Floods are a devastating environmental hazard that requires rigorous spatial susceptibility assessments to guide effective risk mitigation. This study presents an interpretable flood susceptibility mapping framework that couples Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP). This approach offers precise insights into how individual features influence probabilities, providing a level of interpretability beyond what XGBoost’s overall variable importance can achieve. The flood extent was derived from Sentinel-1 SAR imagery, and a set of conditioning factors was used for model construction in the Karkheh Basin, Iran. The XGBoost model demonstrated high predictive performance, with an Area Under the Curve (AUC) of 0.89. SHAP analysis identified slope, altitude, and the normalized difference vegetation index (NDVI) as the most influential factors in determining flood susceptibility. SHAP dependence plots elucidated the functional relationships and interactions between the top predictors and flood susceptibility. They quantitatively confirmed that flood susceptibility increases at lower altitudes and decreases on steeper slopes. In contrast, the stream power index (SPI) exhibited a highly conditional, non-linear relationship, where its effect on flood risk depended on interactions with other topographic factors. This study demonstrates that integrating SHAP with machine learning models in flood risk assessments can enhance stakeholder confidence and promote broader model adoption. By providing transparent and post-hoc explanations of model predictions, the SHAP framework effectively bridges the gap between complex ‘black-box’ models and practical, actionable disaster management planning.

Keywords