Rainfall-Stratified Explainable Machine Learning for Quantifying Nonlinear Drivers of Waterlogging Severity: A Case Study in Shanghai, China

Urban flooding poses escalating threats to high-density cities, yet the nonlinear mechanisms linking rainfall characteristics and urban morphology to waterlogging severity remain poorly understood. This study proposes a rainfall-stratified explainable machine learning framework to distinguish deep from shallow inundation at the block scale, taking Shanghai as a case study. Four models (XGBoost, random forest, SVM, and logistic regression) were compared via nested cross-validation and Bayesian optimization, with XGBoost identified as the optimal model. Three physically distinct rainfall dimensions and multi-dimensional urban morphological indicators were incorporated as predictive features. SHAP-based attribution and PDP were employed to unveil the driving mechanisms behind inundation severity, characterizing scenario-dependent shifts in driver dominance and nonlinear threshold effects. Urban morphology primarily governs spatial risk under non-extreme rainfall, with building shape coefficient (BSC) remaining the primary driver overall. Meteorologically, waterlogging severity surges beyond critical thresholds for maximum hourly rainfall (>18.40 mm/h) and total volume (>139 mm), while duration exhibits an inverted U-shaped response. Morphologically, a high BSC (>0.39 m−1) is consistently associated with elevated deep inundation probability, whereas higher SDBV (>54,155 m3) and greater DR (>582 m) are associated with a severity-attenuating effect. These findings provide threshold-driven insights for integrating morphological resilience into urban renewal and sustainable flood adaptation strategies in high-density metropolises.