A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Environmental Engineering and Management

Thesis (M. Sc.) - Asian Institute of Technology, 2026

This study aims to predict ground-level PM₂.₅ concentrations in the Lower Mekong River region by exploiting the high temporal resolution of the Geostationary Environment Monitoring Spectrometer (GEMS). Satellite remote exploration provides an important alternative to the supplement lack of ground monitoring networks in Southeast Asia. Accurate monitoring of ground-level particulate matter PM₂.₅ in Southeast Asia is often challenging by the retrieval biases and seasonal and diurnal noises in geostationary satellite retrievals. This study addresses these challenges by developing two-stage machine learning framework designed to enhance the utility of GEMS data over the Lower Mekong Region.In the first stage, an XGBoost-based calibration model was implemented to refine GEMS Aerosol Optical Depth (AOD) across three primary spectral bands (350 nm, 440 nm, and 550 nm). By integrating ground-based observations from AERONET and PANDORA networks, the model successfully mitigated systematic dual-biases, correcting overestimation at low AOD levels and underestimation during high AOD loads. The calibration significantly improved retrieval accuracy, raising R 2 values from an initial range of 0.49{u2013}0.59 to 0.92{u2013}0.95. Furthermore, the framework effectively neutralized diurnal variations and early-morning noise, ensuring a physically consistent aerosol signal.The second stage XGBoost model leveraged these calibrated AOD products as primary predictors for estimating ground-level PM₂.₅ concentration. The model demonstrated sophisticated predictive logic, particularly utilizing the UV-spectrum AOD and auxiliary indices sensitive to biomass burning aerosols, which are prevalent in the study area. Validation via 10-fold cross-validation revealed an overall R 2 of 0.87, a dramatic improvement over the operational GEMS Level 4 PM₂.₅ product{u2019}s performance (R 2 = 0.43) in the same region. These results highlight the critical necessity of localized, multi-stage modeling to overcome the limitations of global operational products. This research provides a expandable architecture for incorporating multiple ground observation and satellite datasets from various operating entities and data quality, to enable high spatio-temporal resolution assessment of air quality management in Southeast Asia.