Urban pluvial flood susceptibility assessment for the central Lin’an District in Hangzhou, China, using the GNNWLR model with SHAP interpretability

The study focuses on the central urban area of Lin’an District, Hangzhou, China. This work employs a Geographically Neural Network Weighted Logistic Regression (GNNWLR) model integrated with SHapley Additive exPlanations (SHAP) to evaluate urban pluvial flood susceptibility. Twelve predictors derived from topography, land cover, the Normalized Difference Vegetation Index (NDVI), and population density were integrated into a unified grid dataset. To address the class imbalance caused by limited flood samples, a hybrid resampling approach combining Generative Adversarial Network (GAN)–based oversampling and spatial undersampling was applied. Model performance was compared with Geographically Weighted Logistic Regression (GWLR) and XGBoost to examine improvements in representing spatial non-stationarity and nonlinear relationships. The GNNWLR model achieved superior accuracy (Accuracy = 0.847, AUC = 0.928, Recall = 0.886), outperforming GWLR and XGBoost. Impervious Surface Fraction (ISF), Road Network Density (RND), and NDVI were identified as key drivers, with NDVI acting as a mitigating factor. Terrain indicators such as Height Above Nearest Drainage (HAND) and Topographic Wetness Index (TWI) modulated spatial flood patterns, where lower HAND and higher TWI corresponded to higher risk. High-susceptibility zones were concentrated in central and northeastern areas, providing interpretable, spatially explicit insights for sustainable flood management and sponge city planning. While demonstrated for Lin’an District, the proposed framework is conceptually transferable to other urban areas with appropriate local calibration. • A spatially adaptive, interpretable framework integrating GNNWLR and SHAP was developed for urban pluvial flood susceptibility assessment. • The model achieved high predictive accuracy (Accuracy = 0.865, AUC = 0.931, Recall = 0.886). • Spatial heterogeneity, nonlinear relationships captured; class imbalance solved via GAN and spatial undersampling. • ISF and RND were identified as dominant drivers in built-up areas, while vegetation mitigated flood risk in hilly zones. • The results provide spatially explicit insights into flood mechanisms, supporting sustainable urban planning and flood management.