Abstract: Based on the daily PM2.5 data of China in 2018, a high-precision PM2.5 concentration estimation model was constructed using random forest method, and the temporal and spatial applicability of the model was verified at seasonal and regional scales. Further, the importance of each influencing factor to the change of PM2.5 concentration was systematically explained using the feature importance method. Finally, the comprehensive influence of the interaction of different influencing factors on PM2.5 concentration change was explored using the partial dependence technique. The results show that: (1) Compared with the multiple linear regression model and the extreme gradient ascending tree model, the random forest model based on multi-source data has the highest accuracy, which not only can accurately simulate the PM2.5 concentration, but also has good applicability at the seasonal and regional scales. (2) According to the ranking results of model feature importance, the factors that had significant impact on the average daily PM2.5 concentration in 2018 were mainly global factors such as space-time and atmospheric boundary layer height, which indicated that the prevention and control of air pollution should follow the PM2.5 transmission mechanism, and regional joint prevention and control should be strengthened in air pollution prevention. (3) The partial dependent interaction effect study shows that the combination of temperature, relative humidity, annual cumulative day, latitude, temperature and atmospheric boundary layer height has a significant impact on PM2.5 concentration change, indicating that to improve air quality should start from the perspective of multi-factor collaborative governance.

Key words: atmospheric remote sensing, random forest, PM2.5, spatio-temporal correlation