Particulate matter air quality assessment using integrated surface, satellite, and meteorological products: 2. A neural network approach

TLDR

Satellite‑derived aerosol products have improved sparse surface monitoring, but they capture the entire aerosol column rather than just the surface‑level aerosols that affect human health. The study proposes an artificial neural network framework to reduce uncertainty in surface PM estimation from satellite data. The ANN combines 3 years of MODIS aerosol optical thickness at 0.55 µm with rapid‑update meteorological analyses to estimate surface PM2.5 over the southeast United States. The ANN achieved higher correlations (R = 0.74 hourly, R = 0.78 daily) than simple or multiple regression, demonstrating its potential for operational air‑quality monitoring, though improvement varies by site and season.

Abstract

In recent years, sparse, surface‐based air quality monitoring has been improved by using wide‐swath, satellite‐derived aerosol products. However, satellites are sensitive to the entire aerosol column, not only the aerosol near the surface that impacts human health. In part 1 of this series, we used multiple regression to demonstrate how inclusion of meteorological analyses can help constrain the surface level proportion of the aerosol profile and improve the estimate of surface PM2.5. Here, instead of multiple regression technique, we describe an artificial neural network (ANN) framework that reduces the uncertainty of surface PM estimation from satellite data. We use 3 years of MODIS aerosol optical thickness data at 0.55 μ m and meteorological analyses from the rapid update cycle to estimate surface level PM2.5 over the southeast United States (EPA region 4). As compared to regression coefficients obtained through simple correlation (R = 0.60) or multiple regression (R = 0.68) techniques, the ANN derives hourly PM2.5 data that compare with observations with R = 0.74. For estimating daily mean PM2.5, the ANN techniques results in correlation of R = 0.78. Although the degree of improvement varies over different sites and seasons, this study demonstrates the potential for using ANN for operational air quality monitoring.