Flood susceptibility assessment in Seti Gandaki river basin using an integrated gradients approach

Accurate flood susceptibility mapping in mountainous regions remains challenging due to complex terrain, nonlinear interactions, and variable environmental drivers. This study introduces a novel, high-resolution framework for flood risk assessment in the Seti Gandaki Basin, Nepal, integrating advanced deep learning with multi-dimensional geospatial analysis. The methodology follows a systematic five-phase process: (1) rigorous selection of geomorphological, hydrological, and environmental features, (2) geospatial preprocessing and standardization, (3) statistically-informed sampling using feature-specific threshold analysis, (4) development and training of a deep neural network with dropout and batch normalization, and (5) spatial prediction and comprehensive validation. The model, trained on 175,000 spatially stratified flood and non-flood samples, achieved outstanding predictive performance (Accuracy: 94.6%, Precision: 92.8%, Recall: 95.1%, ROC-AUC: 98.9%). Integrated Gradients feature attribution identified slope, TWI, rainfall, and distance to rivers as dominant flood drivers, while NDVI-LULC analysis quantified the protective role of vegetation. High-resolution flood susceptibility and uncertainty maps were produced, and a boundary-masked comparative analysis with the Analytical Hierarchy Process (AHP) revealed strong methodological convergence (91.6% continuous agreement within $${\pm }$$ 0.5 tolerance; 85.7% categorical agreement within $${\pm }$$ 1 class). By combining data-driven insights with expert knowledge, this framework provides a robust, interpretable, and transferable decision-support tool for flood risk management in complex Himalayan watersheds, offering a scalable approach for disaster risk reduction and climate adaptation in similar mountainous environments.