MACHINE-LEARNING-BASED INVESTIGATION OF THE VARIABLES AFFECTING LIGHTNING OVER THE U.S. SOUTHERN GREAT PLAINS AND THE SOUTHEASTERN SOUTH AMERICA IN SUMMER

Abstract

Lightning is a key component of convective systems and an indicator of storm intensity and atmospheric instability. This study applies a machine learning (ML) framework to investigate the environmental and cloud macro-physical and micro-physical factors controlling lightning activity across two climatically distinct regions: the U.S. Southern Great Plains (SGP) and the Southeastern South America (SESA). By integrating ground-based, satellite, and reanalysis datasets, it aims to quantify lightning predictability and the underlying physical processes driving it.Over the SGP, continuous Atmospheric Radiation Measurement (ARM) observations allow high-resolution analysis of convective cloud structure. The random forest (RF) model identifies cloud thickness, rain rate, and convective available potential energy (CAPE) as dominant predictors of lightning occurrence, confirming the roles of thermodynamic instability and cloud vertical development. A positive relationship between CAPE and the intra-cloud (IC) flash fraction suggests that stronger updrafts elevate charge centers and favor IC discharges. Extending the same framework to the SESA, the RF model achieves high overall accuracy but reduced Recall, or true positive rate (TPR), for intense lightning events, highlighting challenges posed by class imbalance and environmental diversity. Rain rate and CAPE remain the leading predictors. These findings are broadly consistent with those from the Amazon Basin (Allen et al., 2024), where aerosol and microphysical effects on lightning were also found to depend strongly on the thermodynamic background. In particular, enhancements in cloud depth, IWC, and lightning activity under polluted conditions were most pronounced at low-to-moderate CAPE levels, but diminished once strong convection was already established. This suggests that certain environmental or aerosol-related influences become evident only under specific CAPE regimes, a behavior similarly reflected in the SESA results. Analyses of CloudSat observations reveal that higher CAPE corresponds to greater ice water content (IWC) and higher cloud-top and centroid altitudes, supporting a link between instability, glaciation depth, and lightning frequency. The positive relationship between CAPE and the IC flash fraction, first identified over the SGP, is further validated in the SESA, where both IC fraction and IC height increase with CAPE but plateau at high instability. Collectively, these results demonstrate that combining ML with vertically resolved satellite observations provides a powerful framework for understanding and predicting lightning variability across regions and advancing its representation in weather and climate models.