A Multi-Stage Deep Learning Framework for Daily PM2.5 Estimation at 100 m Resolution Across the Contiguous United States, 2000–2024 (A Quarter Century of Data)

Abstract

Accurate high-resolution estimation of fine particulate matter (PM2.5) remains challenging because of sparse monitoring networks and missing satellite observations. We developed a multistage deep learning framework to generate daily PM2.5 concentrations at 100 m resolution across the contiguous United States (CONUS) from 2000 to 2024. The framework first reconstructed missing satellite aerosol optical depth (AOD) using a U-Net-based encoder–decoder informed by reanalysis data, refined temporal dependencies using a bidirectional long short-term memory network, and downscaled reconstructed AOD to 100 m using terrain information. Daily PM2.5 was subsequently estimated using a multistream deep learning architecture integrating reconstructed AOD, meteorological, spatiotemporal, and geospatial predictors. Evaluation using strict site-level data partitioning yielded strong predictive performance (R2 = 0.82, RMSE = 2.85 μg/m3, MAE = 1.84 μg/m3), with high spatial (R2 = 0.94) and temporal (R2 = 0.78) performance. The framework generated spatially continuous daily PM2.5 surfaces across more than 766 million 100 m grid cells while capturing broad spatial gradients and fine-scale variability. These long-term, high-resolution estimates provide an exposure surface suitable for epidemiological, environmental justice, and air-pollution assessment applications.