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

Siyu Li ^* Peter Knippertz Michael Kunz Jannik Wilhelm Julian Quinting

• Institute for Meteorology and Climate Research, Faculty of Physics, Karlsruhe Institute of Technology (KIT), Karlsruhe, Baden-Wurttemberg, Germany

This study investigates the use of a convolutional neural network (CNN) to predict the daily hail-affected area in Germany estimated from radar-based hail footprints during the period 2005 to 2019. The machine learning (ML) model leverages a combination of 18 different thermodynamic and dynamic, convection-related parameters derived from ERA5 reanalyses. From these, seven parameters have been identified to be the most important predictors. The ML model clearly outperforms two basic reference forecasts based on climatology and persistence. The Heidke Skill Score (HSS) for large hail-affected areas, for example, reaches values of up to 0.66. Lowest prediction skill is obtained for days with lower values of the product of convective available potential energy with bulk wind shear (CAPESHEAR) or when the hailstorms are isolated. Sensitivity analyses with gradient-weighted class activation mapping contribute to better understand the performance of the ML model predictions. Based on those analyses, CAPESHEAR stands out as the primary predictor. Given the very low computational resources needed for the predictions once the model has been trained, this study offers promising results for daily hail predictions using ML models based solely on convective environmental parameters and opens avenues for further improvement and refinement.