A multi-model framework for interpreting hydrometeorological flood drivers in complex orographic and data scarce regions

Khyber Pakhtunkhwa (KPK) is one of Pakistan’s most flood vulnerable provinces, with complex topography and highly data scarce region, yet no province scale machine learning (ML) based flood susceptibility map exists for it. This study addresses this critical gap by developing an explainable stacked ensemble ML framework integrating twenty geo-environmental conditioning factors for the data scarce region and a multi model framework for interpreting flood drivers in similar other regions. Seven individual classifiers and a legacy stacked ensemble benchmark were evaluated using both random stratified and geographically grouped 100-km block cross-validation. XGBoost was selected before map production because it achieved the highest mean ROC-AUC under the spatial-block criterion (0.930 ± 0.031) and the stacked ensemble ranked second (0.926 ± 0.032). The random-CV XGBoost AUC was 0.955, demonstrating that random resampling gives a more optimistic estimate than geographically grouped evaluation. The Very High class covers 7,856.63 km², and 7.88% of the mapped area has a model score above 0.75. SHapley Additive exPlanations (SHAP) identified Land Use Land Cover, Elevation, and Annual Rainfall as dominant flood drivers. The resulting continuous probability map and five class susceptibility maps pinpoint the Peshawar Valley and Kabul to Swat River confluences as Very High risk zones. This is the first province level, SHAP interpreted and ensemble based flood susceptibility study for KPK, providing a transparent and reproducible decision support tool for disaster risk management.