Particulate matter with diameter less than 1 μm (PM1) has been found to be closely associated with air quality, climate changes, and even adverse human health. However, a large gap in our knowledge concerning the large-scale distribution and variability of PM1 remains, which is expected to be bridged with advanced remote-sensing techniques. In this study, a hybrid model called principal component analysis-general regression neural network (PCA-GRNN) is developed to estimate hourly PM1 concentrations from Himawari-8 aerosol optical depth in combination with coincident ground-based PM1 measurements in China. Results indicate that the hourly estimated PM1 concentrations from satellite agree well with the measured values at national scale, with R2 of 0.65, root-mean-square error (RMSE) of 22.0 μg/m3 and mean absolute error (MAE) of 13.8 μg/m3. On daily and monthly time scales, R2 increases to 0.70 and 0.81, respectively. Spatially, highly polluted regions of PM1 are largely located in the North China Plain and Northeast China, in accordance with the distribution of industrialisation and urbanisation. In terms of diurnal variability, PM1 concentration tends to peak in rush hours during the daytime. PM1 exhibits distinct seasonality with winter having the largest concentration (31.5±3.5 μg/m3), largely due to peak combustion emissions. We further attempt to estimate PM2.5 and PM10 with the proposed method and find that the accuracies of the proposed model for PM1 and PM2.5 estimation are significantly higher than that of PM10. Our findings suggest that geostationary data is one of the promising data to estimate fine particle concentration on large spatial scale.
Keywords: AOD; Himawari; Neural network; Particulate matter; Remote sensing.
Copyright © 2018 Elsevier Ltd. All rights reserved.