AbstractAbstract
Lightning endangers people and infrastructure and is a major source of wildfire ignitions in the western United States. This study builds upon a satellite-based machine learning model, LightningCast, by incorporating a radar-based predictor of near-term lightning probability. While numerous studies have merged satellite and radar predictors for lightning prediction, few have investigated the impact of using both predictors in tandem from these two remote sensing sources. Here, we focus on reflectivity observations at −10°C (Ref10), which have been widely used by forecasters to nowcast lightning occurrence. Using ablation experiments, we investigated the change in performance when using Ref10 1) in regions with radar coverage, 2) in regions without radar coverage, and 3) for first-flash events. Satellite predictors were provided by the
GOES-16
Advanced Baseline Imager (ABI), while the Ref10 was computed from the Multi-Radar Multi-Sensor (MRMS) system. The target was created from the Geostationary Lightning Mapper data aboard
GOES-16
. Overall, we found that including the Ref10 predictor improves predictions significantly where radar coverage is valid while not significantly reducing performance in regions outside of radar coverage. Furthermore, the lead time to first-flash events is slightly improved overall but significantly improved in certain situations. From the ablation experiments, we determined that ABI predictors improve lead time to first-flash events on the order of 5–10 min, compared to a Ref10-only experiment. The combined ABI+Ref10 experiment achieved average lead times of 20–30 min for first-flash events at the most-skillful forecast probability thresholds of 25%–40%, the highest of all tested experiments. These results indicate that it is possible for models to leverage the strengths of multimodal observations for short-term lightning prediction.
Significance StatementThis research quantifies the ability of artificial intelligence methods to properly leverage the advantages of both satellite and radar observations to improve lightning nowcasting in developing thunderstorms (i.e., before the first lightning flash occurs) and in ongoing thunderstorms. This provides forecasters with better probabilistic guidance for interests such as airports, outdoor event managers, mariners, and the general public.
