A rapid flood inundation integrated surrogate model based on a CNN-LSTM-ATT deep learning framework

The Pajiang River Basin, China The Pajiang River Basin is characterized by scarce hydrological data and frequent flash flood events, requiring efficient simulation for timely prediction and effective mitigation. To address these challenges, this study develops a rapid flood inundation integrated surrogate model (ISM) that incorporates physical-based hydrological-hydrodynamic models with deep learning algorithms. The ISM includes an improved TOPMODEL (3S-TOPMODEL), the LISFLOOD-FP hydrodynamic model, and a CNN-LSTM-Attention (CNN-LSTM-ATT) deep learning framework. The 3S-TOPMODEL simulates interactions between surface water, soil water, and groundwater. It demonstrates better performance than the traditional TOPMODEL in calibration and validation, particularly for peak flows. The hydrological-hydrodynamic module of the ISM, which couples 3S-TOPMODEL with LISFLOOD-FP, achieves high accuracy in predicting flood depth and extent in the Pajiang River Basin. The CNN-LSTM-ATT algorithm is embedded within the ISM, accelerating flood simulations by reducing computational time from approximately one hour to just 5.4–6.79 s. Compared with baseline models (CNN-LSTM and 3D-CNN), CNN-LSTM-ATT provides higher accuracy, improved spatial generalization, and greater robustness, maintaining superior performance even with limited training data. Overall, the ISM provides both mechanistic hydrological understanding and a high-speed, reliable tool for real-time flood forecasting, supporting effective flood risk management and decision-making in the Pajiang River Basin. • A surrogate hydrological-hydrodynamic model is developed using deep learning methods. • The enhanced model includes runoff generation with surface, soil, and groundwater interactions. • The integrated surrogate model performs well with less data than the baseline models.