An accurate understanding of river dynamics is critical for preserving ecological diversity, supporting economic activities, and addressing environmental challenges amidst climate change. To tackle these challenges, advanced modeling tools are essential to grasp historical river dynamics and thus project future scenarios effectively. This study introduces a hybrid approach that integrates physics-based models and deep learning to reconstruct long-term water levels at high frequency, applied to the Seine Estuary in Normandy, France. Deep learning models refine the outcomes of this hydrodynamic model, by integrating diverse observational and meteorological datasets. This multi-process framework effectively captures spatial and temporal variability across hydrodynamic zones, using Bidirectional Long Short-Term Memory (BiLSTM) networks to enhance predictive accuracy. The study underscores the critical role of multiple processes—hydrodynamic modeling, observational constraints, and meteorological influences—in shaping reconstruction quality. Results demonstrate that the hybrid approach outperforms standalone physics-based and traditional deep learning models, reducing RMSE by approximately 58% compared to the physics-based model alone. During major storm events since 2017, this approach has consistently improved accuracy by 50–65% in RMSE, demonstrating its robustness under extreme conditions. Reconstruction accuracy is primarily driven by the relevance of hydrodynamic physics-based outputs and observational data to the target station, while meteorological influences playing a secondary role due to their coarse resolution. These findings highlight the potential of multi-process hybrid models for broader hydrological applications, enabling more reliable reconstructions and projections of river dynamics.

An accurate understanding of river dynamics is critical to preserve ecological diversity, supporting economic activities, and addressing environmental challenges amidst climate change. To tackle these challenges, advanced modeling tools are essential to grasp historical river dynamics and thus project future scenarios effectively. This study introduces a hybrid approach that integrates physics-based models and deep learning to reconstruct long-term water level, applied for the Seine Estuary in Normandy, France. Deep learning models initially enhance the well-established physics-based model by refining its input data to provide more reliable hydrodynamic simulations as a baseline. The outcomes of this physics-based model, combined with diverse data sources, then feeds the relative neural network to enhance overall performance. This successive refinement optimizes the utilization of complex spatial and temporal variability observed across extensive modeling areas, leveraging Bidirectional Long Short-Term Memory neural networks to enhance the quality of both models. The study thus highlights the explainability and significance of datasets in driving the output accuracy across various hydrodynamic zones. The hybrid approach outperforms standalone physics-based and traditional deep learning models, achieving roughly 58% reduction in RMSE compared to the physics-based model. Insights reveal that reconstruction accuracy mainly relies on physics-based outcomes and observational data, with their relevance to the target station. During major storm events since 2017, this hybrid model has consistently enhanced accuracy by 50-65% in RMSE, highlighting its superior performance under challenging conditions. This approach holds promise for application to other rivers and broader hydrological challenges.