Air Pollution Exposure Based on Nighttime Light Remote Sensing and Multi-source Geographic Data in Beijing

Air pollution is a significant environmental problem that affects human health. As shown in a number or existing studies, exposure to air pollution has various adverse health effects (Krewski and Rainham, 2007; Jing and Wei, 2012; Kan et al., 2012; Straif et al., 2013), including respiratory and cardiovascular diseases. The air quality index (AQI) used in China indicates the overall level of air pollution and is based on the level of six air pollutants: SO₂, NO₂, PM₁₀, PM_2.5, CO and O₃. In research on environmental pollution in which only population distribution and environmental pollution are monitored, there are significant differences in spatial distribution (Caplin et al., 2019; Rivas et al., 2017; Benjamín et al., 2019). Since exposure to environmental pollution refers to the exposure of individuals or populations to environmental factors (Chen and Zhang, 2004), the investigation of population exposure to air pollution involves exploring people’s contact with air pollution. At present, commonly used research methods of air pollution exposure are based on air pollutant concentration, population exposure intensity and population-weighted exposure risk assessment (Zhao, 2018). Among the described methods, studies that use the air quality concentration to estimate exposure are based on the simplifying assumption that the population is evenly distributed, and thus, an approximation will be given (Liu et al., 2004). Population exposure intensity is a method of assessing pollution intensity through raster operations after the spatialization of pollutant concentration and population distribution. For instance, several scholars have drawn hierarchical population density maps, calculated the PM₁₀ concentration of each grid, and superimposed the two to investigate the exposure of the Chongqing population to particulate pollution (Wang et al., 2008). The population-weighted exposure concentration is the result of the population and the concentration of a pollutant in the spatial unit of a certain spatial scale. In calculating the concentration, the influence of the population distribution is taken into account to describe the exposure intensity of the pollutant to the population in the grid (Hystad et al., 2001), which is an evaluation index of exposure intensity (Zou et al., 2016). Studies have shown that long-term exposure to air pollution has a strong correlation with several diseases (Beelen et al., 2014), and the concept of cumulative air pollution exposure intensity has been proposed, which is the total exposure over a period (Xie, 2016). The total exposure intensity of an area is evaluated by calculating the cumulative exposure intensity (Zhang et al., 2013).

In previous research, the significance of nighttime light remote sensing data for various applications has been reported. Nighttime light remote sensing data reflects several aspects of human activity and are a significant factor in human activity monitoring research (Yang et al., 2011). In addition to town lights, the lights displayed in the image also include nocturnal fishing boats, natural gas burning, forest fires and other nighttime light sources. Nightlight data were widely used in socioeconomic parameter estimation (Li et al., 2013; Li, 2018), urban expansion progress (Dong et al., 2017; Lou et al., 2019), fishery monitoring (Paulino et al., 2017; Zhang, 2017; Wu and Guan, 2019), energy (He et al., 2013; Zhao et al., 2018) and other fields (Li and Li, 2015; Levin et al., 2020). Multi-source data, including Luojia-1 nighttime light remote sensing data, were used to estimate population distributions in the present study. The new type of remote sensing image data provided by Luojia-1 have a higher spatial resolution (130 m) compared with Defense Meteorological Program (DMSP)/Operational Line-Scan System (OLS) and National Polar-Orbiting (NPP)/Visible Infrared Imaging Radiometer Suite (VIIRS) data, which have been extensively adopted for population spatialization research in the past. As shown by studies mostly based on DMSP/OLS Data and NPP-VIIRS (Bagan and Yamagata, 2015; Lu et al., 2021) data, remotely sensed nightlight data and population data have a high degree of correlation, indicating that such data types can be used to simulate and estimate population distribution. Notably, there is a scarcity of research in which Luojia-1 remote sensing data were used. China published the population data of its seventh census in 2021, and compared with other population data, the seventh census accurately reflects the number, structure, and distribution of China’s population. There has been no research on air pollution exposure based on data from the seventh census. At present, cities set up air quality monitoring stations and use the data from several locations to evaluate the air pollution of an entire city. At the same time, the AQI representing pollution level only represents the pollutants contained in the air within a certain range and cannot reflect the hazard level for the population based on the population distribution. Considering that air pollution with the same AQI is more harmful in densely populated areas than in sparsely populated areas, its air pollution needs to evaluated in combination with the population distribution.

The research methods for the simulation of the spatial and temporal distribution of air pollution can be mainly divided into two categories: the traditional statistical model and the method of machine learning. Traditional statistical models include linear regression such as analysis of the relationship between AOD (Aerosol Optical Depth) obtained by MODIS (Moderate-resolution Imaging Spectroradiometer) and ground hourly PM_2.5 (Wang et al., 2003). Further, several researchers have constructed a physical model of PM₁₀ inversion based on MERIS (Medium Resolution Imaging Spectrometer) (Rohen et al., 2011). The other research method is machine learning, including random forest model (Kamińska, 2018), support vector machine (Duong et al., 2020), deep learning method (Wu et al., 2012) and others. In terms of predicting air quality conditions in the face of large amounts of data, machine learning has a higher accuracy than traditional regression models (Zhan et al., 2018).

In the present study, the daily air pollution data of Beijing in 2018 were used to spatialize the AQI of Beijing. In addition, Luojia-1 satellite nighttime light remote sensing data and China’s seventh census data were combined with auxiliary data to spatialize the population of Beijing. Since the latest nighttime light remote sensing data released by the Luojia-1 satellite were from 2018, the district-level population data in the Beijing Statistical Yearbook in 2018 (Beijing Municipal Bureau of Statistics, 2019) were used to revise the township-level population data in the seventh census, so as to calculate the township-level population data of Beijing in 2018. Subsequently, based on the population-weighted pollution exposure level (PWEL) and intensity, the air pollution exposure in Beijing in 2018 was analyzed. Investigation of the current air pollution exposure problems can help to conduct a more time-efficient assessment of air pollution in light as population distribution data are updated to address air pollution problems more effectively. As one of the most important mega-cities in China, the high-precision population inversion data obtained based on Luojia-1 can help us to study the population exposure to pollution in Beijing. The findings of the study have important significance for understanding the regional distribution characteristics of air pollution exposure among populations in large, densely populated mega-cities in China and for developing differentiated regional population management policies

The location of Beijing in China is shown in Fig. 1. As the political, cultural, international exchange, and scientific and technological innovation center of China. Beijing is a world-famous ancient capital and international city that attracts a large number of domestic and foreign tourists. The air quality of Beijing is considerably important to the lives of the local residents and the overall image of the city. In 2021, WHO issued new air quality guidelines, establishing new guidance value levels for the concentrations of PM_2.5, PM₁₀, ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide (WHO, 2021). As as an example, the average guideline for PM_2.5 in the new guidelines is 5 μg/m³. Despite the environmental quality of Beijing has improved in recent years. according to the 2021 Bulletin on the State of China’s Ecological Environment provided by the Ministry of Ecology and Environment of China, China’s PM_2.5 average in 2021 was 30 μg/m³, ranking 22nd among the most polluted countries in the world, and being six times higher than the WHO requirement of 5 μg/m³ (Ministry of Ecology and Environment of China, 2022). The average PM_2.5 in Beijing in 2021 was 33 μg/m³, which was higher than the average in China. However, the air quality guidelines state that global deaths due to air pollution exposure can be reduced by 80% if countries can meet the annual air quality guideline level of 5 μg/m³. As such, investigation of the exposure level and intensity of Beijing’s air pollution is of considerable significance.

Figure 1.  Location of Beijing, China

The nighttime light remote sensing data used in the present study was derived from the Luojia-1 satellite and downloaded from the website of Luojia-1 (http://59.175.109.173:8888/app/login.html). Successfully launched on June 2, 2018, Luojia-1 is the world’s first professional nighttime light remote sensing satellite jointly developed by Wuhan University and related research institutions. The satellite provides data with improved spatial resolution (130 m) and light intensity recognition accuracy compared with mainstream global nighttime light data (Zhong and Liu, 2019). Due to data transmission and other issues, the downloaded data must undergo radiance conversion before use.

In the present study, points of interest such as the distribution of various facilities, buildings, and road points were used (Table 1). To investigate the relationship to population distribution, the points of interest were classified into six categories: shopping, leisure and entertainment, catering, real estate, medical care, and education.

Hourly air quality data were obtained from the national air quality release platform (https://air.cnemc.cn:18007/), which includes 34 stations in Beijing, as well as the air quality monitoring data of 9 stations in Hebei, 24 stations in Tianjin and 3 stations in Inner Mongolia surrounding Beijing. The time span of the data was from January 1, 2018, to December 31, 2018. Sites with no data were ignored and the hourly data of the remaining sites were averaged to calculate the daily average data of PM_2.5, PM₁₀, NO₂, CO, O₃, SO₂ and AQI data.

The district-level population data of Beijing in 2018 were obtained from the Beijing Statistical Yearbook (Beijing Municipal Bureau of Statistics, 2019). The township-level population data of Beijing were obtained from the Seventh Census of China released by the China International Bureau of Statistics (http://www.stats.gov.cn/). For the Seventh Population Census of the People’s Republic of China, surveys and statistics were collected on the number of permanent residents in various regions. The permanent population refers to the population who have lived in an area for over half a year. The subjects of the census were natural persons who resided within the territory of China at the time of the census and Chinese citizens who had not settled outside of China. Foreigners who were in the territory of China for a short period of time were not included.

To investigate the air pollution exposure in Beijing in 2018, the air quality and population of each grid within the study area were required. Since the existing air monitoring station data only reflected the air pollution situation near the monitoring point, the AQI data and population data needed to be spatially interpolated.

Monthly air monitoring station data in Beijing and the surrounding cities were imported into ArcGIS10.8, and the kriging interpolation tool was used to obtain monthly air quality interpolation images for Beijing. Taking the circumscribed rectangle of the Beijing vector map as the range, a fishing net with a side length of 200 m was created and the center point was used to extract the PM_2.5, PM₁₀, NO₂, CO, O₃, SO₂ data of each point in each month in Beijing.

The concentrations of PM_2.5, PM₁₀, NO₂, CO, O₃, and SO₂ at each site were used as independent variables, and AQI was used as the dependent variable to construct and train a random forest model. The random forest model was established to divide the data into training set and validation set in a ratio of 7:3. When building the model, first the number of trees needed to participate in the modeling needed to be determined first and then the parameters of each tree needed to be adjusted. Notably, the number of trees in the model needed to be determined, followed by the maximum depth of the trees. Set the number of trees to 1200 and maximum depth being set to 20. The larger the number of trees, the better the effect of the model, but the greater the amount of computation and memory required. As such Python was used to build a loop structure to determine the appropriate value of number of trees. Subsequently, the PM_2.5, PM₁₀, NO₂, CO, O₃, SO₂ data of the fishing nets of each month were substituted into the constructed random forest model to predict the AQI value of each fishing net. The described method was used to spatially process the air monitoring station data in Beijing in 2018 and obtain the monthly AQI spatial distribution raster map.

The population data of each township from the seventh census data of all townships in Beijing were input into the vector attribute table of Beijing township-level administrative divisions according to the township. At the same time, the Beijing point of interest data were divided into six categories: shopping, medical care, education, real estate, restaurant, leisure, and entertainment. The spatial relationships between the township vector diagram and various points of interest in Beijing were established and the numbers of various points of interest in each township were obtained. Du et al. (2021) suggested that 200 m resolution grids are often used in municipal units, thus a 200 m × 200 m grid was established to count the total number of various points of interest data in each grid. After exporting the various points of interest, average nighttime light remote sensing data and total population of each township, SPSS 24 was used to perform regression analysis, using the number of various points of interest and average nighttime light remote sensing data value as independent variables, and the total population as the dependent variable. After excluding the less significant points of interest data, the numbers of points of interest in the three categories of medical care, education, and real estate and average nighttime light were selected as independent variables to construct a regression model:

where POP_k represents the number of permanent residents in the kth township of the seventh census data. Variable l_k represents the number of education-related points of interest in township k; variable m_k represents the number of real estate-related points of interest in township k; variable n_k represents the number of medical-related points of interest in township k and variable p_k is the average nighttime light value of township k. The numbers a, b, c, d are the regression coefficients corresponding to l, m, n, and p, respectively. After bringing the data into SPSS 24 for calculation, the final equation was:

An observation can be made from the output results that the final input variables of the constructed model were the various points of interest of each township and the average nighttime light remote sensing data value. According to the model, the preliminary population value could be obtained. Substituting the statistical population number and the preliminary population value obtained by the model into the following formula for correction derives the population number on each 200 m × 200 m could be derived:

where POP′_kij is the population of township k of the grid cell in the row i and column j; P_k is the statistical population number of the seventh census of the township numbered k. POP_k is the result of Eq. (2), which means the number of permanent residents in the kth township. M_l is the district-level population in Beijing. N_l is the district-level population of the seventh census of Beijing.

Using population and AQI distribution data, the air pollution exposure situation in Beijing in 2018 was calculated from the population-weighted air pollution exposure level.

Because the spatial distribution of population is not completely consistent with the distribution of pollutants, the spatial distribution of the AQI does not reflect the real exposure risk of residents. Therefore, population weighting was used for exposure risk assessment. The calculation formula of population-weighted air pollution exposure level PWEL is as follows:

where C_ij is the AQI in the row i and column j of the grid. We used the ArcGIS10.8 regional statistical tools and grid calculator were used to calculate the population distribution map obtained before and the monthly AQI distribution map.

The population exposure risk assessment model proposed by Kousa et al. (2002) was used to calculate the monthly exposure intensity of the polluted population in Beijing in 2018. The calculation formula is as follows:

where E_p is the pollution exposure intensity of grid p; and c_p is the AQI of grid p.

The monthly test set of the AQI was used to perform batch value extraction to point operations on the previously obtained prediction raster images. The extracted AQI values were compared and analyzed with those obtained from the original monitoring stations to verify the model prediction results. The following indicators were calculated: R², root mean square error (RMSE), and mean absolute error (MAE) (Xu et al., 2019):

where ${y} $ is the monitored value; $ \hat{{y}} $ is the predicted value; $ \overline{y} $ is the average value of the observed value; and N is the number of samples in the test set.

where RMSE is the square root of the mean of the square of all of the error, O_t and P_t are the observations and the predicted values of a variable t. MAE is the average value of absolute error, reflecting the actual situation of the predicted value error.

Fig. 2 shows the final distribution map of Beijing’s permanent population in 2018, where red means higher population density, and green means lower population density. Resident population in Beijing was mainly distributed in the central area of Beijing. Other urban areas also had residential clusters, but such clusters were relatively small.

Figure 2.  Population spatialization map of Beijing in 2018

Eq. (5) was used to calculate the population-weighted air pollution exposure levels of different districts with different AQI in Beijing, and the monthly time series are shown in Fig. 3. The pollution exposure level of Beijing’s population in each month of 2018 shows that the air pollution exposure levels in various regions had similar trends, being highest in March and lowest in September. The trend is similar to that of the AQI spatialization results.

Figure 3.  Monthly changes of air quality index (AQI) of different districts of Beijing in 2018

Fig. 4 shows the monthly change of AQI in different months of Beijing in 2018. From the AQI spatialized raster map, we can see that the AQI value in the southeastern part of Beijing is relatively high, which means that Daxing, Tongzhou, and Fangshan districts are areas with more serious air pollution. The AQI value was higher in the northwestern urban areas in April–August. The peak of the AQI in Beijing in 2018 occurred in November.

Figure 4.  The monthly AQI (air quality index) of Beijing in 2018

Fig. 5 exhibits the air pollution exposure risk of Beijing in each month of 2018. The air pollution in various districts in Beijing had roughly the same trend in terms of the seasonal changes. The air pollution was severe from November to December in winter and less severe from July to September in summer. Overall, the air pollution exposure problem of Beijing was more serious in March and November and less serious in September. The southern area of Beijing had high pollution exposure levels in March and November, while the pollution exposure levels of the densely populated central area of Beijing, which ranked the highest in permanent population density areas, were higher from May to July. For changes of air pollution exposure risk of Beijing over time, the difference was not large and the trend is similar to that of the spatial distribution of the AQI: the highest AQI concentrations in March and November were higher than those in other months, and the highest concentration in September was also at the lowest annual level. The highest pollution concentration occurred in the southern area of Beijing. Comparing the pollution exposure distribution map with the resident population distribution map, an observation can be made that the exposure intensity had a certain correlation with the degree of population concentration. The population concentration area had greater exposure intensity, while the sparsely populated area was less exposed to pollution.

Figure 5.  The monthly air pollution exposure risk of Beijing in 2018

Among the air quality data of 70 monitoring sites, data from 63 sites were selected as the training set and data from 7 sites were selected as the verification set for random forest training. The air quality test set of the 12 months of 2018 was compared with the actual observation data for accuracy verification, and the values of R², RMSE and MAE were calculated.

After substituting the test set for the 12 months of 2018 into the relevant equation, the verification results were obtained, as shown in Table 2. The data show that the accuracy of the model was high and the AQI data could be spatialized accurately.

The obtained pollution exposure intensity in Fig. 5 were divided through the natural breakpoint method into three categories: low risk, medium risk, and high risk. The proportions of population and area involved in the three classes of risk are shown in Table 3. In 2018, the monthly proportion of areas for each pollution exposure risk changed little over time, while difference was significant between three pollution exposure risk. Most of the permanent residents in Beijing were belong to the medium and high air pollution exposure risk, and population subjected to low pollution exposure risk was relatively small. In 2018, the average monthly proportion of population in low pollution exposure risk was about 23%, and that involved in medium pollution exposure risk was about 35%. The proportion of the population involved in high pollution exposure risk was about 43%. The proportion of the area involved in different exposure intensities also fluctuated from month to month slightly. Most areas in Beijing were exposed to low-risk pollution, and a small part was exposed to high-risk pollution.

A random forest-based machine learning method to extrapolate the air quality of various locations based on the air quality data of air monitoring stations in Beijing. As can be seen in Fig. 5 and Fig. 6, high exposure risk areas were concentrated in densely populated central urban areas such as Dongcheng District, Xicheng District and Chaoyang District, indicating showed that the spatial distribution of exposure risk in Beijing is highly consistent with the spatial distribution of population density in Fig. 2.

Figure 6.  Population of each district in Beijing in 2018. Only townships with high air pollution exposure risk are marked on the map

To further determine the approximate location of the high exposure risk, the latest township zoning vector diagram of Beijing was used to calculate regional statistics on the monthly population pollution exposure intensity. The ten townships with the most serious pollution exposure in each month were respectively counted. Among them, seven towns that recured in each month were marked in Fig. 7. The populations of the seven townships with the most frequent occurrence are shown in Table 4, wherein red means higher population density, and green means lower population density. Townships with more serious exposure risk all had a large population and the population density within such blocks were relatively high, which led to high pollution exposure risk.

Figure 7.  Population of each township in Beijing. Only townships with high air pollution exposure risk are marked

In this study, land use data of Beijing and the nighttime light remote sensing data of Luojia-1 were used calculate the average nighttime light remote sensing data value of each grid determine the number of points of interest throughout the grid, and finally establish a stepwise regression equation with demographic and township-level population data of Beijing derived from the seventh census of China to obtain the number of permanent residents in each grid. In most previous studies on air pollution exposure in Beijing and the surrounding areas using air pollution exposure models based on population weighting, on low resolution population spatialization data were used.

As an example, researchers investigated the air pollution exposure of Beijing, Tianjin and Hebei based on LandScan global population distribution data (https://landscan.ornl.gov/) developed by the US Department of Energy’s Oak Ridge National Laboratory (ORNL) at a spatial resolution of 1000 m × 1000 m (Wang et al., 2021). Several other researchers used 1000 m × 1000 m spatialized population data based on DMSP/OLS global night-time steady light from the US National Geophysical Data Centre to study the air pollution exposure of Beijing (Zhang and Hu, 2018). These studies are only accurate to the district or municipal level for areas of high pollution exposure risk.

In the present study, population grid data with a resolution of 200 m × 200 m in Beijing were obtained from the nighttime light data of Luojia-1 and the latest township-level population data of the 7th Census. As such, higher precision population pollution exposure analysis data were obtained, and pollution exposure could be accurately determined at the township-level.

Additionally, air quality data and population distribution data in Beijing in 2018 were analyzed through the PWEL formula and the pollution intensity formula. From the analysis of air pollution on a temporal scale, the monthly changes in 16 regions of Beijing during 2018 were analyzed according to the pollution exposure levels, and similar trends were identified. The highest values occurred in March, while the lowest values occurred in September. The AQI of Beijing was highest in winter, followed by spring and autumn, with the lowest AQI in summer. Several researchers have confirmed (Liu et al., 2021) that due to seasonal differences in pollutant emissions and meteorological conditions, air pollution in China undergoes seasonal variations. The most severe air pollution tends to occur in winter, followed by spring and autumn, with the lowest risk in summer, which is also consistent with the seasonal variation in the AQI in Beijing.

By analyzing air pollution exposure spatially in Beijing at a spatial scale, the AQI was found to be higher in the southern areas of Beijing in 2018, which also meant that air pollution was more severe in such areas. The results have shown that the permanent population clusters of Beijing were mainly concentrated in the central part of Beijing, namely Dongcheng District, Xicheng District and the surrounding urban areas like Chaoyang District. At the township scale, nine of the ten townships with the most serious pollution exposure problems in Beijing had a population of 200 000 or more, a figure much higher than the median population of townships in Beijing of 51 741. A strong relationship between population density and air exposure has been demonstrated in previous research (Beelen et al., 2013), and the results of the present study confirm that densely populated areas in Beijing have a higher risk of exposure to pollution. The results also show that exposure intensity was higher in densely populated areas of Beijing and exposure to pollution was generally lower in sparsely populated areas, which also meant that the individual townships with higher populations were at much higher risk of exposure to pollution than other areas.

This paper used land use data, point of interest data, the township-level seventh census data of China, and Luojia-1 nighttime light remote sensing data to map the spatial distribution of the population in Beijing with high resolution. By using the latest census data and night-time light remote sensing data, the population could be spatialized more accurately, and a 200 m × 200 m high-resolution spatialization of the population distribution of Beijing in 2018 could be obtained for pollution exposure risk analysis. The air pollution concentration data of 67 monitoring stations in Beijing and the surrounding cities in 2018 were used, including SO₂, NO₂, PM₁₀, PM_2.5, CO and O₃, and the concentration of air pollutants in Beijing was obtained by interpolating the concentration of the six air pollutants.

Firstly, the population distribution in Beijing is highly heterogeneous, with the central city being extremely densely populated, much more than the suburbs. Secondly, by using higher resolution spatialized population data, the high-risk areas could be pinpointed to the township level, and the townships with the highest exposure to pollution were found to have much higher population densities than the other townships. The government should strengthen air pollution control in the more polluted areas of southern Beijing and regulate illegal emissions from factories, enterprises, and vehicles. For areas with higher exposure risks in urban areas with lower overall pollutant concentrations, the government should pay more attention to the risk of high pollution exposure due to population density. At the same time, the government should take care to reduce the risk of exposure to pollution by minimizing the concentration of people and atmospheric pollution in months of high exposure and in areas of high exposure. Population densities and air pollution emissions should be controlled to reduce air pollution exposure in months and areas with high exposure risks. The layout of green landscapes in the urban area should be planned based on such results to reduce the concentration of pollutants in densely populated areas. Such efforts can help prevent residents from being subjected to high pollution exposure levels and intensity.