Improvements to LightningCast to Capture Lightning Initiation

LightningCast is a deep-learning AI model (a U-Net) that predicts the probability of lightning within the next 60 minutes using GOES-R ABI data at one timestamp as inputs. Core inputs to the model are the 0.64-µm reflectance (red portion of the visible, 0.5-km resolution), the 1.6-µm reflectance (shortwave infrared, 1-km resolution), the 10.3-µm brightness temperature (longwave infrared, 2-km resolution), and the 12.3-µm brightness temperature (longwave infrared, 2-km resolution) with the truth dataset derived from the gridded Geostationary Lightning Mapper (GLM) flash extent density (FED) product.

Although LightningCast has been shown to provide users with approximately 20 minutes of lead time, it still may miss lightning initiation events, which are important to predict for public safety. To improve the prediction of these events, the truth dataset is expanded to include ground-based Earth Networks Total Lightning Network (ENTLN) data as ENTLN data may capture lightning not observed with GLM. ENTLN and GLM data are combined in two different ways: 1) a centroid-gridded product of the flash locations is produced and 2) ENTLN flash locations are incorporated into the GLM FED product. Results for the two methods at different resolutions are compared with each other as well as with LightningCast on a subset of data (e.g., 30-60 days) to inform us the best approach to apply to the entire dataset.