Evaluating Machine Learning Methods for PM_2.5 Estimation of GEMS Satellite AOD Data

Abstract. Rapid urbanization and industrialization affected the air quality in the Philippines. Fine particulate matter (PM_2.5) are of particular concern due to their health, environmental, as well as climate effects. Due to the lack of active and available air quality monitoring in the Philippines, air quality monitoring and mitigation cannot be performed. Satellite air quality data can be utilized to provide extensive spatial and temporal coverage. In this study, aerosol optical depth (AOD) data from the Geostationary Environment Monitoring Spectrometer (GEMS) onboard the GEO-COMPSAT-2B satellite was used to estimate PM_2.5 and compared with data from a ground monitoring station in Manila, Philippines along with meteorological data from the European Centre for Medium- Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5). Random forest (RD), support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost) were evaluated for their accuracy in predicting ground-level PM_2.5. SVM achieved the highest accuracy (R²: 0.998) followed by RF (R²: 0.997), and then XGBoost (R²: 0.673). SHAP analysis showed that wind speed has the highest contribution in predicting PM2.5 This study shows that satellite air quality data can be used for ground-level PM_2.5 estimation.