Flood susceptibility mapping in arid urban areas using SHAP-enhanced stacked ensemble learning: A case study of...

Flooding is an escalating hazard in arid and rapidly urbanizing environments such as Jeddah, Saudi Arabia, where the lack of historical flood records and sparse monitoring systems challenge effective risk prediction. To address this gap, this study aims to develop an accurate and interpretable flood susceptibility-mapping framework tailored to data-scarce urban settings. The research integrates a stacked ensemble model-comprising machine learning: XGBoost, CatBoost, and Histogram-based Gradient Boosting (HGB)-with SHapley Additive exPlanations (SHAP) to enhance prediction accuracy and model transparency. Random Forest was excluded from the final model stack due to inferior classification performance. A diverse set of geospatial inputs, including digital elevation model, slope, flow direction, Curve Number, topographic indices, and LULC (from ESRI Sentinel-2) were used as predictors. Furthermore, 92 and 198 flooded and non-flooded points were used for model validation. The model achieved strong predictive performance (AUC = 0.92, Accuracy = 0.82) on the validation set. In the absence of official flood records, model outputs were intersected with road network data to identify 395 road points in highly susceptible zones. Although these points do not represent a formal validation dataset-due to the general lack of detailed flood event records in the region, particularly in relation to infrastructure-they provide a valuable proxy for identifying flood-prone road segments. SHAP explainability analysis revealed that TRI, TPI, and distance to rivers were the most globally influential features, while Curve Number and LULC were key drivers of high-risk predictions. The model mapped 139 km² (8.7 %) of the area as very high flood susceptibility and 325 km² (20.3 %) as high susceptibility, outperforming individual learners. These results confirm that stacked ensemble learning, paired with explainable AI and creative validation strategies, can produce reliable flood susceptibility maps even in data-constrained contexts. This framework offers a transferable and scalable solution for flood risk assessment in similar arid and urbanizing environments.