High-resolution estimation of ambient sulfate concentration over Taiwan Island using a novel ensemble machine-learning model

Heavy loadings of sulfate aerosol trigger haze formation and pose great damage to human health in Taiwan Island. Nevertheless, high-resolution spatiotemporal variation of ambient sulfate across Taiwan Island still remained unknown because of the scarce monitoring sites. Thus, we developed a novel ensemble model named extreme gradient boosting coupled with geographically and temporally weighted regression (XGBoost-GTWR) to predict the high-resolution sulfate concentration (0.05°) based on satellite data, assimilated meteorology, and the output of chemical transport models (CTMs). The result suggested that XGBoost-GTWR model outperformed other five models in predicting the sulfate concentration with the highest R value (R = 0.58) and the lowest relative mean square error (RMSE = 1.96 μg/m). Besides, the transferability of the XGBoost-GTWR model was also validated based on the ground-level sulfate data in 2019. The result suggested that the R value of the extrapolation equation (0.53) did not show notable decrease compared with the 10-fold cross-validation result (0.58), indicating that the model was robust to predict the sulfate concentration. The ambient sulfate concentration in Taiwan Island displayed featured spatial variation with the highest one in Southwest Taiwan and the lowest one in Northeast Taiwan, respectively. It was assumed that the higher anthropogenic emission combined with the adverse meteorological condition led to the higher sulfate level in the southwestern coastal region. The ambient sulfate concentration exhibited significantly seasonal variation with the highest value in spring (5.65 ± 0.84 μg/m), followed by those in winter (5.45 ± 1.25 μg/m) and autumn (4.60 ± 0.80 μg/m), and the lowest one in summer (3.80 ± 0.65 μg/m). The higher sulfate concentration in spring was mainly contributed by the dense biomass burning and scarce rainfall amount. The present study develops a novel model to capture the high-resolution sulfate map and provides basic data for effective regulations of air pollution and epidemiological studies.