Geographia Technica, Vol 21(2), Special Issue: Artificial Intelligence Applications in Geography, 2026, pp. 134-157
AN AI-BASED FRAMEWORK FOR ESTIMATING PM2.5 USING SATELLITE AEROSOL OPTICAL DEPTH AND METEOROLOGICAL DATA IN A COASTAL INDUSTRIAL REGION OF THAILAND
Maliwan THABTHONG, Teerawong LAOSUWAN 
, Satith SANGPRADID , Yannawut UTTARUK , Jenjira PIAMDEE , Piyatida AWICHIN , Titipong PHOOPHATHONG , Maharaja SINGHARAJ
ABSTRACT: Monitoring fine particulate matter (PM2.5) in coastal industrial areas remains difficult, mainly because ground-based monitoring stations are sparse and unevenly distributed. In this study, we developed a practical AI-based framework to estimate daily PM2.5 concentrations by combining satellite-derived Aerosol Optical Depth (AOD) with commonly available reanalysis-based meteorological data. The analysis focused on Rayong Province, Thailand, an industrial coastal region where atmospheric processes and pollution sources are highly variable. Daily PM2.5 measurements from monitoring stations were merged with MODIS/MAIAC AOD and meteorological variables obtained from ERA5 and ERA5-Land. The meteorological inputs included air temperature, relative humidity, wind speed, and boundary layer height (BLH). Five machine learning models were tested, including Multiple Linear Regression (MLR), Support Vector Regression (SVR), Artificial Neural Networks (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Feature selection was carried out using Recursive Feature Elimination with Cross-Validation and Mutual Information, and model performance was evaluated using five-fold cross-validation. Among the tested models, Random Forest showed the best overall performance, achieving an R2 value of 0.70, followed by XGBoost and ANN. The results also indicate that nonlinear models consistently performed better than linear regression, suggesting that the relationship between PM2.5 and atmospheric variables is not linear. Wind speed emerged as the most influential predictor, with relative humidity and BLH also playing important roles. In particular, the PM2.5–meteorology relationships exhibited clear threshold behavior under low wind speed conditions. Overall, the proposed GeoAI-based framework provides a practical and transferable approach for spatial PM2.5 estimation in complex coastal industrial environments, offering an effective solution for air quality assessment in regions with sparse ground monitoring networks.
Keywords: PM2.5 estimation; Aerosol Optical Depth; Machine learning; Spatial air pollution; Coastal industrial region.
Full article here