Flood susceptibility mapping is critical for urban planning in semi-arid regions exposed to high-intensity seasonal rainfall and rapid urban growth. Conventional approaches often involve subjective weighting and limited interpretability. This study proposes a GIS-based weight-level integration approach linking interpretable machine learning with the Analytic Hierarchy Process (AHP). Four tree-based models were trained using standard and spatial cross-validation; SHapley Additive exPlanations (SHAP) quantified factor contributions, enabling recalibration of expert-derived weights. CatBoost demonstrated the strongest generalization, achieving 85.19% standard and 73.30% spatial accuracy. ROC-AUC analysis showed that AHP + SHAP and standalone AHP attained the highest discriminative performance (AUC = 0.964), outperforming conventional map-level integration (AUC = 0.958) and standalone CatBoost (AUC = 0.859). The framework proved robust to weight perturbations, unlike map-level integration which produces polarized patterns. Weight-level integration yields balanced spatial distributions and reduces false alarms, providing a transparent scheme for data-constrained urban contexts to support informed environmental decision-making.