Can Storm Morphology Be Used To Improve Severe Hail Detection? A Comparison Using A Convolutional Neural Network Applied To Dual-Polarization Radar

Severe convective storms (SCS) result in considerable financial losses on an annual basis across the globe. In the United States and Europe, the financial impact of SCS can reach billions of dollars annually. Hail is a significant contributor to these losses. In 2022, France experienced its most costly year on record, with hail-related damages reaching $5 billion. The effects of these storms were visible on damaged vehicles and roofs, broken solar panels, and devastated agricultural fields.

The real-time detection of hailstorms is of paramount importance for the mitigation of potential damages and the adaptation of disaster responses in affected communities. Furthermore, an effective hail detection technique is invaluable for reconstructing past events and quantifying risk, particularly for risk carriers (insurers). Dual-polarization radars continue to represent the foundation for the development of precise real-time hail detection. The existing hail detection algorithms employ either vertical profiles of the reflectivity field or polarimetric variables to perform a prediction. Such techniques exploit either the radar data in a vertical column of air or the pixel values of different dual-polarization variables. They are, by design, column- or pixel-based methods. Nevertheless, numerous studies have demonstrated the existence of specific spatial structures within radar data that may be associated with the occurrence and the formation of hail (Kumjian et. al, 2014 ; Dennis and Kumjian, 2017). The majority of current hail detection techniques are unable to extract information embedded in the storm's morphology, and thus fail to identify such potential signatures. Recently, studies using deep learning, and particularly convolutional neural networks (CNNs), have sparked interest in capturing relevant information in the parent storm's structure that would help discriminate hail storms from other storms (Wang et al., 2022; Forcadell et al., 2024).

In this study, we employ convolutional neural networks (CNNs) applied to dual-polarization radar data in order to detect severe hail on the ground (≥ 2 cm). To this end, a quality-controlled dataset of radar samples is constructed in conjunction with verified hail reports on the ground from the European Severe Weather Database (ESWD) in France. Additionally, a dataset of negative instances, comprising instances of rain or small hail (up to 2 cm), was constructed for the purposes of the study. A shallow CNN is trained for the task of severe hail detection and is demonstrated to outperform radar-based hail proxies (POH, POSH, MESH) on multiple metrics. The application of a feature importance study and attribution methods enables the retrieval of valuable insights into the CNN's decision-making process and its capacity to utilise the storm's morphology. The reliability of the CNN's predictions is also investigated and discussed.

References:

Dennis, E. J., and M. R. Kumjian, 2017: The Impact of Vertical Wind Shear on Hail Growth in Simulated Supercells. J. Atmos. Sci., 74, 641–663, doi: 10.1175/JAS-D-16-0066.1.

Forcadell, V., C. Augros, O. Caumont, K. Dedieu, M. Ouradou, C. David, J. Figueras i Ventura, O. Laurantin and H. Al-Sakka, 2024 : Severe Hail Detection with C-band Dual-polarisation Radars using Convolutional Neural Networks, EGUsphere, doi : 10.5194/egusphere-2024-1336

Kumjian, M. R., A. P. Khain, N. Benmoshe, E. Ilotoviz, A. V. Ryzhkov, and V. T. J. Phillips, 2014: The Anatomy and Physics of ZDR Columns: Investigating a Polarimetric Radar Signature with a Spectral Bin Microphysical Model. J. Appl. Meteor. Climatol., 53, 1820–1843, doi: 10.1175/JAMC-D-13-0354.1.

Wang, P., W. Lv, C. Wang and J. Hou, 2018: Hail Storms Recognition Based on Convolutional Neural Network, 2018 13th World Congress on Intelligent Control and Automation (WCICA), Changsha, China, 2018, pp. 1703-1708, doi: 10.1109/WCICA.2018.8630701.