William Koshak (ST11) is co-author on an article recently accepted for publication in Earth and Space Science, with title “Classification of GLM Flashes Using Random Forests”. The Geostationary Lightning Mapper (GLM) detects ground and cloud flashes, but there is not yet available a GLM product that distinguishes between the two flash types.
This work employs a Machine Learning technique (i.e., a Random Forest algorithm) in an attempt to determine flash type. It examines the spatial and temporal characteristics of the flash optical emissions to make the flash type distinction. The results show that the Random Forest model distinguishes flash type moderately well, with 83% probability of detection, 71% percent correct, 46% false alarm rate, 29% false alarm ratio, and 65% critical success index.
There is interest in continuing this work effort to further improve these performance numbers, because space-based detection of ground flashes has direct application to many lightning-caused impacts (e.g., wildfire ignition, human injury/death, property/crop damage, power outages, insurance claims).
Read the paper at: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021EA001861.
Feature importance for each lightning flash feature in the random forests model. MGA = maximum group area and MNEG = maximum number of events per group.