Analysis of Primary Air Pollutants’ Spatiotemporal Distributions Based on Satellite Imagery and Machine-Learning Techniques

Abstract

Accurate monitoring of air pollution is crucial to human health and the global environment. In this research, the various multispectral satellite data, including MODIS AOD/SR, Landsat 8 OLI, and Sentinel-2, together with the two most commonly used machine-learning models, viz. multi-layer backpropagation neural network (MLBPN) and random forest (RF), have been employed to analyze the spatiotemporal distributions of the primary air pollutant from 2019 to 2022 in Guanzhong Region, China. In the conducted experiments, the RF-based model, using the MODIS AOD data, has generally demonstrated the “optimal” estimation performance for the ground-surface concentrations of the primary air-pollutants. Then, the “optimal” estimation model has been employed to analyze the spatiotemporal distribution of the various air pollutants—in terms of temporal distribution, the annual average concentrations of PM_2.5, PM₁₀, NO₂, and SO₂ in the research area showed a decreasing trend from 2019 to 2022, while the annual average concentration of CO remained relatively stable and the annual average concentration of O₃ slightly increased; in terms of the spatial distribution, the air pollution presents a gradual increase from west to east in the research area, with the distribution of higher concentrations in the center of the built-up areas and lower in the surrounding rural areas. The proposed estimation model and spatiotemporal analysis can provide reliable methodologies and data support for the further study of the air pollution characteristics in the research area.

1. Introduction

The swift progression of the global economy, coupled with rapid urban expansion and escalating industrial production, has amplified the strain on the atmospheric environment. The consequential atmospheric pollution can not only disrupt the delicate balance of ecosystems and contribute to global climate change but can also pose significant risks to human health [1]. The traditional ground-based observation methods are only able to provide a limited amount of station monitoring data, which cannot comprehensively represent the characteristics regarding the continuous spatial distribution and the temporal changing trends of the primary air pollutants in expansive spatial areas. Characterized by their high timeliness, extensive coverage, and superior resolution, satellite remote-sensing technologies have emerged as a promising solution to compensate for the limitations of the ground-monitoring methods, which have thereby increasingly been leveraged for monitoring atmospheric pollution.

Previous research has mainly relied on limited ground-based monitoring station data and geographic information system (GIS) technology to obtain the spatial distribution of air pollutants. Guo et al. [2] analyzed the temporal–spatial variation characteristics of six air pollutants in 366 major cities in China from 2015 to 2017 using data on the daily average mass concentrations of air pollutants monitored at about 1500 stations published by the Ministry of Environmental Protection (MEP). Han et al. [3] analyzed the spatiotemporal variations in PM_2.5 in South Korea among 49 pollution events by using Kriging interpolation and statistical analyses based on the observations from 462 air quality monitoring stations (AQMS) in South Korea from 2015 to 2020. Wang et al. [4] utilized the air pollutant concentration data from 23 monitoring stations in Nanchang City from 2017 to 2020 and obtained the spatial distribution maps of air pollutants by Kriging interpolation to explore the spatiotemporal changes in urban air pollution. However, the limited monitoring stations and spatial coverage make it difficult to depict the spatial distribution of air pollutants comprehensively and accurately. Since the 1960s, with the continuous development of satellite remote-sensing technology, some scholars have begun to use this technology to study the spatial distribution of air pollutants. Yuan et al. [5] analyzed the aerosol spatiotemporal distribution characteristics of Hangzhou using the moderate-resolution imaging spectroradiometer (MODIS) 3 km AOD products from 2012 to 2020, and found that MODIS AOD data can be used as an important basis for atmospheric studies in Hangzhou.

Aerosol research is a critical aspect of advancing contemporary atmospheric quality monitoring technology. Aerosol optical depth (AOD) serves as a measure of aerosol extinction in the vertical direction, closely linked to near-surface atmospheric particulate matter [6,7]. Wang et al. [8] studied the relationship between column aerosol optical thickness (AOT) and PM_2.5 mass in the United States and demonstrated that MODIS AOT can be used quantitatively to estimate air quality. Engel-Cox et al. [9] conducted correlation research on MODIS AOD and ground-based particulate matter concentration values at urban and regional scales and found that the correlations of MODIS AOD with ground-based particulate matter were better. Su et al., Qin et al., and Zhang et al. found that fine particulate matter concentration correlated more strongly with AOD through statistical modelling of satellite-observed aerosol optical thickness data and ground-based particulate matter observations, with a better estimation accuracy [10,11,12]. Meanwhile, considering that AOD can be estimated from top-of-atmosphere reflectance (TOA), some researchers have proposed to utilize TOA to estimate PM_2.5 concentration directly and proved its effectiveness and applicability [13,14]. Satellite estimation of air pollution fundamentally revolves around establishing the spectral information–pollutant concentration relationship. Directly employing reflectance data for atmospheric pollutant estimation not only streamlines the remote-sensing estimation process but also mitigates cumulative errors in AOD estimation procedures [15].

Air pollutant distribution is influenced by various factors such as meteorological conditions, emission sources, underlying surfaces, and physicochemical processes, exhibiting strong nonlinear characteristics. Consequently, machine-learning methods suited for solving nonlinear mapping problems have emerged as a crucial approach for the continuous spatial estimation of air pollutants. For instance, Gupta et al. [16] estimated the PM_2.5 concentration in the southeastern United States using artificial neural networks together with MODIS AOD data and meteorological data, demonstrating the potential of artificial neural networks in air quality monitoring. Yang et al. [17] estimated the spatial continuous distribution of O₃ in the Beijing–Tianjin–Hebei Region based on Landsat 8 reflectance data and backpropagation neural networks, and the coefficient of determination (R²) between the estimated and observed values achieved was 0.9, yielding satisfactory results. Furthermore, machine-learning models such as random forest (RF), support vector machine (SVM), and deep learning (DL) have also found extensive applications [18,19,20].

Although extensive research explores satellite remote sensing and machine learning for air pollutant estimation, most studies have concentrated on individual models and a limited number of pollutants, such as PM_2.5/PM₁₀ or O₃, and relied primarily on one kind of satellite data. However, there has been a relative scarcity of research on the comprehensive performance assessment and spatiotemporal analysis of a wider range of air pollutants based on multiple machine-learning models and satellite data [21].

In this research, the various publicly available multispectral satellite data, including MODIS AOD/SR, Landsat 8 OLI, and Sentinel-2, together with the two most commonly used and well-established machine-learning models, viz. the multi-layer backpropagation neural network (MLBPN) and random forest (RF), have been employed to establish the estimation model of the six primary air pollutants’ concentrations, including particulate matter PM_2.5 and PM₁₀, as well as the trace gases O₃, CO, NO₂, and SO₂. The “optimal” estimation model and satellite data have been selected by comparative analysis to obtain high spatial resolution concentration distribution maps to analyze the spatiotemporal distribution of air pollutants in the research area. The results can provide a scientific basis for air pollution characterization and decision support for the formulation of relevant environmental protection measures.

2. Materials and Methods

2.1. Research Area

The Guanzhong Region, showed in Figure 1, situated in the central part of Shaanxi Province, China, which spans approximately 106°56′~110°22′ E longitude and 33°39′~35°52′ N latitude, has been selected as the research area because of its typical plain topography and because it is one of the most severely polluted air regions, with representative cities such as Xi’an, Xianyang, and Weinan consistently ranking among the bottom 20 of 168 key cities of China. Bounded by the Qinling Mountains to the south and the Loess Plateau to the north, it stretches about 360 km from east to west and has an average elevation of around 500 m. The terrain gradually descends from west to east. The region exhibits a warm temperate continental monsoon climate, characterized by an average annual temperature of 12 °C to 14 °C and an annual average rainfall of 530 to 750 mm. Dominant prevailing winds throughout the year are from the northeast, followed by secondary winds from the southwest. The average relative humidity ranges between 60% and 70%. Encompassing five cities—Xi’an, Baoji, Xianyang, Weinan, and Tongchuan—the Guanzhong Region spans a total area of 55,623 square kilometers and is classified as one of China’s four major geographical divisions, the northern region.

2.2. Data Sources

A substantial volume of high-quality data stands as a prerequisite for effective machine-learning applications. As illustrated in

Table 1. Description of the data employed in the research.

Category	Variable	Unit	Temporal Resolution	Spatial Resolution	Data Source
Satellite remote-sensing images	MCD19A2	Band	1–2 Day	1000 m	GEE, USGS
	MOD09GA	Band	1 Day	500 m
	Landsat 8	Band	16 Day	15 m, 30 m
	Sentinel-2	Band	5 Day,10 Day	10 m, 20 m, 60 m
Air quality data	PM2.5	μg/m3	1 h		CNEMC
	PM10	μg/m3
	SO2	μg/m3
	NO2	μg/m3
	O3	μg/m3
	CO	mg/m3
Meteorological data	Temperature	°C	3 h		CMA-NOAA
	Relative humidity	%
	Atmospheric pressure	Pa
	Wind direction	◦
	Wind speed	m/s

, the data utilized in this research comprised station data published by the China National Environmental Monitoring Center (CNEMC) and satellite imagery data from the Google Earth Engine (GEE, https://earthengine.google.com, accessed on 16 July 2023) platform and the United States Geological Survey (USGS, https://earthexplorer.usgs.gov/, accessed on 16 July 2023) including MODIS AOD/SR, Landsat 8 OLI, and Sentinel-2, as well as meteorological data from ground-based monitoring stations operated by the China Meteorological Administration (CMA) and the National Oceanic and Atmospheric Administration (NOAA) in the United States. All of these datasets are openly accessible.

2.2.1. MODIS Data

The moderate-resolution imaging spectroradiometer (MODIS) is equipped on both the Terra and Aqua satellites. It captures data in 36 spectral bands covering the spectral range from 0.4 μm (visible light) to 14.4 μm (thermal infrared) with spatial resolutions ranging from 250 to 1000 m. These instruments provide complete spectral coverage and revisit the Chinese mainland area 1–2 times per day [22]. This research utilized the daily land aerosol optical depth (AOD) data MCD19A2 and daily surface reflectance (SR) data MOD09GA.

2.2.2. Landsat 8 Data

The Landsat 8 satellite was launched in February 2013 and began offering free data in May of the same year. It carries two sensors: the operational land imager (OLI) and the thermal infrared sensor (TIRS) [23]. The OLI sensor comprises 9 spectral bands, with a spatial resolution of 15 m for panchromatic bands and 30 m for other bands. The imaging swath width is 185 × 185 km, with a revisit cycle of 16 days. In this research, we utilized band data from Landsat 8 Collection 2 Tier 1 Raw Scenes products.

2.2.3. Sentinel-2 Data

Sentinel-2 is an Earth observation mission from the Copernicus program by the European Space Agency, primarily aimed at observing the Earth’s surface to provide various remote-sensing services such as agricultural monitoring, emergencies management, land-cover classification, or water quality. This mission consists of two identical satellites, Sentinel-2A and Sentinel-2B, equipped with 13 spectral bands ranging from the visible to the shortwave infrared spectrum. The spatial resolutions are 10 m, 20 m, and 60 m, respectively. The revisit period for each satellite is 10 days [24]. In this research, we utilized surface reflectance data from Sentinel-2A.

2.2.4. Air Quality Data

The China National Environmental Monitoring Center (CNEMC) is a specialized institution under the Ministry of Ecology and Environment. It is responsible for nationwide environmental monitoring across various domains, including air, water, ecology, soil, coastal areas, noise, and pollution sources. Presently, CNEMC provides real-time monitoring concentration data for PM_2.5, PM₁₀, O₃, CO, NO₂, and SO₂. These datasets serve as the primary foundation for estimating surface air pollutant concentrations based on remote sensing and can be collected from the CNEMC platform (http://www.cnemc.cn, accessed on 16 July 2023). CNEMC operates over 1500 monitoring stations throughout mainland China, including 41 stations in the research area (Figure 1).

2.2.5. Meteorological Data

The National Oceanic and Atmospheric Administration (NOAA) is the governing body responsible for meteorological operations in the United States. In 2013, the China Meteorological Administration (CMA) and NOAA entered into a collaborative agreement, authorizing NOAA to publicly release meteorological data for mainland China on the Internet (https://gis.ncdc.noaa.gov/maps/ncei/cdo/hourly, accessed on 16 July 2023). These data are updated every three hours. This research utilized atmospheric factors including atmospheric pressure (PRS), relative humidity (RH), temperature (T), wind direction (WD), and wind speed (WS) [25]. Particularly, the calculation of relative humidity was obtained using the following formula [26]:

$$T_d=\frac{b\gamma \left(T,RH\right)}{a-\gamma \left(T,RH\right)},$$

(1)

$$\gamma \left(T,RH\right)=\frac{aT}{b+T}+ln\left(RH/100\right),$$

(2)

where T and dew point temperature (T_d) are in Celsius, RH is in percent, ln represents the natural logarithm, the constant a is 17.27, and the constant b is 237.7 °C.

CMA-NOAA has more than 300 cooperative monitoring stations in mainland China, including four stations in the research area (Figure 1).

2.3. Data Preprocessing

Due to the wide variety of data sources utilized in this research, significant differences exist in data types and resolutions. Therefore, preprocessing and spatiotemporal matching of the aforementioned research data are imperative.

(1) Regarding outliers in land stations, we referenced the “Technical Regulation on Ambient Air Quality Index (on trial)” and conducted outlier removal procedures accordingly. Quality control tools provided by Google Earth Engine (GEE) were employed to eliminate satellite data samples significantly affected by clouds. With the MODIS data, MCD19A2 removed clouds based on the AOD_QA quality control band (Bits8-11 = “0000”), while MOD09GA removed clouds based on the state_1 km band (Bits0-2 = “000” and Bit10 = “0”). Landsat 8 data removed clouds based on the QA_PIXEL quality control band (Bits3-4 = 00), and Sentinel-2 data used the QA60 band for quality control, retaining data with QA60 = 0 [27]. High-quality satellite remote-sensing images of the research area were obtained through mosaicking, projection transformation, resampling, reprojection, and cropping. Kriging interpolation was applied to meteorological station data to acquire spatial meteorological data in the research area.

(2) The above data require spatial and temporal matching due to differences in observation time, temporal resolution and spatial resolution. Regarding the different temporal resolutions of the data, the imaging time properties of each satellite are the time reference for time matching. The imaging time properties of each satellite are regarded as the temporal reference to match air quality data and meteorological data. For MODIS satellite products that provide daily averaged data, the air quality and meteorological data are processed into daily averages and matched accordingly. For Landsat 8 and Sentinel-2 satellite products that provide instantaneous data, the air quality and meteorological data are matched using linear interpolation and a time buffer interval of ±1 h. The near analysis algorithm (NAA) has been employed for spatial matching to integrate the CEMC observations and NOAA datasets. As for the normalized difference vegetation index (NDVI), it was calculated by the NIR and red bands according to Formula (3) in MODIS SR, Landsat 8 OLI, and Sentinel-2, while it was matched by the NDVI band of MOD13A2 product in MODIS AOD.

$$NDVI=\frac{NIR-Red}{NIR+Red},$$

(3)

(3) Given the differences in data scales and measurement units, normalization of the data is necessary to minimize model errors. This research employed the min–max normalization method:

$$I_n=\frac{I-I_{min}}{I_{max}-I_{min}},$$

(4)

where I represents the original feature data, $I_{min}$ and $I_{max}$ represent the minimum and maximum values of the feature data, respectively, and $I_n$ represents the normalized feature data.

2.4. Methods

As shown in Figure 2, machine-learning methods were employed to explore and establish the relationships between internal features to achieve spatial estimation after completing the data collection and preprocessing. This research employed multi-layer backpropagation neural network (MLBPN) and random forest (RF) models, which have demonstrated a favorable performance in air pollutant estimation, as evidenced by both this experiment and previous literature. In order to strike a balance between training the model effectively and thoroughly evaluating its performance, the dataset was partitioned into training and testing sets using an 80/20 split [28]. Furthermore, given the temporal properties of data and the large data volume, cross-validation was performed using the “Hold-out Method”. All the experiments were conducted on a Windows 11 Professional 64-bit operating system with 12th Gen Intel(R) Core(TM)i7-12700KF CPU configuration and NVIDIA GeForce RTX 3070Ti GPU configuration.

2.4.1. Multi-Layer Backpropagation Neural Network (MLBPN)

The multi-layer backpropagation neural network (MLBPN) model is a type of multi-layer feedforward neural network trained using the backpropagation algorithm, which demonstrates clear advantages in addressing nonlinear system problems. The MLBPN-based model consists of an input layer, several hidden layers, and an output layer [29]. Upon providing a set of learning samples to the MLBPN-based model, activation values propagate from the input layer through the intermediate hidden layers to the output layer, acquiring the network’s input response at the output layer neurons. Subsequently, the connection weight values are incrementally adjusted layer by layer, from the output layer through each intermediate layer, following the direction to minimize the target output and actual error, ultimately returning to the input layer. In Figure 2, x_i represents the input value of this neural network (i = 1, 2, …, m), and Y represents the output value. $w_{ij}^l$ represents the connection weight between node j in hidden layer l-1 and node i in hidden layer l. $b_i^l$ represents the bias of node i in layer l. The transfer between different hidden layers utilizes the Tansig function equation, while the Purelin function equation is employed between the hidden layer and the output layer [30,31,32]. They are expressed as follows:

$$tansig\left(x\right)=\frac{2}{1+e^{-2x}}-1,$$

(5)

$$Purelin\left(x\right)=x,$$

(6)

The optimal number of nodes for the hidden layers in the model is set to [15,15]. The Levenberg–Marquardt (LM) algorithm is utilized for neural network training using the trainlm function, with the maximum training epochs set to 500, the network training goal set to 0.001, and a learning rate of 0.1.

2.4.2. Random Forest (RF)

The random forest (RF) algorithm is an optimization of the decision tree method, constructing multiple independent decision trees and averaging their outcomes for fitting. This approach effectively addresses the issues of low accuracy and overfitting commonly associated with individual decision trees, thereby enhancing the generalization ability of the algorithm. The decision trees in a random forest are binary trees, following the recursive top-down splitting pattern, and nodes are split based on the principle of minimizing node impurity. In this research, the TreeBagger function from the Matlab R2021b software was employed for remote-sensing estimation of surface air pollutant concentrations. During the model optimization phase, adjustments were made to enhance the estimation accuracy by setting the number of decision trees to 100 and the minimum leaf size to 5.

2.5. Validation

In order to better evaluate the model, four statistical indicators, the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean error (ME), were chosen for comparative analysis in this research [33]. Their calculation formulas are as follows, respectively:

$$R^2=\frac{cov\left(P_G-P_S\right)}{\sqrt{var\left(P_G\right)var\left(P_S\right)}},$$

(7)

$$RMSE=\sqrt{\frac{\sum _1^N{\left(P_G-P_S\right)}^2}{N}},$$

(8)

$$MAE=\frac{1}{N}\sum _1^N\left|\left(P_G-P_S\right)\right|,$$

(9)

$$ME=\frac{1}{N}\sum _1^N\left(P_G-P_S\right),$$

(10)

where $P_G$ denotes the pollutant concentration estimation value, $P_S$ denotes the pollutant concentration measured value, and $N$ denotes the number of samples.

3. Results

3.1. Model Performance

The comparison of the estimation performances of the RF and MLBPN models for the concentrations of air pollutants, based on the MODIS AOD data with a total of 15,577 samples in the research area from 2019 to 2022, is shown in

Table 2. Estimation performances of the RF and MLBPN models using MODIS AOD data.

Model	RF				MLBPN
Accuracy evaluation	R2	RMSE	MAE	ME	R2	RMSE	MAE	ME
		(μg/m3)	(μg/m3)	(μg/m3)		(μg/m3)	(μg/m3)	(μg/m3)
PM2.5	0.90	11.67	7.87	0.20	0.80	16.36	11.55	−0.08
PM10	0.87	23.99	16.39	0.05	0.67	35.86	26.02	−0.06
O3	0.92	12.26	9.09	−0.17	0.84	16.63	12.95	0.02
NO2	0.73	9.20	6.84	0.11	0.54	12.02	8.94	−0.21
CO	0.75	180	130	2	0.66	210	150	3
SO2	0.52	3.69	2.54	0.15	0.36	4.55	3.09	0.07

. Compared to the MLBPN-based model, the R² of the RF-based model improves by about 0.1 for PM_2.5, O₃, and CO and 0.2 for PM₁₀, NO₂, and SO₂; at the same time, the RMSE and the MAE are significantly decreased. These values indicate that the RF-based model has a better estimation performance than the MLBPN-based model. In the context of the RF-based model, the R² values for PM_2.5, PM₁₀, and O₃ are around 0.9; for NO₂ and CO exceed 0.7; and for SO₂ surpass 0.5, which indicates that the MODIS AOD data can accurately estimate the continuous spatial distribution of air pollutants.

The comparison of the estimation performances of RF and MLBPN models for the concentrations of air pollutants, based on the MODIS SR data with a total of 7821 samples in the research area from 2019 to 2022, is shown in

Table 3. Estimation performances of the RF and MLBPN models using MODIS SR data.

Model	RF				MLBPN
Accuracy evaluation	R2	RMSE	MAE	ME	R2	RMSE	MAE	ME
		(μg/m3)	(μg/m3)	(μg/m3)		(μg/m3)	(μg/m3)	(μg/m3)
PM2.5	0.77	11.46	7.84	0.13	0.66	14.00	9.98	−0.06
PM10	0.72	25.57	16.71	0.08	0.65	29.28	21.49	−1.79
O3	0.82	14.09	10.88	0.15	0.75	16.64	12.99	−0.35
NO2	0.61	10.39	8.08	−0.16	0.52	11.05	8.63	−0.53
CO	0.58	170	130	3	0.50	180	130	2
SO2	0.31	4.06	2.63	0.12	0.30	4.44	2.88	−0.08

. Compared to the MLBPN-based model, the R² values of the RF-based model for PM_2.5, PM₁₀, O₃, NO₂, and CO improve by about 0.1; simultaneously, the RMSE and the MAE are significantly decreased. These values indicate that the RF-based model has a better estimation performance than the MLBPN-based model. For the RF-based model, the R² values for PM_2.5 and O₃ are around 0.8; PM₁₀ exceeds 0.7; NO₂ and CO are around 0.6; and SO₂ surpasses 0.3. These results indicate that the MODIS SR data can effectively estimate the continuous spatial distribution of PM_2.5, PM₁₀, O₃, NO₂, and CO.

The comparison of the estimation performances of the RF and MLBPN models for the concentrations of air pollutants, based on the Landsat 8 OLI data with a total of 1181 samples in the research area from 2019 to 2022, is shown in

Table 4. Estimation performances of the RF and MLBPN models using Landsat 8 OLI data.

Model	RF				MLBPN
Accuracy evaluation	R2	RMSE	MAE	ME	R2	RMSE	MAE	ME
		(μg/m3)	(μg/m3)	(μg/m3)		(μg/m3)	(μg/m3)	(μg/m3)
PM2.5	0.86	16.83	10.83	−1.69	0.88	13.30	9.60	0.35
PM10	0.71	30.52	22.41	−3.08	0.80	27.95	21.52	0.09
O3	0.85	16.16	11.51	−0.88	0.89	15.3	11.87	−2.05
NO2	0.69	9.33	7.23	0.37	0.64	10.94	8.54	−0.91
CO	0.72	200	150	−4	0.76	200	150	−10
SO2	0.57	4.07	2.87	−0.09	0.59	4.03	2.94	−0.17

. It can be seen that the difference in accuracy between these two models is relatively close, but on the whole, the MLBPN-based model performs better than the RF-based model. For the MLBPN-based model, the R² of PM_2.5 and O₃ is close to 0.9; that of PM₁₀ and CO is close to 0.8; and that of NO₂ and SO₂ is around 0.6, which indicates that the Landsat 8 OLI data can accurately estimate the continuous spatial distribution of air pollutants.

The comparison of the estimation performances of RF and MLBPN models for the concentrations of air pollutants, based on the Sentinel-2 data with a total of 7933 samples in the research area from 2019 to 2022, is shown in

Table 5. Estimation performances of the RF and MLBPN models using Sentinel-2 data.

Model	RF				MLBPN
Accuracy evaluation	R2	RMSE	MAE	ME	R2	RMSE	MAE	ME
		(μg/m3)	(μg/m3)	(μg/m3)		(μg/m3)	(μg/m3)	(μg/m3)
PM2.5	0.90	12.52	8.20	−0.30	0.76	18.34	13.55	0.60
PM10	0.91	25.99	17.73	−0.49	0.76	33.47	25.49	−0.66
O3	0.92	12.48	9.08	0.61	0.83	18.08	13.97	−0.30
NO2	0.77	9.36	6.50	0.11	0.58	12.63	9.16	−0.98
CO	0.79	160	110	1	0.62	210	160	8
SO2	0.65	3.26	2.23	0.07	0.40	4.13	3.04	−0.02

. Compared to the MLBPN-based model, the R² of the RF-based model improves about 0.1 for O₃, 0.15 for PM_2.5 and PM₁₀, 0.2 for NO₂ and CO, and 0.25 for SO₂; at the same time, the RMSE and the MAE are significantly decreased. These values indicate that the RF-based model has a better estimation performance than the MLBPN-based model. The RF-based model demonstrates a strong performance with R² values exceeding 0.9 for PM_2.5, PM₁₀, and O₃; close to 0.8 for NO₂ and CO; and larger than 0.6 for SO₂, which indicates that the Sentinel-2 data can accurately estimate the continuous spatial distribution of air pollutants.

In summary, the RF-based model has revealed a better estimation performance using the satellite imagery from MODIS AOD/SR and Sentinel-2, and for the MLBPN-based model, a better estimation performance can be obtained from the use of Landsat 8 OLI. Scatterplots between the monitored values and the values estimated by the RF-based model using MODIS AOD data for air pollutants are presented in Figure 3. Similar scatterplots between the monitored values and the values estimated by the corresponding optimal model using MODIS SR, Landsat 8 OLI, and Sentinel-2 data are provided in Figure A1, Figure A2 and Figure A3 in Appendix A.

3.2. Spatiotemporal Analysis

Although the RF-based model with Sentinel-2 revealed a slightly better performance for the estimation of PM₁₀, NO₂, CO, and SO₂ than that with MODIS AOD, such differences are not significant. Combined with the consideration that the MODIS satellite has a high temporal resolution, covering the entire research area and revealing a better performance for the PM_2.5 and O₃ estimations, the MODIS AOD data have been confirmed as the basic satellite data for spatiotemporal analysis in the research area.

3.2.1. Temporal Distribution Characteristics

The annual average concentrations of air pollutants in the research area from 2019 to 2022 display various trends in Figure 4a: PM_2.5, PM₁₀, NO₂, and SO₂ display a consistent decrease, whereas O₃ and CO display a trend of decreasing followed by increasing. The annual average concentrations of PM_2.5 and PM₁₀ decreased by approximately 6 μg/m³, with a decrease of 19% and 10%, respectively. The decreases in NO₂ and SO₂ were 8% and 10%, respectively. CO concentration decreased by 0.02 mg/m³ in 2021, followed by an increase of 0.03 mg/m³ in 2022. However, the annual concentration of CO was relatively stable and remained below 0.6 mg/m³. O₃ concentration decreased by 3.4 μg/m³ in 2020, followed by an increase of 12.4 μg/m³ in 2021, with an overall increase of 11%, and finally remained around 83 μg/m³.

The monthly concentrations of air pollutants in the research area in 2021 are shown in Figure 4b. PM_2.5, PM₁₀, NO₂, CO, and SO₂ concentrations show a “U”-shaped monthly trend, with the lowest pollutant concentrations in summer and the highest in winter. The reasons are analyzed: the increase in pollutant concentrations is primarily attributed to the heating season from November to March, involving substantial burning of coal, biomass, and biofuels, and an inverse temperature and high atmospheric pressure, causing pollutants to accumulate in the lower atmosphere with the movement of the descending airflow. On the contrary, the decrease in pollutant concentrations is primarily attributed to lower atmospheric pressure and higher temperatures from May to September, creating atmospheric instability and favoring pollutant dispersion, and increased rainfall during this period contributes to air purification. Notably, PM₁₀ exhibited a high concentration anomaly in May, attributed to frequent sandstorms caused by cold air and strong winds, increasing dust and particulate matter levels, and worsening pollution.

O₃ concentration shows the opposite trend to the other pollutants, peaking in summer, followed by spring, autumn, and winter. This is because near-surface O₃ primarily arises from precursor elements such as volatile organic compounds (VOCs), carbon monoxide (CO), and nitrogen oxides (NOx) through complex photochemical reactions heavily influenced by sunlight and radiation intensity.

3.2.2. Spatial Distribution Characteristics

The annual-scale spatial distribution maps of air pollutants in the research area from 2019 to 2022, estimated by MODIS AOD data, showcase a consistent trend: a gradual increase from west to east, with the distribution of higher concentrations in the center of the built-up areas and lower in the surrounding rural areas (as depicted in Figure 5).

The high-pollution areas are mainly concentrated in the Guanzhong urban agglomeration, especially in the southern part of Xianyang, the northern part of Xi’an, the entire Weinan region, the central part of Baoji, and parts of southern Tongchuan. These areas are located in the lower elevation basin, bordered by the Loess Plateau in the north and the Qinling Mountains in the south, forming a trumpet-shaped topography with a higher west and lower east. East and northeast winds prevail in Guanzhong all year round, and the airflow enters from the eastern entrance along the Weihe Plain, making the concentration of air pollutants higher in the low-altitude areas within the basin. The low-pollution areas are mainly in the surrounding rural areas, with high-altitude mountain ranges, high vegetation cover, low human activities, and low industrial pollution.

Notably, PM_2.5, PM_10, NO₂, CO, and SO₂ show the highest concentration distribution in the urban areas of Xi’an and the adjoining regions bordering Shanxi Province and Weinan City, while O₃ shows the lowest concentration distribution. This is because the meteorological conditions for the formation of O₃ pollution are low humidity, low pressure, high temperature, and low wind, which is the opposite to the correlation between the other pollutants and meteorological factors.

Figure 6 illustrates the spatial distribution of the monthly average concentration of PM_2.5 in the research area in 2021. The spatial distribution of monthly average concentrations of other air pollutants can be found in Figure A4, Figure A5, Figure A6, Figure A7 and Figure A8 in Appendix B. It can be found that the monthly spatial distribution of PM_2.5, PM₁₀, NO₂, CO, and SO₂ is consistent with the annual spatial distribution. Meanwhile, pollutant concentrations are low from May to September and high in other months. The spatial distribution of O₃ shows distinct fluctuations: concentrations are high in the center area and low in the peripheral area from March to October, whereas this pattern is reversed in other months. This variation reflects the complexity of O₃ pollution with seasonal and meteorological conditions.

4. Discussion

In this research, RF- and MLBPN-based models have been respectively proposed for the estimation of the ground surface concentrations of PM_2.5, PM₁₀, O₃, NO₂, CO, and SO₂ from the various satellite imagery of MODIS AOD, MODIS SR, Landsat 8 OLI, and Sentinel-2. The conducted experiments demonstrate that (a) except for the Landsat 8 OLI data, the RF-based model revealed a better estimation performance than the MLBPN-based model using the satellite imagery from MODIS AOD/SR and Sentinel-2; (b) the results estimated from the MODIS AOD, Sentinel-2, and Landsat 8 OLI satellite data revealed significantly higher accuracies than MODIS SR data; (c) the RF-based model combined with the MODIS AOD data revealed the best performance for the estimation of PM_2.5 and O₃—the R² and RMSE reached 0.90 and 11.67 μg/m³ for PM_2.5, and 0.92 and 12.26 μg/m³ for O₃, respectively; while the RF-based model combined with the Sentinel-2 or MODIS AOD data achieved a comparable performance for PM₁₀, NO₂, CO, and SO₂—the R² and RMSE reached ca. 0.90 and 25 μg/m³ for PM₁₀, ca. 0.75 and 9.3 μg/m³ for NO₂, ca. 0.77 and 170 mg/m³ for CO, and ca. 0.6 and 3.5 μg/m³ for SO₂, respectively.

Considering the estimation performance and the temporal resolution of the satellite data, the RF-based model together with the MODIS AOD data can be treated as the optimal combination of “model and data”—for PM_2.5, PM₁₀, and O₃, the estimation results with the R² of ca. 0.9 are convincing; for NO₂ and CO, the results with the R² close to 0.8 are still reliable; while for the SO₂, the R² reaches 0.65, and hence the estimation results can only be used under some kinds of specific conditions. The estimation accuracy of trace gases, especially SO₂, is relatively lower than that of particulate matter. This may be due to the fact that the spectra range of the satellites used in the research did not include the ultraviolet (UV) band, which is crucial for the estimation of some trace gases, especially SO₂.

Based on the discussions mentioned above, the RF-based model with the MODIS AOD data was employed for the analysis of the spatiotemporal distributions in the Guanzhong Region, China. In terms of annual scale changes, the average concentrations of PM_2.5, PM₁₀, NO₂, and SO₂ show a decreasing trend, while CO remains relatively stable, and O₃ shows an increasing trend from 2019 to 2022. In terms of monthly scale changes, the concentrations of PM_2.5, PM₁₀, NO₂, CO, and SO₂ in 2021 showed a “U”-shaped trend, with the lowest pollutant concentrations in the summer and the highest in the winter, and the O₃ concentration showed the opposite trend. The spatial distribution of air pollution presents a gradual increase from west to east in the research area, with the distribution of higher concentrations in the center of the built-up areas and lower in the surrounding rural areas.

One particular aspect of the research is that we have also added meteorological data in addition to satellite remote-sensing data. The Pearson correlation coefficient heatmap based on MODIS AOD data, as shown in Figure 7, demonstrates a strong correlation between meteorological data and air pollutant concentrations.

Additionally,

Table 6. Estimation performances with meteorological parameters as independent variables based on the RF model and MODIS AOD data.

	Meteorological Parameters				No Meteorological Parameters
Accuracy evaluation	R2	RMSE	MAE	ME	R2	RMSE	MAE	ME
		(μg/m3)	(μg/m3)	(μg/m3)		(μg/m3)	(μg/m3)	(μg/m3)
PM2.5	0.86	16.83	10.83	−1.69	0.57	24.71	17.16	0.16
PM10	0.71	30.52	22.41	−3.08	0.31	49.48	34.55	−0.07
O3	0.85	16.16	11.51	−0.88	0.45	31.33	24.44	−0.38
NO2	0.69	9.33	7.23	0.37	0.24	15.71	11.54	−0.4
CO	0.72	200	150	−4	0.42	260	195	4
SO2	0.57	4.07	2.87	−0.09	0.19	5.02	3.58	0.09

presents the results of a comparative experiment conducted using meteorological parameters as independent variables based on the RF model and MODIS AOD data. It indicates a significant improvement in model accuracy when meteorological data are included, with the R² increasing by approximately 0.3 for PM_2.5 and CO, and approximately 0.4 for PM₁₀, O₃, NO₂, and SO₂.

However, the publicly available meteorological data from the National Oceanic and Atmospheric Administration consisted of only four stations within the research area. The limited meteorological data could have affected the generic nature of the sample data for model training and thereby could reduce the estimation accuracy and reliability in some kinds of specific situations. Hence, non-public meteorological data should be considered in further studies for a more detailed and accurate estimation of air pollutant concentrations.

5. Conclusions

This research compared the estimation results generated by different models from diverse types of satellite data, and the “optimal” model and satellite data were identified to provide one practical method for air pollutant concentration estimation. Furthermore, an analysis of spatiotemporal variations in the various primary air pollutants was conducted, which proved that the air quality tends to be better in mountainous areas than in urban centers, highlighting significant implications for public health. This discovery suggests that individuals may engage in more outdoor activities in mountainous regions to enhance opportunities for aerobic exercise, thereby promoting both physical and mental well-being. Additionally, it indicates that the air quality in the summer tends to be better than the other three seasons, suggesting the possibility of relaxing certain environmental restrictive policies, such as the bans on burning straw during this season.

To sum up, the conducted research can provide decision support for the development of relevant environmental protection measures to improve the air environmental quality and foster socioeconomic development in the Guanzhong Region.