Abstract: Long-term wind observation data from 142 meteorological stations, historical apple wind disaster data, apple growth stages data, etc. were used in this study. Firstly, the optimal machine learning model has been built to extend long-term extreme wind speed. Then the thresholds of different levels of wind disasters in different apple growth stages were determined. The spatial-temporal distribution characteristics and a comprehensive risk assessment of apple wind disasters were carried out. The results showed that the random forest model had the highest accuracy and showed good performance in extending the extreme wind speed time series which is suitable for Hebei province. The thresholds of apple wind disasters were extreme wind speed≥9.1m·s⁻¹ in flowering to fruit set, and extreme wind speed≥7.9m·s⁻¹ in fruit development to mature. Furthermore, 4 levels of wind disasters in different apple growth stages were determined and the validation were in good agreement with the historical records. The overall trends of apple wind disasters were first decreasing and then increasing. In the northwest Hebei province and eastern Cangzhou, the frequencies of wind disasters were higher than in other regions. The high and the severe comprehensive risk areas of apple wind disaster were scattered which accounted for nearly 20% of Hebei and mainly in eastern Zhangjiakou, southeast Hengshui, middle Hengshui, and eastern Shijiazhuang.

Key words: Apple, Wind disasters, Machine learning, Risk assessment