1. Introduction
PM
2.5 is a fine particulate matter (PM), which has an aerodynamic hollow diameter ≤2.5 μm, and is one of the main air pollutants [
1]. To address increasing air pollution, the World Health Organization has recommended lowering the average annual PM
2.5 concentrations from 10 µg/m
3 to 5 µg/m
3 [
2]. Long-term exposure to elevated PM
2.5 concentrations is established to cause health problems and increase mortality risk [
3,
4,
5,
6,
7,
8]. However, recent studies have found that even low PM
2.5 concentrations that meet the air-quality standards can increase the risk of premature death [
9]. Thus, considerable attention has been devoted to investigating PM
2.5 sources, health effects, and long-term monitoring of its variation [
10,
11,
12]. While satellite remote sensing has collected much data on daytime PM
2.5 concentrations, research at night is scarce. The lack of nighttime PM
2.5 data hampers research on its continuous temporal and spatial variation. Continuous monitoring of both day and nighttime PM
2.5 concentrations will help to better understand its formation and dissipation [
13,
14].
Strong correlations were found between PM
2.5 concentrations and aerosol optical depth (AOD). Hence, many AOD-PM
2.5 models have been constructed to estimate ground PM
2.5 concentrations [
15,
16,
17,
18]. For example, in order to create an ensemble model of the PM
2.5 concentration in the London region, Yazdi et al. [
19] employed the AOD products from the Moderate Resolution Imaging Spectroradiometer (MODIS), land-cover type, and meteorological data as the inputs. The model comprised K-nearest neighbor, gradient-boosting machine, and random forest (RF). Using the Top-of-Atmospheric Reflectance (TOAR) and meteorological data, Feng et al. [
20] calculated the PM
2.5 concentrations in the BTH region. Directly using TOAR products avoids the intermediate process of AOD inversion and improves the accuracy. The characterization of TOAR at night requires accurate lunar irradiance at the top of the atmosphere. However, lunar irradiance varies dramatically depending on the variable of the moon phase and the geometry of the Sun/Earth/Moon [
21]. At present, there is still a lack of high-precision lunar irradiance models and nighttime radiation transmission models to accurately describe the nighttime radiation transmission process. Therefore, it is difficult to obtain accurate TOAR and AOD at night, which limits their applications in estimating PM
2.5 concentrations.
The successful launch of the Suomi National Polar-Orbit Partnership (NPP) satellite has provided remote sensing data support for night monitoring of the atmosphere [
22]. The first study of nighttime aerosols using a Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) proposed obtaining nighttime aerosol optical depth from the difference between the observed urban light sources and dark background radiation [
23]. Based on this study [
23], McHardy et al. [
24] proposed a spatial standard deviation (spatial derivative) for urban lighting to estimate nighttime atmospheric aerosols. This method does not require calculating the radiation of the dark background area without an artificial light source. Next, Zhang et al. [
25] followed up on the methods of McHardy et al. using surface diffuse reflection to characterize the spatial variation in artificial light sources and then estimated nighttime AOD. More recently, a different study established a multiple linear regression (MLR) model with VIIRS/DNB, atmospheric water vapor content, air pressure, and PM
2.5 concentration in the Atlanta region based on the radiative transfer theory [
26]. Thus, VIIRS/DNB-derived techniques demonstrate that nighttime PM
2.5 concentration can be measured. Subsequently, Zhao et al. [
27] developed a nighttime PM
2.5 concentration inversion model based on the back propagation (BP) neural network using VIIRS/DNB radiance combined with meteorological variables (including relative humidity and temperature). The correlation coefficient of validation across 12 Beijing sites was 0.91, further verifying the feasibility of VIIRS/DNB radiance as a method to estimate nighttime PM
2.5 concentrations. Moreover, Fu et al. [
28] established a mixed-effect model (MEM) between the VIIRS/DNB radiance and PM
2.5 concentration, and then tested it in both the light-intensive urban areas and light-sparse suburbs of Beijing. The correlation coefficients of the model were 0.95 ± 0.02 in the metropolitan area and 0.88 ± 0.06 in the suburbs. Furthermore, VIIRS/DNB radiance was found capable of estimating PM
2.5 concentrations in sparsely lit areas. Based on the relationship between PM
2.5 concentration and AOD, Erkin et al. [
29] proposed a hybrid adaptive deep-learning algorithm. The input-independent variables were VIIRS/DNB
AOD, meteorological data, and auxiliary data, and the output was the PM
2.5 concentration. This study focused on the changes in the PM
2.5 concentration caused by fireworks in cities before, during, and after the Spring Festival. The results showed that the rapid increase in PM
2.5 concentration caused by the fireworks affected the monthly average concentration of PM
2.5 during the Spring Festival and that in the period after the Spring Festival.
Multiple linear regression, the mixed effect model, and various machine-learning methods have been involved in the study of nighttime PM2.5 concentration estimation, but only data with a large moon phase angle are used to reduce the influence of moonlight. Therefore, this study selected VIIRS/DNB radiance under all moon phase angles and incorporated meteorological data to establish a multi-parameter comprehensive random forest model. Our aim was to validate this model for estimating the nighttime PM2.5 concentration in the BTH region.
4. Discussion
This study estimated the nighttime PM
2.5 concentrations in the BTH region during 2021 using a combination of VIIRS/DNB radiance and PM
2.5 concentration data from ground stations. Based on the VIIRS/DNB radiance and moon phase angles combined with meteorological factors (SP, PBLH, AT, RH, and WS) as independent variables, we established several PM
2.5 concentration estimation models, including the NightPMES, MLR, and DNN. Our results show that the NightPMES model performed significantly better than the MLR or DNN models. The NightPMES model had an R
2, RMSE, and MAE of 0.82, 16.67, and 10.20, respectively. Additionally, the NightPMES model had better performance than most previous models (
Table 5). The nighttime PM
2.5 concentration obtained with this model effectively characterized the spatial patterns of the nighttime atmosphere in the BTH region. Moreover, we obtained more information to monitor the nighttime atmospheric environment.
The possible errors of the proposed NightPMES model were analyzed. First, this study used the threshold method to remove clouds, and there may be residual cloud pixels, especially for thin clouds, as the reflectivity of clouds is usually higher than the surface, and its transmittance is lower than PM2.5. The residual cloud pixels will enhance the reflection of moonlight radiation when there is moonlight, while it will reduce the upward radiation of ground light when there is only ground light. Therefore, it will be mistaken for the contribution of PM2.5. Second, there are obvious spatial differences between the ground observation and satellite observation. Usually, the ground observation represents a single point, while the satellite observation is the average of a certain range (spatial resolution of 750 m). Therefore, there will be a smooth effect, which leads to the overestimation of PM2.5 when the concentration is low. Third, the spatial resolution of the meteorological factors used in this study is coarse (0.5° × 0.625°), resulting in the spatial variation of the meteorological factors being smoothed. The meteorological data with smooth spatial changes also have a mean effect, which leads to errors in the PM2.5 concentration estimation.
Various complex relationship models for estimating the nighttime PM
2.5 concentration can be obtained via combining satellite data with different base models. A model that ignored the influence of moonlight (including only data with phase angle > 120°) performed better than the NightPMES model (
Figure 8). This outcome indicates that the estimation model of nighttime PM
2.5 concentration is sensitive to lunar illumination. The moon phase angle magnitude reflects the influence of lunar radiation; therefore, a lunar irradiance model is necessary for quantifying this influence in subsequent research [
45,
46].
There were four distinct seasons in the BTH region, and we observed clear spatiotemporal variations to the nighttime PM
2.5 concentrations (
Figure 6). Seasonally, winter saw greater PM
2.5 concentrations, whereas summer saw lower levels. Spatially, Handan, Xingtai, and Shijiazhuang in the southeast had higher concentrations, whereas Zhangjiakou and Chengde in the northwest had lower concentrations. With our seasonal model, we analyzed the influence of the seasonality of the nighttime PM
2.5 concentration in summer and winter, providing a reference for further research. Additionally, based on the analysis of daytime and nighttime PM
2.5 concentrations, we found that PM
2.5 pollution can be better understood if nighttime data are included. Overall, the NightPMES model provides a reliable estimation of the nighttime PM
2.5 concentration in the BTH region and should help improve PM
2.5 monitoring efforts.
Although this study proposed a better model for nighttime PM2.5 concentration monitoring, it has several limitations. The NightPMES model was only tested in the BTH region, so to validate it further, studies with larger spatial scales and extended time series are needed. In addition, although we included several variables that may influence nighttime PM2.5 concentration, we did not consider many other factors, such as NDVI, precipitation, elevation, population, and economy. Adding more variables into our model will possibly increase the estimation accuracy. Furthermore, we mainly concentrated on urban areas, thus limiting model applicability. Increasing the number of suburban sites in future research will address this issue. Finally, the VIIRS/DNB measures of nighttime radiation mainly originate from moonlight and ground light. Therefore, the contribution of moonlight must be quantified. To optimize the monitoring and estimation of nighttime PM2.5 concentration, a moon irradiance model can be introduced into the NightPMES model.
Satellites equipped with low-light sensors, such as FY-3E and the Sustainable Development Science Satellite 1 (SDGSAT-1), have been launched recently. These sensors have higher temporal and radiation resolution, providing an excellent opportunity for monitoring the nighttime atmospheric environment.
5. Conclusions
Here, we generated a NightPMES model for estimating the nighttime PM2.5 concentration in the BTH region using DNB radiance, meteorological elements, and moon phase angle. After experimental verification, we arrived at four conclusions.
First, the NightPMES model (R2 = 0.82, RMSE = 16.67, and MAE = 10.20) performed better than the MLR or DNN models.
Second, when all moon phase angles were included, we increased the available data by approximately 60% from previous studies. Hence, the NightPMES model was capable of obtaining daily estimations for PM2.5 concentration.
Third, significant seasonal differences in the nighttime PM2.5 concentrations are explained by the seasonal characteristics of DNB radiation (highest in winter and lowest in summer), along with obvious seasonal meteorological conditions.
Fourth, the dynamics of PM2.5 concentration can be analyzed more accurately when the nighttime PM2.5 concentration (estimated by the NightPMES model) is combined with the daytime PM2.5 concentration.
To conclude, pollution monitoring should take advantage of the many satellites equipped with low-light imagers that have been or will be launched. This study validates the use of low-light remote sensing data to estimate nighttime PM2.5 concentration and provides a basis for applying satellite data to nighttime atmospheric environmental monitoring.