LightningCast v1 vs v2
Version 1 of the NOAA/CIMSS LightningCast model uses GOES-R ABI images to predict the probability of lightning in the next 60 minutes at any given location. It is being transitioned to NOAA/NESDIS operations. LightningCast v2 has been developed and is being evaluated at NOAA’s Hazardous Weather Testbed. Version 2 incorporates MRMS Reflectivity at -10oC, which is a well-known product used to help determine the ice content in convection. We’ve found that version 2 improves short-term lightning predictions across the contiguous U.S. (CONUS), with a very small reduction in performance in regions outside CONUS. Here are some examples.
Dallas / Forth Worth Metro
On April 29, the Dallas / Fort Worth metro region was socked in with dense mid- and high-level cloud cover. There was very little contrast or texture in the cloud tops in the ABI imagery. Thus, LightningCast v1 had low probabilities until convective cloud features began to poke up from the thick cloud canopy and the cloud-top brightness temperature began to cool. However, LightningCast v2 had moderate-to-high probabilities much sooner, owing to the Reflectivity -10oC predictor.
In two convective areas southwest of Forth Worth and over Fort Worth proper, LightningCast v2 provided 17 minutes and 5 minutes of additional lead time to the first flashes detected by GLM, respectively, compared to version 1. When plotted as a line graph over TCU’s Amon G. Carter Stadium, the version 2 probability shows a clear uptick 5 minutes before version 1. Later, version 2 remains higher than version 1 during another burst of lightning.
Pennsylvania
Meanwhile, Pennsylvania was also socked in with dense cloud cover on the same day. Some shallow convection remained hidden to version 1, whereas version 2 had higher probabilities over the electrified region. While this was a difficult case due to the marginal nature of the convection, version 2 still provided more reliable guidance.
Mississippi Valley
In the central Mississippi Valley, LightningCast version 2 correctly had higher probabilities in southern Illinois, western Kentucky, eastern Kansas, and eastern Arkansas compared to version 1. It correctly had lower probabilities in western Indiana, as well. All are areas of adequate radar coverage.
We’ve demonstrated that the radar predictor adds a lot of value over CONUS, while the model has in general learned to rely on satellite inputs where radar coverage is absent. Version 2 isn’t better in every instance, but overall, it provides superior guidance.