Exposure and Inequality of PM2.5 Pollution to Chinese Population: A Case Study of 31 Provincial Capital Cities from 2000 to 2016

Fine particulate matter (PM2.5) exposure has been linked to numerous adverse health effects, with some disadvantaged subgroups bearing a disproportionate exposure burden. Few studies have been conducted to estimate the exposure and inequality of different subgroups due to a lack of adequate characterization of disparities in exposure to air pollutants in urban areas, and a mechanistic understanding of the causes of these exposure inequalities. Based on a long-term series of PM2.5 concentrations, this study analyzed the spatial and temporal characteristics of PM2.5 in 31 provincial capital cities of China from 2000 to 2016 using the coefficient of variation and trend analyses. A health risk assessment of human exposure to PM2.5 from 2000 to 2016 was then undertaken. A cumulative population-weighted average concentration method was applied to investigate exposures and inequality for education level, job category, age, gender and income population subgroups. The relationships between socioeconomic factors and PM2.5 exposure concentrations were quantified using the geographically and temporally weighted regression model (GTWR). Results indicate that the PM2.5 concentrations in most of the capital cities in the study experienced an increasing trend at a rate of 0.98 μg m−3 per year from 2000 to 2016. The proportion of the population exposed to high PM2.5 (above 35 μg m−3) increased annually, mainly due to the increase of population migrating into north, east, south and central China. The higher educated, older, higher income and urban secondary industry share (SIS) subgroups suffered from the most significant environmental inequality, respectively. The per capita GDP, population size, and the share of the secondary industry played an essential role in unequal exposure to PM2.5.


Introduction
Long-term exposure to high concentrations of urban air pollution not only introduces negative effects on human health, but also threatens sustainable development of the economy and society [1][2][3][4][5][6]. In particular, inequality in exposure to air pollution has recently become an important and timely issue because the health risk of some effects associated with worse conditions of environmental stressors such as air pollution, noise or excessive heat is high for socioeconomically disadvantaged groups [7,8]. This is particularly important in provincial capital cities with strong socioeconomic and air quality gradients, where a local action plan on air pollution can reduce or unintentionally exacerbate inequalities [9,10].
A quantified study on the spatio-temporal variation of the population exposed to different PM 2.5 concentrations is essential to investigate the inequality of PM 2.5 exposure in different subgroups and to screen out the main drivers [11]. PM 2.5 concentration data are a prerequisite for the study to be conducted, and the main methods currently include retrieval of Aerosol Optical Depth as well as actual monitoring [12,13]. In recent years, many comprehensive exposure models have been used to simulate the distribution of populations exposed to different levels of air pollution globally and regionally [14,15]. These models explain the population exposure caused by air pollution to a certain extent but cannot analyze the impact of PM 2.5 contribution of population change affected by pollution risk. In addition, PM 2.5 concentrations were significantly related to many socioeconomic factors [7,16,17]. Various spatial econometric method approaches (e.g., spatial lag model, spatial error model, or dynamic spatial panel data model) have been used to explore the relationship between PM 2.5 concentration and socioeconomic factors [18][19][20]. However, these studies only consider the time dimension and space dimension separately. It is necessary to consider both time and space dimensions. The geographically and temporally weighted regression (GTWR) model incorporates the time effect based on the geographically weighted regression (GWR) model, and thereby the possibility of assessing parametric heterogeneity both in time and space dimensions is provided [21]. Previous studies have shown that the GTWR model has more explanatory power for the spatial and temporal changes in the relationship between air pollution and socioeconomic factors [16,22,23].
Large numbers of studies over the past several decades have been conducted to investigate the inequality of exposure to air pollution among the subgroups of distinct ages, income and ethnicity, which can potentially contribute to environmental health disparities [7,[24][25][26][27][28][29][30][31]. For example, based on a 1 km 2 model, PM 2.5 and NO 2 concentrations in Massachusetts over eight years and Census demographic data, Rosofsky firstly quantified inequality between sociodemographic groups exposed to PM 2.5 and NO 2 in Massachusetts [29]. They found that certain vulnerable populations, such as blacks, Hispanics and economically deprived people, suffered from inequitable exposures and effects because they were more likely to live in areas with higher concentrations of air pollutants. Bell and Ebisu also investigated the inequality of environmental exposures to PM 2.5 components that differed by race or ethnicity, age, and socioeconomic status (SES). The results indicated Hispanics, young persons, and lower SES had the highest exposures, respectively. Other researchers also show low-income earners or minorities are exposed to relatively high concentrations of air pollutants [8,32]. Little research on exposure inequality has been carried out in China. Sun et al. found that the poorer people were more likely to suffer from air pollution exposure inequality than the richer people because they invested less in self-protection products [30]. Similarly, the research on disproportionate exposure to PM 2.5 conducted in Beijing showed that for the highest 1-h concentration, older people (age ≥ 60) and residents with tertiary education were disproportionately exposed to the most PM 2.5 [28].
However, previous studies have mostly assessed short-term exposure for the entire population, estimated inequality focused on race, ages and income subgroups at a single city scale, ignored the inequality caused by the type of work, gender and education level. In addition, few researchers investigated the drivers of changes in the spatio-temporal distribution of exposure inequality to air pollutants. Therefore, in this study, PM 2.5 pollution spatio-temporal trends in 31 provincial capital cities in China from 2000 to 2016 were investigated. We used the Exposed Population Contribution Analysis Model (EPCAM) to quantitatively study the populations exposed to different PM 2.5 concentrations. Then, we combined PM 2.5 concentration with demographic data to estimate cumulative population weighted average concentrations (CPWAC) in different subgroups of the capital cities and compared the exposure inequality of these subgroups. Additionally, the GTWR model was constructed to quantify the impact of the key socioeconomic factors on PM 2.5 pollution exposure inequality in time and space dimensions. These findings can provide scientific methods for air pollution exposure and long-term health risk assessments.

Study Areas
This study analyzed the PM 2.5 pollution data and socioeconomic data in 31 provincial capital cities of China. These cities are divided into seven regions, including the central, east, north, northeast, northwest, south and southwest China ( Figure 1).

PM 2.5 Concentration
This study analyzed the annual mean concentrations of PM 2.5 in each provincial capital city from 2000 to 2016. In January 2013, the China National Environmental Monitoring Centre started publishing real-time hourly concentrations of PM 2.5 monitoring data [33,34]. Therefore, this study processed the 0.01 • × 0.01 • annual mean global GWR-adjusted PM 2.5 estimates to obtain the annual average PM 2.5 concentrations of 31 provincial capital cities in China from 2000 to 2013 using the zonal statistics tool in ArcMap provided by ArcGIS Desktop 10.6. The estimates were calculated by Donkelaar et al., who estimated global PM 2.5 concentrations using information from satellite-, simulation-and monitor-based sources by applying a Geographically Weighted Regression (GWR) to global geophysically based satellite-derived PM 2.5 estimates. The resultant PM 2.5 estimates were highly consistent (R 2 = 0.81; the slope of 0.82) with out-of-sample cross-validated PM 2.5 concentrations from monitors [35].
The annual mean levels of PM 2.5 from 2014 to 2016 were collected from the National Urban Air Quality Real-time Publishing Platform.

Socioeconomic Data
The socioeconomic data and the population data were collected from the National Bureau of Statistics of the People's Republic of China (NBSPRC) and the Easy Professional Superior (EPS) data platform. Populations were considered as the sum of people in a given city. Therefore, to assess exposure, we only used the population at the city level corresponding to the mean PM 2.5 concentrations. The main socioeconomic data include urban area (UA), natural population growth rate, total urban population (UP), population density (PD), urban secondary industry share (SIS), real gross domestic product per capita (GDPPC), education attainment, urban per capita disposable income (UPCDI), rural per capita disposable income (RPCDI), job category, age and gender. We categorized population characteristics into the groups shown in Table 1. The sources and download links of the above multi-source datasets are provided in Table 2.

Trend Analysis
Trend analysis is commonly used in temporal dynamic analyses to explore interannual variation characteristics. In this study, a long sequence of PM 2.5 change trends was quantitatively analyzed based on the trend analysis method. Tendencies of PM 2.5 variations can be determined as follows: where PM 2.5 is the PM 2.5 concentrations in each city, n is the timespan, and i is the time unit.

Coefficient of Variation
The spatial heterogeneity of observed concentrations in each city was investigated using a coefficient of variation (CV), and it was used to describe the degree of the spatial variations of air pollutant concentrations in Chinese capital cities [36], expressed by: where STD and x represent standard deviation and mean value.

Contribution Analysis
The increased population exposed to the PM 2.5 risk in each provincial capital city of China was derived from population migration and natural growth [37], which could be calculated by using the population amount and annual natural population growth rate (α i ), and defined as follows: where ∆P represents the change of population that is exposed to PM 2.5 pollution, Pn and Pm represent the population migration and natural growth, respectively. α i represents the annual natural population growth rate.

Cumulative Population Weighted Average Concentrations (CPWAC)
The cumulative annual mass concentrations of air pollutants in the 31 provincial capital cities were represented by the cumulative annual population weighted average concentrations (CPWAC) in all years within each city.
where C is the mass concentration of the air pollutants. Pop is the amount of population. The suffixes i represent different years. n is the number of all years within each city.

Geographically and Temporally Weighted Regression (GTWR) Model
In this study, the impact of socioeconomic factors on PM 2.5 concentration was simulated by Geographically and Temporally Weighted Regression model (GTWR). The GTWR captures spatio-temporal heterogeneity based on a weighting matrix referencing both spatial and temporal dimensions [38]. In this study, a GTWR model was fitted using the following structure: where PM 2.5i is the annual surface PM 2.5 concentration of sample i; (µ i , υ i , t i ) is the longitude, dimension and time of i. β 0 (µ i , υ i , t i ) represents the intercepts of i. β 1 -β 6 are the locationtime-specific slopes for UR, UP, GDPPC, SIS, PD, and UPCDI, respectively. The procedure of GTWR model calculation was performed in the GWmodel S (BETA) software (http: //gwmodel.whu.edu.cn (accessed on 30 May 2021)), where the type of spatial kernel function selected during the model implementation was Bisquare, the type of bandwidth search was AICc, and the type of distance criterion was CRS type. Meanwhile, the spatiotemporal heterogeneous characteristics of the independent variables were examined with the OLS model in the advanced settings of the GTWR model.  [20]. Specifically, the "Air Pollution Prevention and Control Action Plan" set definite targets for decreasing PM 2.5 concentrations in most eastern regions and other heavily polluted regions. Additionally, the changing trends were different from city to city, with strong spatial variability (Appendix A). Spatially, in comparison to other cities, towns in eastern and northern China had greater magnitudes of decreases in PM 2.5 concentrations (

The Populations Exposed to Different PM2.5 Concentrations
Based on population data statistics, we quantified the temporal changes in the proportions of populations exposed to different PM2.5 concentrations in provincial capital cities of China from 2000 to 2016 ( Figure 3). In general, the proportion of the total population exposed to PM2.5 less than 35 μg m −3 decreased from 45.14% in 2000 to 6.44% in 2008 and 6.35% in 2016, whereas the proportion exposed to PM2.5 over 35 μg m −3 increased from 54.86% in 2000 to 93.56% in 2008 and 93.65% in 2016. These results showed that an increasing fraction of the population suffered from increasingly serious PM2.5 pollution and that fewer people experienced light PM2.5 pollution. Therefore, over the past 17 years, an in-

The Populations Exposed to Different PM 2.5 Concentrations
Based on population data statistics, we quantified the temporal changes in the proportions of populations exposed to different PM 2.5 concentrations in provincial capital cities of China from 2000 to 2016 ( Figure 3). In general, the proportion of the total population exposed to PM 2.5 less than 35 µg m −3 decreased from 45.14% in 2000 to 6.44% in 2008 and 6.35% in 2016, whereas the proportion exposed to PM 2.5 over 35 µg m −3 increased from 54.86% in 2000 to 93.56% in 2008 and 93.65% in 2016. These results showed that an increasing fraction of the population suffered from increasingly serious PM 2.5 pollution and that fewer people experienced light PM 2.5 pollution. Therefore, over the past 17 years, an increasing proportion of the population has mainly concentrated in those areas with concentrations of 35-75 µg m −3 of PM 2.5 . In summary, the most significant change was the increasing population susceptible to experiencing high PM 2.5 concentrations, and a declining proportion of the population experiencing low PM 2.5 concentrations. In addition, the population exposed to PM 2.5 over 35 µg m −3 will increase following the increase of PM 2.5 pollution in the future years ( Figure 3).  From 2000 to 2016, a total of 3561.62 million people were exposed to PM2.5 over 35 μ m −3 (i.e., potential risk) while only 406.55 million were exposed to PM2.5 less than 35 μ m −3 in 31 provincial capital cities of China. All populations of Beijing, Tianjin, Taiyua Shijiazhuang, Chengdu, Chongqing, Wuhan, Zhengzhou, Xi'an, Hangzhou, Nanjing, an Jinan were exposed to the potential risk. More than 90% of the population in the provinc capital cities of Hefei (95%) and Shanghai (92%) in east China, Urumqi (95%) in northwe China, Guangzhou (95%) in south China, and Changsha (95%) in central China, were e posed to the potential risk. In contrast, the population exposed to the potential risk Hohhot, Xining and Yinchuan located in northwest China was less than 30%. In additio all populations of Haikou in south China along with Kunming and Lhasa in southwe China were exposed to PM2.5 less than 35 μg m −3 . The population exposed to the potent risk was increased in all provincial capital cities except Chengdu during the study period  From 2000 to 2016, a total of 3561.62 million people were exposed to PM 2.5 over 35 µg m −3 (i.e., potential risk) while only 406.55 million were exposed to PM 2.5 less than 35 µg m −3 in 31 provincial capital cities of China. All populations of Beijing, Tianjin, Taiyuan, Shijiazhuang, Chengdu, Chongqing, Wuhan, Zhengzhou, Xi'an, Hangzhou, Nanjing, and Jinan were exposed to the potential risk. More than 90% of the population in the provincial capital cities of Hefei (95%) and Shanghai (92%) in east China, Urumqi (95%) in northwest China, Guangzhou (95%) in south China, and Changsha (95%) in central China, were exposed to the potential risk. In contrast, the population exposed to the potential risk in Hohhot, Xining and Yinchuan located in northwest China was less than 30%. In addition, all populations of Haikou in south China along with Kunming and Lhasa in southwest China were exposed to PM 2.5 less than 35 µg m −3 . The population exposed to the potential risk was increased in all provincial capital cities except Chengdu during the study periods. The exposed population of most provincial capital cities increased in three different periods, and the average population changes in 2000-2005, 2005-2010 and 2010-2016 were 2.90 million, 0.73 million and 0.47 million, respectively ( Figure A2, Appendix A). These results showed that the population exposed to PM 2.5 over 35 µg m −3 increased, but its growth rates declined. Spatially, the population exposure was mainly found in the east (57.04 million in 2016) and north (51.84 million in 2016).
The increased population exposed to the PM 2.5 risk in each city was derived from population migration and natural growth. As shown in Figure 4, the contributions of exposed population were increased, mainly due to the increase of migration of population in north, east, south and central China. Specifically, population sizes of two Chinese northern cities, Beijing and Tianjin, as well as one eastern city, Shanghai, in 2016 have increased by 8.16, 7.79 and 5.61 million, respectively, compared to 2000, with over 90% of population movement derived from migration. Likewise, from 2000 to 2016, the average increments of population exposed to PM 2.5 in 10 other cities exceed 1,000,000, including three western cities (Yinchuan, Urumqi and Xi'an), three eastern cities (Hefei, Hangzhou and Nanjing), three central cities (Zhengzhou, Changsha and Wuhan) and one southern city (Guangzhou), and above 60% of this increased population was from migration. These findings indicate that migration could promote the rise of the population exposed to PM 2.5 and increasing population size would aggravate PM 2.5 pollution in urban areas. In contrast, the contribution of the migrant population to the exposed population played a negative role in Harbin and Changchun in northeast China, Lanzhou, Xining and Hohhot in northwest China and Chongqing in southwest China (Figure 4). movement derived from migration. Likewise, from 2000 to 2016, the average increments of population exposed to PM2.5 in 10 other cities exceed 1,000,000, including three western cities (Yinchuan, Urumqi and Xi'an), three eastern cities (Hefei, Hangzhou and Nanjing), three central cities (Zhengzhou, Changsha and Wuhan) and one southern city (Guangzhou), and above 60% of this increased population was from migration. These findings indicate that migration could promote the rise of the population exposed to PM2.5 and increasing population size would aggravate PM2.5 pollution in urban areas. In contrast, the contribution of the migrant population to the exposed population played a negative role in Harbin and Changchun in northeast China, Lanzhou, Xining and Hohhot in northwest China and Chongqing in southwest China (Figure 4).

Exposure Inequality
The cumulative annual population weighted exposure concentrations to PM2.5 for each subgroup were calculated ( Figure 5). For all subgroups, the cumulative annual population weighted exposure concentrations of SIS subgroups (58.31 μg m −3 ) were the highest, while RPCDI subgroups (48.49 μg m −3 ) experienced the lowest exposure to PM2.5. Spe-

Exposure Inequality
The cumulative annual population weighted exposure concentrations to PM 2.5 for each subgroup were calculated ( Figure 5). For all subgroups, the cumulative annual population weighted exposure concentrations of SIS subgroups (58. 31   For the education subgroups, the estimated average exposure for people with tertiary education was the highest (58.36 μg m −3 ), followed by those with secondary (55.41 μg m −3 ) and primary (54.97 μg m −3 ) education (Figure 5a). This indicated that the exposure levels increased with education level improvement. However, among the job category subgroups, the exposure level for the professionals category (54.10 μg m −3 ) was the highest and less than the average for capital cities (2.67 μg m −3 lower). The population weighted exposure concentrations of PATH, CVS, PINBI and PAI in the subgroup of job categories to PM2.5 had similar exposure levels, except PAA categories, in which the average annual population weighted exposure concentrations were 53.69, 53.58, 53.21 and 53.15 μg m −3 , respectively (Figure 5b). For the age subgroups, the adults (20-59, 53.31 μg m −3 ) and older people (age ≥ 60, 53.81 μg m −3 ) suffered from greater exposure to PM2.5. Additionally, among age subgroups, the exposure concentrations of adolescents (5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19) were the lowest and slightly lower than the average of capital cities (4.42 μg m −3 lower) during the study period. The exposure concentrations of young children (age ≤ 4, 52.86 μg m −3 ) were also slightly lower than the average exposure level (Figure 5c). Differences in gender were indistinguishable, and the exposure concentrations of women were slightly higher than men's (0.11 μg m −3 higher) (Figure 5d).
The highest GDPPC (over 60,000 Yuan) subgroups were found to have the highest average exposure to PM2.5 concentration among all subgroups, with 5.01 μg m −3 higher than the capital cities annual average for PM2.5 exposures, while the lowest GDPPC category groups were exposed to the lowest PM2.5 concentrations (Figure 5e). This means that the population weighted exposure generally increased with an increasing of GDPPC. A similar inequality was found in income groups. Residents with an annual income of more than RMB 40,000 in urban areas had the highest average exposures, while those who earned more than RMB 10,000 per year in rural areas had the highest ones (Figure 5f and g). Annual average population weighted exposures to PM2.5 concentration for the highest income of UPCDI and RPCDI groups (greater than RMB 40,000 and RMB 10,000) were 0.03 and 2.40 μg m −3 , which were more than the capital cities average. Among the SIS subgroups, the lower SIS groups (less than 50%) had the lowest exposure levels (56.04 μg m −3 ) and were lower than the city average (0.73 μg m −3 lower) during the study period For the education subgroups, the estimated average exposure for people with tertiary education was the highest (58.36 µg m −3 ), followed by those with secondary (55.41 µg m −3 ) and primary (54.97 µg m −3 ) education (Figure 5a). This indicated that the exposure levels increased with education level improvement. However, among the job category subgroups, the exposure level for the professionals category (54.10 µg m −3 ) was the highest and less than the average for capital cities (2.67 µg m −3 lower). The population weighted exposure concentrations of PATH, CVS, PINBI and PAI in the subgroup of job categories to PM 2.5 had similar exposure levels, except PAA categories, in which the average annual population weighted exposure concentrations were 53.69, 53.58, 53.21 and 53.15 µg m −3 , respectively (Figure 5b). For the age subgroups, the adults (20-59, 53.31 µg m −3 ) and older people (age ≥ 60, 53.81 µg m −3 ) suffered from greater exposure to PM 2.5 . Additionally, among age subgroups, the exposure concentrations of adolescents (5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19) were the lowest and slightly lower than the average of capital cities (4.42 µg m −3 lower) during the study period. The exposure concentrations of young children (age ≤ 4, 52.86 µg m −3 ) were also slightly lower than the average exposure level (Figure 5c). Differences in gender were indistinguishable, and the exposure concentrations of women were slightly higher than men's (0.11 µg m −3 higher) (Figure 5d).
The highest GDPPC (over 60,000 Yuan) subgroups were found to have the highest average exposure to PM 2.5 concentration among all subgroups, with 5.01 µg m −3 higher than the capital cities annual average for PM 2.5 exposures, while the lowest GDPPC category groups were exposed to the lowest PM 2.5 concentrations (Figure 5e). This means that the population weighted exposure generally increased with an increasing of GDPPC. A similar inequality was found in income groups. Residents with an annual income of more than RMB 40,000 in urban areas had the highest average exposures, while those who earned more than RMB 10,000 per year in rural areas had the highest ones (Figure 5f,g). Annual average population weighted exposures to PM 2.5 concentration for the highest income of UPCDI and RPCDI groups (greater than RMB 40,000 and RMB 10,000) were 0.03 and 2.40 µg m −3 , which were more than the capital cities average. Among the SIS subgroups, the lower SIS groups (less than 50%) had the lowest exposure levels (56.04 µg m −3 ) and were lower than the city average (0.73 µg m −3 lower) during the study period (Figure 5h).

Economic Effectiveness of Exposure Inequality
PM 2.5 exposure was significantly related to many socioeconomic factors [10,17,24]. GTWR model (see Appendix A for model assessment process) was used to study the spatial heterogeneity of the relationship between socio-economic factors and PM 2.5 concentration. The results from the GTWR model showed that all variables, except for urban area, were significantly related to PM 2.5 concentrations. Among all the variables, urban secondary industry share (0.279) was the most important predictor, and population density (0.022) was the second most important predictor ( Figure 6). It was indicated that the secondary industry share and population contributed to the change of urban PM 2.5 concentration. PM2.5 exposure was significantly related to many socioeconomic factors [10,17,24]. GTWR model (see Appendix A for model assessment process) was used to study the spatial heterogeneity of the relationship between socio-economic factors and PM2.5 concentration. The results from the GTWR model showed that all variables, except for urban area, were significantly related to PM2.5 concentrations. Among all the variables, urban secondary industry share (0.279) was the most important predictor, and population density (0.022) was the second most important predictor ( Figure 6). It was indicated that the secondary industry share and population contributed to the change of urban PM2.5 concentration. According to Figure 7, the impacts of the urban secondary industry share, population density, urban population and per capita disposable income on the urban PM2.5 concentration in most years were positive, while urban area and per capita GDP had negative effects on PM2.5 concentration, that is, if the share of secondary industry, population size and population density increase, the change rate of urban PM2.5 emissions will be greater, making the urban PM2.5 concentration increase. In addition, the trend of the coefficient of the share of the secondary industry (Figure 7b), population density ( Figure 7c) and per capita disposable income (Figure 7f) gradually decreased from 2005 to 2016. This indicated that their impacts on PM2.5 concentration were getting lower, which was mainly attributed to China's urban population policy and industrial structure adjustment. On the contrary, According to Figure 7, the impacts of the urban secondary industry share, population density, urban population and per capita disposable income on the urban PM 2.5 concentration in most years were positive, while urban area and per capita GDP had negative effects on PM 2.5 concentration, that is, if the share of secondary industry, population size and population density increase, the change rate of urban PM 2.5 emissions will be greater, making the urban PM 2.5 concentration increase. In addition, the trend of the coefficient of the share of the secondary industry (Figure 7b), population density ( Figure 7c) and per capita disposable income (Figure 7f) gradually decreased from 2005 to 2016. This indicated that their impacts on PM 2.5 concentration were getting lower, which was mainly attributed to China's urban population policy and industrial structure adjustment. On the contrary, sometimes the increased urban population (Figure 7d) and per capita GDP (Figure 7e) have a positive contribution to PM 2.5 concentration, because, with the further process of industrialization and urbanization, the proportion of secondary industry in GDP has increased, resulting in increased demand for energy consumption [39].  Figure 8 shows the spatial distribution of different driver factor coefficients estimated by the GTWR model. There was significant spatial heterogeneity in coefficients of each socioeconomic factor except urban area. The highest coefficients were urban population (0.066), population density (0.062), GDP per capita (0.0003), secondary industry share (0.769) and per capita disposable income (0.005) estimated by GTWR were distributed in northwest, central, south, and north provincial capitals, respectively. This shows that population size and density are important driving factors for PM 2.5 concentration in cities in the northwest and central China, while secondary industry share and per capita income play an important role in PM 2.5 concentration in cities in northern China. Specifically, the urban population and population density promoted the increase of PM 2.5 emission, with estimated coefficients of 0.014 and 0.018, respectively. Compared with other provincial capital cities, cities in the north, east and southwest of China had a higher urban population and population density coefficient, which indicated that the more densely populated the more important influence on PM 2.5 concentration in a certain area. Higher population density means more energy consumption and pollution emissions, as well as a decrease of green space associated with densification that can reduce the city's capacity to mitigate air pollution [20,40,41].
ban second industry share. Our results showed that, except for Lanzhou (−0.14), Xining (−0.18), Yinchuan (−0.12), Hefei (−0.60), Kunming (−0.50), Guiyang (−0.29), Urumqi (−0.57) and Lhasa (−0.33), the share of the secondary industry had a significant positive impact on PM2.5 concentration. Spatially, compared with other provincial capital cities, the share of the secondary industry had a more significant impact on PM2.5 concentration in cities in the north, central and eastern China (Table A2, Appendix A). In the process of industrialization and urbanization, the secondary industry accounted for about 40% of GDP, which led to a surge in energy consumption, and at present, the main fuel consumption in most cities is fossil fuels, thus, a large amount of fossil energy consumption produces a large amount of PM2.5 emissions and dust emissions [26,39,42].

Differences in Spatio-Temporal Distribution of PM2.5
This study analyzed the annual average PM2.5 spatio-temporal distribution during 2000-2016 in 31 provincial capital cities of China. The higher PM2.5 concentrations primarily occurred in northern, eastern, and central China. The highest annual mean PM2.5 concentrations occurred in Shijiazhuang, and the annual mean PM2.5 concentrations exceed 75 μg m −3 during 2000-2016, except in 2000 and 2001, which was mainly produced by fossil fuel combustion, industrial emissions, and floating dust from building sites. Coalburning is the most important pollution source, accounting for 28.5% of local emitted Among all the driving factors, per capita GDP played an important role, and over 60% of provincial capital cities per capita GDP had a negative impact on urban PM 2.5 concentration. Especially, for economically developed regions, such as the Beijing-Tianjin-Hebei region (BTH) and the Pearl River Delta region (PRD), the per capita GDP had a negative impact on the urban PM 2.5 concentration. On the contrary, for the cities with low economic levels, such as Xining, Yinchuan, Lanzhou, Harbin, Changchun, Shenyang and Lhasa, the per capita GDP had a significant positive impact on the urban PM 2.5 concentration. The results showed that the spatial difference in PM 2.5 concentration would cause regional economic inequality. In contrast, per capita disposable income had a positive impact on urban PM 2.5 concentration in many cities, except for Xi'an (−0.0001), Lhasa (−0.0001), Fuzhou (−0.0003), Haikou (−0.0004), Nanning (−0.0008) and Guangzhou (−0.001). Another important driving factor that can affect urban PM 2.5 concentration is urban second industry share. Our results showed that, except for Lanzhou (−0.14), Xining (−0.18), Yinchuan (−0.12), Hefei (−0.60), Kunming (−0.50), Guiyang (−0.29), Urumqi (−0.57) and Lhasa (−0.33), the share of the secondary industry had a significant positive impact on PM 2.5 concentration. Spatially, compared with other provincial capital cities, the share of the secondary industry had a more significant impact on PM 2.5 concentration in cities in the north, central and eastern China (Table A2, Appendix A). In the process of industrialization and urbanization, the secondary industry accounted for about 40% of GDP, which led to a surge in energy consumption, and at present, the main fuel consumption in most cities is fossil fuels, thus, a large amount of fossil energy consumption produces a large amount of PM 2.5 emissions and dust emissions [26,39,42].

Differences in Spatio-Temporal Distribution of PM 2.5
This study analyzed the annual average PM 2.5 spatio-temporal distribution during 2000-2016 in 31 provincial capital cities of China. The higher PM 2.5 concentrations pri-marily occurred in northern, eastern, and central China. The highest annual mean PM 2.5 concentrations occurred in Shijiazhuang, and the annual mean PM 2.5 concentrations exceed 75 µg m −3 during 2000-2016, except in 2000 and 2001, which was mainly produced by fossil fuel combustion, industrial emissions, and floating dust from building sites. Coalburning is the most important pollution source, accounting for 28.5% of local emitted PM 2.5 , while PM 2.5 mass concentrations from industrial emissions and floating dust accounted for 25.2% and 22.5% of local emitted PM 2.5 in Shijiazhuang, respectively [43]. Additionally, unfavorable meteorological conditions such as weak winds and vertical diffusion and low mixing layer height are the main factors leading to long-term severe air pollution at the low atmosphere in the North China Plain, especially Shijiazhuang and Jinan [44]. The lowest annual mean PM 2.5 concentrations occurred in Lhasa (14.46 + 5.19 µg m −3 ) which is lower than 35 µg m −3 from 2000 to 2016, mainly due to fewer coal-based industries or more favorable meteorological conditions for pollution dispersion than other areas, resulting in these places having lower levels of air pollution [45]. The change rates of PM 2.5 concentration in three terms of 2000-2005, 2005-2010 and 2010-2016 were 31.41%, −0.08% and −0.16% respectively, which showed that the focus and measures exerted on air pollution control in recent years may have been very effective [20]. Specifically, the "Air Pollution Prevention and Control Action Plan" set definite targets for decreases in PM 2.5 concentrations for most eastern regions and other heavily polluted regions.

Contribution of Population Mobility to Urban PM 2.5 Pollution
Previous studies have shown that with the rapid urbanization and industrialization in China, the population exposed to the potential risks of PM 2.5 was sharply increasing [46,47]. Our results show that over the past 17 years, an increasing proportion of the population was mainly concentrated in areas with PM 2.5 concentrations of 35-75 µg m −3 . In addition, more and more people in rural area have migrated to cities, especially where infrastructures are well-developed and economically developed cities are located, such as Beijing, Shanghai, Tianjin and Guangzhou. Chinese urbanization strategy has aimed to increase the urban population from 52% in 2012 to 60% in 2020, which means that at least 100 million people have migrated to cities [48]. This will lead to further urban expansion and larger urban populations imposing additional pressure on urban air quality and PM 2.5 concentrations than before. Meanwhile, a huge amount of population migration from rural to urban means more energy consumption and pollution emissions, as well as a decrease in the area of green space associated with densification that can weaken the city's capacity to mitigate air pollution [20,40,41]. Therefore, more reasonable local regulations are needed to be conducted, such as strict limitation of industrial air pollutant emissions, upgrading the gasoline quality ahead or together with improvements in vehicle emission standards [10,45]. The census data used in the assessment of PM 2.5 exposure are based on the assumption that the population distribution is stationary. In reality, people will be in different places and exposed to different PM 2.5 environments across different times [49]. Although uncertainty in static-based assessments has been discussed in previous studies, this uncertainty is smaller in the assessment of PM 2.5 exposure for long time series.

Differences in Exposure and Inequality
Exposure inequality to air pollution means that some impacts related to air pollution are relatively high for socially disadvantaged groups. For the education levels subgroup, the exposure concentrations for people with a tertiary education were the highest, while people with a primary education experienced the lowest levels. This can be explained by the fact that over 80% of the lower education population inhabited the southwest areas with relatively low PM 2.5 concentration. On the other hand, the higher education groups were often engaged in high-level occupations and had a high income, so they can afford excessive prices and live in the central urban areas where good infrastructure, convenient transportation, good educational and medical institutions are provided, as they are exposed to higher PM 2.5 concentration. By contrast, residents with low education and low income usually live around the central areas because they cannot afford high housing prices [28,50]. For job category subgroups, public government employees and tertiary industry service personnel, due to the friendly office environment and advanced air purification equipment, had a lower risk of exposure to PM 2.5 concentrations than workers and professional technicians who had been engaged in the manufacturing industry for a long time.
Among the age subgroups, children and older people were the most unfairly exposed population, which was consistent with the result found by Ouyang et al. [28]. For GDPPC and SIS subgroups, people with higher per capita GDP and located in industry-led industries were more likely to be exposed to high PM 2.5 risks. These are mainly because economic development in most Chinese provincial capital cities tend to rely on coal-based secondary industries, and coal is the dominating energy source for the secondary industry in China. As a result, the fast growth of the secondary industry caused a huge increase of coal consumption. As Hao and Liu pointed out, the huge coal consumption caused by the development of secondary industry in recent years was an important reason for the rise of urban PM 2.5 concentrations [19]. For income subgroups, including urban and rural residents, the cumulative population weighted exposure generally increased with an increasing per capita disposable income. For urban residents, higher income people (UPCDI >25) were mainly concentrated in the north and east areas or lived in urban core areas. The distribution characterization of higher income people was in line with that of annual PM 2.5 concentration, which resulted in the most disproportionate exposure for higher income residents. In contrast, for rural residents, the high-income group was mainly from the population who migrated from rural to urban areas. Most of them were mainly engaged in construction and industrial industries, and construction dust and industrial emissions were the main sources of particulate pollution [51,52].

Implications and Limitations
This research examined the characteristics of PM 2.5 pollution exposure and explored the inequality of PM 2.5 exposure in different population subgroups, and discussed the driving factors of unequal exposure in 31 Chinese capital cities from 2000 to 2016. This study demonstrated the importance of multiple temporal scales. It provided evidence for residential exposures to PM 2.5 , which is helpful for the government to identify the exposed subgroups and to take protective measures for these residents. However, there are also some limitations in our research. Firstly, the data of PM 2.5 concentration used in this study were from 2000 to 2016, while population census data were investigated in 2000 and 2010.
To study the exposure inequality of PM 2.5 to different population subgroups, we only used 2000 and 2010 census data. Although the period of census data does not match that of PM 2.5 concentration, the results of exposure inequality for PM 2.5 are small. Secondly, the trends of PM 2.5 concentration and population were changing with time, so the exposure level and exposure inequality calculated in this study may not represent future exposures. Thirdly, when investigating the impacts of driving factors on PM 2.5 exposure, this study did not consider the impacts of natural variances.

Conclusions
In this study, we found that the average annual PM 2.5 over the whole study of capital cities experienced a minor increase at a rate of 0.94 µg m −3 per year, except Lanzhou and a small CV with 0.16 from 2000 to 2013. However, a majority of the capital cities, 74%, saw a sharp decrease at a rate of 2.86 µg m −3 per year from 2013 to 2016. According to the variation in temporal scales, an increasing number of people suffered from increasingly serious PM 2.5 pollution and fewer people experienced less PM 2.5 pollution. The contributions of increased exposure population are mainly caused by the increase of migrating population in north, east, south and central China. Additionally, the exposure and inequality among different subgroups showed distinct differences. Overall, older people, higher income and higher SIS subgroups suffered from the greater environmental inequality, respectively. At the same time, gender and job category groups had similar exposure levels. In addition, the cumulative population-weighted exposure and inequality level increased with the increasing education level. Finally, we established nonlinear and the GTWR model to analyze the economic effectiveness of exposure inequality. The results showed that per capita GDP, population size and the share of the secondary industry played an important role in exposure inequality to PM 2.5 .   Figure A1 shows the annual averages and spatial distribution of PM 2.5 concentrations in 31 provinces capital cities during 2000-2016. The higher concentrations of PM 2.5 primarily occurred in northern, eastern, and central China, which mainly resulted from anthropogenic emissions including coal and biomass burning, industrial and motor vehicles emissions, etc. All of the provincial capital cities exceeded WHO air quality guidelines (AQG), and only 3.23%, 3.23%, and 12.90% of the cities met WHO interim target-3 (IT-3, 10-15 µg/m 3 ), interim target-2 (IT-2, 15-25 µg/m 3 ) and IT-1 thresholds. These results indicate that only 19.25% of the urban population was exposed to annual mean PM 2.5 concentrations less than the WHO IT-1 limit of 35 µg/m 3 , and 80.65% of the urban population was exposed to annual mean PM 2.5 concentrations exceeding the WHO IT-1 threshold. Our research shows that the population exposed to PM 2.5 over 35 µg/m 3 is still increasing; however, its growth rate is becoming slower and slower. Spatially, the population exposure was mainly found in East (57.04 million in 2016) and North (51.84 million in 2016) China, and the increased population exposure from 2000 to 2016 was primarily due to the increase in the exposed population of East China.

Appendix A.3. GTWR Model Evaluation Process
To better understand the influence of socio-economic factors on urban PM 2.5 concentration, the GTWR model was used to further study the spatial heterogeneity of the relationship between socio-economic factors and PM 2.5 concentration. To examine the existence of multicollinearity between the explanatory variables before conducting the regression analysis, this study performed multicollinearity tests for each explanatory variable with SPSS 25.0 software, which showed that the variance inflation factors (VIF) for UR, UP, GDPPC, SIS, UPCDI and PD were 2.44, 1.73, 1.17, 1.06, 2.57 and 2.52, respectively. All of these are less than 10, indicating that there was no significant effect of multicollinearity between the explanatory variables, which allows for the regression analysis in the next step. The GTWR model global regression results show that the GTWR has a bandwidth of 0.1150, with a residual sum of squares of 33,007.9, an AICc of 3822.93, a spatio-temporal distance ratio of 0.3594, and a goodness-of-fit (R 2 Adjusted) of 0.8316. These results indicated that the GTWR model uses fewer parameters to obtain regression results closer to the valid values (Tables A1 and A4).