PM2.5 (fine particulate matter with aerodynamics diameter <2.5 μm) is the most important component of air pollutants, and has a significant impact on the atmospheric environment and human health. Using satellite remote sensing aerosol optical depth (AOD) to explore the hourly ground PM2.5 distribution is very helpful for PM2.5 pollution control. In this study, Himawari-8 AOD, meteorological factors, geographic information, and a new deep forest model were used to construct an AOD-PM2.5 estimation model in China. Hourly cross-validation results indicated that estimated PM2.5 values were consistent with the site observation values, with an R2 range of 0.82-0.91 and root mean square error (RMSE) of 8.79-14.72 μg/m³, among which the model performance reached the optimum value between 13:00 and 15:00 Beijing time (R2 > 0.9). Analysis of the correlation coefficient between important features and PM2.5 showed that the model performance was related to AOD and affected by meteorological factors, particularly the boundary layer height. Deep forest can detect diurnal variations in pollutant concentrations, which were higher in the morning, peaked at 10:00-11:00, and then began to decline. High-resolution PM2.5 concentrations derived from the deep forest model revealed that some cities in China are seriously polluted, such as Xi 'an, Wuhan, and Chengdu. Our model can also capture the direction of PM2.5, which conforms to the wind field. The results indicated that due to the combined effect of wind and mountains, some areas in China experience PM2.5 pollution accumulation during spring and winter. We need to be vigilant because these areas with high PM2.5 concentrations typically occur near cities.
Keywords: AOD; Himawari-8; Machine learning; PM(2.5); Wind field.
Copyright © 2022 Elsevier Ltd. All rights reserved.