Deciphering Multi-Scale Impacts of Urban Morphology on Flooding for Climate-Resilient Planning: An Explainable AI Approach

Urban flooding is shaped by urban morphology, but the inferred relationships can change with the spatial units used to represent flooding and urban form. Existing studies often report results for a selected spatial configuration, leaving unclear whether predictive performance and identified dominant predictors remain robust when analytical scale, grid placement, and spatially separated validation are varied. Using the 22 May 2020 Guangzhou storm as an event-specific case, this study evaluates the robustness of multi-scale morphology–flood associations to these spatial analytical choices. The analysis integrated 119 unique official waterlogging locations with 67 geocoded social-media locations. After removing four cross-source matches within 100 m, 182 unique observations were used to construct a kernel density response surface and examine 1–5 km analytical grids. Spatial autocorrelation, repeated nested geographic cross-validation of XGBoost, out-of-fold SHAP attribution, and accumulated local effects (ALEs) were used to quantify scale-dependent patterns. Global Moran’s I increased from 0.125 at 1 km to 0.524 at 4 km and decreased to 0.479 at 5 km (all permutation p < 0.001). Mean spatially validated R2 ranged from 0.483 to 0.610, with the highest R2 and lowest RMSE at 4 km, although residual spatial autocorrelation remained. Road density was the largest individual SHAP contributor at every scale (35.54–40.89%). ALE indicated broad positive associations for road density, building density, and impervious surface ratio and a negative association for elevation, without supporting universal sharp thresholds. These event-specific results show that analytical scale and grid placement should be reported explicitly when morphology-based evidence is used for flood screening and climate-resilient planning.