Machine Learning Based Investigation of the Variables Affecting Summertime Lightning over Southeastern South America

Machine learning (ML) is used to explore complex relationships between meteorological variables that affect summertime lightning occurrence over southeastern South America (SA) (20°-40° S, 40°-60° W). Lightning occurrence is examined on clean (VIIRS daily AOT < 25 % percentile) and polluted (VIIRS daily AOT > 75 % percentile) days using reanalysis data from ERA-5 and IMERG. Meteorological variables that are uncorrelated are then used via ML to predict lightning occurrence and extreme lightning events. Preliminary results suggest accuracy rates > 95 % for most cases including cases with a mixture of land-ocean and clean-polluted events. CAPE and rain rate were found to be the best predictors of lightning in the model. The model performance over this region is better than that achieved when a similar model was used over the south-central U.S. The causes of this difference will be discussed in the context of possible differences in the electrification production mechanism. Finally, additional analysis will be conducted by using the MERRA-2 product that provides typing information. The goal of this research will be to determine the impact of aerosol type (if any) on the intensity of deep convection and lightning.