Abstract:

The development of a classification product for hail occurrence probability and hail size can improve the accuracy of hail identification. Based on 92 hail observation records from Henan Province in 2022, together with radar data and sounding data, nine characteristic parameters were selected: composite reflectivity（CR）, height difference between 55dBZ base reflectivity and 0°C level （H₀）, height difference between 45dBZ base reflectivity and −20°C level （H₋₂₀）, vertical integrated liquid （VIL）, vertical integrated liquid density（VILD）, echo top height （ET）, differential reflectivity （ZDR）, specific differential phase（KDP） and correlation coefficient （CC）. By integrating fuzzy logic with a Semi−supervised Learning algorithm based on a weak K−nearest neighbor classifier (referred to as the FL−ST−KNN model), the probability of hail occurrence and hail size grades were identified, thereby further reducing the impacts of hail on buildings, agricultural production, and human safety. The results show that the FL−ST−KNN model achieved an accuracy of 83% on the test set (20% of the dataset). Its precision was 80% and its recall was 83%, indicating high reliability in identifying majority−class samples. Moreover, the F1−score approached the excellent threshold of 80%, demonstrating that the proposed model performs well in identifying both hail occurrence probability and hail size.

Key words: Dual?polarization radar, Hail probability, Fuzzy logic, Semi?supervised machine learning