A Tropical Cyclone Risk Prediction Framework using Flood Susceptibility and Tree-based Machine Learning Models: County-level Direct Economic Loss Prediction in Guangdong Province
Jian Yang,Sixiao Chen,Yanan Tang,Ping Lu,Sen Lin,Zhongdong Duan,Jinping Ou
DOI: https://doi.org/10.1016/j.ijdrr.2024.104955
IF: 4.842
2024-11-06
International Journal of Disaster Risk Reduction
Abstract:Tropical cyclones (TCs), characterized by strong winds, heavy rainfall, storm surges, and flooding, have caused significant economic losses and fatalities in coastal regions globally. However, existing TC risk prediction frameworks often fail to adequately account for the direct impacts of flooding. In this study, we propose integrating flood susceptibility, a critical component of flood early warning systems, into TC risk prediction frameworks. Focusing on Guangdong Province, we employ four tree-based machine learning (ML) models (random forest, extreme gradient boosting, light gradient boosting machine, and categorical boosting) to predict county-level direct economic losses (DELs) based on flood susceptibility, oceanographic-meteorological data, and vulnerability data. These ML models are trained and tested on a dataset of 896 samples, achieving high prediction accuracies, with Pearson correlation coefficients exceeding 0.81 between the predicted and observed DEL values. Among the four models, the light gradient boosting machine demonstrates the best performance, achieving the highest values of R and R 2 , and the lowest values of MSE, MAE, and MedAE. The integration of flood susceptibility is validated by comparing it with traditional methods that directly incorporate environmental factors. Furthermore, the proposed TC risk prediction framework is applied to forecast the impacts of Super Typhoon Mangkhut in 2018, illustrating its potential ability for "real-time" TC risk assessments. These "real-time" DEL predictions not only estimate potential losses but also facilitate timely interventions, thereby enhancing the practical value of the model for disaster prevention and response.
geosciences, multidisciplinary,meteorology & atmospheric sciences,water resources