang, C., Wang, D., Wang, F., Xu, Q., 2023. A PANN-Based Grid Downscaling Technology and Its Application in Landslide and Flood Modeling. Remote Sensing 15, 5075. https://doi.org/10.3390/rs15205075
[3] Xu, Q., Shi, Y., Bamber, J.L., Ouyang, C., Zhu, X.X., 2024. Large-scale flood modeling and forecasting with FloodCast. Water Research 264, 122162. https://doi.org/10.1016/j.watres.2024.122162
How to cite: zhang, B., xu, Q., and ouyang, C.: Runoff Forecasting in Unmeasured Catchments and Rapid Flash Flood Prediction Based on Deep Learning., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4387, https://doi.org/10.5194/egusphere-egu25-4387, 2025.
Machine learning (ML) models have transformed our ability to perform reasonably-accurate, large-scale river discharge modeling, opening new opportunities for global prediction in ungauged basins. These ML models are data-hungry, and results have conclusively shown that ML techniques do best when a single ML model is trained on all basins in the dataset. This is contrary to inuitions from the hydrological sciences, where individual basin calibration traditionally provides the best forecasts.
We bridge this gap between intuitions from traditional ML and hydrology by pre-training a single global model on basins in the worldwide Caravan dataset (~6000 basins), and then fine-tune that model on individual basins. This is a well-known practice within ML, and for us serves the purpose of producing models aimed at high-quality local prediction problems while still capturing the advantages of large-sample training. We show that this leads to a significant skill improvement.
We have also conducted analysis of geophysical and hydrological regimes that benefit most from fine-tuning. These results point to how flood forecasting and water management agencies and operators can expect to fine-tune large, pretrained models on their own local data, which may be proprietary and not part of large, global training datasets.
This work illustrates how local agencies like national hydromet agencies or flood forecasting agencies might be able to leverage machine learning based hydrological forecast models while also maximizing the value and information of local data by tailoring large, pretrained models to their own local context. This is an important step in allowing local agencies to take ownership of these global models, and directly incorporate local hydrological understanding to improve performance.
How to cite: Ryd, E. and Nearing, G.: Fine Flood Forecasts: Calibrating global machine learning flood forecast models at the basin level through fine-tuning., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13027, https://doi.org/10.5194/egusphere-egu25-13027, 2025.
Sea level rise and increasing coastal flood risks demand the development of accurate and efficient coastal risk models capable of generating large ensembles of projections to support robust adaptation strategies. The latest IPCC report emphasizes the importance of projecting storm surge changes and their associated uncertainties, alongside mean sea level rise. However, the high computational cost of storm surge simulations continues to limit the feasibility of generating large ensembles.
Artificial Intelligence (AI) is emerging as a promising alternative to simulate storm surge scenarios with significantly reduced computational costs. Despite recent advancements, key challenges remain in accurately representing extreme events and ensuring robust model extrapolation under changing climate conditions. While AI-based surrogate models have been proposed in the literature, gaps persist in understanding their performance limits for extreme events in future scenarios, hindering their application in climate adaptation planning.
To address these challenges, we developed a deep learning (DL) surrogate model of the physics-based Global Tide and Surge Model (GTSM). The DL model is trained using reanalysis data (ERA5) and historical scenarios from the Coupled Model Intercomparison Project Phase 6 (CMIP6) High Resolution Model Intercomparison Project (HighResMIP). Our analysis focuses on the DL model's performance in simulating extreme storm surge events, validated against GTSM outputs for both historical reanalysis and future projections, with a case study along the New York coastline.
To enhance the surrogate model’s performance for extreme events, we explore various loss functions, including a customized quantile loss function, and test alternative DL architectures with different input configurations. Results demonstrate that the quantile loss improves the model's accuracy for extremes compared to standard loss functions such as mean square error. Additionally, fine-tuning DL models with specific Global Climate Model forcing fields improves the alignment of AI-predicted storm surge trajectories with GTSM outputs, even under diverse spatiotemporal resolutions and model setups.
These findings highlight the critical importance of selecting appropriate loss functions and training datasets to ensure robust performance over extreme events and projected future scenarios. Our globally applicable framework, relying solely on open-source data, offers a promising pathway to scalable and efficient storm surge projections, with implications for robust coastal adaptation planning.
How to cite: Longo, E., Ficchì, A., Muis, S., Verlaan, M., and Castelletti, A.: Projecting storm surge extremes with a deep learning surrogate model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9088, https://doi.org/10.5194/egusphere-egu25-9088, 2025.
Accurate soil moisture prediction is increasingly important due to its critical role in water resource management and agricultural sustainability under global climate change. While machine learning models have achieved high accuracy in soil moisture prediction, their ability to generalize to different environmental and meteorological conditions remains a significant challenge. Existing models often perform poorly when applied to conditions that differ from their training data, highlighting the need for approaches that improve generalization while effectively capturing underlying soil moisture dynamics.
In this study, we propose a novel soil moisture prediction model that combines self-supervised learning with a Transformer architecture. The performance of the model was compared with the widely used Long Short-Term Memory (LSTM)-based approach to evaluate its ability to generalize. The proposed model outperformed the baseline in tasks such as capturing extreme soil dryness, adapting to unobserved meteorological humidity conditions, and forecasting soil moisture dynamics at untrained depths. Further analysis revealed that the model’s success stems from its capability to learn comprehensive representations of underlying soil moisture processes. These results highlight the potential of advanced deep learning methods to improve prediction and our process understanding of soil hydrology in a changing climate.
How to cite: Wang, L., Shi, L., and Jiang, S.: Improving generalization of soil moisture prediction using self-supervised learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6294, https://doi.org/10.5194/egusphere-egu25-6294, 2025.
The increasing complexity of water pollution and its impact on aquatic ecosystems necessitates the accurate prediction of water pollutant loads for effective river management. Total Organic Carbon (TOC), a key indicator of organic pollution levels, is central to assessing ecosystem health and informing water treatment strategies. However, conventional process-based modeling methods, while capable of providing precise water quality predictions, require extensive input data and significant computational resources, limiting their practical application. Consequently, alternative modeling approaches, particularly those leveraging artificial intelligence, have been explored. Recent advancements in deep learning have improved predictive modeling in environmental sciences. These approaches have showed effectiveness in hydrological applications, such as streamflow forecasting, by capturing complex nonlinear relationships within environmental systems. Despite these advancements, a notable limitation of these models is their difficulty in maintaining physical consistency, specifically in adhering to the principle of mass balance—a fundamental concept in both hydrology and water quality modeling. In this study, we evaluate a Mass-Conserving Long Short-Term Memory network integrated with QUAL2E kinetics (MC-LSTM-QUAL2E) to predict TOC loads in river systems. By incorporating representations of decay and reaeration processes within a mass-conserving neural network framework, this model combines data-driven prediction capabilities with the requirements of physical consistency. A key component of this framework is the trash cell, designed to simulate TOC transformations based on QUAL2E dynamics. Within the trash cell, TOC decay and reaeration are modeled using parameters kdecay and kreaeration, which are determined by environmental variables such as temperature, pH, dissolved oxygen, total nitrogen, and total phosphorus. The QUAL2E module updates the trash state at each timestep to account for TOC losses due to decay and gains from reaeration, ensuring mass conservation. The MC-LSTM-QUAL2E model was compared to a conventional LSTM model using environmental variables, including temperature, pH, dissolved oxygen, and nutrient levels, as inputs. The analysis used data from 2012 to 2020, with the period from 2012 to 2017 designated for training and 2018 to 2020 for tests. Model performance was assessed using metrics such as Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Root Mean Square Error-observations standard deviation Ratio (RSR), and Percent Bias (PBIAS). By maintaining mass balance and incorporating QUAL2E dynamics, the model provides reliable predictions of TOC loads in river systems and offers insights into associated biochemical and hydrological processes.
How to cite: Jeong, H., Lee, B., Lee, Y., and Lee, S.: Predicting total organic carbon loads in river using a mass-conserving LSTM integrated with QUAL2E kinetics, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7531, https://doi.org/10.5194/egusphere-egu25-7531, 2025.
Deep learning methods in hydrology have traditionally focused on deterministic models, limiting their ability to quantify prediction uncertainty. Recent advances in generative modeling have opened new possibilities for probabilistic modelling in various applied fields, including hydrological forecasting (Jahangir & Quilty, 2024). These models learn to represent underlying probability distributions using neural networks, enabling uncertainty quantification through sampling in a very flexible framework.
In this submission we introduce vLSTM, a variational extension of the traditional long short-term memory (LSTM) architecture that quantifies predictive uncertainty by adding noise sampled from a learned multivariate Gaussian distribution to perturb the model’s hidden state. The vLSTM preserves the traditional LSTM’s state-space dynamics while introducing a probabilistic component that enables uncertainty quantification through sampling. Unlike mixed-density networks (MDNs) which directly model the distribution of the target variable, vLSTM’s uncertainty is obtained by perturbations to the hidden state, providing a novel approach to probabilistic prediction. In rainfall-runoff modeling, vLSTM offers a different mechanism for uncertainty quantification to the well established MDN models (Klotz et al., 2022). This approach enriches the existing toolkit of uncertainty methods in deep learning while maintaining the simplicity of sampling for probabilistic predictions.
To rigorously evaluate probabilistic predictions across different model architectures, we develop new information-theoretic metrics that capture key aspects of how uncertainty is handled by a particular model. These include the average prediction entropy H(X), which quantifies model confidence, and average relative entropy DKL(pq), which measures the average alignment between the predicted distribution of a model and a target, among others. The proposed metrics take advantage of non-parametric estimators for Information Theory which have been implemented in the easy to use UNITE toolbox (https://github.com/manuel-alvarez-chaves/unite_toolbox). By expressing these metrics in compatible units of bits (or nats), we enable direct comparisons between different uncertainty measures. We apply these metrics to our newly introduced vLSTM and the existing MDN models to show strengths and weaknesses of each approach. This information-theoretic framework provides a unified language for analyzing and understanding predictive uncertainty in probabilistic models.
References
- Jahangir, M. S., & Quilty, J. (2024). Generative deep learning for probabilistic streamflow forecasting: Conditional variational auto-encoder. Journal of Hydrology, 629, 130498. https://doi.org/10.1016/j.jhydrol.2023.130498
- Klotz, D., Kratzert, F., Gauch, M., Keefe Sampson, A., Brandstetter, J., Klambauer, G., Hochreiter, S., & Nearing, G. (2022). Uncertainty estimation with deep learning for rainfall–runoff modeling. Hydrology and Earth System Sciences, 26(6), 1673–1693. https://doi.org/10.5194/hess-26-1673-2022
How to cite: Alvarez Chaves, M., Gupta, H., Ehret, U., and Guthke, A.: Evaluating uncertainty in probabilistic deep learning models using Information Theory, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7205, https://doi.org/10.5194/egusphere-egu25-7205, 2025.
Hydrological models are vital in river discharge to issue timely flood warnings and mitigate hydrological risks. Recently, advanced techniques in deep learning have significantly enhanced flood prediction by improving the accuracy and efficiency of forecasts, enabling more reliable early warnings and decision-making in flood risk management. Nevertheless, current applications of deep learning methods are still more restricted to local-scale models or in the best case on selected river points at a global scale. Many studies also lack spatial and topological information for training deep learning models, which can limit their generalization ability when applied to large regions with heterogeneous hydrological conditions. In addition, the lack of probabilistic forecasting impedes the quantification of uncertainty in flood predictions. Here we present the Artificial Intelligence Flood Awareness System (AIFAS) for probabilistic global river discharge forecasting. AIFAS is a generative AI model that is trained with long-term historical reanalysis data and can provide grid-based global river discharge forecasting at 0.05°. At the core of our model are the built-in vision module upon state space model (SSM) [1] and the diffusion-based loss function [2]. The vision SSM allows the model to connect the routing of the channel networks globally, while the diffusion loss generates ensembles of stochastic river discharge forecasts. We evaluate the AIFAS forecast skill against other state-of-the-art deep learning models, such as Google LSTM [3], climatology baseline, persistence baseline, and operational GloFAS forecasts [4]. The impact of different hydrometeorological products that drive AIFAS performance on different forecasting lead times will also be discussed. Our results show that the new forecasting system achieves reliable predictions of extreme flood events across different return periods and lead times.
References
[1] Gu, A., and Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.
[2] Ho, J., Jain, A., and Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems, 33, 6840-6851, 2020.
[3] Nearing, G., Cohen, D., Dube, V. et al. Global prediction of extreme floods in ungauged watersheds. Nature 627, 559–563 2024.
[4] Harrigan, S., Zsoter, E., Cloke, H., Salamon, P., and Prudhomme, C.: Daily ensemble river discharge reforecasts and real-time forecasts from the operational Global Flood Awareness System, Hydrol. Earth Syst. Sci., 27, 1–19, 2023.
How to cite: Shams Eddin, M. H., Zhang, Y., Kollet, S., and Gall, J.: AIFAS: Probabilistic Global River Discharge Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13844, https://doi.org/10.5194/egusphere-egu25-13844, 2025.
This study evaluates the capacity of a Long Short-Term Memory (LSTM) model trained on a diverse river discharge dataset from over 4,000 USGS gauges across the United States with the aim to generate extremely long stochastic discharge simulations.
The LSTM model (Kratzert et al., 2022) was trained using 30 years of NLDAS v2 forcings, which were split into 10-year periods for training, validation, and testing respectively. Sixty percent of the gauges had a Nash Sutcliffe Efficiency (NSE) larger than 0.4 in the validation period, and ten percent had an NSE larger than 0.8, which was considered sufficient to proceed with applying the model using stochastic precipitation.
The stochastic simulations are evaluated in terms of the model’s ability to capture peak discharges. The stochastic return period (RP) curves were evaluated against those from the historical time period and the observed discharge. For most of the gauges, the stochastic RP curves are in line with the historical RP curves, and for all of the gauges, the stochastic RP curves discharge of the extreme return period extend far beyond the discharge of the historical time period, showing the capacity of the model to extrapolate beyond the training dataset.
This capacity, which is usually lacking in single-basin trained models, most likely results from training on a large dataset with a wide range of climatic conditions and variability as suggested by Kratzert et al. (2024). These findings underscore the robustness and versatility of the LSTM model in long-term stochastic discharge simulations, highlighting its potential for broader hydrological applications.
Kratzert, F., Gauch, M., Nearing, G., & Klotz, D. (2022). NeuralHydrology — A Python library for Deep Learning research in hydrology. Journal of Open Source Software, 7(71), 4050. https://doi.org/10.21105/joss.04050
Kratzert, F., Gauch, M., Klotz, D., and Nearing, G. (2024). HESS Opinions: Never train an LSTM on a single basin. Hydrology and Earth System Sciences (HESS), Volume 28, Issue 17, published on September 12, 2024. https://doi.org/10.5194/hess-28-4187-2024.
How to cite: Jankowfsky, S., Gokul, K., Li, S., Hilberts, A., and Assteerawatt, A.: Evaluation of LSTM Model for Stochastic Discharge Simulation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15152, https://doi.org/10.5194/egusphere-egu25-15152, 2025.
Long Short-Term Memory (LSTM) networks have become popular for streamflow prediction in hydrological systems due to their ability to model sequential data. However, their reliance on lumped catchment representation and climate summaries often limits their capacity to capture spatial heterogeneity in rainfall patterns and evolving rainfall trends, both of which are critical for hydrological consistency. This study explores the limitations of LSTM-based streamflow predictions by employing a distributed conceptual hydrological model, SIMHYD, coupled with Muskingum-Cunge routing, to generate synthetic datasets representing diverse hydroclimatic conditions. These datasets are designed to replicate rainfall-runoff dynamics across selected catchments from all 18 ecoregions in CAMELS-US and key Indian river basins, providing a robust testbed for evaluating model performance under controlled conditions. The pre-trained LSTM model is tested against synthetic discharge data, enabling direct comparisons to assess its ability to simulate realistic hydrological responses. Performance is evaluated using multiple metrics, including Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Percent Bias (PBIAS), and mean peak timing errors, to identify systematic deviations. Results reveal that LSTM models struggle with spatially variable and temporally shifting rainfall patterns, leading to inaccuracies in peak flow timing, magnitude, and overall discharge volumes. These issues highlight vulnerabilities in current LSTM-based flood forecasting systems, particularly in their ability to generalize across diverse climatic conditions and regions. This study also characterizes specific failure pathways, such as underestimation of extreme events and poor temporal coherence in hydrographs, which are critical for operational forecasting. By diagnosing these limitations, the study provides a framework for integrating process-based hydrological knowledge with data-driven techniques to improve model robustness. The findings underscore the importance of using synthetic datasets and diverse diagnostic tools to evaluate and enhance the reliability of LSTM-based models, paving the way for hybrid approaches capable of addressing the complexities of real-world hydrological systems.
How to cite: Dubey, S., Bhasme, P., and Bhatia, U.: Characterizing possible failure modes: Insights from LSTM-Based Streamflow Predictions, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-935, https://doi.org/10.5194/egusphere-egu25-935, 2025.
Accurate streamflow prediction is critical for flood forecasting and water resource management, particularly in data-scarce regions like Central Asia (CA), where traditional hydrological models struggle due to insufficient discharge data. Deep learning models, such as Long Short-Term Memory (LSTM), have demonstrated the potential for global hydrologic regionalization by leveraging both climate data and catchment characteristics. We used a transfer learning (TL) approach to improve streamflow predictions by first pretraining LSTM models on catchments from data-rich regions like Switzerland, Scotland, and British Columbia (source regions). These deep learning models were then fine-tuned on the data scarce target region (CA basins). This approach leverages the knowledge gained from the source regions to adapt the model to the target region, enhancing prediction accuracy despite the data scarcity in CA. Incorporating lagged streamflow alongside ERA-5 climate data boosted prediction accuracy, particularly in snowmelt and glaciers influenced basins like Switzerland (median NSE=0.707 to 0.837), British Columbia (median NSE= 0.775 to 0.923) and CA (median NSE=0.693 to 0.798). K-Means algorithm was applied to categorize catchments from four global locations into five clusters (labeled 0–4) based on their specific attributes. The predictive performance of fine-tuned LSTM model has significantly enhanced when leveraging a pre-trained model with cluster 2, as demonstrated by higher median metrics (NSE=0.958, KGE=0.905, RMSE=10.723, MSE=115.055) compared to both the locally trained model (NSE=0.851, KGE=0.792, RMSE=20.377, MSE=415.579) and individual basin-based training approaches (NSE=0.69, KGE=0.692, RMSE=25.563, MSE=676.110). These results highlight the effectiveness of pretraining the LSTM model on diverse clusters (0, 1, 2, and 4) before fine-tuning on the target region (cluster 3). Moreover, pretraining the LSTM model with clusters 0 and 4 resulted in enhanced performance by increasing the number of basins, whereas the impact was minimal or even declined when using clusters 1 and 2, as well as when all basins from the four clusters were included. These findings demonstrate the feasibility of transfer learning in addressing data scarcity challenges and underscore the importance of diverse and high-quality training data in developing robust, regionalized hydrological models. This approach bridges the gap between data-rich and data-scarce regions, offering a pathway to improved flood prediction and water resource management.
How to cite: Hassan, J., Rowan, J., and Mukherjee, N.: Exploring the Transferability of Knowledge In Deep Learning-Based Streamflow Models Across Global Catchments, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11240, https://doi.org/10.5194/egusphere-egu25-11240, 2025.
Accurately forecasting streamflow is essential for effectively managing water resources. High-quality operational forecasts allow us to prepare for extreme weather events, optimize hydropower generation, and minimize the impact of human development on the natural environment. However, streamflow forecasts are inherently limited by the quality and availability of upstream weather sources. The weather forecasts that drive hydrological modeling vary in their temporal resolutions and are prone to outages, such as the ECMWF data outage in November of 2023.
Here, we present HydroForecast Short Term 3 (ST-3), a state-of-the-art probabilistic deep learning model for medium-term (10-day) streamflow forecasts. ST-3 combines long short-term memory architecture with Boolean tensors representing data availability and dense embeddings for processing of the information in these tensors. This architecture allows for a training routine that implements data augmentation to synthesize varying amounts of availability of weather inputs. The result is a model that 1) makes accurate forecasts even in the case of an upstream data outage, 2) achieves higher accuracy by leveraging data of varying temporal resolutions including regional weather inputs with shorter lead times than the most common medium term weather inputs, and 3) generates individual forecast traces for each individual weather source, facilitating inference across regions where weather data availability is limited.
Initial results across CAMELS sites in North America indicate that the incorporation of near-term high resolution weather data increases early horizon forecast KGE by nearly 0.25 with meaningful improvements in metrics seen across our customers’ operational sites. Validation metrics across individual weather sources, as well as model interrogation through integrated gradients highlights a high level of fidelity in the model’s learned physical relationships across forecast scenarios.
How to cite: Lambl, D., Topp, S., Butcher, P., Elkurdy, M., Reed, L., and Sampson, A. K.: Increasing the Accuracy and Resilience of Streamflow Forecasts through Data Augmentation and High Resolution Weather Inputs, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12110, https://doi.org/10.5194/egusphere-egu25-12110, 2025.
Accurate streamflow estimation is crucial for effective water resource management and flood forecasting. However, physics-based hydrological models fail to respond promptly to rapid hydrological events due to lack efficiency in model calibration and computing time for large-scale catchment , while existing deep learning models tend to neglect the physical processes of runoff transfer, failing to account for the spatial and temporal dependencies inherent in runoff dynamics. In this study, we propose a topological process-based model that integrates Graph Attention Networks (GAT) to capture the spatial topology of runoff transfer and Long Short-Term Memory (LSTM) networks to simulate the temporal transfer between upstream and downstream runoff. The model was applied to the Yangtze River Basin which is the largest river basin in China to predict streamflow at 10 km spatial resolution. Validation results show that our model achieves a median Nash-Sutcliffe Efficiency (NSE) value of 0.783 at secondary outlet stations across the basin and effectively simulates the streamflow peak due to flooding. Additionally, the model is capable of simulating the spatial distribution of daily streamflow for an entire year within 10 seconds, providing a significant computational speedup compared to physical process-based river confluence models. This work represents a step towards more efficient and responsive prediction of extreme hydrological events using deep learning model.
How to cite: Wang, X., Zhang, X., and Zhang, Y.: Application of Attention-Based Graph Neural Networks for Spatial Distribution Prediction of Streamflow, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19190, https://doi.org/10.5194/egusphere-egu25-19190, 2025.
Recent advances in catchment hydrology [Kratzert et al., 2019–2021] demonstrate the superiority of LSTMs over traditional conceptual models for streamflow prediction in large-sample datasets. LSTMs achieve better streamflow accuracies by leveraging information from diverse hydrological behaviors. These models are enriched with static catchment attributes, which, when combined with meteorological drivers, play a critical role in streamflow formation. Augmenting LSTMs with these attributes further enhances their performance compared to vanilla LSTMs, underscoring the importance of these attributes for accurate streamflow predictions. Building on this, a recent study [Bassi et al., 2024] employed a conditional autoencoder to reveal that most of the relevant catchment information for streamflow prediction can be distilled into two features, with a third feature being beneficial for particularly challenging catchments. In this work, we directly derive a minimal set of catchment features from known attributes by passing them through the encoder and subsequently comparing streamflow predictions against state-of-the-art benchmarks [Kratzert et al., 2021]. Our findings indicate that while the intrinsic dimension of 26 commonly used attributes is four, only two features suffice for accurate streamflow prediction. This aligns closely with the findings of Bassi et al. (2024), suggesting that nearly all relevant information for streamflow prediction is encapsulated in known attributes. Finally, we provide an interpretation of these two machine-learning-derived features using information theory techniques.
How to cite: Bassi, A. and Albert, C.: Leveraging Machine Learning to Uncover and Interpret Relevant Catchment Features , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19716, https://doi.org/10.5194/egusphere-egu25-19716, 2025.
In the last decade, the realization that certain deep learning (DL) architectures are particularly well-suited to the simulation and prediction of hydrologic systems and their characteristic memory-influenced dynamics has led to remarkable rise in DL-centered hydrologic research and applications. Numerous new datasets, computational and open software resources, and progress in related fields such as numerical weather prediction have also bolstered this growth. Advances in DL for hydrologic forecasting research and operations is likely the most eye-catching and intuitive use case, but DL methods are now also making inroads into more process-intensive hydrologic modeling contexts, and among groups that have been skeptical of their potential suitability despite performance-related headlines. Nevertheless, even in the forecasting context, and despite offering new strategies and concepts to resolve long-standing hurdles in hydrologic process-based modeling efforts, the uptake of DL-based systems in many public-facing services and applications has been slow.
This presentation provides perspective on the ways in which DL techniques are garnering interest in traditionally process-oriented modeling arenas -- from flood and drought forecasting to watershed studies to hydroclimate risk modeling – and on sources of hesitancy. Clear pathways, momentum and motivations for DL approaches to supplant process-based models exist in some applications, whereas in others, governing interests and constraints appear likely to restrict DL innovations to narrower niches. Concerns over explainability have been a common topic, but less discussed questions about fitness or adequacy for purpose and institutional requirements can also be influential. Drawing from relevant hydrologic modeling programs, projects and initiatives in the US and elsewhere, we aim to provide a real-world status update on the advancing frontier of deep learning in applied hydrologic science and practice.
How to cite: Wood, A., Read, L., Nearing, G., Mai, J., Frans, C., Clark, M., and Pappenberger, F.: On the advancing frontier of deep learning in hydrology: a hydrologic applications perspective, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14223, https://doi.org/10.5194/egusphere-egu25-14223, 2025.
Measuring river surface velocity enables river discharge estimation, a fundamental task for hydrologists, environmental scientists, and water resource managers. While traditional image-based velocimetry methods are often effective, they struggle to produce complete velocity fields under complex environmental conditions. Poor lighting, reflective glare, lack of visible surface features, or excessive turbulence can all result in regions where feature tracking fails, leading to gaps in the resolved velocity field. Addressing these gaps through the reconstruction of missing velocity measurements is an important research challenge. Recently, researchers have employed deep learning to address various hydrology problems, demonstrating promising improvements. In this work, we propose a neural operator-based model to address the challenge of missing velocities in river surface velocimetry. Specifically, our model is based on the Fourier neural operator with a graph-enhanced lifting layer. It is trained on the river surface velocimetry reconstruction task using a self-supervised paradigm. Once trained, it can be used to infer missing velocities in unseen samples. Experiments conducted on a dataset collected from a river in the Netherlands demonstrate our approach’s ability to accurately infill missing surface velocities, even when faced with large amounts of missing data. We attribute this robustness to the neural operator’s ability to learn continuous functions, which enhances our model’s capacity for high-level feature representation and extraction. Our findings suggest that the reconstructed velocity fields produced by our model can act as reliable ground truth data for deep learning-based methods. In the future, we aim to improve our model’s performance and generalization by incorporating additional data collected from a wider range of rivers and under varying environmental conditions.
How to cite: Chen, X., Winsemius, H., and Taormina, R.: Graph-enhanced Neural Operator for Missing Velocities Infilling in River Surface Velocimetry, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1644, https://doi.org/10.5194/egusphere-egu25-1644, 2025.
Historical records observed at hydrological stations are scarce in many regions, bringing significant challenges to the hydrological predictions for these regions. Transfer learning (TL), increasingly applied in hydrology, leverages knowledge from data-rich catchments (sources) to enhance predictions in data-scarce catchments (targets), providing new insight into data-scarce region predictions. Most existing TL approaches pre-train models using large meteoro-hydrological datasets to improve overall generalizability to target catchments. However, the predictive performance for specific catchments would be constrained due to irrelevant source data inputs and the lack of effective source fusion strategies. To address these challenges, this study proposes the Dual-Source Adaptive Fusion TL Network (DSAF-Net), which utilizes a pre-trained dual-branch feature extraction module (DBFE) to extract knowledge from two carefully selected source catchments, minimizing noise and redundancy associated with larger datasets. A cross-attention fusion module is then incorporated to dynamically identify key knowledge of the target catchment and adaptively fuse complementary information. This fusion module is embedded after each layer in the DBFE to enhance multi-level feature integration. Results demonstrate that DSAF-Net achieves superior prediction accuracy to single-source TL and large dataset TL strategies. These findings highlight the potential of DSAF-Net to advance hydrological forecasting and support water resource management in data-scarce regions.
How to cite: Gao, Y. and Liang, D.: Dual-Source Adaptive-Fusion Transfer Learning for Hydrological Forecasting in Data-scarce Catchments, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2616, https://doi.org/10.5194/egusphere-egu25-2616, 2025.
The increasing frequency of extreme hydrological events, such as floods and droughts, poses significant challenges for operators of drinking water reservoirs in maintaining a balance between water supply and demand. While the security of supply typically requires high water levels to meet consumer demands throughout the year, ensuring flood protection, on the contrary, necessitates that reservoir storage is kept partially free to accommodate high inflows. Accurate inflow forecasting is essential for making risk-based operational decisions, including the timely release of water from drinking water reservoirs to mitigate flood risks. While deep learning approaches, particularly Long Short-Term Memory (LSTM) networks, have become prevalent in rainfall-runoff modeling, most existing studies focus on small, homogeneous datasets limited to single hydrological basins. This study leverages the newly published CAMELS-DE dataset to develop a regionally trained and finetuned LSTM model encompassing 1,582 catchments across Germany. We apply this regional model to five small catchments upstream of drinking water reservoirs and compare its performance against basin-specific LSTM models. Our findings demonstrate that the regionally trained LSTM model significantly improves the accuracy of inflow estimates, especially when finetuned to our target catchments. This is highlighting its potential for enhancing reservoir management strategies in the face of climate change.
How to cite: Johnen, G., Niemann, A., Nistahl, P., Dolich, A., and Hutwalker, A.: Forecasting Reservoir Inflows Using Regionally Trained and Finetuned LSTM Models: A Case Study with CAMELS-DE, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2761, https://doi.org/10.5194/egusphere-egu25-2761, 2025.
Deep learning models such as the Long Short Term Memory Network (LSTM) are capable of representing rainfall-runoff relationships and outperform classical hydrological models in gauged and ungauged settings (Kratzert et al., 2018). Previous studies have shown that combining multiple precipitation data in a single LSTM significantly improves the accuracy of simulated runoff, as the neural network learns to combine temporal and spatial patterns of inputs (Kratzert et al., 2021). However, every operational runoff forecasting setting requires meteorological forecasts. Nearing et al. (2024) have developed a global runoff forecast model based on an LSTM, with an ECMWF forecasting product as additional input over the forecast horizon. Compared with observed or reanalysis meteorological input data, forecasting products generally have a lower accuracy, with different reliabilities between various forecasting products. This is where the synergies of several meteorological forecasts combined with historical observational and reanalysis data can be used in a single deep learning model.
This study investigates how well LSTMs can predict runoff when trained on (1) multiple archived meteorological forecasts and (2) a combination of multiple archived meteorological forecasts and reanalysis data. All meteorological input data are aggregated to the catchments of the LamaH-CE dataset (Klingler, Schulz and Herrnegger, 2021). Runoff predictions are evaluated for a 24 hours forecasting horizon. Preliminary analyses indicate that the coupling of reanalysis data and forecasting products from different sources improves the accuracy of operational runoff forecasting, suggesting that the model is able to learn and adjust real-time biases in forecasting data.
Klingler, C., Schulz, K. and Herrnegger, M. (2021) ‘LamaH-CE: LArge-SaMple DAta for Hydrology and Environmental Sciences for Central Europe’, Earth System Science Data, 13(9), pp. 4529–4565. DOI: 10.5194/essd-13-4529-2021.
Kratzert, F., Klotz, D., Brenner, C., Schulz, K. and Herrnegger, M. (2018) ‘Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks’, Hydrology and Earth System Sciences, 22(11), pp. 6005–6022. DOI: 10.5194/hess-22-6005-2018.
Kratzert, F., Klotz, D., Hochreiter, S. and Nearing, G.S. (2021) ‘A note on leveraging synergy in multiple meteorological data sets with deep learning for rainfall–runoff modeling’, Hydrology and Earth System Sciences, 25(5), pp. 2685–2703. DOI: 10.5194/hess-25-2685-2021.
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. and Matias, Y. (2024) ‘Global prediction of extreme floods in ungauged watersheds’, Nature, 627(8004), pp. 559–563. DOI: 10.1038/s41586-024-07145-1.
How to cite: Konold, O., Feigl, M., Klingler, C., and Schulz, K.: From multiple meteorological forecasts to river runoff: Learning and adjusting real-time biases to enhance predictions, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4247, https://doi.org/10.5194/egusphere-egu25-4247, 2025.
Accurate water level prediction is essential for flood risk management, water resources management, inland water transportation and climate resilience. Traditional statistical methods, such as autoregressive models, and physically-based hydrological simulations, have been widely used in water level forecasting. However, these approaches often struggle to capture complex, dynamic, and nonlinear interactions in a hydrological system, particularly those affected by climate change. In recent years, machine learning models have emerged as a promising alternative, offering improved predictive accuracy and adaptability across varying environmental conditions. A special type of such models is the Graph Neural Network (GNN), which focuses especially on the reproduction of spatial dependencies, and hence it can be employed to capture the spatial dynamic of the hydrologic/hydraulic system, by treating hydrological networks as graph structures (e.g. nodes as gauges). Going one step further, GNN models can be combined with sequence-based machine learning techniques, such as the Long short-term memory (LSTM) neural network, to capture simultaneously the spatial and temporal dynamics of the system. In this work, we develop and assess a series of advanced hybrid-graph structured machine learning models (such as GNN-LSTM) to make hydrometric predictions across a long river channel. The developed models will be assessed on the basis of alternative performance metrics and against a series of traditional benchmark statistical and machine learning models such as ARIMA and LSTM respectively. As a test case, we exploit data from 19 water level gauges in the Red River of the North, which spans 885 km, serving the natural boundary between North Dakota and Minnesota and has experienced several severe historical flood events.
How to cite: Koutsos, G., Kossieris, P., Thomopoulou, V., and Makropoulos, C.: Leveraging Graph Neural Networks for water level prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4500, https://doi.org/10.5194/egusphere-egu25-4500, 2025.
River level predicting underpins the management of water resource projects, steers navigational activities in rivers, and protects the lives and properties of riverside communities, etc. Traditionally, hydrological-hydraulic coupled models have been at the forefront of simulating and predicting river levels, achieving notable success. Despite their utility, these models encounter limitations due to the exhaustive demand for various data types—often difficult to obtain—and the ambiguity in determining downstream boundary conditions for the hydraulic model. Responding to these limitations, this study utilizes Long Short-Term Memory (LSTM) model, a deep learning technique, to predict river levels using upstream discharges. Three approaches were used to further enhance the accuracy and reliability of our model. Firstly, we incorporated historical water level data at or downsteam of the predicted station as input, secondly, we classified the datasets based on physical principles, and thirdly, we employed data augmentation techniques. These methods were evaluated within the Jingjiang-Dongting river-lake system in China. It achieves high prediction accuracy of water level and can mitigate the impact of input inaccuracies. The incorporation of water level data as input and the Classification-Enhanced LSTM model that segregates the input data according to rising and recession trends of water level,significantly improve prediction accuracy under extreme water level conditions compared with other deep learning approaches. The proposed model uses easily accessible data to predict water levels, offering enhanced robustness and new strategies for improving prediction accuracy under extreme conditions. It is applicable for predicting water levels at any hydrological station along a river and can enhance the prediction accuracy of hydraulic models by proving a robust downstream boundary condition.
How to cite: Luo, J.: Classification-Enhanced LSTM Model for Predicting River Water Levels, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4683, https://doi.org/10.5194/egusphere-egu25-4683, 2025.
This study evaluates a new approach to improving streamflow forecasting with deep learning: it focuses on the novel application of a dynamic multimodal feature fusion mechanism that adapts fusion operations based on the data's characteristics. Two baseline Long Short-Term Memory (LSTM) architectures are used, applying two dynamic fusion methods: dynamic operation-level fusion and attention-based fusion, to combine heterogeneous and multisource hydrometeorological data. The models are used for univariate (single flow gauge) and multivariate (multi-gauge) streamflow forecasting approaches. Applying these four approaches to the Severn Basin in the UK, known for long medium- to high-flow periods and shorter low-flow intervals, shows that the dynamic operation-level fusion consistently improved over the attention-based fusion in key performance metrics. In the multivariate case, Nash-Sutcliffe Efficiency (NSE) improved by 1.43%, Mean Absolute Error (MAE) decreased by 1.73%, Mean Absolute Scaled Error (MASE) dropped by 1.82%, and high-peak MAE decreased by 3.36%. For the univariate case, NSE improved by 1.44%, MAE decreased by 4.02%, MASE dropped by 3.89%, and high-peak MAE improved by 2.8%. In addition, multivariate models were considerably faster than univariate models, with training and inference times reduced by 74.57% and 73.81%, respectively. The multivariate models showed a 2.75% increase in NSE and a 72.04% decrease in MASE, indicating they captured better the hydrologic variability than the univariate models. Conversely, univariate models had a 20.59% lower MAE, a 21.17% lower high-peak MAE, and greater stability as indicated by tighter interquartile ranges, suggesting better error minimisation and more reliable predictions. Notably, in two river stations all models underperformed due to rapid flow variability and flashy hydrological responses in smaller catchment areas, suggesting in the future the use of higher-resolution climatic data. Overall, the study shows the potential of new dynamic multimodal fusion techniques, navigating the operational trade-offs between speed, stability, and accuracy across multi and uni-variate training strategies in streamflow forecasting. Nonetheless, the need for an optimal operational balance remains, suggesting further refinement of fusion techniques and focusing on minimising uncertainty.
How to cite: Theodosiadou, K., Rodding Kjeldsen, T., and Barnes, A.: Improving high-flow forecasting using dynamic multimodal feature fusion, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7020, https://doi.org/10.5194/egusphere-egu25-7020, 2025.
Accurate nutrient predictions are crucial for river water quality management. While deep learning (DL) has shown promise in various Earth science applications, challenges such as data scarcity and limited interpretability hinder its use in river nutrient predictions. Building on insights into the physical dynamics of nutrients, this research investigates how incorporating extreme weather indices as additional input data, which are often overlooked in current DL-based nutrient prediction, could affect model performance. Additionally, we aim to improve model interpretability by developing hybrid DL-physical structures and identify the optimal structure for predicting nutrient indicators.
The study proposes an assessment workflow and demonstrates its application by predicting dissolved inorganic nitrogen (DIN) and soluble reactive phosphorus (SRP) concentrations at the outlet of the Salmons Brook catchment, UK, where nutrient observations are scarce. The workflow includes two key decisions: selecting the input dataset and defining the DL-physical hybrid structure, each with two options. Comparing multiple predictions generated from all decision combinations enables the evaluation of the impacts of extreme weather events and different hybrid structures.
The simulations demonstrate that incorporating extreme weather indices as additional inputs enhanced performance for both nutrient indicators, particularly in capturing extreme values. Overall, the choice of input dataset had a greater impact on the simulations than the hybrid structure, highlighting the importance of careful input selection and preprocessing in DL model development. Integrating results from a physical model into a DL model can improve simulation interpretability by introducing nutrient-related physical processes. In addition to the hybrid structure, incorporating insights into the physical behaviour of nutrients further enhances the interpretability of DL-based predictions, which is crucial for gaining the trust of domain experts, especially when validating results.
How to cite: Tang, J., Liu, L., Chun, K., and Mijic, A.: Enhancing River Nutrient Predictions with Extreme Weather Indices and DL-Physical Hybrid Structures for Improved Interpretability, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7094, https://doi.org/10.5194/egusphere-egu25-7094, 2025.
Deep learning (DL)-based hydrological models, particularly those using Long Short-Term Memory (LSTM) networks, typically require large datasets for effective training. In the context of large-scale rainfall-runoff modeling, dataset size can refer to either the number of watersheds or the length of the training period. While it is well established that training a regional model across more watersheds improves performance (Kratzert et al., 2024), the benefits of extending the training period are less clear.
Empirical evidence from studies such as Boulmaiz et al. (2020) and Gauch et al. (2021) suggests that longer training periods enhance LSTM performance in rainfall-runoff modeling. This improvement is attributed to the need for extensive datasets to ensure proper model convergence and the ability to capture a wide range of hydrological conditions and events. However, these studies neglected the influence of data recency (or data recentness), which is critical for operational applications that forecast current and future hydrological conditions. In the context of climate change and anthropogenic interventions, the assumption of stationarity (i.e., that historical patterns reliably represent future conditions) may no longer hold for hydrological systems (Shen et al., 2022). Consequently, the selection of training periods should account for potential non-stationarity, as more recent data may better reflect current rainfall-runoff dynamics. Intriguingly, Shen et al. (2022) found that calibrating hydrologic models to the latest data is a superior approach compared to using old data, and completely discarding the oldest data can even improve the performance in streamflow prediction.
This study aims to address two research questions: (1) As the number of watersheds increases, is it still necessary to train LSTM models on decades of historical observations? (2) Can LSTM models achieve comparable performance using shorter training periods focused on more recent data? Specifically, we examine whether models trained on recent data outperform those trained on older data and explore how different temporal partitions of historical records affect predictive skill.
This study leverages a comprehensive dataset comprising streamflow records from over 1,300 watersheds across North America, representing diverse climatic and hydrological regimes, with streamflow data spanning 1950 to 2023. Training periods are designed to isolate the effects of temporal data recency while keeping period lengths consistent. This approach enables a systematic comparison of model performance using exclusively older (e.g., pre-1980) versus exclusively recent data (e.g., post-1980). This research provides evidence-based recommendations for selecting training data while balancing computational costs, data availability, and prediction accuracy.
References
Boulmaiz, T., Guermoui, M., and Boutaghane, H.: Impact of training data size on the LSTM performances for rainfall–runoff modeling, Model Earth Syst Environ, 6, 2153–2164, https://doi.org/10.1007/S40808-020-00830-W/FIGURES/9, 2020.
Gauch, M., Mai, J., and Lin, J.: The proper care and feeding of CAMELS: How limited training data affects streamflow prediction, Environmental Modelling and Software, 135, https://doi.org/10.1016/j.envsoft.2020.104926, 2021.
Kratzert, F., Gauch, M., Klotz, D., and Nearing, G.: HESS Opinions: Never train an LSTM on a single basin, Hydrology and Earth System Science, https://doi.org/10.5194/hess-2023-275, 2024.
Shen, H., Tolson, B. A., and Mai, J.: Time to Update the Split-Sample Approach in Hydrological Model Calibration, Water Resour Res, 58, e2021WR031523, https://doi.org/10.1029/2021WR031523, 2022.
How to cite: Yu, Q. and Tolson, B.: Empirical Evidence of the Importance of Data Recency in LSTM-Based Rainfall-Runoff Modeling , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7468, https://doi.org/10.5194/egusphere-egu25-7468, 2025.
Integrating deep learning techniques into hydrology has opened a new way to improve urban flood modeling, with various solutions being developed to address urban flood problems driven by climate change and urbanization. However, predicting urban inundation in near real-time for large urban areas remains challenging due to computational demands and limited data availability. This work proposes a deep learning-based super-resolution framework that enhances the spatial resolution of process-based urban flood modeling outputs using convolutional neural networks (CNNs) while improving computational efficiency. This study investigates the interaction between deep learning model architecture and the underlying physical processes to improve prediction accuracy and robustness in urban pluvial flood mapping. The methodology will be applied to various urban flood scenarios, including extreme rainfall events and hurricane-induced flooding, and its performance will be evaluated through quantitative indicators and sensitivity analyses. The applicability and scalability of this model will also be discussed. In particular, strategies to enhance model reliability and integrate additional hydrological information under extreme conditions will be explored. The study will further address uncertainty estimation in deep learning-based super-resolution models and scalability challenges associated with super-resolution approaches for large-scale flood simulations. The findings aim to demonstrate the potential of deep learning as an innovative tool in hydrological modeling and to enable more effective flood risk management strategies.
How to cite: Choi, H., Woo, H., Kim, M., Ryu, H., Lee, J., Lee, S., and Noh, S. J.: Developing a Deep Learning-Based Super-Resolution Urban Flood Model: Towards Scalable and Reliable Hydrological Predictions, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8045, https://doi.org/10.5194/egusphere-egu25-8045, 2025.
Machine learning is increasingly important for rainfall–runoff modelling. In particular, the community started to widely adopt the Long Short-Term Memory (LSTM) network. One of the most important established best practices in this context is to train the LSTMs on a large number of diverse basins (Kratzert et al., 2019; 2024). Intuitively, the reason for adopting this practice is that training deep learning models on small and homogeneous data sets (e.g., data from only a single hydrological basin) leads to poor generalization behavior — especially for high-flows.
To examine this behavior, Kratzert et al. (2024) use a theoretical maximum prediction limit for LSTMs. This theoretical limit is computed as the L1 norm (i.e., the sum of the absolute values of each vector component) of the learned weight vector that relates the hidden states to the estimated streamflow. Hence, for random vectors we could simply obtain larger theoretical limits by increasing the size of the network (i.e., the number of parameters). However, since LSTMs are trained using gradient descent, this relationship is more intricate.
This contribution explores the relationship between the theoretical limit and the network size. In particular, we will look at how increasing the network size in untrained models increases the prediction limit and contrast it to the scaling behavior of trained models.
How to cite: Klotz, D., Baste, S., Loritz, R., Gauch, M., and Kratzert, F.: The relationship between theoretical maximum prediction limits of the LSTM and network size, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10650, https://doi.org/10.5194/egusphere-egu25-10650, 2025.
Effective discharge forecasting is critical in operational hydrology. This study explores novel methods to improve forecast accuracy by combining data assimilation techniques and hydrograph decomposition. Traditional rainfall-runoff modeling, including AI-based approaches, typically simulates the entire discharge signal using a single model. However, runoff is generated by multiple processes with contrasting kinetics, which a single-model approach may fail to capture adequately. This study proposes using hydrograph decomposition to separate baseflow and quickflow components, training specific forecasting models for each component individually, and then merging their outputs to reconstruct the total discharge signal. This approach is expected to enhance forecast accuracy for both floods and droughts, identifying long-term dependencies governing baseflow to improve seasonal low-flow forecasts. Experiments will be conducted using a subset of the CAMELS dataset.
How to cite: Saint Fleur, B. E., Gaume, E., and Akil, N.: Improving AI-based discharge forecasting through hydrograph decomposition and data assimilation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11682, https://doi.org/10.5194/egusphere-egu25-11682, 2025.
Parameter regionalization for ungauged basins remains a critical challenge in hydrological modeling. While traditional approaches rely on physical catchment descriptors or spatial proximity, and recent machine learning applications have focused primarily on direct streamflow prediction, there remains significant potential to leverage machine learning for improved parameter transfer strategies. This study explores novel approaches that combine Long Short-Term Memory (LSTM) networks and Random Forest (RF) models to predict basin similarity and optimize parameter transfer for physically-based hydrologic models. Using case studies from British Columbia's Fraser River Basin and Ontario's Great Lakes region, we test multiple methodologies for integrating deep learning with traditional parameter transfer approaches. Our primary benchmark is established through an exhaustive parameter transfer experiment using the Raven hydrological model, where parameters from each potential donor basin were transferred to every possible receiver basin across 10 independent trials. This benchmark represents the best achievable KGE via parameter transfer methods. Our framework employs a regional LSTM model to capture complex streamflow dynamics and characterize basin similarity, then explores various RF-based approaches to predict optimal donor-receiver basin pairs for parameter transfer. These methods are evaluated against both the exhaustive transfer benchmark and emerging machine learning approaches. Results indicate that thoughtfully combining deep learning and random forest techniques can capture nuanced relationships between basin characteristics and hydrological response similarity, advancing the state-of-the-art in parameter regionalization for ungauged basins while maintaining physical interpretability.
How to cite: Meysami, R., Yu, Q., Tolson, B., Shen, H., and Arabzadeh, R.: Learning Basin Similarity Through Combined Deep Learning and Random Forest Approaches for Improved Parameter Transfer in Ungauged Basins, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14004, https://doi.org/10.5194/egusphere-egu25-14004, 2025.
Physics-informed neural networks are an optimization-based approach for solving differential equations and have the potential to significantly speed up the modelling of complex phenomena, which conventionally is achieved via expensive numerical solvers. We present a Physics-Informed Deep Operator Network (DeepONet) framework for solving two-dimensional shallow water equations with variable bed topography under given boundary and initial conditions. While traditional physics-informed neural networks can solve differential equations on meshless grids using prescribed conditions, they require retraining for each new set of initial and boundary conditions. Our approach uses a DeepONet to learn the underlying solution operator rather than individual solutions, which provides an enhanced generalizability, making the DeepONet a feasible candidate for real world applications. The framework combines the advantages of neural networks with physical laws, effectively handling the complexities of varying bed topography and wet-dry transitions. We demonstrate that our DeepONet approach achieves comparable accuracy to classical numerical methods while significantly reducing inference time. In our modelling experiments we investigate the sensitivity of hyperparameter values and network architecture as well as the potential of introducing an additional data loss, emulating the availability of additional observational data on water levels or inundation extent. This acceleration in computation speed makes the method particularly valuable for time-critical applications such as flood forecasting. The results establish physics-informed DeepONets as a promising alternative to traditional numerical solvers for shallow water systems, offering a balance between computational efficiency and solution accuracy.
How to cite: Keppler, R., Koch, J., and Fensholt, R.: Learning shallow water equations with physics-informed Deep Operator Network (DeepONet), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15171, https://doi.org/10.5194/egusphere-egu25-15171, 2025.
Deep learning models have been successfully applied to simulate streamflow in mountain catchments. While these mostly lumped models have demonstrated the ability to learn processes such as snow accumulation and melt that are crucial for streamflow generation in these regions, they still show deficiencies in simulating streamflow during the melting period. This suggests a misrepresentation of melting dynamics encoded within these models. We hypothesize that the sets of lumped meteorological variables (such as air temperature, precipitation, PET) and static attributes currently used to train and drive these models are not sufficient to describe the melting processes.
To enhance the representation of snow and ice-related processes, we thus propose to incorporate additional data on snow and ice cover, such as Snow Covered Area, Snow Water Equivalent, and glacier mass within the respective basin. We assess (1) how much additional value can be extracted from cryosphere data to improve the representation of cryosphere related processes and (2) how the added value varies across different geographies and catchment types. In a lumped Long Short-Term Memory (LSTM) setup covering a large sample of catchments in different European mountainous regions, we compare different data integration methods with respect to their uncertainty reduction for streamflow simulation and their limitations for model applications.
Our findings provide insights into optimizing model configurations and data usage and offer practical guidance for ultimately improving the accuracy of streamflow simulations in mountainous, snow-influenced regions.
How to cite: Frank, C., Bohl, J. P., Brunner, M., Gauch, M., and Höge, M.: Cryosphere Data and Its Value for Deep Learning Hydrological Simulations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15229, https://doi.org/10.5194/egusphere-egu25-15229, 2025.
Extreme hydrological events such as droughts and floods are expected to become more frequent and severe according to climate change projections, making effective water management very important to mitigate environmental and socio-economic impacts. In this context, advanced hydrological modeling tools are essential for understanding and managing water systems. The Soil and Water Assessment Tool (SWAT+), a process-based and semi-distributed eco-hydrological model, has become very popular for simulating hydrological processes and water management scenarios, especially with its improved water allocation and reservoir modules. At the same time, Graph Neural Networks (GNNs), a deep learning model, have shown potential for modeling complex relationships in networked systems. Both SWAT+'s water allocation module and GNNs use graph-like structures to model water systems. The goal of this study is to systematically compare the structural components of these two approaches and provide insights into potential integration.
Using the Upper Isar River Basin's complex water management system as a case study, we examine how SWAT+ and GNNs can be used to model it. We perform a component-wise analysis, focusing on how these models can represent nodes, edges, and attributes in a networked water management system. While this study focuses on structural rather than performance comparisons, we anticipate that our results will highlight the strengths and limitations of each approach. SWAT+ is expected to excel at incorporating domain-specific knowledge and explicitly representing management actions. GNNs could provide advantages in learning complex patterns from data and faster simulations for larger catchments.
The findings could open the way for hybrid approaches that combine traditional hydrological models' strengths with GNNs' learning capabilities. This could lead to more robust and adaptable water management tools to deal with the growing complexity of hydrological systems caused by climate change and human intervention.
How to cite: Wirthensohn, M., Tuo, Y., and Disse, M.: Graph-Based Representations in Hydrological Modeling: Comparing SWAT+ and Graph Neural Networks for Water Management Systems, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15350, https://doi.org/10.5194/egusphere-egu25-15350, 2025.
Accurate reservoir outflow simulation is crucial for modeling streamflow in reservoir-regulated basins. In this study, we introduce a knowledge-guided Long Short-Term Memory model (KG-LSTM) to simulate the outflow of reservoirs-Fengshuba, Xinfengjiang, and Baipenzhu in the Dongjiang River Basin, China. KG-LSTM is built on the standard hyperparameters-optimized-LSTM and the loss function considering reservoir operation knowledge, while traditional reservoir model level pool scheme (LPS) is used as a benchmark model. Model uncertainty is analyzed using the bootstrap method. We then propose a hybrid approach that combines KG-LSTM with the Three-parameter monthly hydrological Model based on the Proportionality Hypothesis (KG-LSTM-TMPH) for streamflow simulation. The propagation of inflow errors to outflow simulations is studied across the three reservoirs. Results indicate that LSTM-based models greatly outperform LPS in all three reservoirs, with KG-LSTM demonstrating superior capability in capture reservoir outflow dynamics compared to the standard LSTM model. In the multi-year regulated Xinfengjiang Reservoir, KG-LSTM improves Nash-Sutcliffe efficiency (NSE) from 0.59 to 0.64, and reduces root mean squared error (RMSE) from 55.59 m³/s to 54.84 m³/s during the testing period. KG-LSTM shows reduced model uncertainty, decreasing the relative width (RW) from 0.55 to 0.51 in the Xinfengjiang Reservoir and from 0.48 to 0.44 in the Baipenzhu Reservoir, while demonstrating limited change in the Fengshuba Reservoir. For streamflow simulation, KG-LSTM-TMPH performs best across all four stations, achieving NSE values of approximately 0.87, 0.88, 0.91, and 0.92 at Longchuan, Heyuan, Lingxia, and Boluo stations, respectively. In the dry season, KG-LSTM-TMPH demonstrates substantial improvement over LSTM-TMPH, increasing R² by +0.11 and reducing RMSE by -4.22 m³/s at Heyuan station. Inflow errors impact outflow most significantly for the Xinfengjiang Reservoir in April and May, for the Fengshuba Reservoir throughout the year (particularly in April, May, July, and August), and for the Baipenzhu Reservoir primarily in July and August. This study enhances reservoir outflow modeling by integrating reservoir operation knowledge with deep learning. The hybrid KG-LSTM-TMPH approach shows practical potential for streamflow simulation in reservoir-regulated basins, offering valuable applications for water resource management.
How to cite: Wang, D.: A Knowledge-guided LSTM reservoir outflow model and its application to streamflow simulation in reservoir-regulated basins, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16156, https://doi.org/10.5194/egusphere-egu25-16156, 2025.
State-of-the-art deep learning models for streamflow prediction (so-called Entity-Aware models, EA) integrate information about physical catchment properties (static features) with climate forcing data (dynamic features) from multiple catchments simultaneously. However, recent studies challenge the notion that this approach truly leverages generalization ability. We explore this issue by conducting experiments running Long-Short Term Memory (LSTM) networks across multiple temporal and spatial in-sample and out-of-sample setups using the CAMELS-US dataset. We compare LSTMs equipped with static features with ablated variants lacking these features. Our findings reveal that the superior performance of EA models is primarily driven by meteorological data, with negligible contributions by static features, particularly in spatial out-of-sample tests. We conclude that EA models cannot generalize to new locations based on provided physical catchment properties. This suggests that current methods of encoding static feature information in our models may need improvement, and that the quality of static features in the hydrologic domain might be limited. We contextualize our results with observations made in the broader deep learning field, which increasingly grapples with the challenges of (lacking) generalization ability in state-of-the-art deep learning models.
How to cite: Heudorfer, B., Gupta, H., and Loritz, R.: Exploring the Limits of Spatial Generalization Ability in Deep Learning Models for Hydrology, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16414, https://doi.org/10.5194/egusphere-egu25-16414, 2025.
Posters virtual: Tue, 29 Apr, 14:00–15:45 | vPoster spot A
EGU25-15731 | ECS | Posters virtual | VPS9
Data-driven models for streamflow regionalization in Krishna River Basin, IndiaTue, 29 Apr, 14:00–15:45 (CEST) | vPA.4
Predicting streamflow in ungauged basins remains a significant challenge in hydrological studies. In recent years, data-driven models have been shown to outperform traditional physics-based models in streamflow prediction for ungauged catchments. However, few studies have examined the potential of such models for predicting streamflow in ungauged basins within India. This study aims to evaluate the performance of two machine learning models, namely Support Vector Regression (SVR) and Random Forest (RF), alongside two deep learning models, Long Short-Term Memory (LSTM) and Bi-LSTM, in the context of streamflow regionalization within the Krishna River Basin in India. Each prediction model is trained using meteorological variables as input features, with streamflow as the output variable. K-means clustering is employed to group selected catchments (based on data availability) into an optimum number of clusters based on spatial proximity and physical similarity. It is assumed that catchments within a cluster share homogeneous characteristics. Regionalization is achieved by sharing model parameters across catchments within the same cluster. For each cluster, one gauged catchment is designated as the donor catchment, while the others are treated as pseudo-ungauged. Each proposed model is trained and tested using the meteorological inputs and streamflow data available at the gauged donor catchment. The trained model for each cluster is then transferred to the remaining receptor catchments within the cluster, where the meteorological variables corresponding to each ungauged catchment are used as inputs. The performance of the models in ungauged catchments is rigorously evaluated by comparing the simulated streamflow against observed streamflow using metrics such as Nash-Sutcliffe Efficiency (NSE), Root Mean Squared Error (RMSE), Coefficient of Determination (R²), and Percentage Bias (PBIAS). This study highlights the advantages of utilizing data-driven methods for streamflow prediction in both gauged and ungauged basins, particularly due to their ability to capture complex, non-linear relationships between meteorological inputs and streamflow generation. The findings of this study are expected to be instrumental in water resources planning and management, flood assessment, and the design of hydraulic structures in the Krishna River Basin.
How to cite: Kaur, S. and Chavan, S.: Data-driven models for streamflow regionalization in Krishna River Basin, India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15731, https://doi.org/10.5194/egusphere-egu25-15731, 2025.
EGU25-18458 | ECS | Posters virtual | VPS9
Can the catchment features influence the performance of the conceptual hydrological and deep learning models? A study using large sample hydrologic dataTue, 29 Apr, 14:00–15:45 (CEST) | vPA.5
The prior literature on hydrologic model performance is dispersed, encompassing a small number of catchments, different methodology, and rarely linking the results to specific catchment characteristics. This study addresses these constraints by systematically attributing model performance to catchment variables in 671 US catchments, providing a formal framework for determining the best models for specific conditions. Daily streamflow estimation was performed using eight process-based (PB) models and three deep learning (DL) models, with performance measured using the Nash-Sutcliffe Efficiency (NSE). The PB models were tested with a variety of optimization techniques, and the most effective approach for each model was chosen based on the number of catchments that exceeded a predetermined performance threshold. Four models were selected as the top performers based on three performance metrics. Further analyses, such as Classification and Regression Tree (CART) and SHAPley, were used to correlate model performance with catchment variables across all models.
The results showed that PB models (GR4J, HBV, and SACSMA) performed well in catchments with low to medium aridity and a high Q/P ratio, indicating quick hydrologic responses. In contrast, the LSTM-based DL model performed well in medium to high aridity regions but had limits in catchments with rapid precipitation responses and low sand percentages. These findings provide a thorough understanding of the links between model performance and catchment descriptors.
Keywords: Process-based models, Deep learning model, CART analysis, SHAPley analysis, catchment characteristics.
How to cite: sourya, D. A., manikanta, V., and rathinasamy, M.: Can the catchment features influence the performance of the conceptual hydrological and deep learning models? A study using large sample hydrologic data , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18458, https://doi.org/10.5194/egusphere-egu25-18458, 2025.
EGU25-18599 | ECS | Posters virtual | VPS9
Downscaling MODIS ET using deep learningTue, 29 Apr, 14:00–15:45 (CEST) | vPA.6
Knowing evapotranspiration (ET) accurately at fine spatial scales is very important. This would improve understanding hydrological processes and contribute to the advancement of water resource management. In this study, we set a framework based on deep learning to downscale Terra Net Evapotranspiration Gap-Filled 8-Day Global 500m dataset, developed and managed by NASA's Earth Observing System. This approach resulted in a scale enhancement of 20 times. The U-Net architecture was used for this purpose. It incorporated MODIS Land Cover Type 1 (LC Type 1) as an auxiliary variable. This was done to account for land cover changes. The study covered a diverse region that encompasses latitudes 28° to 32°N and longitudes 74° to 78°E. A synthetic design of experiments was utilized to systematically generate and evaluate training data, this ensures robust model performance and reliable downscaling outcomes across the heterogeneous terrain of the study area. Model training, validation, and testing were conducted using the 2001–2014 dataset, 2015–2018, and 2019–2023 dataset, respectively. The model showed excellent performance on the testing dataset. The average PSNR was 34.35 dB and the mean SSIM was 0.8517. The U-Net module effectively downscale and enhance the spatial resolution of ET data. The results show ET's spatial and structural features are well preserved. This study shows how deep learning improves climate data spatial resolution. It provides reliable local hydrological and agricultural resources.
How to cite: Jha, S. K. and Gupta, V.: Downscaling MODIS ET using deep learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18599, https://doi.org/10.5194/egusphere-egu25-18599, 2025.
EGU25-580 | ECS | Posters virtual | VPS9
Streamflow simulations using regionalized Long Short-Term Memory (LSTM) neural network models in contrasting climatic conditionsTue, 29 Apr, 14:00–15:45 (CEST) | vPA.26
We investigate the potential of using Long Short-Term Memory (LSTM) neural networks for estimating streamflow in (sub)tropical catchments under contrasting hydroclimatic regimes (semi-arid and humid). We have used 176 Brazilian catchments with at least 30 years of streamflow data and LSTM models with 16 static catchment attributes as input features. We tested different LSTM model configurations to assess their sensitivity to varying input sequence lengths (lookbacks). The primary objective was to explore the hydrological insights offered by LSTM-based streamflow models and compare their performance with the traditional GR4J hydrological model. With this design, we aim to address two research questions: (i) Does the performance of LSTM models depend on catchments' hydroclimatic characteristics? (ii) How effective are LSTM-based models for streamflow simulation in tropical and subtropical catchments under semi-arid and humid conditions? We adopt two modeling approaches: (1) regionalized models trained on catchments within the same hydroclimatic regime and (2) a composite model trained on a heterogeneous sample combining both arid and humid catchments. The findings reveal distinct sensitivities of LSTM models to hydroclimatic conditions. LSTM models exhibit higher sensitivity to the length of input sequences (lookbacks) in humid catchments, with longer sequences yielding better performance. This is attributed to the dominant hydrological processes in humid regions, which are influenced by long-term memory effects such as soil moisture and groundwater storage. Conversely, this sensitivity is not observed in semi-arid catchments, where streamflow dynamics are primarily driven by short-term precipitation events and exhibit less dependence on long-term hydrological processes. Furthermore, the composite model, which combines semi-arid and humid catchments, demonstrates a decrease in performance for semi-arid catchments. This suggests that adding catchments with contrasting hydroclimatic characteristics introduces heterogeneity in the dataset, potentially reducing the model's ability to capture the specific dynamics of semi-arid catchments. Overall, the regionalized LSTM models outperformed the GR4J model in both semi-arid and humid regimes, particularly in humid catchments. Approximately 87% of humid catchments and 50% of semi-arid catchments achieved Kling-Gupta Efficiency (KGE) values above 0.60 during the testing phase of the regionalized LSTM models. These results highlight the potential of LSTM networks for streamflow regionalization, especially in humid regions where long-term hydrological memory plays a critical role. The study underscores the strengths and limitations of LSTM models in tropical and subtropical catchments with contrasting hydroclimatic regimes. The findings suggest that LSTM models could serve as valuable tools for regional hydrological applications, aiding local and regional decision-making processes. Additionally, the results emphasize the importance of tailoring LSTM model configurations to the specific hydrological characteristics of catchments, particularly the choice of input sequence length, to maximize model performance.
How to cite: Maria de Andrade, J., Nóbrega, R., Ribeiro Neto, A., Rico-Ramirez, M., Coxon, G., and Montenegro, S.: Streamflow simulations using regionalized Long Short-Term Memory (LSTM) neural network models in contrasting climatic conditions, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-580, https://doi.org/10.5194/egusphere-egu25-580, 2025.
