Predicting and Explaining Lightning and Lightning‐Ignited Wildfires in Northern California With a Two‐Step Convolutional Neural Network Framework

Abstract

Lightning‐ignited wildfires (LIWs) are a major contributor to burned area in Northern California and have increased substantially in size over recent decades. Predicting LIWs remains challenging because ignition depends on both lightning occurrence and fuel‐meteorology interactions. We develop a two‐step convolutional neural network (CNN) framework to conjunctively predict (a) lightning occurrence from meteorological conditions and (b) wildfire ignition conditional on lightning using both meteorological and fuel‐related predictors. The framework achieves skillful prediction, with testing accuracy of 0.82 for predicting LIWs. Explainability analyses using layer‐wise relevance propagation (LRP) reveal that the two‐step framework separates the dominant controls on lightning‐producing versus fire‐producing environments, highlighting the transition from synoptic‐scale instability and moisture patterns to locally receptive fuel conditions. Composite analyses across prediction outcomes further show that missed LIWs arise more often in “non‐typical” conditions, with high lapse rate but limited mid‐level moisture. Application to the 2020 Sonoma‐Lake‐Napa Unit (LNU) Lightning Complex fires demonstrates that the framework captures the large‐scale circulation favorable for lightning development and local fuel conditions conducive to ignition. These results demonstrate the value of separating lightning production from ignition processes and show how explainable machine learning can diagnose multiscale pathways and model failure modes associated with LIW.