An interpretable multi-hazard machine learning model for county-level loss assessment of tropical cyclones

Reliable loss assessment of tropical cyclones (TCs) is critical for effective disaster emergency response. Existing methods often overlook the combined impacts of multiple hazards associated with TCs, such as wind, rainfall, storm surge, waves, and floods, which can decrease loss estimation accuracy. To address this issue, a novel assessment framework is proposed that integrates these multi-hazard effects to enhance disaster loss modeling. This framework begins by identifying multi-hazard features of TCs, including maximum gust wind (3s), total rainfall, daily rainfall, hourly rainfall, surge heights, significant wave heights, and daily runoff. Using a dataset of 1,341 county-level records, four machine learning algorithms—Categorical Boosting (CatBoost), Transformer, Backpropagation Neural Network (BPNN), and Support Vector Machine (SVM)—are trained and optimized. The best-performing model is applied to assess the impact of feature variables and training samples. Additionally, shapley additive explanations (SHAP) are employed to interpret the model, providing insights into feature importance and relationships among hazards. Results indicate that CatBoost outperforms other algorithms, achieving an accuracy of 0.8196. Incorporating all feature variables results in a maximum performance improvement of 19.06% compared to using single, double, or triple hazards. The model demonstrates strong applicability across coastal and inland regions at the national scale, maintaining an accuracy above 0.79. By integrating SHAP analysis, this approach enhances model interpretability, offering valuable insights into factor contributions and inter-hazard relationships. The proposed framework improves the reliability of loss assessments and addresses the limitations of machine learning "black boxes," supporting more informed and effective disaster response strategies.