Abstract
The growing understanding of PM2.5 compositions has promoted the application of machine learning in exposure estimation. However, data-driven approaches may neglect the necessity for expert knowledge. This study presents a theory-guided approach combining land use regression (LUR), extreme gradient boosting (XGBoost), and Shapley additive explanations (SHAP) to estimate the spatiotemporal variation of ionic PM2.5 in Taiwan. Ground-based monthly observations of eight PM2.5 ions (Na+, NH4+, K+, Mg2+, Ca2+, Cl−, NO3−, and SO42−) from 28 air quality monitoring stations across Taiwan, meteorological data simulated using the Weather Research and Forecasting model, and satellite-derived information (including land use, aerosol characteristics, greenness, and insolation) were collected to develop monthly predictions of ionic PM2.5 from 2019 to 2021. LUR, XGBoost, and SHAP were applied for feature selection, model building, and the evaluation of feature impacts, respectively. The models demonstrated moderate performances for most ionic PM2.5, except for Mg2+ and Cl−, with an independent test set showing adjusted R2 values of 0.69–0.87. Sensitivity analyses were conducted by removing directional selection and incorporating ground-observed anthropogenic factors; however, the model performance was not improved. This study offers insights into the interaction between contributing features and ionic PM2.5 for climate research.
[Display omitted]
•The meteorological features have higher contribution than other features.•The biomass-burning PM2.5 derived from Himawari-8 has clear contribution for ionic PM2.5.•Different types of vegetation result in heterogeneous impacts on ionic PM2.5.•Diverse satellite-derived information makes the model construction flexible.