Regional Disaster Loss Prediction Under Tropical Cyclone Hazard to Support Real-time Risk Forecasting: A Data-driven Approach

Abstract

Rapid prediction of high-probability disaster hotspots during a tropical cyclone is crucial to facilitate proactive mitigation measures. In this study, a data-driven model based on a convolutional neural network (CNN) is developed for Zhejiang Province, China, to provide rapid and fine-resolution regional prediction of tropical cyclone induced losses. The model considers both hazard intensity measures and environmental characteristics as explanatory predictors, and outputs county-level disaster losses. In addition, through a unique design of an intermediate layer in the CNN architecture, the model can also produce grid-level loss predictions with 1km resolution. Such a gridded outcome (1 km2) can further inform disaster hotspots across the entire region of Zhejiang Province (approximately 105,000 km2). The CNN model is trained and calibrated using loss records of 9 severe historical tropical cyclones that impacted Zhejiang during the period of 2012 to 2019. The proposed model, with promising accuracy and resolution, shows evident advantages in time efficiency and computational cost for regional loss predictions compared to physics-based simulations.