Machine Learning Predictions of PM2.5 and Their Applications to Exposure and Health Assessment in the CONUS.

Atmospheric fine particulate matter (PM2.5) is one of the most important risk factors for various respiratory and cardiovascular diseases. We develop a two-phase data-driven framework that uses geostationary Aerosol Optical Depth (AOD) retrievals to predict hourly PM2.5 at 0.01° × 0.01° resolution and support exposure assessment across multiple spatial and temporal contexts. We use a U-Net-like partial convolutional neural network framework inspired by image inpainting to address missing Aerosol Optical Depth (AOD) data from the GOES-R geostationary satellite. To fill large spatiotemporal gaps in the GOES-R record, we implement an iterative gap-filling approach that expands contextual information for successive predictions, maintaining strong agreement with collocated AERONET observations (R = 0.75) while increasing paired observations from 430,603 to 728,115. A random forest is used to predict surface PM2.5 concentrations using our spatiotemporally contiguous gap-filled AOD. The resulting PM2.5 estimates over 2019-2023 in the top 50 most populous counties over the CONUS capture the magnitude and spatial gradient as measured by the EPA Air Quality System (AQS) and PurpleAir low-cost sensor network (daily R 2 = 0.85, mean bias = -0.11 μg/m3). We use these predictions to assess population exposure and estimate PM2.5-attributable mortality, finding that Black, Hispanic, and Indigenous populations experienced 1.7%-5.5% higher PM2.5 exposure than White populations across these counties. Additional wildfire case studies in less densely monitored rural regions demonstrate the framework's utility for characterizing short-term exposure events despite substantial gaps in satellite observations. These findings highlight the value of gap-filled geostationary AOD for high-resolution exposure assessment.