Toward real-time high-resolution fluvial flood forecasting: A robust surrogate approach based on overland flow models

Timely flood prediction is critical for mitigating risks under the growing impacts of climate change. Traditional physics-based hydrodynamic models, while effective at capturing flood dynamics, are limited by high computational demands, restricting real-time applicability. This study presents a hybrid framework that integrates machine learning (ML) with physics-based modeling to enable efficient real-time flood forecasting. Physics-based simulations provide detailed inundation information, while ML models serve as fast surrogate predictors. Applied to the Cambodia floodplain — a region highly prone to seasonal flooding — the surrogate models were trained on outputs from TELEMAC simulations. Explainable AI was employed to interpret model decision-making. Results show that the hybrid approach achieves substantial computational efficiency while preserving accuracy. The best surrogate attained R $=$ 0.97 and KGE $=$ 0.91, reducing simulation time by over 70-fold compared with TELEMAC. Incorporating geographic features such as latitude and longitude further enhanced predictive skill, particularly in flat floodplain settings.