CONUS PM2.5 estimation

CONUS PM2.5 

We estimated daily ground-level PM2.5 concentrations across the contiguous U.S. by integrating EPA AQS reference measurements with PurpleAir sensor data, applying rigorous quality control and geographically weighted regression (GWR) with temperature and humidity adjustments. We incorporated Community Multiscale Air Quality (CMAQ) model outputs to estimate total and nonfire PM2.5, deriving wildfire smoke PM2.5 by differencing the two. Satellite-derived MAIAC AOD, meteorology, land cover, topography, population density, and NOAA smoke plume data were harmonized to 1 km grids through gap-filling and interpolation. To isolate the smoke contribution, we trained two separate random forest models: one to predict total PM₂.₅ in smoke-impacted regions and another to predict background PM2.5 in both smoke and no-smoke regions, identified using NOAA plume polygons and CMAQ smoke ratios. Smoke PM2.5 was then calculated as the difference between these two predictions, and SMOTE oversampling was applied to improve performance for high-concentration events, resulting in approximately 1.68 million smoke-impacted and 2.01 million no-smoke station-day observations.

For the ZIP-level daily data, the 1 km total and nonfire PM2.5 estimates were aggregated to ESRI ZIP Code boundaries using arithmetic mean for all grid points within each ZIP. Daily smoke PM2.5 at the ZIP level was then computed as the difference between the aggregated total and nonfire PM2.5 estimates.