With the increasing attention to air pollution, monitoring atmospheric particulate mass concentration has become a hot research field. In this paper, the scientific data sets of near-surface PM 2.5 mass concentration generated by two popular algorithms (model simulation and machine learning) are compared, quantitatively evaluated the uncertainty of the two data sets by using urban annual average PM 2.5 monitoring data, and the spatial rationality of the two data sets through spatial autocorrelation analysis. Meanwhile, the spatial-temporal evolution trend of PM 2.5 in major pollution areas (Beijing, Tianjin, Hebei, Henan, Shanxi and Shandong) from 2000 to 2018 was also studied by using standard deviation ellipse analysis. The results show that the data set (CHAP) based on machine learning algorithm has high precision and is suitable for regional air quality research, and the data set (vanDonkelaarA) generated by the model simulation algorithm has a reasonable spatial distribution and is more suitable for large-scale and long-term pollution trend analysis. According to the analysis of standard deviation ellipse, the center of standard deviation ellipse in the study area moved to the northeast from 2000 to 2018. Before 2013, the distribution range and annual mean value of PM 2.5 showed an overall trend of increase, and then decreased significantly. It is shown that the main factor contributing to the decrease in PM 2.5 concentration is the implementation of effective control measures. The result provides a reference for the selection of fine particulate matter pollution research data sets in China, and also provides scientific support for the prevention and control of atmospheric fine particulate matter pollution.