Explaining the Role of Lightning Data in Hail Nowcasting

To advance the understanding of the value of lightning data within emerging deep learning applications, specifically focusing on enhancing hail prediction capabilities, we conducted a study to explore the significance of various lightning characteristics, including flash rate, flash/source type, and polarity in discerning between hail-producing and non-hail-producing thunderstorms for short-term nowcasts (0-1 hour horizons). A deep learning model was developed to achieve these objectives by integrating data from three distinct sources. The first data source was a set of Vaisala lightning observations derived from the National Lightning Detection Network (NLDN). The second data source were fields from the Warn on Forecast System (WoFS), representing traditional Numerical Weather Prediction (NWP) information. The third data source was a set of observations obtained from GridRad, providing essential spatiotemporal insights and approximate hail labels using the maximum estimated size of hail (MESH) product. These observations were aligned spatially and temporally to allow the training of a 3D U-Net over a southern Great Plains domain at a 3km grid scale and at 5-min intervals. To shed light on the model's decision-making process, eXplainable Artificial Intelligence (XAI) techniques were employed to determine local and global feature importance of the inputs . These techniques enabled the identification of key contributors to the model's predictions, offering insights into the relative influence of lightning data and other variables. Particularly, we find Inter/Intra-Cloud (IC) flashes to be a powerful predictor of hail that contributes to model skill even in the presence of radar observations. Our findings not only underscore the importance of lightning characteristics in hail prediction but also provide interpretable insights into the model's decision rationale.