Machine learning model for the prediction of hail in Germany

A CNN-based model successfully predicts the daily hail-affected areas in Germany combining reanalysis and radar data

Hailstorms present significant risks in Germany (Figure 1), causing substantial economic damage and necessitating accurate forecasts for timely warnings. In this study IMKTRO scientists from the research groups “Atmospheric Risks”, “Atmospheric Dynamics” and “Meteorological Data Science” investigate the potential of machine learning, more specifically convolutional neural networks (CNNs), in predicting the daily hail-affected area in Germany originally identified from radar data for the period 2005 to 2019.

Figure 1: Accumulated hail-days derived from radar reflectivity of the DWD radar network over Germany from 2005 to 2019

The study initially evaluates 18 thermodynamic and dynamic convection-related parameters as potential predictors for the hail-affected area which are derived from the ERA5 reanalysis data set. Through feature selection techniques seven key predictors are identified. Among these, a parameter including information regarding the energy available to convection and the vertical wind shear (CAPESHEAR) emerges as the most influential factor for hail prediction (Figure 2). “The machine learning model is evaluated against traditional climatology- and persistence-based forecasts, demonstrating superior performance particularly for large hail events” says Siyu Li, lead author of the study.

To explore the model’s decision-making process for its prediction, a method rooted in methods of explainable AI is applied. The results underscore the dominant role of CAPESHEAR in influencing predictions. However, the model exhibits reduced skill in scenarios where CAPESHEAR values are comparably low or when hailstorms occur in isolated events. These findings suggest that while the CNN approach provides a significant improvement over traditional methods, additional regional refinements and severity classifications could further enhance predictive capability.

“One of the key advantages of this ML-based approach is its computational efficiency, making it a promising candidate for operational forecasting applications”, states Siyu Li. “Future research directions could include refining regional perspectives and integrating severity metrics to enhance the model’s applicability in real-world forecasting scenarios.”

This research highlights the potential of machine learning techniques in meteorology and showcases how AI can contribute to improving predictions and understanding of weather hazards. By harnessing environmental convective parameters, this method represents a step forward in providing more accurate and timely hailstorm forecasts.

You can find the article in: https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2025.1527391/full