A machine learning model for the prediction of hail-affected area in Germany

Hailstorms pose significant risks in Germany, calling for accurate forecasts and warnings. This study explores the application of a convolutional neural network (CNN) to predict daily hail-affected areas using radar-based hail footprints from 2005 to 2019. The ML model utilizes 18 thermodynamic and dynamic convection-related parameters derived from ERA5 reanalysis data. Feature selection identifies seven key predictors, with a particular emphasis on the convective available potential energy and bulk wind shear (CAPESHEAR). Model performance is assessed against climatology- and persistence-based reference forecasts, and sensitivity analyses using gradient-weighted class activation mapping (Grad-CAM) are conducted to interpret the predictions. The CNN model significantly outperforms the reference forecasts, achieving a Heidke Skill Score (HSS) of up to 0.66 for large hail-affected areas. However, lower predictive skill is observed on days with weak CAPESHEAR values or when hailstorms are isolated. Sensitivity analysis highlights CAPESHEAR as the dominant predictor influencing model decisions. These findings demonstrate the potential of ML-based hail prediction using only convective environmental parameters. Given its low computational demand once trained, this approach offers a promising tool for operational forecasting. It would be desirable to extend this approach to a more regional perspective and to include information on severity.