SPATIALLY VALIDATED AND EXPLAINABLE HYDRO-GEOAI STACKING ENSEMBLE FOR FLASH-FLOOD SUSCEPTIBILITY MAPPING IN KHUZDAR DISTRICT, BALOCHISTAN, PAKISTAN

Abstract

Khuzdar District in Balochistan, Pakistan, is acutely vulnerable to flash floods due to its rugged mountainous relief, intermontane valley morphology, ephemeral drainage channels, and extreme episodic precipitation. This study presents a spatially validated, explainable Hydro-GeoAI stacking ensemble for mapping flash-flood susceptibility across Khuzdar District. Sixteen conditioning factors encompassing topography, hydrology, satellite remote sensing indices, land use/land cover, lithology, and soil characteristics were compiled from validated regional geospatial archives. To prevent optimistic spatial autocorrelation bias and feature leakage, a spatial block Group KFold cross-validation scheme was employed. Three baseline algorithms, Random Forest, XGBoost, and CatBoost, were benchmarked against a logistic regression meta-learner stacking ensemble. Model performance was evaluated via ROC-AUC, PR-AUC, F1-score, Matthews Correlation Coefficient (MCC), and Brier Score. Feature influences and decision mechanisms were unraveled using Shapley Additive Explanations (SHAP). Random Forest achieved the highest discriminative performance (accuracy: 98.66%, F1-score: 98.69%, ROC-AUC: 0.9993), while CatBoost attained the lowest Brier Score (0.0116). Across all models, Height Above the Nearest Drainage (HAND), event rainfall, antecedent rainfall, and Topographic Wetness Index (TWI) emerged as primary environmental drivers. The resulting susceptibility map provides a scientific screening tool for transportation corridors, settlements, and civil infrastructure planning in data-scarce arid mountain basins.