Integrating explainable ensemble machine learning with hydrodynamic modeling to assess tropical cyclone-induced flood susceptibility

Tropical cyclone (TC)-induced floods have posed catastrophic risks to coastal regions, demanding high-resolution susceptibility mapping for effective risk management. This study proposes a framework that integrates an explainable ensemble machine learning (ML) model with a hydrodynamic model for flood susceptibility mapping. The hydrodynamic model first simulates 78 historical TC events in the Pearl River Delta during 1960-2022, then the simulation results are used to train and test five representative ML models, including Logistic Regression, Multilayer Perceptron, Random Forest, Extreme Gradient Boosting, and Light Gradient Boosting Machine. Subsequently, the five ML algorithms are ensembled to improve the overall performance. The proposed framework can efficiently map flood susceptibility at a 30 m spatial resolution, achieving an excellent predictive performance with an Area Under the Curve (AUC) value of 0.908. The importance of base learners and 17 controlling features is evaluated through Shapley Additive Explanations (SHAP) value analysis and model structure interpretation, which enhances the model’s interpretability and transparency. Among the 17 features, elevation is the most influential, contributing 21.7% to the model’s predictions. It exhibits a negative correlation with the SHAP main effects, with a threshold value of 11.04 m. This model offers accurate, high-resolution susceptibility information, and could be applied to other cities facing the similar TC-induced flooding hazards.