Abstract

Flood risk in semi-arid regions is intensifying due to climate change; however, reliable prediction remains a fundamental challenge in catchments with severe hydrometric data scarcity (station density < 0.2 per 1000 km²). To address this gap, we developed and validated a novel hybrid modeling framework that integrates a physically based hydrological model (Hydrologic Engineering Center - Hydrologic Modeling System, HEC-HMS) with interpretable, Bayesian-optimised machine learning techniques (Random Forest, eXtreme Gradient Boosting, XGBoost). The SHapley Additive exPlanations (SHAP) method was employed to interpret and identify the key drivers of flooding. SHAP analysis revealed that antecedent precipitation and soil permeability are the primary controlling factors in flood occurrence. The validated framework demonstrates robust performance, achieving excellent accuracy in flood event classification (Precision-Recall- Area Under the Curve (PR-AUC) = 0.98) and reducing false alarm rates by 87.5% compared to the standalone physical model. This study provides a practical, validated tool that enhances prediction reliability and delivers actionable insights for climate-resilient flood risk management in data-scarce environments.