Uncovering the Influence of Land Cover and Climate Extremes on Wildfire Smoke PM _2.5 in the United States Using Explainable Artificial Intelligence

Abstract

Fine particulate matter (PM_2.5) from wildfire smoke has emerged as a significant environmental health threat in the United States (U.S.), yet the combined roles of climate extremes and land cover in shaping wildfire smoke PM_2.5 exposure remain poorly understood. Drawing on census tract-level data from 2010 to 2018, we developed an integrative framework that combined time series analysis, multilevel regression, and interpretable machine learning (XGBoost with SHapley Additive exPlanations, SHAP) to examine the drivers of wildfire smoke PM_2.5. Across methods, elevated temperatures, drought conditions, and limited precipitation consistently preceded increases in smoke PM_2.5, particularly in the western and northern U.S. Cold waves and heavy rainfall, in contrast, were associated with modest reductions. The influence of land cover on smoke exposure differed across regions, with forests, wetlands, and grasslands affecting exposure according to local fuel availability and landscape heterogeneity. The XGBoost-SHAP model achieved high predictive performance (R2 = 0.77-0.87) and further revealed that compound drivers such as drought amplification and cold wave suppression became increasingly important in later years. Our findings underscore the need for adaptive, geographically targeted smoke mitigation strategies that account for evolving influence of climatic and land-based factors across space and time.