Improving PM2.5 Estimation from Satellite Aerosol Optical Depth Using Boundary Layer Height and Meteorological Variables

Shreyas Khobragade 

Accurate estimation of surface particulate matter (PM2.5) from satellite observations remains challenging because aerosol optical depth (AOD) represents column-integrated aerosol loading and is strongly influenced by meteorological conditions. This study investigates the relationship between satellite-derived AOD, meteorological variables, and surface PM2.5 concentrations over Mumbai, India, during the period 2021–2025.

A total of 1,248 collocated observations of PM2.5, AOD, boundary layer height (BLH), relative humidity (RH), temperature, wind parameters, and solar radiation were analyzed. Several physically informed predictor variables were developed to account for atmospheric mixing and humidity effects, including a humidity- and boundary-layer-corrected aerosol indicator (AOD_DRY_BLH). Linear regression and Random Forest models were subsequently evaluated using statistical validation, seasonal analysis, temporal analysis, and error assessment.

Raw AOD exhibited a weak relationship with PM2.5 (r = 0.270, R² = 0.073), whereas the engineered variable AOD_DRY_BLH increased explanatory power by approximately 4.3 times relative to raw AOD (r = 0.562, R² = 0.316). Boundary layer height emerged as the strongest controlling factor, followed by relative humidity and atmospheric dryness. The best-performing linear model achieved a test-set R² of 0.449, while the Random Forest model yielded a test-set R² of 0.667 with lower prediction errors (MAE = 12.00 μg m⁻³; RMSE = 16.49 μg m⁻³) than the best-performing linear formulation. Temporal validation demonstrated relatively stable performance across the study period, whereas seasonal analysis revealed stronger predictive skill during the pre-monsoon season and reduced performance during winter and monsoon conditions.

The results demonstrate that incorporating meteorological constraints significantly improves the capability of satellite-derived aerosol observations to represent surface PM2.5 concentrations. The findings highlight the dominant influence of atmospheric mixing and humidity on aerosol–PM2.5 relationships and support the integration of physically informed variables with machine learning approaches for urban air-quality assessment in tropical environments.