Rapid Generation of Urban Flood Susceptibility Maps Using Machine Learning Under Different Rainfall Scenarios

Urban floods increasingly pose threats to transportation systems, residential areas, and public safety, which has created an urgent need for rapid urban flood susceptibility mapping (UFSM) under varying rainfall conditions. In this study, five tree-based machine learning models, namely, random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost), and gradient boosting decision tree (GBDT) models, are evaluated for rapid UFSM generation under 18 rainfall scenarios with different return periods and rainfall patterns. Model performance is assessed on the basis of accuracy (ACC), F1 score (F1), precision (PRE), and recall (REC). Among the five models, the RF model performs best overall, with ACC, F1, PRE, and REC values ranging from 0.97 to 0.99, and is the most robust under the 50- and 100-year rainfall scenarios. The GBDT and CatBoost models perform poorly under low-return-period scenarios, whereas the XGBoost and LightGBM models perform better under high-return-period conditions but remain inferior to the RF model. Feature importance analysis reveals elevation, topographic position index, and curvature as the main controlling factors. These findings provide an efficient framework for rapid UFSM generation and offer useful support for urban flood risk management and planning.