Daily SLP and 500-hPa Z SOM Patterns Explain Lightning Variability in Alaska Artificial Intelligence for the Earth Systems

Lightning-ignited fires accounted for 97.5% of the area burned in Alaska from 2004 to 2023. While lightning parameterization schemes have been developed to project future changes or to provide near-term forecasts of lightning activity, skillful subseasonal to seasonal (S2S) predictions—especially at high latitudes—remain limited. In Alaska, effective wildland fire management would greatly benefit from reliable S2S forecasts of lightning. Toward this goal, our study investigates the predictability of daily lightning intensity based on the internal variability of synoptic-scale atmospheric dynamics. We use the self-organizing map (SOM) to extract daily weather patterns from observed sea level pressure and 500-hPa geopotential height fields. These patterns are analyzed in relation to known lightning-favorable conditions and compared with observations from the Alaska Lightning Detection Network. The random forest model is used as a tool for exploratory data analysis. The SOM patterns are incorporated into random forest models, along with established lightning proxies, to classify daily lightning intensity. Analyses of the permutation feature importance, ablation tests, and feature correlations reveal that the predictive skill from the SOM patterns is physically linked to lightning through local atmospheric processes. The random forest models achieve area under the receiver operating characteristic curve (AUROC) scores ranging from 0.60 to 0.84, depending on model complexity and region defined by predictive service area (PSA) divisions. This diagnostic study demonstrates that internal large-scale climate variability holds useful information for predicting daily lightning intensity, suggesting a promising pathway toward developing skillful S2S lightning forecasts for Alaska. Grant no. NA23OAR4690390