Satellite and Reanalysis-Based Estimation of Surface PM2.5 Concentrations Using Bayesian Averaging and Random Forest Models

Abstract

Air pollution characterized by elevated fine particulate matter (PM₂.₅) is a critical environmental and public health challenge in rapidly industrializing regions where ground-based monitoring networks remain sparse. This study evaluates and compares two predictive models - a probabilistic linear ensemble via Bayesian Model Averaging (BMA) and a non-parametric machine learning algorithm using Random Forest (RF) - to estimate surface PM₂.₅ concentrations across Binh Duong province, located in the North of Ho Chi Minh City, Vietnam during the dry seasons of 2024 - 2025. The models integrate multi-source satellite observations, including MODIS Aerosol Optical Depth (AOD), Sentinel-5P TROPOMI tropospheric pollutants (NO₂, SO₂, and CO), and ERA5 meteorological reanalysis variables. Model parameters and predictive robustness were optimized and evaluated using repeated 5-fold cross-validation. The cross-validation results revealed that the Random Forest model outperformed the BMA model, achieving an R² of 0.703, MAE of 4.594 ± 0.48 µg/m³, RMSE of 5.608 ± 0.62 µg/m³, and MAPE of 39.621%. In comparison, the BMA model yielded a validation R² of 0.626, MAE of 4.839 µg/m³, RMSE of 5.992 µg/m³, and MAPE of 40.852%. The moderate MAPE values (~40%) indicate that while the framework is not suitable for strict regulatory compliance enforcement, it serves as an effective screening and spatial assessment tool. Satellite-derived surface maps successfully reproduced regional spatial gradients, highlighting pronounced PM₂.₅ concentrations in southern industrial corridors (Di An, Thuan An, and Thu Dau Mot) contrasting with lower levels in northern rural districts. These findings demonstrate the practical utility of integrating satellite remote sensing with machine learning for dry-season air quality screening and environmental management in data-limited developing regions.