Reliable, long-term information on the chemical composition of fine particulate matter (PM 2.5 ) is essential for air quality management and health risk assessment, yet such data are rarely available with both high spatial and temporal resolution across large geographic areas. We developed a spatiotemporal modeling framework using the deep forest algorithm to estimate daily, gap-free PM 2.5 component concentrations at 1-km resolution across the western United States from 2002 to 2019. By integrating data from ground-level chemical speciation networks, satellite-derived PM 2.5 mass concentrations, CMAQ simulations, and multiple auxiliary predictors, we estimated daily concentrations of sulfate (SO 4 2− ), nitrate (NO 3 − ), organic carbon (OC), elemental carbon (EC), and mineral dust (DUST). Model performance showed strong agreement with observations, with cross-validation R 2 values of 0.81, 0.89, 0.75, 0.66, and 0.75, and root-mean-squared errors (RMSEs) of 0.30, 0.59, 0.26, 1.52, and 0.59 μg/m 3 , respectively. The estimated component sum captured the expected fraction of total PM 2.5 , closely tracking observed total mass with accuracy comparable to the observed component sum (R 2 = 0.91 vs. 0.92; RMSE = 2.66 μg/m 3 ; slope = 0.66 for both). Our high-resolution estimates revealed marked spatial variability: SO 4 2− , NO 3 − , EC, and OC were elevated in urban areas—particularly the Los Angeles–Long Beach–Anaheim region—while DUST was concentrated along the southern California–western Arizona border. The model also captured sharp increases in OC, NO 3 − , and EC during wildfire events compared to non-fire periods. Over the long term, urban populations experienced 1.5–2 times higher exposure to SO 4 2− , NO 3 − , EC, and OC than rural populations, with similar DUST exposure levels. All five components showed declining trends over the study period, especially in the southwestern U.S., with total concentrations falling from 12.32 μg/m 3 in 2002 to 9.03 μg/m 3 in 2019, primarily due to reductions in OC and NO 3 − . The long-term modeling framework and dataset can support attainment planning, source-specific mitigation, wildfire smoke operations, and health impact assessments. • Analyzed spatiotemporal trends of PM 2.5 component exposures in the western U.S. (2002-2019). • Developed hybrid models for PM 2.5 components at high spatial and temporal resolution. • Deep-forest models with localized feature importance yielded accurate, interpretable results. • Total PM 2.5 mass outperformed AOD as a predictor in component modeling. • Models captured both day-to-day and localized variations in PM 2.5 components.
He et al. (Sat,) studied this question.