High-accuracy reconstruction of water levels through integrating physics-based and data-driven models: application to the Seine estuary

An accurate understanding of river dynamics is vital for preserving ecosystems, supporting economic activities, and tackling climate challenges. This study presents a hybrid modelling approach that combines physics-based models with deep learning to reconstruct high-frequency, long-term water levels in the Seine Estuary, Normandy, France. Deep learning – specifically bidirectional long short-term memory networks – enhances hydrodynamic model outputs by integrating diverse observational and meteorological data. The multi-process framework effectively captures spatial and temporal variability across hydrodynamic zones. Results show the hybrid model outperforms standalone physics-based and deep learning models, reducing RMSE by approximately 58% and improving accuracy by 50–65% during major storm events since 2017. Reconstruction quality is mainly influenced by the relevance of hydrodynamic outputs and observational data, while meteorological inputs play a lesser role due to their coarse resolution. These findings demonstrate the robustness and potential of hybrid models for accurate, reliable reconstructions and projections in complex river systems.