A Verification of Hail Proxies with Insurance Data

Information about the hail threat of a thunderstorm is typically limited to rather indirect data from remote sensing or reanalysis. Ground observations of maximum hail diameters provide a more accurate assessment but often suffer from other problems, such as limited or non-uniform coverage. Despite these shortcomings, the above data sources are commonly used for nowcasting of hail producing storms and in hail climatologies. However, only very few studies have actually compared the skill of these different proxies. One reason for this is the lack of ground truth data which could be used to verify whether damaging hail was falling. This is because of the described issues with hail reports, and the fact that insurance datasets, which could provide a more reliable confirmation of hail damage, are usually not made available for research.

To fill this gap, this study uses a 5-year dataset of crop damage claims of a German agricultural insurance. These insurance claims cover large parts of Germany and most accurately reflect hail damage during the growing season of crops from May to August, which is also the time of year with the strongest thunderstorm activity. The damage claims are used to verify and compare common proxies for hail, which include the radar-based (I) column maximum reflectivity and (II) hail tracks using the more refined TRACE3D tracking algorithm, (III) lightning density, (IV) European Severe Weather Database hail reports, (V) observed overshooting tops from geostationary satellites, and (VI) microwave imager hail signatures from polar-orbiting satellites. Their skill in predicting damaging hail is assessed by categorical verification with probability of detection, false-alarm rate, and Heidke Skill Score, including a sensitivity analysis to varying thresholds.

Preliminary results based on 30 events in 2014 indicate that none of the proxies alone can predict hail damage with high accuracy. However, all of them show at least some skill, except for the microwave imager. The radar-based predictors show the largest skill on average. If these findings can be confirmed over more cases, while also including null cases, this would support the use of proxies I-V for hail climatologies but discourage nowcasts of hail for an individual thunderstorm with one of the proxies alone.