[Spatio-temporal Evolution of PM2.5 Concentration During 2000-2019 in China].

An ensemble estimation model of PM2.5 concentration was proposed on the basis of extreme gradient boosting, gradient boosting, random forest model, and stacking model fusion technology. Measured PM2.5 data, MERRA-2 AOD and PM2.5 reanalysis data, meteorological parameters, and night light data sets were used. On this basis, the spatiotemporal evolution features of PM2.5 concentration in China during 2000-2019 were analyzed at monthly, seasonal, and annual temporal scales. The results showed that:① Monthly PM2.5 concentration in China from 2000-2019 can be estimated reliably by the ensemble model. ② PM2.5 annual concentration changed from rapid increase to remaining stable and then changed to significant decline from 2000-2019, with turning points in 2007 and 2014. The monthly variation of PM2.5 concentration showed a U shape that first decreased then increased, with the minimum value in July and the maximum value in December. ③ Natural geographic conditions and human activities laid the foundation for the annual spatial pattern change of PM2.5 concentration in China, and the main trend of monthly spatial pattern change of PM2.5 concentration was determined by meteorological conditions. ④ At an annual scale, the national PM2.5 concentration average center of standard deviation ellipse moved eastward from 2000-2014 and westward from 2014-2018. At a monthly scale, the average center shifted to the west from January to March, moved northward then southward from April to September, and shifted to the east from September to December.

[PM2.5 Concentration Influencing Factors in China Based on the Random Forest Model].

In this paper, aerosol optical depth (AOD), elevation (DEM), annual precipitation (PRE), annual average temperature (TEM), annual average wind speed (WS), population density (POP), gross domestic product density (GDP), and normalized difference vegetation index (NDVI) were selected as factors influencing PM2.5 concentration. The random forest model, order of feature importance, and partial dependency plots were applied to investigate these factors and their regional differences in PM2.5 spatial pattern. The results showed that:① The random forest model was more accurate than multiple regression, generalized additive, and back propagation neural network models in estimating PM2.5 concentration, which can be applied to quantifying PM2.5 influencing factors. ② PM2.5 concentration initially increased and then remained stable with increases in AOD, POP, and GDP, and initially decreased and then stabilized with increases in PRE, WS, and NDVI. The responses of DEM and TEM to PM2.5 concentration changed from decline to ascend and then changed to decline again. ③ AOD had the largest influence on PM2.5 annual concentrations with a spatial influencing magnitude of 37.96%, whereas PRE had the least influence with a merely individual spatial influencing magnitude of 5.75%. ④ The relationships between PM2.5 pollution and influencing variables vary with geography and thus exhibit significant spatial heterogeneity. The same factor had different spatial influencing magnitudes on PM2.5 annual concentrations in seven geographical subareas. AOD had the greatest influence on PM2.5 concentration in the south of China, with the least influence in the northeast.