Integrating AOD-based PM2.5 Retrieval, Land Use Regression, and Machine Learning to Estimate the Spatiotemporal Variation of Ions in PM2.5 across Taiwan

KANG LO, Yee-Lin Wu, Tang-Huang Lin, Chang-Fu Wu, National Taiwan University

     Abstract Number: 179
     Working Group: Aerosols Spanning Spatial Scales: Measurement Networks to Models and Satellites

Abstract
The impact of various composition of ambient air particulate matter on the ecosystem and human health is one of the leading global issues. Mounting researches have engaged in modeling spatiotemporal variation on the components in the fine particles (PM_2.5). While being compared to the elements in PM_2.5 composition, however, the spatiotemporal prediction of ions in PM_2.5 which are crucial factors in the formation of secondary aerosols still remains limited. In this study, aerosol optical depth (AOD)-retrieved PM_2.5, land use regression (LUR), and machine learning algorithm, Extreme Gradient Boosting (XGBoost), were integrated to estimate the spatiotemporal variation of water-soluble ions (WSI) in PM_2.5. Monthly average PM_2.5 WSIs (SO₄²⁻, NO₃⁻, NH₄⁺, Na⁺, Cl⁻, K⁺, Ca²⁺, and Mg²⁺) were obtained from 31 air quality monitoring stations of Taiwan EPA from 2019 to 2021. Several datasets, including land use, road information, MODIS AOD, Himawari-8 AOD, Himawari-8 AOD-retrieved PM_2.5, MODIS Normalized Difference Vegetation Index (NDVI), elevation, stationary emission sources, demographic data, meteorological data, and distribution of temples, were collected as the features in the models. The results demonstrated that these models had good performance on predicting external datasets of SO₄^2- (adjusted R² = 0.77), NO₃⁻ (adjusted R² = 0.84), NH₄⁺ (adjusted R² = 0.83), Na⁺ (adjusted R² = 0.76), Cl⁻ (adjusted R² = 0.59), K⁺ (adjusted R² = 0.64), and Ca²⁺ (adjusted R² = 0.68). The model had relatively poor performance on predicting external datasets of Mg²⁺ (adjusted R² <0.5). These findings suggest that the current approach could be a proper method to predict the spatiotemporal variation of WSIs in PM_2.5.