Estimation of global ground-level PM10 concentrations using a stacking model

Abstract

Atmospheric environmental pollution has gained significant attention recently. Inhalable particulate matter (PM10) impacts societal development and poses serious health threats. Consequently, PM10 has become a primary focus of air pollutant research. In this article, we propose a stacking model incorporating a back propagation neural network and extremely randomized trees (Stacking-BP-ET model). This model combined aerosol optical depth products with auxiliary factors, including meteorology, elevation and population distribution, to construct a global PM10 dataset with a 1 km spatial resolution from 2015 to 2021. The product had an out-of-station and out-of-year spatiotemporal cross-validation coefficient of determination (R2) of 0.833, and the mean absolute error (MAE) and root mean square error (RMSE) were 6.411 and 14.071 μg/m3, respectively. High PM10 concentrations were found in areas with frequent dust storm events such as Mongolia, the Arabian Peninsula and Northwest China and wildfire disasters such as the west coast of the United States and Australia. The PM10 concentrations in these areas had shown an increasing trend in recent years as the disaster frequency has increased. Overall, the global PM10 dataset may be useful for future large-scale studies of air pollution.

Keywords

• PM10
• aerosol optical depth
• Stacking-BP-ET