Flood Inundation Area Prediction Under Climate Change Scenarios by Integrating Hydrological and Hydraulic Models with a Hybrid Deep Learning Framework

Flood inundation modeling in low-gradient monsoon floodplains requires a physically consistent representation of rainfall–runoff–inundation processes. This study develops a hybrid modeling framework that integrates a coupled hydrological–hydraulic model (HEC-HMS–HEC-RAS) with a deep learning-based LSTM–U-Net surrogate to represent temporal hydrological memory and spatial inundation patterns. The framework is applied to the Upper Songkhram River Basin in northeastern Thailand, a storage-dominated floodplain strongly influenced by monsoon hydrology. The hydrological model demonstrated strong validation performance (NSE = 0.896, KGE = 0.827, R2 = 0.909), while hydraulic simulations showed high spatial agreement with satellite-derived inundation maps (F1 = 0.876, Kappa = 0.873). Trained on hydraulically simulated discharge–inundation pairs, the LSTM–U-Net model successfully reproduced two-dimensional flood patterns across independent flood events (mean F1 = 0.838, IoU = 0.721), with prediction errors mainly occurring along shallow floodplain margins. Future projections under CMIP6 SSP2-4.5 and SSP5-8.5 indicate clear changes in flood-season discharge, with stronger increases under SSP5-8.5, whereas maximum inundation extent shows more moderate changes (≤21%), reflecting nonlinear floodplain response in low-gradient systems. The proposed framework preserves hydrological–hydraulic consistency while supporting future flood inundation projection, climate-informed flood risk assessment, and adaptation planning.