Urban waterlogging is an escalating challenge under rapid urbanization and climate change, yet accurate spatial prediction remains hindered by nonlinear drivers and spatial heterogeneity. This study proposes a spatially interpretable machine learning framework by integrating remote sensing and geospatial data with hybrid modeling. Using recorded waterlogging locations in Beijing, we constructed a balanced dataset with topographic, hydrological, land cover, and proximity-based predictors. Four machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and eXtreme Gradient Boosting (XGBoost) —were evaluated, with XGBoost achieving the best classification performance (area under the curve (AUC) = 0. 913 ± 0. 055). To enhance spatial interpretability, two hybrid strategies were further developed: (1) XGBoostMGWR, in which XGBoost serves as the primary predictor and MGWR corrects its spatially structured residuals, thereby improving spatial explanatory power; and (2) MGWRXGBoost, where MGWR first models spatially varying effects and XGBoost subsequently fits the residuals to refine predictive performance. Results from spatially blocked five-fold cross-validation show that MGWRXGBoost provides the best probabilistic accuracy (Brier = 0. 289 ± 0. 039) and the highest area under the precision recall (PR-AUC) (0. 576), with substantially higher specificity (0. 734) and a spatially stable local R 2 pattern; therefore, it was selected for final risk mapping. The proposed framework enables high-resolution, spatially explicit risk mapping and offers practical support for drainage planning, green infrastructure prioritization, and adaptive flood governance. Beyond Beijing, this approach shows strong potential for improving resilience in other data-scarce urban environments facing intensifying flood risks.
Yi Tang (2026) studied this question.