Estimating Hourly PM2.5 Concentration from Satellite-Measured Top-of-Atmosphere Reflectance by Using a Machine Learning Algorithm

As a major indicator of atmospheric pollution, the mass concentration of aerosol particulate with dynamic diameter less than 2.5 mm (PM_2.5) has been estimated from satellite, modeled by chemical transport models and their hybrids. Overall, they provide reasonable gross amount and bulk features of variations, but the accuracy is still rather poor at any particular location and time due largely to (1) far more variables influencing its value than what can be observed and (2) complex and even nonlinear interactions among the variables that joint affect PM_2.5.

Different from many previous studies, which used satellite-derived aerosol optical depth (AOD) as the primary predictor to estimate the daily mean ground-level PM_2.5concentrations (denoted as AOD-PM_2.5model), the current study developed a statistical model at a national-scale of China based on an ensemble machine learning algorithm by directly incorporating the satellite measured top-of-atmosphere (TOA) reflectance and meteorological parameters (denoted as Ref-PM_2.5model). The cross-validation indicated that PM_2.5concentrations estimated by this method have a comparable accuracy to the method using the satellite retrieved AOD, but the former has a relatively stronger predictive power on the spatial and temporal coverage than the latter. By using the Ref-PM_2.5model, the annual and seasonal distribution of ground-level PM_2.5concentrations over three major megalopolises over China were estimated and analyzed. This study proposed an alternative means of estimating surface PM_2.5concentration from satellite by using the satellite TOA reflectance directly. The approach can circumvent multiple sources of uncertainties incurred by any deterministic methods, or generally referred to as physical approach. The two are, however, complimentary and some synergy will be explored to improve both. Besides, the high capacity of Ref-PM_2.5model at hourly scale could help to prove the information on the diurnal cycle of PM_2.5concentrations and improve our understanding the formation processes of regional PM_2.5pollution episodes and the PM_2.5evolution.