Using Deep Learning to Predict Cloud-to-Ground Lightning in the Western United States

Lightning is a major source of summer wildfire ignition in the western United States. However, future projections of lightning are uncertain since lightning is not directly simulated by global climate models (GCMs). Previous studies have projected future lightning occurrence by using statistical relationships with meteorological variables that are simulated by GCMs. However, these approaches rely on parameterizing lightning from local conditions without awareness of large-scale weather patterns, which can augment successful prediction of small-scale weather phenomena. Here, we employ convolutional neural networks (CNNs), which are a type of automated image classification, to predict the occurrence of lightning across the western United States based on large-scale meteorological variables for the period of 1995 to 2022. Individually trained CNN models at each grid cell show high skill (AUC >0.9) at predicting lightning days in parts of the interior Southwest where summertime lightning is more common, but lower skill (AUC <0.6) in Pacific coastal areas where lightning is relatively rare. We then use explainable AI techniques (Layer-wise relevance propagation and Shapley additive explanations) to investigate the regional importance of individual predictor variables to successful lightning predictions. Results show that total-column water vapor, 700-500 hPa lapse rates, and 500 hPa relative humidity are substantially more important than other predictors in many grid cells, while large-scale circulation patterns are only important outside the North American monsoon core. In future work, we will also apply these grid-based CNNs to output from GCM projections to quantify future lightning occurrence across the western United States, which can inform changes to lightning-caused wildfire risk.