We estimated the monthly satellite-based PM2.5 data from 2004 to 2013 at 0.1∘ resolution in our previous work (Ma et al., 2016). Briefly, we developed a two-stage statistical model using MODIS Collection 6 AOD and assimilated meteorology, land use data, and ground-monitored PM2.5 concentrations in 2013. The overall model cross-validation R2 (coefficient of determination) was 0.79 (daily estimates) for the model year. Since ground-monitored data before 2013 have been lacking and therefore it is not possible to develop statistical models before 2013 to estimate historical PM2.5 concentrations. Thus, the historical PM2.5 concentrations (2004–2012) were then estimated using the model developed based on the 2013 model. Two methods were used to validate the accuracy of historical estimates. First, we compared the historical estimate monitoring data from Hong Kong and Taiwan before 2013. Second, we estimated PM2.5 concentrations in the first half of 2014 using the 2013 model and compared them with the ground measurements to evaluate the accuracy of PM2.5 estimates beyond the model year, which can represent the accuracy of historical estimates. Validation results indicated that it accurately predicted PM2.5 concentrations with little bias at the monthly level (R2=0.73, slope =0.91).

For PM2.5 concentrations from 2004 to 2013, we used the abovementioned satellite-based PM2.5 dataset, which was estimated using the model developed in 2013. First, this dataset has shown high accuracy and has been widely used in environmental epidemiological (Liu et al., 2016a; Wang et al., 2018a), health impact (Liu et al., 2017b; Wang et al., 2018b), and social economic impact (Chen and Jin, 2019; Yang and Zhang, 2018) studies in China. Second, a recent study has shown that the historical hindcast ability of the annual model decreased when hindcast year was long before the model year (Xiao et al., 2018). Therefore, we did not use the models of 2014 to 2017 to estimate the hindcast PM2.5.

Unlike historical estimates from 2004 to 2012, we have sufficient ground-monitored PM2.5 data to develop statistical models after 2013, which allows us to estimate daily PM2.5 concentrations accurately. Therefore, we developed a separate PM2.5–AOD statistical model for each year of 2014–2017 to estimate the spatially resolved (0.1∘ resolution) PM2.5 concentrations. To keep satellite PM2.5 estimates of 2014–2017 consistent with our previous satellite PM2.5 dataset, we used the same method as described in our previous study (Ma et al., 2016). The data, model development, and model validation are briefly introduced as follows.

The data used in this study include ground-monitored PM2.5 concentrations (µg m-3), Aqua MODIS Collection 6 Dark Target (DT) AOD and Deep Blue (DB) AOD data, planetary boundary layer height (PBLH, 100 m), wind speed (WS, m s-1) at 10 m above the ground, mean relative humidity in PBL (RH_PBLH,  %), surface pressure (PS, hPa), precipitation of the previous day (Precip_Lag1; mm), MODIS active fire spots, urban cover ( %), and forest cover ( %). ground-monitored PM2.5 data were collected from the China Environmental Monitoring Center (CEMC), environmental protection agencies of Hong Kong and Taiwan. Figure 1 shows the ground PM2.5 monitors used in this study. AOD data were downloaded from the Level 1 and Atmospheric Archive and Distribution System (https://ladsweb.modaps.eosdis.nasa.gov/, last access: 29 March 2019). Meteorological data were extracted from Goddard Earth Observing System Data Assimilation System GEOS-5 Forward Processing (GEOS 5-FP) meteorological data (ftp://rain.ucis.dal.ca, last access: 29 March 2019). MODIS fire spots were from the NASA Fire Information for Resource Management System (https://earthdata.nasa.gov/earth-observation-data/near-real-time/firms, last access: 29 March 2019). Land use information were downloaded from Resource and Environment Data Cloud Platform of Chinese Academy of Science (http://www.resdc.cn/data.aspx?DATAID=184, last access: 29 March 2019).

Previous studies have shown the data quality issue of ground PM2.5 measurements from the CEMC network (Liu et al., 2016b; Rohde and Muller, 2015). We performed the data screening procedure before model fitting. Abnormal values (extreme high or extreme low values for a site compared with its neighboring sites, repeated values for continuous hours, etc.) were deleted before model fitting. We required at least 20 hourly records to calculate the daily average PM2.5 concentrations. DT and DB AOD were combined using an inverse variance weighting method to improve the spatial coverage of AOD data (Ma et al., 2016). These combined AOD data have shown good consistency (R2=0.8, mean bias =0.07) with ground AOD measurements from the Aerosol Robotic Network (AERONET) (Ma et al., 2016). All data were assigned to a predefined 0.1∘ grid. Then all of the variables were matched by grid cell and day of the year (DOY) for model fitting.

A two-stage statistical model was developed for each year separately from 2014 to 2017. The first-stage linear mixed effects (LME) model included day-specific random intercepts and slopes for AOD, season-specific random slopes for meteorological variables, and fixed slope for precipitation and fire spots. The model structure of first-stage model is shown as follows: PM2.5,st=μ+μ′+β1+β1′AODst+β2+β2′WSst+β3+β3′PBLHst+β4+β4′PSst+β5+β5′RH_PBLHst+β6Precip_Lag1st+β7Fire_spotsst+ε1,stμ′β1′∼N0,0,Ψ11+ε2,sjβ2′β3′β4′β5′∼N0,0,0,0,Ψ2, where PM2.5,st is ground PM2.5 measurements at grid cell s on DOY t; AODst is DT-DB merged AOD; WSst, PBLHst, PSst, RH_PBLHst, and Precip_Lag1st are meteorological variables; Fire_spotsst is the fire count; μ and μ′ are the fixed and day-specific random intercepts, respectively; β1–β7 are fixed slopes; β1′ is the day-specific random slope for AOD; β2′–β5′ are the season-specific random slopes for meteorological variables; ε1,st is the error term at grid cell s on DOY t;ε2,sj is the error term at grid cell s in season j; Ψ1 and Ψ2 are the variance–covariance matrices for the day- and season-specific random effects, respectively. The first-stage model was fitted for each province separately. We created a buffer zone for each province to include data with at least 3000 data records and at least 300 d. We averaged overlapped predictions from neighboring provinces to generate a smooth national PM2.5 concentration surface.

The second-stage generalized additive model (GAM) established the relationship between the residuals of the first-stage model and smooth terms of geographical coordinates, forest and urban cover. PM2.5_residst=μ0+s(X,Y)s+s(ForestCover)s2+s(UrbanCover)s+εst, where PM2.5_residst is the residual of the first-stage model at grid cell s on DOY t, μ0 is the intercept, s(X,Y)s is the smooth term of the coordinates of the centroid of grid cell s, s(ForestCover)s and s(UrbanCover)s are the smooth functions of forest cover and urban area for grid cell s, and εst is the error term.

To evaluate the model over-fitting, 10-fold cross-validation (CV) was used; that is, the model could have better prediction performance in the model fitting dataset than the data, which are not included model fitting. In 10-fold CV, all samples in the model dataset are randomly and equally divided into 10 subsets. One subset was used as testing samples and the rest of the subsets are used to fit the model. This process was repeated for 10 rounds until each subset was used for testing for once. Statistical indicators of coefficient of determination (R2), mean prediction error (MPE), and root mean squared prediction error (RMSE) were calculated and compared between model fitting and CV to assess model performance and over-fitting.

Monthly mean PM2.5 concentrations for each grid cell were calculated to perform the time series analysis. Following our previous study (Ma et al., 2016), we required at least six daily PM2.5 predictions in each month to calculate the monthly mean PM2.5. We deseasonalized the monthly PM2.5 time series by calculating the monthly PM2.5 anomaly time series for each grid cell to remove the seasonal effect. The PM2.5 trend for each grid cell was calculated using least squares regression (Weatherhead et al., 1998): (PM2.5)anomaly,s,m=(PM2.5)s,m-(PM2.5)s,j‾3m=1,2,3,…,M;j=1,2,3,…,12,(PM2.5)anomaly,s,m=μ+β×m+ε;4m=1,2,3,…,M, where (PM2.5)anomaly,s,m is the PM2.5 anomaly at grid cell s for month m during the calculating period; (PM2.5)s,m is the estimated PM2.5 concentration at grid cell s for month m; m is the month index and M is the total number of months during the calculating period (2004–2017, M=168); (PM2.5)s,j‾ is the 14-year average PM2.5 concentration of the month to which month m belongs (j=1 for January, j=2 for February, … , etc.); μ is the intercept; β is the slope, which is also the trend of PM2.5 (µg m-3 month-1); and ε is the error term. The annual PM2.5 trend (µg m-3 year-1) is 12×β. A t test was used to obtain the statistical significance of the trends. This method has been successfully applied to trend analyses of monthly mean PM2.5 and AOD anomaly time-series data (Hsu et al., 2012; Boys et al., 2014; Zhang and Reid, 2010; Xue et al., 2019). We analyzed the PM2.5 trend for different periods to examine the effects of air pollution control policies on PM2.5 pollution improvement.

Table S1 in the Supplement summarized the statistics of all variables for the modeling dataset from 2014 to 2017. Overall, there are 95 649, 110 805, 113 490, and 123 652 matchups for the model fitting datasets for years of 2014, 2015, 2016, and 2017, respectively. The average PM2.5 concentration decreases year by year, from 65.66 µg m-3 in 2014 to 48.32 µg m-3 in 2017. Correspondingly, the average AOD also shows a decreasing trend from 0.67 in 2014 to 0.50 in 2017.

Figure 2 shows the model fitting and cross-validation results for each year's model. The model fitting R2 ranges from 0.75 (2015) to 0.80 (2017) and CV R2 ranges from 0.72 (2015) to 0.77 (2017), which is similar to the 2013 model (0.82 for model fitting and 0.79 for CV) developed in our previous study (Ma et al., 2016). The model prediction accuracy is different among years, which is consistent with previous studies. Hu et al. (2014) studied the 10-year spatial and temporal trends of PM2.5 concentrations in the southeastern US from 2001 to 2010. They developed a separate two-stage statistical model for each year and found the CV R2 ranged from 0.62 in 2009 to 0.78 in 2005 and 2006. Kloog et al. (2011, 2012) conducted two studies in the northeastern US and also found that the validation R2 varied among years. Compared to the model fitting R2, the CV R2 only decreases by 0.02 in 2016 and 0.03 in 2014, 2015, and 2017, showing that our models were not substantially over-fitted. For the monthly mean concentrations calculated from at least six daily PM2.5 predictions, the validation R2 values range from 0.75 to 0.81 (Fig. 3). The results show that the overall prediction accuracy of the models from 2014 to 2017 is satisfying.

The fixed effects, model fitting, and CV results of the first-stage LME model for each province are shown in Tables S2–S5. AOD is the only variable that was statistically significant in all provincial models for all years (p<0.05). Wind speed, relative humidity, precipitation, and fire spots were significant in most provincial models. The CV R2 varies for different provinces and different years. The CV R2 values range from 0.61 in Xinjiang to 0.77 in Heilongjiang for 2014, from 0.34 in Xinjiang to 0.76 in Hebei for 2015, from 0.44 in Tibet to 0.77 in Jiangsu for 2016, and from 0.38 in Xinjiang to 0.79 in Sichuan for 2017. We also fitted a first-stage LME model for all of China. Results show that the overall CV R2 values for the first-stage LME model dropped to 0.57, 0.52, 0.56, and 0.54, for 2014, 2015, 2016, and 2017, respectively. Therefore, fitting the first-stage model for each province separately can greatly improve the prediction accuracy.

A potential source of uncertainties in statistical models is the uneven spatial distribution of ground PM2.5 monitors. The CEMC air quality network mainly covers large urban centers with very limited site coverage in rural areas, especially in western part of the country. Since it requires a large amount of ground-measured PM2.5 data to develop satellite-based statistical model, this bias cannot be avoided. Despite this limitation, high model performances, which are much better than those using the scaling method, have been achieved in this study and previous similar studies (Zheng et al., 2016; Huang et al., 2018; Xue et al., 2019). For example, Geng et al. (2015) estimated long-term PM2.5 concentrations in China using a scaling method and found the validation R2 of PM2.5 predictions was 0.72 compared to the 5-month averaged ground PM2.5 concentrations for January–May 2013. A global study of PM2.5 estimates combining scaling and statistical methods shows that their validation R2 of long-term average PM2.5 was 0.67 for their first-stage scaling method (van Donkelaar et al., 2016).

Figure 4 shows that spatial distribution characteristics of annual mean PM2.5 concentrations are similar among the years from 2004 to 2017. The most polluted area was the North China Plain (including the south of the Jingjinji region, Henan, and Shandong Provinces), which was also the largest polluted area. The Sichuan Basin (including eastern Sichuan and western Chongqing) is another polluted area. The cleanest areas were mainly located in Tibet, Hainan, Taiwan, Yunnan, and the north of Inner Mongolia. The spatial distributions of satellite-derived PM2.5 concentrations from 2013 to 2017 are consistent with the spatial characteristics of ground-monitored PM2.5 (Fig. S2).

Figure 5 shows the spatial distributions of PM2.5 trends and significance levels in China from 2004 to 2017. Overall, the PM2.5 pollution level of most areas in China showed a decreasing trend (p<0.05). Figure 6 and Table 2 show that the overall trends of 2004–2017 for all of China and the Jingjinji, Yangtze River Delta (YRD), and Pearl River Delta (PRD) regions were -1.27, -1.55, -1.60, and -1.27 µg m-3 year-1 (all p<0.001), respectively. Back to Fig. 4, we can see that the decrease in PM2.5 mainly happened after 2013. PM2.5 concentrations showed an obvious increase from 2004 to 2007. The area with PM2.5 concentrations higher than 100 µg m-3 continuously expanded during this period. From 2008 to 2013, the pollution levels plateaued in most areas. After 2013, the PM2.5 concentrations obviously decreased.

To assess the effect of ECER policy during the 11th FYP, we calculated the trends of PM2.5 for 2005–2010, 2004–2007, and 2007–2010 for each grid cell (Fig. 7).

Compared to the base year (2005) of the 11th FYP period, the overall PM2.5 pollution of 2010 did not show obvious change. Some of the area showed decreasing trends (Fig. 7a) but the trends were insignificant (Fig. 7b). Some regions (Shandong, Henan, and Jiangsu provinces and northeastern China) showed a slight increasing trend (∼1–2 µg m-3 year-1, p<0.001). Overall, the trends for all of China and the Jingjinji, YRD, and PRD regions were all insignificant (0.41, 0.26, 0.61, and -1.26 µg m-3 year-1, and all p>0.1) during the 11th FYP period.

However, when separating this period into two periods, we can see that before 2007, the PM2.5 concentrations generally had significant increasing trends (Fig. 7c, d), especially in the south of the Jingjinji region and in Henan, Shandong, and Hubei provinces. The overall trends for all of China and the Jingjinji region are 1.88 (p<0.001) and 3.14 µg m-3 year-1 (p<0.005) (Table 2). The trends for YRD and PRD regions are insignificant. During the 10th FYP period, China missed the emission control goals. The emission of sulfur dioxide increased by ∼28 % (Xue et al., 2014; Schreifels et al., 2012). The 11th FYP for National Economic and Social Development of China released in 2006 proposed the ECER goals. However, China did not achieve the annual goal in 2006. These could explain the increasing trend of PM2.5 during 2004–2007.

After that, China released the Comprehensive Working Plan on ECER (http://www.gov.cn/zwgk/2007-06/03/content_634545.htm, last access: 29 March 2019) in 2007 to strengthen the ECER measures. Major control measures included (Schreifels et al., 2012) implementing flue gas desulfurization for coal-fired power plants, closing inefficient and backward production centers, implementing energy conservation projects, increasing the pollution levy for SO2 emission, recommending baghouse dust filters for industrial soot and dust emission control, etc. As a result, great achievements had been made at the end of 11th FYP (Schreifels et al., 2012; Zhou et al., 2015): total emission of SO2 decreased by ∼14 % compared to the level of 1995; approximately 86 % of the power plants were installed with desulfurization facilities in 2010 compared to 14 % in 2005; nearly 80 GW of small coal-fired power units were closed during 2006–2010; soot emission of coal-fired power plants in 2010 was reduced by 55.6 % compared with that in 2005, etc.

Due to these control measures, the increasing trend of PM2.5 pollution was suppressed after 2007. PM2.5 concentrations of central and southern China decreased significantly, with highest trend of around -9 µg m-3 year-1 (Fig. 7e, f), p<0.01). The south of Jingjinji region and Henan, Shandong, and Hubei provinces, which had significantly increased before 2007, showed insignificant trends (Fig. 7f, p>0.05). Table 2 shows that the overall PM2.5 trend for all of China was -0.56 µg m-3 year-1 with marginal significance (p=0.053). Overall trends for the Jingjinji and YRD regions were not significant during the latter half of the 11th FYP period. And PM2.5 concentrations in the PRD region had a big drop (-4.81 µg m-3 year-1, p<0.001). Results show that although air pollution control policies of the 11th FYP were not designed for PM2.5 prevention and control, they still had co-benefits on PM2.5 pollution control. There were two main reasons. First, SO2 is the precursor gas of sulfate. Previous studies have shown that sulfate was the major component of PM2.5 during the 11th FYP period (Li et al., 2009, 2010; Pathak et al., 2009). The reduction of SO2 could therefore contribute to the suppression of increasing PM2.5 pollution. Second, the control of industrial dust and soot, which include a portion of primary PM2.5 (Yao et al., 2009), also contributed to the PM2.5 pollution reduction.

Figure 8a and b show that most of the areas of China show a significant decreasing trend during the 12th FYP period. PM2.5 concentrations of all of China, Jingjinji, and YRD dropped by 2.89, 3.63, and 3.33 µg m-3 year-1 (p<0.001). When considering the years from 2010 to 2013, although the overall trend of all of China was -1.03 µg m-3 year-1 (p<0.05, Table 2), the decreasing trend mainly happened in Xinjiang and central Inner Mongolia. The deserts in Xinjiang and Inner Mongolia are the major sources of dust pollution in China. A recent study showed that dust is the largest contributor to PM2.5 over this region (Philip et al., 2014). The change in natural dust in desert areas may be the major contributor to the decreasing trend of PM2.5 during 2010–2013. Most of the polluted area in China did not show obvious change (Fig. 8c and d). As we mentioned above, The ECER policy during the 12th FYP period was basically the extension of the policy in the 11th FYP, which mainly focused on emissions reduction. Due to the development of the social economy, the ECER policy has shown limitations in PM2.5 reduction. PM2.5 is a kind of composite pollutant and its constituents includes primary particles and secondary particles such as sulfate, nitrate, ammonium, organic carbon, elemental carbon, etc. With the deepening of SO2 and industrial dust and soot emission reduction, their contributions to PM2.5 pollution control would be reduced, although the 12th FYP on Environmental Protection also proposed a 10 % reduction of NOx from 2010 to 2015. However, along with economic growth in China, the benefits of emission control for a single pollutant could be offset by increased energy usage. Considering the complicated PM2.5 compositions, comprehensive and coordinated control measures for multiple pollutants are urgently needed.

Therefore, China issued the 12th FYP on APPC-KR in late 2012, which is the first special plan for air pollution prevention and control and focused on air quality improvement. APPC-KR proposed a series of key projects which included 477 SO2 treatment projects, 755 NOx treatment projects, 10 073 industrial soot and dust treatment projects, 1311 VOC treatment projects in key industrial sectors, 281 vapor recovery projects for oil and gas, 188 yellow-sticker vehicle elimination projects, 192 fugitive dust comprehensive treatment projects, and 122 capacity building projects. An English translation of APPC-KR and its key projects has been prepared by the Clean Air Alliance of China (CAAC) and can be found elsewhere (http://www.cleanairchina.org/product/6347.html, last access: 29 March 2019) (CAAC, 2013c, a).

In addition, in 2012, China issued a new air quality standard, i.e., the National Ambient Air Quality Standard of China (NAAQS) (GB 3095-2012). Compared with the former NAAQS (GB 3095-1996) issued in 1996, this new standard incorporated PM2.5 as a major control pollutant. According to GB 3095-2012, the Level 1 annual mean standard of PM2.5 is 15 µg m-3, which is assigned for protecting the air quality of natural reserves and scenic areas and is equivalent to the World Health Organization (WHO) Air Quality Interim Target-3 (IT-3). The Level 2 standard of 35 µg m-3 is designated for residential, cultural, industrial, and commercial areas, which is equivalent to WHO Air Quality Interim Target-1 (IT-1). Meanwhile, a comprehensive real-time air quality monitoring network covering 74 major Chinese cities was established in late 2012.

The implementation of APPC-KR, together with the implementation of APPC-AP starting from 2013 (shown in the following section), led to dramatic drops in PM2.5 concentrations in China after 2013. Table 3 shows PM2.5 concentration improvement goals and final accomplishments for key regions (see Fig. S1) of the 12th FYP on APPC-KR calculated from satellite PM2.5. Results show that all key regions accomplished the goals except for Yinchuan. The changes in population-weighted averages also show similar results. Overall, the 12th FYP on APPC-KR accomplished its air pollution control goals. And the decrease in PM2.5 concentrations was mainly attributable to the decrease after 2013.

China issued the APPC-AP (2013–2017) in late 2013, which further strengthened the air pollution control measures and air quality improvement goals. The air pollution control measures included 10 categories:

increase effort for comprehensive pollution control, reduce emissions of multi-pollutants;

Detailed measures of the APPC-AP can be found in English translation at http://www.cleanairchina.org/product/6349.html (last access: 29 March 2019) (CAAC, 2013b). To ensure that APPC-AP goals could be accomplished, China adopted a temporary measure in 2017, i.e., the intensified supervision for air pollution control in Jingjinji and the surrounding area (http://www.gov.cn/hudong/2017-07/14/content_5210588.htm, last access: 29 March 2019). There had been great achievements at the end of 2017. For example (Zheng et al., 2018), 71 % of the power plants met the ultralow emission levels; the average efficiency of coal-fired power units decreased from 321 g.c.e. kWh-1 in 2013 to 309 g.c.e. kWh-1 in 2017; non-methane volatile organic compound (NMVOC) emissions were cut down by 30 % through the implementation of a leak detection and repair (LDAR) program for the petrochemical industry; all coal boilers smaller than 7 MW in urban areas were shut down; and all “yellow label” vehicles (referring to gasoline and diesel vehicles that fail to meet Euro 1 and Euro 3 standards, respectively) were eliminated by the end of 2017, to name a few.

The implementation of APPC-AP, together with the 12th FYP on APPC-KR, had led to a dramatic drop in PM2.5 concentrations from 2013 to 2017 (Fig. 8e and f). PM2.5 trends of 2013–2017 for all of China, Jingjinji, YRD, and PRD regions were -4.27, -6.77, -6.36, and -2.11 µg m-3 year-1 (all p<0.05), respectively (Table 2). This is comparable to a recent study (Silver et al., 2018), which found that the median trend in annual mean PM2.5 concentration across all ground air pollution monitoring stations is -3.4 µg m-3 year-1 from 2015 to 2017. Table 4 shows PM2.5 concentration improvement goals and final accomplishments for APPC-AP. The goals required that PM2.5 concentrations in Jingjinji, YRD, and PRD regions in 2017 should decrease by 25 %, 20 %, and 15 % compared to 2012, and the annual mean PM2.5 of Beijing should reach around 60 µg m-3. Since there were no ground measurements in 2012, the Ministry of Ecology and Environment (MEE) of China used 2013 as the base year to assess the performance of APPC-AP (http://www.mee.gov.cn/gkml/sthjbgw/stbgth/201806/t20180601_442262.htm, last access: 29 March 2019). To maintain consistency with the official performance assessment, we also used 2013 as the base year. Results show that the arithmetic average of satellite PM2.5 concentrations for Jingjinji, YRD, and PRD regions decreased by 36.9 %, 37.1 %, and 14.0 %, respectively, and annual mean PM2.5 of Beijing was 44.67 µg m-3 in 2017. From the view of satellite, Jingjinji, YRD, and Beijing accomplished their goals, and PRD was very close to the goal. However, the pollution level was still higher than WHO Air Quality IT-1 level and NAAQS (GB 3095-2012) Level 2 annual PM2.5 standards (both 35 µg m-3).

According to the official results of APPC-AP performance assessment (Table 4), PM2.5 of Jingjinji, YRD, and PRD regions decreased by 39.6 %, 34.3 %, and 27.7 %, respectively. And annual mean PM2.5 of Beijing was 58 µg m-3 in 2017. Compared to the arithmetic average satellite PM2.5, the populations weighted average results (Table 4) are closer to the official results. The main reason is that official performance assessment used ground measurements. However, the spatial distribution of ground monitors is uneven. Most of the sites are distributed in populated urban areas and only a few are located in rural areas. Compared to ground monitors, satellite remote sensing has more comprehensive spatial coverage. Figure S3 shows the spatial distribution of satellite and ground PM2.5 concentrations of 2017 in Beijing. It can be seen that the ground monitors are clustered in polluted urban centers. The cleaner north and northwest of Beijing have few sites. Thus the population-weighted results of satellite PM2.5 are closer to the official results, but still have differences. Since satellites have better spatial coverage than ground monitors, satellite PM2.5 can better represent the spatial variation of PM2.5 pollution. The population-weighted average satellite PM2.5 can better represent the health impact of PM2.5 pollution. When using ground monitors to calculate the regional mean concentrations, the weights of area and population for each site should be considered.