An integrated framework combining interpretability and exploratory dependency analysis for flood susceptibility assessment in the Beijing–Tianjin–Hebei region

Flood susceptibility in rapidly urbanizing regions is shaped by terrain, infrastructure, land cover and climate variability, yet modelled spatial associations remain difficult to interpret. We assessed urban inundation susceptibility in the Beijing–Tianjin–Hebei (BTH) region using 895 Amap inundation locations, 895 sampled background locations and 17 conditioning factors. After feature screening, eight machine-learning algorithms were compared. Random Forest was selected by validation ROC-AUC (0.9476), closely followed by Extra Trees (0.9469). Its independent-test AUC was 0.9447 (95% CI 0.9218–0.9645), with accuracy of 0.8743, precision of 0.8807, recall of 0.8659 and F1 score of 0.8732. Lower scores occurred in the northwestern mountains and higher scores in the central, southern and southeastern plains, reflecting the distribution of reported urban inundation. SHAP and generalized additive models identified land cover and road-related variables as leading predictors and described their nonlinear contributions. LiNGAM summarised joint dependencies among predictors and spatially cross-fitted scores. All 15 residual-independence tests rejected independence, limiting the graph to exploratory interpretation. Future maps varied five precipitation predictors under three SSP pathways while holding 2020 land cover and other static factors fixed. The four 30-year climate windows are illustrative snapshots; 94.72–100% of future cells fell outside the central training range (1st–99th percentiles) for at least one continuous predictor. The maps estimate relative susceptibility to reported urban inundation, not annual flood probabilities or flood risk.