ProbSevere LightningCast: A Deep-Learning Model for Satellite-Based Lightning Nowcasting

• Published in: Weather and forecasting Vol. 37; no. 7; pp. 1239 - 1257
• Main Authors: Cintineo, John L., Pavolonis, Michael J., Sieglaff, Justin M.
• Format: Journal Article
• Language: English
• Published: Boston American Meteorological Society 01.07.2022
• Subjects: Artificial neural networks; Computer architecture; Deep learning; Domains; Geostationary satellites; Image segmentation; Lead time; Lightning; Lightning activity; Lightning strikes; Machine learning; Mathematical models; Meteorological conditions; Meteorological satellites; Modelling; Neural networks; Nowcasting; Performance degradation; Satellite imagery; Satellites; Short wave radiation; Storms; Synchronous satellites; Windows (intervals)
• ISSN: 0882-8156; 1520-0434
• DOI: 10.1175/WAF-D-22-0019.1

Abstract Lightning strikes pose a hazard to human life and property, and can be difficult to forecast in a timely manner. In this study, a satellite-based machine learning model was developed to provide objective, short-term, location-specific probabilistic guidance for next-hour lightning activity. Using a convolutional neural network architecture designed for semantic segmentation, the model was trained using GOES-16 visible, shortwave infrared, and longwave infrared bands from the Advanced Baseline Imager (ABI). Next-hour GOES-16 Geostationary Lightning Mapper data were used as the truth or target data. The model, known as LightningCast, was trained over the GOES-16 ABI contiguous United States (CONUS) domain. However, the model is shown to generalize to GOES-16 full disk regions that are outside of the CONUS. LightningCast provides predictions for developing and advecting storms, regardless of solar illumination and meteorological conditions. LightningCast, which frequently provides 20 min or more of lead time to new lightning activity, learned salient features consistent with the scientific understanding of the relationships between lightning and satellite imagery interpretation. We also demonstrate that despite being trained on data from a single geostationary satellite domain (GOES-East), the model can be applied to other satellites (e.g., GOES-West) with comparable specifications and without substantial degradation in performance. LightningCast objectively transforms large volumes of satellite imagery into objective, actionable information. Potential application areas are also highlighted. Significance Statement The outcome of this research is a model that spatially forecasts lightning occurrence in a 0–60-min time window, using only images of clouds from the GOES-R Advanced Baseline Imager. This model has the potential to provide early alerts for developing and approaching hazardous conditions.