1. Introduction
Aerosols have extensive impacts on our climate and our environment [
1], and tropospheric aerosols (also known as particulate matter (PM)), in particular, can cause adverse effects on public health [
2]. Epidemiologic studies indicate strong links between the concentrations of PM with aerodynamic diameters of less than 10 μm and less than 2.5 μm (PM
10 or PM
2.5, respectively) with public morbidity, respiratory-related mortality and cardiovascular diseases [
3,
4,
5,
6,
7,
8]. The concentration of PM has become an important index of air pollution and has gained more and more attention from the administrations and organizations of environmental protection, public health and science around the world. Both the European Union (1999) and the United States have set air quality standards that dictate strict limits on PM concentrations in the ambient air.
In recent years, with the rapid development of industrialization and urbanization, PM has become the primary air pollutant in most major cities in China [
9]. This pollution not only threatens people’s health, but also causes decreases atmospheric visibility and degrades city scenery [
10]. Therefore, the Chinese government has enacted ambient air quality standards [
11], which limit the values of PM
2.5 concentrations and set air pollution classification rules.
During the daytime, aerosol characteristics, such as aerosol optical thickness (AOT), have been used to monitor the degree of PM concentrations. The relationship between AOT and PM concentrations has been thoroughly researched [
12,
13,
14,
15,
16,
17]. However, studies on nighttime PM pollution monitoring are very rare.
Since the 1970s, the Defense Meteorological Satellite Program-Operational Linescan System (DMSP-OLS) sensors have gathered meteorological data, which is archived by the National Oceanic and Atmospheric Administration (NOAA). DMSP imagery has a spatial resolution of 2.7 km and consists of two spectral bands and one thermal band. However, DMSP images are widely known not for their initial purpose, but for their ability to capture nighttime images of light on the Earth’s surface. Global population density and economic activity are clearly visible from space using this nighttime light imagery.
The DMSP-OLS sensors have a unique capacity for detecting faint light, and many research topics have been studied using this nighttime light imagery. A major application is the mapping of gross domestic product (GDP) and economic activity on global and regional scales [
18,
19,
20,
21,
22,
23]. Public lighting is a valuable indicator of a country’s economic condition and is directly reflected in nighttime light. Moreover, because larger populations need more public lighting, nighttime light also reflects population density [
24,
25]. Mapping human settlements is also possible because most nighttime light is emitted from human settlements [
26,
27,
28]. Furthermore, a great number of applications involving DMSP-OLS data exist in other areas of study, such as carbon cycling [
29], fishing boat mapping [
30], energy consumption [
31], security evaluation [
32] and ecological evaluation [
33].
In this study, we explored the relationship between daily DMSP nighttime light images and daily PM2.5 concentrations in Beijing, China. We created a back propagation (BP) neural-network model to estimate PM2.5 concentrations at night.
This article is organized as follows:
Section 2 describes our study area and data; presents the analysis of the nighttime light responses to PM
2.5 concentrations from both the spatial and temporal perspectives;
Section 3 verifies the findings from this study, and discusses future work; and
Section 4 summarizes the discoveries of this study.
3. Results and Discussions
The architecture of the BP neural-network was 4-6-1, and the learning algorithm was a gradient descent BP algorithm, with the learning rate of 0.2 and a momentum coefficient of 0.2. The activation function was a sigmoid tangent function in the input layer and hidden layer and a sigmoid logistic function in the output layer. In total, 23 groups of data were used for training, five groups were used for verification purposes, and five groups were used for testing.
The
R2 value of the correlation between the BP network model predictions and the RH_corrected PM
2.5 concentrations was 0.975. The RMSE was 26.26 μg/m
3 and the MBE value was −1.806 μg/m
3, with a corresponding average PM
2.5 concentration of 155.07 μg/m
3. The P
av value was 0.796 and the IA value was 0.988. The predicted PM
2.5 concentrations matched the RH_corrected PM
2.5 concentrations well, as shown in
Figure 5.
Figure 5.
Correlation of the relative humidity corrected (RH_corrected) PM2.5 concentrations and the PM2.5 concentrations predicated by the BP neural-network.
Figure 5.
Correlation of the relative humidity corrected (RH_corrected) PM2.5 concentrations and the PM2.5 concentrations predicated by the BP neural-network.
The BP network model using nighttime light data performs well in estimating the PM2.5 pollution degree of Beijing. The best results are obtained via the careful selection of regions in the DMSP-OLS nighttime light data.
Because the temporal difference between the 2013 Landsat 8 OLI-TIRS images and the 2013 nighttime light data is slight, we can compare these data sets to estimate the spatial consistency of urban areas. The lit areas of the city cover a larger area than the urban built-up areas (see Landsat 8 image). In addition, water bodies and urban forests are also illuminated in the DMSP-OLS data. This “blooming” phenomenon was also reported by other researchers [
27,
56]. The light intensity of the suburbs relative to the urban core contains information on the ground air’s extinction coefficient. Furthermore, the extinction coefficient can be used to estimate the particulate matter content in the air. Thus, the estimation of PM
2.5 concentrations is feasible using DMSP-OLS data.
DMSP-OLS sensors capture images every day, but the daily imagery is affected by sensor noise, atmospheric and moonlight variations. The National Oceanic and Atmospheric Administration (NOAA) has attempted to minimize the sensor noise. By using DMSP-OLS data from nights with little or no moonlight, we have minimized the effects of the along-scan gain and bidirectional reflectance distribution function (BRDF) algorithms. Thus, the daily nighttime light data are comparable in terms of the effects of atmospheric variations.
The DMSP-OLS data provide an effective method for estimating extinction coefficient variations in ground air with PM pollution. PM pollution reduces the transmission capability of light, resulting in differences in spatial distribution of nighttime light data (
Figure 6 [
57]). We presumed that area d was a light-emitting region and that areas b and c were not. The upward luminous flux of areas b and c was radiated from area d. This process is complex and nonlinear. To solve this problem, the BP neural-network was selected to estimate the degree of PM
2.5 pollution.
Figure 6.
The sketch map (
left) describes the position of a point light source and the surface to compute the luminous flux. Point O has a point light source with a light intensity of 1.0 unit (I
0). The distance from point O to point H is 1076 meters (The aerosol scale height of Beijing in winter is 1076 meters [
57]). A horizontal plane passes through point H and is perpendicular to the Z-axis; this plane is termed the Z = 1076 m Plane. Point P is on the Z = 1076 m Plane. The luminous flux through point P and parallel to the Z-axis is a function of two parameters. One is the distance from point P to point H; the other is the extinction coefficient of atmosphere. The line chart (
right) describes the function for a different situation. The X-axis is the distance from point P to point H, and the Y-axis is the upward luminous flux through point P and parallel to the Z-axis. The label e is the extinction coefficient of atmosphere. Area d was a light-emitting region, and areas b and c were not. The upward luminous flux of areas b and c was radiated from area d.
Figure 6.
The sketch map (
left) describes the position of a point light source and the surface to compute the luminous flux. Point O has a point light source with a light intensity of 1.0 unit (I
0). The distance from point O to point H is 1076 meters (The aerosol scale height of Beijing in winter is 1076 meters [
57]). A horizontal plane passes through point H and is perpendicular to the Z-axis; this plane is termed the Z = 1076 m Plane. Point P is on the Z = 1076 m Plane. The luminous flux through point P and parallel to the Z-axis is a function of two parameters. One is the distance from point P to point H; the other is the extinction coefficient of atmosphere. The line chart (
right) describes the function for a different situation. The X-axis is the distance from point P to point H, and the Y-axis is the upward luminous flux through point P and parallel to the Z-axis. The label e is the extinction coefficient of atmosphere. Area d was a light-emitting region, and areas b and c were not. The upward luminous flux of areas b and c was radiated from area d.

The primary data set of this research is DMSP imagery that provides information on the ground air’s extinction coefficient. We proposed an effective method to split DMSP imagery, intentionally highlighting the differences in the spatial distribution of NTL data. Future studies should focus on the quantitative relationship between PM pollution and DMSP-OLS data throughout the lunar cycle. New sources of high spatial resolution nighttime images, such as the NPP-VIIRS Nighttime Light Data [
58,
59,
60], the EROS-B commercial satellite data [
61] and aerial photography [
62], may allow for better estimation methods and results.
Meteorological conditions are a significant driver of local ambient air pollution concentrations, especially wind speed and relative humidity. Wind can disperse particulate matter but cannot influence the extinction coefficient. Therefore, relative humidity was used to correct the daily “dry” PM2.5 concentrations. DMSP-OLS data can be obtained only in clear and cloudless days. Other days were not included in this study. Nighttime inversion conditions and upper air meteorological conditions (turbulence, etc.) may impact the light transmission from the ground to satellite, and we will consider these factors in the future studies.
4. Conclusions
By selecting Beijing, the largest and fastest developing city in the world, as the study area, this research synthesizes RS and GIS techniques and estimates the PM2.5 pollution degree of the ground-level atmosphere. This study proposed a simple monitoring method of nighttime PM2.5 concentrations.
The innovation of this research lies in the selection of spatial regions in the data set. Based on the land use type to the west of Beijing, the distribution characteristics of the land and the light scattering from urban to suburban, four regions were defined from which to extract the nighttime light data, instead of constructing various indices from the entire DMSP-OLS imagery data set. The relative numerical ratio of the data extracted from these four regions reflects the extinction coefficient of the atmosphere, and this extinction coefficient can be used to retrieve the aerosol concentration.
Our method and results can provide guidance for developing and implementing environmental conservation planning. Furthermore, these data can assist government agencies in determining PM pollution control areas, initiating regulation projects, and undertaking nighttime environmental purifying measures.