Public-service exposure-oriented coastal flood susceptibility assessment and priority zone identification in the Pearl River Estuary

Flooding in rapidly urbanizing estuarine lowlands threatens not only exposed land but also the continuity of transport, healthcare, and education services. Conventional coastal flood susceptibility (CFS) mapping identifies flood-prone environments but rarely distinguishes whether these areas contain critical service receptors. This study develops an explainable machine-learning framework coupling CFS assessment with public-service exposure and priority-zone identification in the Pearl River Estuary. A 250 m grid dataset was constructed from flood records, environmental variables, and exposure receptors including population, roads, hospitals, clinics and schools. Four tree-based models were evaluated using spatial block validation, and CatBoost achieved the best overall performance, with an AUC of 0.870, PR-AUC of 0.878, F1-score of 0.785 and MCC of 0.570. High and Very high CFS zones accounted for 43.1% of the study area, while the observed flood occurrence ratio increased from 8.5% to 94.6% across susceptibility classes. SHapley Additive exPlanations (SHAP) identified extreme rainfall, hydrological proximity, wetland-related lowlands, built-up surfaces and coastal meteorological forcing as key drivers. Exposure analysis revealed a clear susceptibility–exposure mismatch. Refined priority zones covered only 346.56 km² but contained 1.49 million people, 6,055 km of roads, 32 hospitals, 137 clinics and 331 schools. These findings shift CFS mapping toward service-oriented spatial prioritization for coastal flood adaptation.