Under the combined pressures of intensified extreme rainfall and accelerating impervious urban expansion, pluvial flooding has emerged as a dominant threat to urban safety and sustainability. Conventional flood-susceptibility models have faced challenges in handling highly sparse, long-tailed target distributions and in providing physical interpretability, which has limited the fine-scale delineation of flood-prone cells and the development of differentiated mitigation strategies. To address this issue, an integrated GeoAI-based framework was developed to systematically links urban surface characteristics with socio-hydrological processes for advancing flood-risk governance. The proposed framework synthesizes 25 natural and socio-economic variables to holistically capture flood-generation mechanisms across diverse city contexts. Through a two-stage feature distillation process, the ten most critical drivers shaping flood susceptibility in each city were identified. These drives underpin a novel zero-inflated convolutional self-attention network (ZI-Geo-CNN), which generated high-resolution susceptibility maps for six major Chinese cities with exceptional accuracy ( and . Post‑hoc analysis using Shapley Additive Explanations (SHAP) quantified each driver’s relative contribution, revealing universal controls alongside economy–infrastructure couplings. Based on shared and differentiated patterns of factor importance across cities, this study compared dominant patterns across cities and discussed several indicative adaptation directions. Overall, the framework breaks the accuracy–interpretability trade-off for sparse, long-tailed flood data and furnishes a replicable GeoAI workflow that can be applied consistently across cities through city-specific training, calibration, and interpretation, thereby providing an evidence-informed basis for resilient drainage planning under non-stationary climates.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号