Ensemble learning-driven hail identification model using FY-4A satellite data

This study utilises manual hail records from ground meteorological stations and hail suppression operation points in Guizhou Province during March to August from 2021 to 2023. It analyses the spatio-temporal characteristics of hail distribution in Guizhou and constructs a hail identification method based on an ensemble of machine learning algorithms. First, analysis of hail data reveals a pronounced west-high, east-low distribution pattern across Guizhou Province, with small-sized hail (diameter ≤5 mm) predominating in recent years. Subsequently, FY-4A satellite data corresponding to the hail observation periods were acquired. Through automated grid file matching, spatio-temporal alignment, and regional cropping, a hail identification dataset was constructed. Finally, an ensemble learning model was constructed using five machine learning algorithms—Random Forest, Logistic Regression, Support Vector Classification, Decision Tree, and XGBoost—to model and evaluate hail Identification performance in Guizhou. Experimental results demonstrate that the Stacking ensemble learning approach effectively identifies hail weather processes. Compared to a single optimal machine learning algorithm, it achieves improvements of approximately 4.3%, 6.3%, and 12.3% in the ACC, POD, and CSI, respectively, while reducing the FAR by approximately 10.2%. This approach balances low false alarm rates with high Identification accuracy. These findings provide valuable reference for hail forecasting and early warning systems suppression in Guizhou.