LAUSR

Abstract Atmospheric particulate matter (PM) that have particle diameter less than 2.5 μm (PM2.5) are hazardous to public health whose concentration has been either measured on the ground or inferred from satellite-retrieved aerosol optical depth (AOD). The latter is subject to numerous sources of errors, making the satellite retrievals of PM2.5 highly uncertain. This study developed an ensemble machine-learning (ML) algorithm for estimating PM2.5 concentration directly from Advanced Himawari Imager satellite measured top-of-the-atmosphere (TOA) reflectances in 2016 integrated with meteorological parameters. The algorithm is demonstrated to perform well across China with high accuracies at different temporal scales. The model has an overall cross-validation coefficient of determination (R2) of 0.86 and a root-mean-square error (RMSE) of 17.3 μg m−3 for hourly PM2.5 concentration estimation. Such accuracies of the estimation on PM2.5 concentration by using TOA reflectance directly are comparable with those of the common methods on estimating PM2.5 concentration by using satellite-derived AODs, but the former has a relatively stronger predictive power relating to spatial-temporal coverages than the latter. Annual and seasonal variations of PM2.5 concentration over three major the developed regions in China are estimated using the model and analyzed. The relatively stronger predictive ability of developed model in this study may help provide information about the diurnal cycle of PM2.5 concentrations as well as aid in monitoring the processes of regional pollution episodes and the evolution of PM2.5 concentration.