Assessment of long-term particulate nitrate air pollution and its health risk in China.

Air pollution is a major environmental and public health challenge in China and the Chinese government has implemented a series of strict air quality policies. However, particulate nitrate (NO3 -) concentration remains high or even increases at monitoring sites despite the total PM2.5 concentration has decreased. Unfortunately, it has been difficult to estimate NO3 - concentration across China due to the lack of a PM2.5 speciation monitoring network. Here, we use a machine learning model incorporating ground measurements and satellite data to characterize the spatiotemporal patterns of NO3 -, thereby understanding the disease burden associated with long-term NO3 - exposure in China. Our results show that existing air pollution control policies are effective, but increased NO3 - of traffic emissions offset reduced NO3 - of industrial emissions. In 2018, the national mean mortality burden attributable to NO3 - was as high as 0.68 million, indicating that targeted regulations are needed to control NO3 - pollution in China.

Introduction

Ambient fine particulate matter (PM2.5, or particulate matter with an aerodynamic diameter of less than 2.5 μm) pollution has been identified as a leading environmental risk factor for premature death worldwide (Cohen et al., 2017; Burnett et al., 2018). For example, long-term PM2.5 exposures in China are estimated to account for 30.8 million adult premature deaths during 2000–2016 (Liang et al., 2020). In response to improving air quality, the Chinese government has promulgated a series of strict clean air policies and effectively reduced PM2.5 pollution (Xiao et al., 2018). However, PM2.5 is a complex mixture of many constituents such as sulfate (SO4−), nitrate (NO3−), organic carbon, elemental carbon, mineral dust, and sea salt. Previous studies suggested that various PM2.5 components may pose different adverse effects on human health (WHO, 2007; Park et al., 2018). Numerous population-based epidemiological studies have provided evidence that long-term and short-term exposures to NO3− are associated with increased risks of mortality, hospital admissions, emergency department visits of cardiovascular diseases, respiratory diseases, and total non-accidental diseases (Peng et al., 2009; Ostro et al., 2009, 2015; Cao et al., 2012; Son et al., 2012; Liu and Zhang, 2015; Chung et al., 2015; Lin et al., 2016). In addition, recent studies based on human cell lines and mouse experiments have found that water-soluble inorganics of particulate matters including NO3− could rapidly penetrate the lung surfactant barrier to the alveolar region and perturb gene expression in the lungs, as well as resulting in lung injuries and metabolite changes in serum samples of mice (Zhao et al., 2019).

There is an urgent need to estimate NO3− in China and quantify the mortality burden associated with its long-term exposures. In recent years, NO3− is becoming the dominant component of PM2.5 during haze events in China (Xu et al., 2019; Feng et al., 2021; Shao et al., 2018; Kong et al., 2020) despite NO3− precursor, NOX (NO2 + NO), has decreased by about 20% under proactive policies (Zheng et al., 2018). This is due to the fact that NO3− formation is not only affected by NOX but also the availability of SO4− precursor, SO2, since NO3− and SO4− are closely connected by the participation of bases (mainly ammonia, NH3) (Stelson et al., 1984). SO2 has dropped by 62% from 2010 to 2017 because desulphurization is one of China’s major air pollution control policies (Zheng et al., 2018). Before the major reduction of SO2, NH3 neutralized SO2 first because ammonium sulfate ((NH4)2SO4) has better stability than ammonium nitrate (NH4NO3) (Liu et al., 2012). After the reduction, more free NH3 becomes available to form NH4NO3 and increases NO3− formation yields. Another possible explanation for the increasing NO3− is that urban NH3 sources from vehicle emissions are expanding due to the rapidly growing domestic vehicle fleet (Liu et al., 2013; Chang et al., 2016; Sun et al., 2017). As reported by the National Bureau of Statistics of China, the civilian vehicle population increased from 31.6 million in 2005 to 232.3 million in 2018 (Figure S1).

Although NO3− is of elevated toxicity and its fraction in PM2.5 is growing, the health impact of long-term NO3− pollution in China has not been investigated due to the lack of accurate exposure estimates. To date, only a few studies utilized chemical transport models (CTMs) predicted China’s NO3− concentration at the national level (Xing et al., 2015; Geng et al., 2019; Zhai et al., 2021; Liu et al., 2017a). However, the performance of CTMs heavily relies on emission inventories, which usually bring large uncertainties in developing countries where emission information is inaccurate (Li et al., 2017). Recently, a prediction model was built based on the GEOS-Chem atmospheric chemistry model to simulate seasonal NO3− concentration in China from 2013–2018 with a correlation coefficient (r) value of 0.82 in winter (Zhai et al., 2021). This NO3− model can provide predictions in large-scale geographical zones such as North China but is limited in providing information on NO3− over smaller regions (i.e., cities and roads) limited by its coarse resolution of 0.5 × 0.625°. However, the concentration field at finer resolution is important for identifying anthropogenic drivers of changes in NO3−. For example, it is necessary to estimate NO3− along major traffic corridors because NO3− formation is likely associated with expanding vehicle emissions (Fan et al., 2020). In addition, the model cannot predict NO3− over Northwest China, where the source of air pollution is gradually shifting from natural dust storms to anthropogenic industrial emissions (Turap et al., 2019). Moreover, the model only predicts NO3− after implementing APPCAP in 2013, thereby failing to evaluate the effectiveness of early air quality regulations, including those related to fuel quality standards. Therefore, a national-wide, high-resolution, long-term prediction model is required to estimate the spatiotemporal patterns of NO3− for policy assessment and advance the effort on understanding the mortality burden associated with long-term NO3− exposures in China.

Assessing the trend of historical air pollution across China without a ground monitoring network is challenging. Machine learning models that incorporate ground measurements, satellite remote sensing data, and other supporting information are one of the emerging solutions (Xiao et al., 2018). We employed a random forest algorithm (Breiman, 2001), a machine learning model, driven by satellite-retrieved aerosol data to build a NO3− prediction model. The out-of-bag (OOB) bootstrap sampling strategy of the random forest algorithm was used to develop and validate the model (Meng et al., 2018b). This method allows a flexible number of predictors to estimate relationships between NO3− and various impact factors such as the spatial distribution of road length, which is closely related to NO3− concentration but is not incorporated in previously published NO3− models (Xing et al., 2015; Geng et al., 2019; Zhai et al., 2021; Liu et al., 2017a; Li et al., 2021). Specifically, we incorporated aerosol optical depth (AOD), a measure of the amount of aerosol present in the atmosphere, retrieved by the Multi-angle Imaging SpectroRadiometer (MISR) onboard Terra satellite to provide unique information on PM2.5 constituents over land (Kahn et al., 2005; Chau et al., 2020). Compared with existing NO3− models that only involved ground observations from several monitoring sites over short periods, we collected daily mean NO3− concentration from 88 urban and rural monitoring sites across China from 2005–2018, including high-quality measurements from Hong Kong and Taiwan PM2.5 speciation supersites (Lin et al., 2008; Huang et al., 2014). Detailed information on ground NO3− observations used in model development is provided in Table S1.

Here, we applied the newly developed NO3− model to address the following research questions: (1) What is the temporal trend of NO3− concentration in China from 2005–2018? (2) What are the long-term spatial variations in the NO3− concentration? And (3) How large is the mortality burden due to the NO3− exposure? We compared the trend of NO3− over four city clusters in China, including Beijing–Tianjin–Hebei (BTH), the biggest urbanized megalopolis in North China with a population of 110 million in 2020; Yangtze River Delta (YRD), the fastest growing and the most developed region in East China; Pearl River Delta (PRD), the most densely urbanized region in South China; and Cheng-Yu (CY), the fourth powerhouse that drives China’s economic growth. We then estimated the spatial and temporal characteristics of all-cause excess mortality attributable to NO3− exposure in China based on published analyses. The detailed methodology is provided in the STAR Methods section.

Results and discussion

High-performance NO3− concentration model

Our study domain covers mainland China, Taiwan, Hong Kong, and Macao, as demonstrated in Figure 1. The circles represent the location of monitoring sites color-coded by daily mean NO3− concentration. In general, NO3− was high in populous city clusters. Taking the BTH region as an example, Beijing’s daily mean NO3− concentration was about 16 μg m−3, which was more than four times higher than that of developed countries (Ostro et al., 2015), indicating estimating NO3− in China is urgently needed. Using ground observations to train a random forest model driven by satellite data and other supporting information, we developed a high-performance model to predict daily NO3− concentration in China at 10-km resolution. Specifically, this model considered atmospheric compositions (i.e., NO2), meteorological conditions (i.e., temperature), emission inventories (i.e., NH3), vegetation cover (i.e., Normalized Difference Vegetation Index, NDVI), and road length (Figure S2). MISR AOD data that is unique in characterizing the shape, size, and light-absorbing property of PM2.5 constituents were incorporated to help identify complicated spatiotemporal variance of NO3−, which CTMs can simulate but with large uncertainties (Xing et al., 2015; Geng et al., 2019, 2020; Friberg et al., 2016; Liu et al., 2007; Franklin et al., 2017; Meng et al., 2018a).

Figure 1

Map of study domain, the spatial distribution of ground monitoring sites, and averaged daily mean NO3- concentrations from 2005–2018.

The circles show the location of ground monitoring sites for measuring NO3- concentrations. The color of these circles represents the value of the calculated daily mean NO3- concentration with a color bar located at the top left corner. Four city clusters (i.e., Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD), Pearl River Delta (PRD), Cheng-Yu (CY)), and provinces of China are labeled by red and black lines, respectively.

Figure S3 shows that model-calculated monthly mean NO3− concentrations are in good agreement with observations collected from monitoring sites as the slope of regression lines is close to unity with intercepts close to zero. The model’s OOB cross-validation (CV) R2 value is 0.88 (Figure S3A), indicating little prediction bias. We conducted spatial and temporal CVs to test whether the model could make reliable predictions where or when ground observations were unavailable (Figures S3B and S3C). In the spatial CV test, ground NO3− observations were divided into groups based on the monitoring site location. NO3− predictions were generated by a random forest model trained with data elsewhere. Similarly, the temporal CV was tested with data divided by the time when ground NO3− was measured.

Figure 2 displays time-series comparisons between predicted and measured monthly mean NO3− concentrations in six geographical regions in China. Only a few monitoring sites documented NO3− before 2013, such as the one in Xi’an, located in Northwest China. NO3− level of Xi’an was high in winter and low in summer (red line in Figure 2A). Our model successfully predicts this seasonal cycle and extreme values (black line in Figure 2A). Compared with high-quality observations in Taiwan supersites, our model accurately simulates the decreasing trend of NO3− in Pingtung (Figure 2B). Starting from 2013, ground monitoring sites of NO3− have become available in many other regions in China. We present the comparison result of four sites, including Harbin in the Northeast, Beijing in the North, Chengdu in the West, and Shanghai in the East (Figures 2C–2F). We find that Beijing is the only city where NO3− concentration has decreased since the implementation of APPCAP (Figure 2D). The NO3− level was stable in Harbin and Shanghai after 2014 (Figures 2C and 2F). In Chengdu, NO3− concentration even increased to 30 μg m−3 in January 2018 (Figure 2E), which doubled the WHO’s standard on 24-h total PM2.5 (15 μg m−3), suggesting reducing NO3− is crucially important for promoting public health in China. Interestingly, NO3− was unusually low in Chengdu in winter 2016, likely related to abnormal meteorological conditions (Zhai et al., 2021; Zhang et al., 2019b; Zhou et al., 2019). Overall, Figure 2 indicates that our model has the potential to accurately predict monthly NO3− concentration over different geographical regions. Compared with recent studies predicting NO3− in China (Zhai et al., 2021; Li et al., 2021), our model exhibits higher accuracy with a longer temporal and broader spatial coverage.

Figure 2

Predicted and observed monthly mean NO3- concentrations at six monitoring sites of different geographical regions in China.

In each figure, the red line represents monthly mean NO3- observations while model-calculated NO3- predictions are shown as the black line. Two sites are selected to display NO3- levels over 2005–2012 (pre APPCAP), including the site located at (A) Xi’an in the Northwest and (B) Pingtung in the South.

(C)–(F) show results from 2013–2018 (during APPCAP) of sites located at (C) Harbin in the Northeast, (D) Beijing in the North, (E) Chengdu in the West, and (F) Shanghai in the East.

Temporal trend of NO3− concentration from 2005–2018

Based on the newly developed NO3− model, we estimated the temporal trend of China’s national annual mean NO3− concentration as depicted in the red line of Figure 3. The linear trend of NO3− was calculated over two periods, 2005–2012 (pre APPCAP) and 2013–2018 (during APPCAP), shown as gray lines in Figure 3. During the pre APPCAP, the national mean NO3− concentration peaked at 9.2 μg m−3 in 2006 and decreased to 8.4 μg m−3 in 2012. This reduction was likely due to the Chinese government releasing the 11th Five Year Plan (2006–2011), effectively reducing emissions from industrial facilities (Jin et al., 2016). In addition, two vehicle NOX emission standards were established, including National III released in 2007 and National IV released in 2011, to control on-road vehicle emissions (Wu et al., 2017). During APPCAP, the national mean NO3− slightly decreased from 9.3 μg m−3 to 7.6 μg m−3 with a linear trend of −0.3 μg m−3 from 2013–2018. The decreasing trend was contributed by the 12th Five Year Plan (2012–2017), and APPCAP promoted accelerating denitrification and required all coal-fired units except circulating fluidized bed boilers to install denitrification facilities (2013). However, the rapidly growing vehicle population (Figure S1) increased NO3− generated from traffic emissions, offsetting reduced NO3− from industrial emissions (Fan et al., 2020). Moreover, the rate of the decreasing NO3− from coal combustions slowed down gradually due to mitigation actions were the most effective at an early stage (Zhong et al., 2021). The minimum national mean NO3− appeared in 2016 at 7.4 μg m−3 was likely influenced by favorable meteorological conditions that suppressed NO3− formation (Zhang et al., 2019b). Compared to other PM2.5 constituents like SO4−, NO3− is much more sensitive to temperature. For example, thermodynamic formation of NH4NO3 favors low temperatures (Zhai et al., 2021; Zhou et al., 2019). Winter 2016 was much warmer than normal (Zhang et al., 2019b), which may contribute to the low NO3− level. Overall, the national mean NO3− was relatively stable with a range of less than 2 μg m−3 during the whole study period and even slightly increased in recent years, which is consistent with findings of existing studies (Zhai et al., 2021; Zhao et al., 2022).

Figure 3

Annual mean NO3- concentrations of China and four city clusters during 2005–2018.

The trends are calculated based on the newly developed NO3- prediction model. The red line shows the estimated national annual mean NO3- concentrations. The four black lines represent the population-weighted annual mean NO3- concentrations of four city clusters. The gray lines are calculated linear trends over 2005–2012 (pre APPCAP) and 2013–2018 (during APPCAP).

To understand the temporal variation of NO3− across different populated regions, we calculated the annual mean population-weighted NO3− of four major city clusters (BTH, YRD, PRD, and CY) shown as black lines in Figure 3. BTH (dashed line) was the most polluted cluster during the entire study period as BTH had not only numerous coal-fired power plants but also a large number of vehicles, which are major sources of generating NOx (Zhao et al., 2012). The annual mean NO3− of BTH ranged from 11.1–15.4 μg m−3 with a linear trend of −0.23 μg m−3 before implementing APPCAP. Note that BTH is the only cluster with a decreasing NO3− trend, contributed by strict air control policies during Beijing Olympic Games and local vehicle regulations (Chen et al., 2013; van der A et al., 2017). For example, a ban on high emission cars was enforced in Beijing but not nationwide. During APPCAP, the NO3− level of BTH substantially decreased from 15.9 μg m−3 in 2013 to 12.2 μg m−3 in 2015, but became stable with a range of 11.7–11.9 μg m−3 after 2015, dominating the temporal trend of the national NO3− level (red line). This tendency was most likely influenced by a significant increase in traffic emissions (Fan et al., 2020). Also, BTH had extensive agricultural activities that provided abundant NH3 to form NO3−, especially when SO2 emission reductions significantly increased over the BTH after 2013 (Liu et al., 2018). The NO3− level of YRD and CY clusters was similar, ranging from 8.0–11.3 μg m−3 during the study period with increasing trends (+0.06 μg m−3 and +0.18 μg m−3) from 2005–2012, and slightly decreasing trends (−0.23 μg m−3 and −0.06 μg m−3) from 2013–2018 despite APPCAP set ambitious goals in these clusters (Zhang et al., 2019a). The lowest NO3− level was found in PRD where climate temperature is much higher than the other three clusters that may relatively restrain NO3− formation (Zhai et al., 2021; Zhou et al., 2019). The trend of NO3− in PRD changed from +0.11 μg m−3 (pre APPCAP) to −0.22 μg m−3 (during APPCAP), with NO3− concentration ranging between 5.7 and 8.1 μg m−3.

Long-term spatial variations in NO3− concentration

The spatial distribution of annual mean NO3− concentration in China during 2005–2018 is mapped in Figure 4. Interestingly, the general spatial distribution of NO3− is similar to the distribution of estimated agricultural NH3 emissions in China (Zhang et al., 2018), showing that regulation of NH3 is an important component of NO3− reduction (Wu et al., 2016). In addition, high NO3− concentration (colored in red) occurred over three types of regions. First, NO3− was especially high along major transportation corridors such as national highways of China, which has not been reported in previous studies (Xing et al., 2015; Geng et al., 2017, 2019; Zhai et al., 2021; Li et al., 2021). For example, the level of NO3− in BTH generally ranged from 5 to 10 μg m−3 but increased by more than twice over the upper section of G4 Beijing-Hong Kong and Macau Expressway (thick yellow line). In Shandong province, NO3− rose to 15 μg m−3 along G20 Qingdao-Yinchuan Expressway (thick blue line) while the background NO3− was about 10 μg m−3. Similarly, NO3− was higher than 15 μg m−3 over cities with dense road networks (i.e., Xi’an, Zhengzhou) or large vehicle populations (i.e., Chengdu and Chongqing). These elevated NO3− levels suggest that control of vehicle emissions would be the key to reducing NO3− pollution in China (Fan et al., 2020).

Figure 4

Spatial distribution of annual mean NO3- concentrations in China, 2005–2018.

The thick lines in the right-side figure refer to the section of G4 Beijing-Hong Kong and Macau Expressway (yellow) and G20 Qingdao-Yinchuan Expressway (blue).

Second, NO3− pollution was severe in areas with many coal combustions and extensive industry facilities (i.e., Shandong and Henan). Taking Shandong as an example, a province in North China where coal burning is commonly used for household heating (Shen et al., 2020), NO3− was always above 10 μg m−3. It was estimated that 67% of NOx in China was emitted from coal combustions (Xu et al., 2000). Additionally, in many cities of Shandong such as Jinan, nearly 40% of the gross domestic product was contributed by typical industries including coal-fired power plants (Cui et al., 2015). Therefore, regulations on improving combustion efficiency and constructing pollution control facilities in coal-fired power plants are necessary to lower NO3− concentration.

Third, NO3− was generally high over major city clusters (i.e., BTH, YRD, PRD, and CY) with a level of 10–15 μg m−3 (Figure 4). This is related to the fact that vehicle populations of these regions grew exponentially, offsetting benefits from regulations on industry facilities and fuel quality (Fan et al., 2020). Although Beijing’s License Lottery has been implemented since 2011 to limit the number of cars, the civilian vehicle population of Beijing has increased by 3.6 million from 2005–2018, as reported by the National Bureau of Statistics of China. In Chengdu and Chongqing where do not have central heating but vehicle usage is the highest among cities in China. In sparsely populated regions (i.e., Tibetan Plateau), NO3− was lower than 3 μg m−3 due to favorable weather conditions (i.e., strong wind) and scarce industrial activities.

Figure S4 compares the spatial distribution of seasonal averaged NO3− during the study period. The highest NO3− level was observed in winter mainly because China’s central heating system is coal-burning based, making coal combustion the dominant source of NO3− during the heating season (Zhao et al., 2020). Populous provinces in North China with central heating (i.e., Hebei, Henan, and Shandong) experienced high NO3− at about 20 μg m−3 (colored in dark red). In addition to anthropogenic factors, meteorological conditions such as low temperature raised NO3− in winter. NO3− of spring and autumn was similar, except NO3− was higher in autumn over North China because central heating started in the region in late October or early November. The lowest NO3− was observed in summer months with a concentration generally under 5 μg m−3 (colored in blue) across China but increased to 8–10 μg m−3 over provinces with many power plants and along traffic corridors, confirming again industrial and vehicle emissions make significant contributions to NO3− pollution in China.

Mortality burden attributable to long-term NO3− in China

The mortality burden of long-term NO3− exposure in China has not been previously investigated though the burden of total PM2.5 has been extensively studied. It is estimated that long-term PM2.5 exposures accounted for an annual burden ranging from 1.5 to 2.2 million across the adult population in China (Liang et al., 2020). The heavy health burden exhibited strong spatial variations, with high attributable deaths concentrated in populated areas such as Beijing (Liu et al., 2017b), and developing inner provinces like Hebei because of the relocation of more polluting industries into these regions (Xie et al., 2016). However, none of these studies explored the long-term mortality burden from NO3− due to the lack of a nationwide NO3− exposure assessment. With the newly developed NO3− model, for the first time, we estimated the mortality burden of long-term NO3− exposure in China. We utilized a pooled relative risk (RR) reported by a meta-analysis (Yang et al., 2019) summarizing the health effect of NO3− from three cohort studies, including estimates from Medicare Cohort, California Teachers Study conducted in the United States, and a cohort of 2.4 million people in Canada. For an interquartile range (IQR = 2.3 μg m−3) increase in NO3− level, the pooled RR increases by 2.6% (95% CI: 0.88%, 4.34%) (Yang et al., 2019).

Figure 5A displays the annual mean premature death rate (counts per 10,000 persons per year) attributed to NO3− in each province of China in the first (2005) and last (2018) year of the study period and the year APPCAP was implemented (2013). During the study period, the highest death rate occurred in 2013 with a national annual mean of 7.8 per 10,000 persons. Regionally, the death rate of BTH, Henan, and Shandong was more than twice the national average, at about 20 per 10,000 persons. Low death rates were observed in Tibet, Yunnan, Hainan, and Taiwan where rates were below 5 per 10,000 persons. In 2018, the national mean death rate dropped back to 6.5 per 10,000 persons, equal to the death rate level in 2005. The reduction was mainly contributed by improved provincial death rates in Hebei, Henan, and Shandong where the rate decreased by nearly 50% after 2013. Comparing the death rate in 2018 with that in 2005, Beijing, Tianjin, Chongqing, and Sichuan were the only regions where the death rate increased by more than 1 per 10,000 persons. This reinforces that NO3− regulations on coal-fired power plants were effective but need more effort to control vehicle emissions.

Figure 5

Annual mean mortality burden of long-term NO3- exposure in China.

(A) compares the annual mean premature death rate of NO3- (per 10,000 persons per year) in each province of China in 2005, 2013, and 2018. The color of the map refers to the upper left color bar. The national annual mean death rate for each year is shown in parentheses. (B) shows the national mean annual NO3- concentration (red dotted line, refers to the right y axis) and its corresponding premature death number (purple bars, refers to the left y axis) from 2005–2018.

Figure 5B shows the national annual mean NO3− concentration (red dotted line) and the total number of adult premature deaths attributable to the NO3− exposure (purple bars) during the study period. Overall, long-term NO3− pollution brought 9.6 (95%, CI: 3.6–15.3) million deaths, accounting for about 1/3 of the death number caused by the total PM2.5 pollution (Liang et al., 2020). Note that the national annual mean NO3− concentration was around 8 μg m−3, only accounted for up to 25% of PM2.5 concentration (Liang et al., 2020), indicating NO3− is of potential disproportionally high toxicity compared to the toxicity of total PM2.5 (Ostro et al., 2015). During APPCAP, a reduction of 20.4% in the national mean NO3− concentration was observed from 2013–2016, and correspondingly, the premature death number decreased from 0.77 (95%, CI: 0.28–1.23) million to 0.65 (95%, CI: 0.23–1.04) million. However, the death number increased to 0.68 (95%, CI: 0.24–1.08) million in 2018, confirming that China’s clean air actions were successful but may need more effort to reduce NO3− air pollution.

Conclusion

In this study, we present a high-performance model to estimate long-term NO3− concentration at 10-km resolution across China from 2005–2018, based on a random forest approach incorporating ground observations, satellite remote sensing data, model reanalysis, and other supporting information. Compared with a typical PM2.5 mass prediction model, we included a unique set of variables tailored for NO3− according to its physical characteristics and formation mechanisms. For example, MISR’s AOD data were used to capture the spherical shape and high light reflectance of fine particles rich in NO3−. Unlike the previous studies with limited spatial coverage and model performance (Zhai et al., 2021; Geng et al., 2019), our model is able to provide spatially resolved high-quality predictions nationwide. The predicted long-term NO3− trend can be applied to investigate the effectiveness of existing emission control policies and regulations relative to reducing PM2.5 air pollution. Also, our model could help refine the health effect assessment of NO3− such as we calculated the mortality burden of NO3− in each province of China. Our findings indicate that relative to its mass fraction in PM2.5, NO3− might represent a greater population health risk in China. These estimates will greatly benefit epidemiological studies on accessing the health effect of PM2.5 chemical compositions in China and other regions. The results presented here reinforce the need to better understand global excess mortality from ambient air pollution.

Limitations of the study

Our NO3− prediction model has a few limitations. First, the spatial and temporal characteristics of ground data used to train the random forest model might introduce sampling bias in the results. We included NO3− observations from available 88 monitoring sites because China does not have a monitoring network of PM2.5 constituents. Luckily, these monitoring sites have covered regions with different population densities and road lengths (Figure S5). In addition, 43 (49%) sites are located in rural areas and 45 (51%) sites are distributed in urban areas. However, most observations available before 2013 were collected from sites in Hong Kong and Taiwan because very few PM2.5 speciation monitoring sites existed in mainland China. Regarding the discussion of city clusters, the different numbers of monitoring sites of different clusters will affect the spatial variation of NO3− prediction (Figure S3B). In the future, we will include more ground observations when they are available to improve the model’s stability and accuracy.

Second, although MISR provides unique information on aerosol microphysical properties, it has a narrow swath that leads to limited spatial coverage. As shown in Figures 4 and S4, NO3− prediction was unavailable over eastern Tibet (colored in white). Gap-filling approaches are needed to fill missing AOD retrievals to improve the model’s spatial coverage. Besides, MISR’s low temporal resolution with global coverage of every nine days is limited to capturing the influence of non-fossil fuel NOx that contributes an important contribution to NO3−. Satellite observations with higher resolutions are needed to detect small burning spots in China and be added to the prediction model to verify the effect of biomass burning on NO3− concentration.

Third, health risk estimates of long-term NO3− exposure in China are not available at the time of this study. The pooled RR was calculated based on three cohort studies conducted either in the United States or Canada, where NO3− concentration was much lower than the level in China. China’s population susceptibility and socioeconomic characteristics were significantly different from these developed countries. As a result, the RR may be limited in fully representing the adverse health effect of NO3− for the Chinese population. We will generate a nation-specific health risk assessment when a cohort study of NO3− in China is available. In addition, only one-year data of age-specific baseline mortality was used in estimating mortality burdens due to the lack of data from other years during our study period. The National Bureau of Statistics of China (http://data.stats.gov.cn/) reported that the population of people 65 years or older increased by 66 million from 2005 to 2018. Therefore, we may have underestimated the mortality burden of NO3− in China of recent years because of the increasing population vulnerability to NO3− pollution (Liang et al., 2020).

STAR★Methods

Key resources table

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Yang Liu (yang.liu@emory.edu).

Materials availability

This study did not generate new unique materials.

Method details

Observation of NO3− concentration

We collected approximately 21,000 ground observational records of daily mean NO3− concentration from 88 monitoring sites, as listed in Table S1. High-quality NO3− data from Hong Kong and Taiwan PM2.5 speciation supersites were also included (Lin et al., 2008; Huang et al., 2014). According to the Finer Resolution Observation and Monitoring Global Land Cover data (http://data.ess.tsinghua.edu.cn/) that provides information on area type, 43 of the 88 (49%) monitoring sites are located in rural areas. Histograms of population density and road length of monitoring sites (Figure S5) show that the ground data has covered regions with different population density and road length levels. Both filter-based and online-based methods are used in these sites to measure NO3− as China does not have a PM2.5 species monitoring network with uniform standards. Previous research reported that NO3− concentration measured by filter method is highly consistent with measurements from the online method in China (Dong et al., 2012), so we used all the available NO3− observations from 2005–2018 to ensure sufficient model training.

Satellite-retrieved AOD

MISR aboard satellite Terra was launched in 1999 to measure upwelling solar radiance in four spectral bands (0.446 μm, 0.558 μm, 0.672 μm, and 0.866 μm) and at nine view angles (0°, ±26.1°, ±45.6°, ±60°, ±70.5°) over a 360 km swath (Diner et al., 1998). This wide range of along-track view angles help the sensor accurately evaluate the surface contribution to the top-of-atmosphere radiances. MISR hence can provide globally continuous retrievals of AOD and information on aerosol microphysical properties over complex land surfaces (Kahn et al., 2005). Our study used the latest Version 23 MISR Level 2 aerosol product (Garay et al., 2020) to distinguish NO3− above bright surfaces with an improved spatial resolution of 4.4 × 4.4-km2. This stable and continuous satellite data make the prediction model reliable over regions where ground measurements are unavailable.

The MISR aerosol product supplies a set of aerosol mixtures that can be used to calculate eight types of aerosol components (#1, #2, #3, #6, #8, #14, #19, #21) based on aerosol’s shape, size distribution, reflection property, and scale height. Detailed information about each aerosol component can be found in Liu et al. (2007). These aerosol components can be used to estimate different AOD types relating to the physical features of NO3−. For example, non-spherical AOD components (i.e., #19) and non-absorption AOD components (i.e., #2, #3, #6) were used in this study as predictors of NO3−. AOD values of small (particle radius <0.35 μm) and large mode (particle radius >0.7 μm) particles provided by the MISR aerosol product were added to the model to reflect the bimodal distribution of NO3−. Since MISR only has a global coverage of 9 days at the equator and 2 days at the poles due to its relatively narrow swath, we used annual mean AOD data from 2005–2018 to develop the NO3− model. We found that MISR AOD data can substantially improve the model’s performance even though the data’s temporal resolution is not high. As our study focus is annual mean NO3− concentration, the temporal resolution of MISR AOD data will not influence our results. All AOD data presented here were processed at the grid cell level of 10-km.

Other variables for model development

Atmospheric composition data of PM2.5 at 10-km spatial resolution is adopted from a published study (Xiao et al., 2018). NO3−, NO, NO2, and HNO3 at 0.75 × 0.75° spatial resolution with a 3-hourly temporal resolution during 2005–2018 were extracted from the latest global reanalysis dataset of atmospheric composition, Copernicus Atmosphere Monitoring Service (CAMS) reanalysis (Inness et al., 2019), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset shows smaller biases than other reanalysis data when comparing data quality with observations (Inness et al., 2019). All CAMS data used in our study were matched to 10-km grid cells at the daily level by the inverse distance weighting (IDW) method. Emission inventory of NH3 also comes from CAMS but at monthly temporal resolution.

Meteorological variables including temperature at 2-meter, humidity at 2-meter, and planetary boundary layer height during 2005–2018 were derived from the Goddard Earth Observing System Model Forward Processing (GEOS-FP) product, which is able to produce real-time meteorological conditions using the most recent validated GEOS system with a spatial resolution of 0.25 × 0.3125° (0.5 × 0.625° before 2013) and a temporal resolution of 3 hours (Apte et al., 2015). Similar to processing atmospheric composition data, all GEOS-FP data were interpolated to 10-km grid cells at the daily level.

Road networks of highway and city expressways in China were obtained from the Institute of Geographic Sciences and Natural Resources Research of the Chinese Academy of Sciences. Only 2014 data was applied to locate road position and length in our random forest model since it has the best data quality. NDVI data from 2005–2018 was extracted from Terra MODIS C6 monthly NDVI product (MOD13A3) at 1 km resolution to define different types of land surfaces.

NO3− model development

A 10 × 10-km2 grid was designed to cover the entire study domain for model development. A total of 100,699 grid cells were within the border of the domain. We calculated the daily mean NO3− concentration in each grid cell during 2005–2018 based on a random forest algorithm trained by available ground observations (Breiman, 2001), a process to build multiple decision trees and merge them to get an accurate and stable prediction of NO3−. A final prediction model is determined by comparing the result of different settings of predictors used to split tree nodes and the number of trees in the forest (Hu et al., 2017). This modeling method has been widely used and shows high predictive abilities to estimate historical air pollution (Chen et al., 2018a; Hu et al., 2017; Geng et al., 2020; Xiao et al., 2018). In this study, the NO3− model was constructed by testing different predictors that can improve model performance (i.e., minimize mean squared error) and relate to NO3− formation or spatial distribution. For example, we choose MISR AODs as predictors because they greatly improve the model performance in predicting NO3− in West China. We select 18 predictors in the final model, including satellite-retrieved AODs and NDVI, model reanalysis of atmospheric composition and meteorological conditions, emission inventory, and road length to achieve the best prediction accuracy. We build 1000 trees in the prediction model to minimize the model’s over-fitting problem (Hu et al., 2017). Information on model variables and their permutation feature importance ranking is provided in Figure S2. Modeling data of NO3−, PM2.5, NO2, humidity and observational data of MISR AOD #19 ranked the top five significant predictors.

Estimation of the mortality burden of NO3−

Mortality burden attributable to long-term NO3− exposure was calculated based on NO3− concentration, a meta-analysis pooled NO3− RR (Yang et al., 2019), population size, and baseline mortality rate, following methods described in the Global Burden of Disease project (Apte et al., 2015) and Liang et al. (2020). All data presented here were linked to 10-km grid cells, then aggregated to the provincial level. To quantify the mortality risk of NO3− pollution in China, the relative risk (RR) over each grid cell i in year j, is defined as:where β is the health effect of long-term NO3− exposure, equal to 0.011 (95%, CI: 0.003-0.018), calculated based on pooled RR from a systematic review (Yang et al., 2019). Detailed information on NO3− mortality RR estimates and associated interquartile range can be found in Yang et al. (2019) x represents the NO3− concentration of each grid cell i in year j, and x is the lowest NO3− concentration estimated during the entire study period representing counterfactual exposure level. The population attributable fraction (PAF) of each grid cell i in year j is determined by:

Then, the PAF of each province was derived using LandScan annual population data (https://landscan.ornl.gov/). Lastly, the annual absolute number of adult death counts attributable to NO3− exposure at the provincial level is calculated using:

Mortality is the estimated premature deaths in province k in year j; PAF is the population attributable fraction; POP and I correspond to the total adult population (≥25-year old) and the baseline mortality rate of adults, respectively. Annual age-specific population and all-cause mortality rate data were obtained from the National Bureau of Statistics of China (http://data.stats.gov.cn/), China Population and Employment Statistics Yearbooks, Census and Statistics Department of the Hong Kong (https://www.censtatd.gov.hk/), Macao (https://www.dsec.gov.mo/), Taiwan (https://www.stat.gov.tw/). Age-specific mortality data were obtained from demographic census statistics (http://data.stats.gov.cn/).

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
Python 3.8.8	Python Software Foundation	https://www.python.org/