Spatiotemporal assessment and structural attribution of urban flood resilience using interpretable machine learning: evidence from the Beijing–Tianjin–Hebei Urban Agglomeration

The Beijing-Tianjin-Hebei Urban Agglomeration (BTHUA) has experienced rapid urbanization and recurrent pluvial flooding under intensifying extreme rainfall. This study assesses flood resilience (FR) from 2012 to 2023 at a 3 km x 3 km grid scale using a Resistance-Recoverability-Adaptability (R-R-A) framework. We construct composite indicators for natural conditions, land-use structure, infrastructure provision, public-service capacity, and socio-economic development, and combine Entropy Weight Method (EWM) and Criteria Importance Through Intercriteria Correlation (CRITIC) weights through a least-squares objective-weight combination. XGBoost is then coupled with SHapley Additive exPlanations (SHAP) to examine the structural contribution and nonlinear sensitivity of candidate indicators to the constructed FR index. Results show that FR increased overall but remained spatially uneven, with higher resilience in northwestern ecological zones and core urban areas and lower resilience in the central-southern plains. Topographic wetness index (TWI), rainfall intensity (RI), public service facility density (PSD), landscape diversity (SHDI), grey infrastructure index (L_GRI), population density (PD), and development index (DI) show heterogeneous associations with FR. The SHAP-based decomposition indicates that topographic constraints remain dominant, while public-service capacity becomes increasingly important in dense urban areas. The findings support differentiated governance strategies for ecological conservation zones, plain detention zones, and dense urban zones.