An improved machine-learning model for lightning-ignited wildfire prediction in Texas

Abstract Lightning-ignited wildfires (LIWs), which predominantly occur in remote regions, pose significant challenges for prediction and mitigation, while contributing disproportionately to burned areas in Texas. Persistent knowledge gaps regarding key ignition drivers and their nonlinear interdependencies hinder the development of targeted prevention strategies and robust early warning systems. To address these limitations, we compiled a statewide dataset spanning 2010–2020 comprising 4775 LIWs and employed an optimized repeated-random undersampling strategy to mitigate class imbalance. Using this dataset, we developed an eXtreme gradient boosting-based machine learning model that integrates meteorological, soil, vegetative, lightning, topographic, and human activity variables to predict LIW probability. The most accurate classifier achieved an out-of-sample prediction accuracy of 85.81%, outperforming the fire weather index and methodologies of random forests and logistic regression. Key drivers of ignition were identified as higher lightning frequency, elevated temperatures, and lower fuel moisture content. Spatial analysis revealed LIW clustering in needleleaf forests of eastern Texas, where maximum fuel loading (indicated by normalized difference vegetation index) and lightning density interacted to create ignition-prone conditions.