Hail poses a significant and growing economic threat to Switzerland. In recent years, both large and small hail have caused substantial damage, with large hail leading to record property losses reported by insurance companies. However, small hail is also not without peril. It can facilitate flooding by clogging drainage systems when it occurs in large quantities. Additionally, it may cause devastating damage to crops, especially in combination with strong winds. Consequently, to better understand the hail hazard, a comprehensive detection of hail of all sizes is needed. Radar-based hail detection in the complex Swiss orography faces significant challenges regarding retrieval quality and visibility due to beam-shielding and ground clutter effects. The recent upgrade of the national weather radar network (Rad4Alp) to polarimetric C-Band Doppler radars has brought significant improvements in both data quality and availability, providing a rich dataset for analysis. Here, we present a new machine learning-based hail size discrimination algorithm using polarimetric radar data trained on a unique dataset of high-density crowdsourced hail reports. The algorithm uses an object-based approach; Individual storms are detected, and a wide range of predictors is extracted from newly created polarimetric radar composites for each storm. The resulting dataset is used to train a random forest for hail size detection, which significantly outperforms the currently operational hail size discrimination algorithm MESHS (maximum expected severe hail size). We further assess the importance of various radar signatures for hail size classification through feature importance analysis of the trained model. The most impactful predictors include both conventionally used quantities, such as echotops and maximum reflectivities, and lesser-known ones like detected ice hail column height and hail differential reflectivity (HDR). Finally, we demonstrate the potential of the generated object-based dataset for hail size nowcasting, indicating its broader utility beyond discrimination.
Aregger et al. (Fri,) studied this question.