Truace
Subscribe

Machine Learning Predictions of PM<sub>2.5</sub> and Their Applications to Exposure and Health Assessment in the CONUS

Atmospheric fine particulate matter (PM 2.5 ) 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 PM 2.5 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…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1611· cite & verify
Machine Learning Predictions of PM<sub>2.5</sub> and Their Applications to Exposure and Health Assessment in the CONUS
Earth from Orbit- Wildfire Smoke Blankets U S (NESDIS 2021-07-15 2021 07 15 WildfireSmokeBlanketsUS UHD) by NOAA. Public domain

In brief

Researchers built a two-phase framework that fills large gaps in GOES-R geostationary satellite aerosol data using an image-inpainting-style neural network, then feeds the completed data to a random forest to predict surface PM2.5 hourly at fine resolution.

The estimates for 2019-2023 in the 50 most populous US counties matched ground monitors and low-cost sensors, and were used to gauge exposure differences by race and ethnicity. How well this holds in rural areas is less settled, since satellite gaps remain substantial there.

Main points

  1. The gap-filling approach raised paired AOD observations against AERONET from 430,603 to 728,115 while keeping R = 0.75.
  2. PM2.5 estimates covered 2019-2023 in the top 50 most populous counties of the contiguous US.
  3. Predictions agreed with EPA AQS and PurpleAir data, with daily R2 of 0.85 and mean bias of -0.11 μg/m3.
  4. Black, Hispanic, and Indigenous populations had 1.7%-5.5% higher exposure than White populations.
  5. Wildfire case studies in sparsely monitored rural regions showed use for short-term exposure events.

The gain

A partial convolutional neural network and random forest produce hourly 0.01-degree PM2.5 estimates that track ground monitors, enabling exposure and mortality assessment and revealing higher exposure among minority groups.

The rundown

The method uses a U-Net-like partial convolutional neural network to address missing AOD data from the GOES-R satellite. An iterative scheme expands the contextual information used in successive predictions to fill large spatiotemporal gaps.

A random forest then converts the contiguous gap-filled AOD into surface PM2.5 concentrations at 0.01° × 0.01° resolution. Validation drew on the EPA Air Quality System and the PurpleAir sensor network.

The authors applied the predictions to population exposure and PM2.5-attributable mortality estimates, noting PM2.5 as a major risk factor for respiratory and cardiovascular disease. Wildfire case studies in rural regions illustrated utility despite satellite gaps.

Sources

  1. Peer-reviewedGeoHealth2026-10-08

The debate