Ensemble machine learning and deep learning framework for flood susceptibility mapping in the transboundary Rapti River Basin

The diverse topography, varying land use, and inconsistent data across political borders render flood-susceptibility mapping especially difficult in transboundary river basins. Heterogeneous topography, variable land use, and discontinuous data across political boundaries make flood-susceptibility mapping particularly challenging in transboundary river basins. In order to identify flood-prone areas in the transboundary Rapti River Basin, this work offers an integrated, interpretable ensemble-based method that combines the Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) models. Open-access geospatial datasets were used to develop twelve flood-conditioning parameters that represent topographic, hydrologic, and land-surface properties. Model performance was assessed using spatially blocked train-test splits and evaluated with receiver-operating characteristic area under the curve, precision-recall metrics, and related indices. The highest accuracy was achieved by XGBoost with an AUC of 0.93, followed by RF with 0.91 and LSTM with 0.88. To ensure interpretability and to overcome limitations of gain-based importance, feature contributions were quantified using SHapley Additive exPlanations (SHAP). The SHAP analysis revealed that slope, elevation, and drainage density are the dominant controls on flood susceptibility, whereas NDVI and land-use/land-cover exert more localized influence in low-relief and urbanized areas. Overall, the proposed workflow enhances transparency, reproducibility, and physical interpretability in flood-susceptibility modeling and provides a robust, scalable tool for data-driven flood-risk management and policy planning in transboundary watersheds.