High-Resolution Daily Spatiotemporal Distribution and Evaluation of Ground-Level Nitrogen Dioxide Concentration in the Beijing–Tianjin–Hebei Region Based on TROPOMI Data

Abstract

This study utilized TROPOMI remote sensing data, MODIS remote sensing data, ground observation data, and other ancillary data to construct a high-resolution spatiotemporal distribution and evaluation of ground-level NO2 concentrations in the Beijing–Tianjin–Hebei (BTH) region using the Geographic Temporal Neural Network Weighted Regression (GTNNWR) model. Through this model, we obtained the daily distribution of ground-level nitrogen dioxide (NO2) concentrations in the Beijing–Tianjin–Hebei region at a resolution of 500 m for the period of 2019–2022. The research results exhibited higher accuracy and more detailed features compared to other models, enabling a more accurate reflection of the spatial distribution and temporal variations of ground-level NO2 concentrations in the region, while retaining more details and trends and excluding the influence of noisy data. Furthermore, we conducted an evaluation analysis considering important events such as public health incidents and the Winter Olympics. The results demonstrated that the GTNNWR model outperformed the Random Forest (RF), Convolutional Neural Network (CNN), and Geographic Neural Network Weighted Regression (GNNWR) models in performance metrics such as R2, RMSE, MAE, and MAPE, showcasing greater reliability when considering spatiotemporal heterogeneity and spatiotemporal non-stationarity. This study provides crucial data support and reference for atmospheric environmental management and pollution prevention and control in the Beijing–Tianjin–Hebei region.

Full text

Citation: Liu, C.; Wu, S.; Dai, Z.;
Wang, Y.; Du, Z.; Liu, X.; Qiu, C.
High-Resolution Daily
Spatiotemporal Distribution and
Evaluation of Ground-Level Nitrogen
Dioxide Concentration in the
Beijing–Tianjin–Hebei Region Based
on TROPOMI Data. Remote Sens.
2023,15, 3878. https://doi.org/
10.3390/rs15153878
Academic Editor: Daniele Bortoli
Received: 6 July 2023
Revised: 26 July 2023
Accepted: 3 August 2023
Published: 4 August 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
remote sensing
Article
High-Resolution Daily Spatiotemporal Distribution and
Evaluation of Ground-Level Nitrogen Dioxide Concentration in
the Beijing–Tianjin–Hebei Region Based on TROPOMI Data
Chunhui Liu 1, Sensen Wu 2,3 , Zhen Dai 4, Yuanyuan Wang 2,5, Zhenhong Du 2,3, Xingyu Liu 1
and Chunxia Qiu 1, *
1College of Geomatics, Xi’an University of Science and Technology, Xi’an 710054, China;
21210061040@stu.xust.edu.cn (C.L.)
2School of Earth Sciences, Zhejiang University, Hangzhou 310027, China
3Zhejiang Provincial Key Laboratory of Geographic Information Science, Hangzhou 310028, China
4China Mobile (Zhejiang) Innovation Research Institute Co., Ltd., Hangzhou 310016, China
5Ocean Academy, Zhejiang University, Zhoushan 316021, China
*Correspondence: 000358@xust.edu.cn
Abstract:
This study utilized TROPOMI remote sensing data, MODIS remote sensing data, ground
observation data, and other ancillary data to construct a high-resolution spatiotemporal distribution
and evaluation of ground-level NO2 concentrations in the Beijing–Tianjin–Hebei (BTH) region using
the Geographic Temporal Neural Network Weighted Regression (GTNNWR) model. Through this
model, we obtained the daily distribution of ground-level nitrogen dioxide (NO2) concentrations in
the Beijing–Tianjin–Hebei region at a resolution of 500 m for the period of 2019–2022. The research
results exhibited higher accuracy and more detailed features compared to other models, enabling
a more accurate reflection of the spatial distribution and temporal variations of ground-level NO2
concentrations in the region, while retaining more details and trends and excluding the influence of
noisy data. Furthermore, we conducted an evaluation analysis considering important events such as
public health incidents and the Winter Olympics. The results demonstrated that the GTNNWR model
outperformed the Random Forest (RF), Convolutional Neural Network (CNN), and Geographic
Neural Network Weighted Regression (GNNWR) models in performance metrics such as R2, RMSE,
MAE, and MAPE, showcasing greater reliability when considering spatiotemporal heterogeneity
and spatiotemporal non-stationarity. This study provides crucial data support and reference for
atmospheric environmental management and pollution prevention and control in the Beijing–Tianjin–
Hebei region.
Keywords:
Beijing–Tianjin–Hebei region; TROPOMI; GTNNWR; ground-level NO2; high resolution
1. Introduction
Air pollution is a complex mixture of pollutants such as particulate matter, ozone,
sulfur dioxide, and nitrogen oxides, posing significant challenges to human health, envi-
ronmental sustainability, and economic development [
1
,
2
]. Nitrogen dioxide (NO2), as one
of the most important air pollution indicators, is a harmful gas mainly produced by the
combustion of fossil fuels, especially from vehicles, power plants, and industrial processes,
as well as from lightning and soil bacteria [3,4]. Moreover, nitrogen dioxide is a precursor
to secondary pollutants such as ozone and particulate matter, which can further exacerbate
the harmful effects of air pollution and are associated with a range of adverse health effects,
including respiratory and cardiovascular diseases [
5
,
6
]. The Beijing–Tianjin–Hebei region
is one of the most polluted areas in China, and monitoring the emission of nitrogen dioxide
has always been a focus of air pollution prevention and control work. Our study aims to
Remote Sens. 2023,15, 3878. https://doi.org/10.3390/rs15153878 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2023,15, 3878 2 of 24
provide a high-resolution spatiotemporal distribution of surface nitrogen dioxide concen-
tration in this region, which can be used to report and evaluate pollution control policies
and measures [7,8].
Due to the harmful effects of nitrogen dioxide, accurate and timely monitoring of its
concentration levels in the atmosphere is increasingly needed. Currently, there are various
methods to measure and estimate ground-level nitrogen dioxide concentration, includ-
ing ground-based monitoring stations, passive samplers, and remote sensing techniques
such as satellite images and unmanned aerial vehicles [
9
,
10
]. Ground-based monitoring
stations provide accurate and continuous measurements of nitrogen dioxide concentra-
tion at specific locations but have limited spatial coverage and may not capture changes
in concentration levels over larger areas. On the other hand, passive samplers can be
used to measure nitrogen dioxide concentration over longer periods and provide more
representative estimates of average concentration levels in specific areas [
11
]. Remote
sensing techniques provide wider spatial coverage and have been widely used to monitor
high-resolution spatiotemporal nitrogen dioxide concentrations. This method relies on
measuring the radiation reflected or emitted by the atmosphere and can be used to infer
nitrogen dioxide concentrations [
12
]. However, remote sensing methods also have limita-
tions, such as the need for calibration and validation with ground-based measurements,
and potential errors due to atmospheric changes and cloud cover [
13
]. Moreover, due to
the resolution limitations of data sources, obtaining information on the distribution of NO2
ground concentration at a higher resolution is also a noteworthy issue. Currently, the most
authoritative results in this field internationally come from the atmospheric environment
remote sensing team led by Dr. Weijing from the University of Maryland and Professor Li
Zhanqing, who developed China’s high-resolution (1 km), high-quality near-surface air
pollution distribution [14,15].
In recent years, there has been increasing interest in developing new methods that
use deep learning and artificial intelligence technologies to estimate nitrogen dioxide
concentrations [
16
,
17
]. These methods can integrate multiple data sources, including
satellite images, ground measurements, and meteorological data, to improve the accu-
racy and reliability of nitrogen dioxide concentration estimates. Among these methods,
convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are some
of the most commonly used deep learning methods. These methods can automatically
extract features from the data and build complex nonlinear models to achieve accurate
predictions of concentration distributions [
18
,
19
]. For example, recent studies have used
CNN models to process remote sensing satellite data to estimate ground-level nitrogen
dioxide concentration distributions [
20
]. The model can automatically learn spatiotemporal
features in remote sensing data and output high-precision concentration distribution im-
ages. Compared with traditional supervised learning methods, this unsupervised learning
method can improve the model’s generalization ability and robustness. In addition to deep
learning methods, there are also artificial intelligence technologies based on physical and
statistical models, such as differential optical absorption spectroscopy (DOAS) and spatial
interpolation techniques. These methods can consider the influence of meteorological and
environmental factors on nitrogen dioxide concentrations to improve estimation accuracy
and reliability [21].
Du et al. proposed a geographically weighted neural network weighted regression
(GNNWR) model, which combines the advantages of ordinary linear regression (OLR)
models and neural network models [
22
,
23
]. By using the learning ability of neural networks
and the local spatial interpretability of spatial weights, the model can handle the spatial het-
erogeneity and complex nonlinearity of regression relationships, has better fitting accuracy
and prediction performance than GWR and neural network models, and has the potential
to solve complex spatial relationships in many fields. However, the GNNWR model faces
challenges in expressing spatiotemporal closeness and constructing optimal weights, which
may lead to insufficient estimation of spatiotemporal non-stationarity. To deal with the
complex nonlinear interactions between time and space, Wu et al. proposed a spatiotempo-

Remote Sens. 2023,15, 3878 3 of 24
ral proximal neural network (STPNN) to accurately generate spatiotemporal distances and
extended it to GNNWR to develop a geospatiotemporal neural network weighted regres-
sion (GTNNWR) model to effectively model spatiotemporal non-stationarity relationships,
and demonstrated that GTNNWR has the potential to handle complex spatiotemporal
non-stationarity in various geographic processes and environmental phenomena in ocean
analysis [24].
Therefore, this study uses the GTNNWR model to integrate the TROPOMI sensor
daily data, MODIS NDVI data, ERA5 meteorological data, and other auxiliary data to
simulate the surface NO2 concentrations from 2019 to 2022 and validate them with the
ground station data to obtain the high-resolution (500 m) daily surface NO2 in the Beijing–
Tianjin–Hebei region from 2019 to 2022 concentration distribution from 2019 to 2022. By
comparing the performance indexes with regression methods such as CNN and GNNWR,
the advantages of the GTNNWR model in the spatiotemporal non-smoothness relationship
of atmospheric pollutants are investigated, and the consistency and differences between the
results and other models in predicting ground-level NO2 concentration data are verified.
The study provides a valuable reference for the spatiotemporal distribution of NO2 in the
Beijing–Tianjin–Hebei region from 2019–2022. In addition, the research results emphasize
the importance of remote sensing technology in NO2 pollution monitoring and evaluation,
and demonstrate the potential of the GTNNWR model in improving the accuracy and
reliability of air pollution evaluation.
2. Materials and Methods
2.1. Study Area
The Beijing–Tianjin–Hebei region (BTH) is one of the most industrialized and densely
populated areas in China [
25
]. The region covers an area of approximately 218,000 square
kilometers and is home to over 100 million people [
26
]. Due to its rapid economic de-
velopment and urbanization, air pollution has become a major public health issue in the
Beijing–Tianjin–Hebei region, where residents are frequently exposed to high levels of
pollutants such as NO2, PM
2.5
, and ozone [
27
]. The sources of air pollution in the region are
complex and include transportation, industrial activities, residential heating, and natural
dust [
28
,
29
]. In recent years, a series of measures have been implemented in the region to
reduce emissions and improve air quality, including the implementation of strict emission
standards, the promotion of clean energy, and the closure of high-polluting enterprises [
30
].
However, despite these efforts, air pollution remains an urgent problem in the region. It is
necessary to continue monitoring air quality comprehensively to evaluate the effectiveness
of these measures and provide a basis for future policy decisions. This research will provide
valuable references for the temporal and spatial patterns of NO2 in the region. Figure 1a
shows the geographical location of the Beijing–Tianjin–Hebei region in China, a region that
covers the political center, important port cities, and three provincial divisions with the
most developed steel industries in China. It is due to the unique industrial structure and
importance of this study area that high-resolution monitoring of its air quality is necessary,
particularly for nitrogen dioxide which is strongly influenced by secondary and tertiary
industries and transport. Figure 1b shows the population distribution in the study area,
which also has a strong correlation to the production of nitrogen dioxide, and it can be seen
that the population is mostly concentrated in cities and urban agglomerations, which can
provide a preliminary reference for the study results. Figure 1c shows the location and
elevation data of air quality monitoring stations, which are regionally heterogeneous and
concentrated in large cities. Elevation has some influence on the propagation of nitrogen
dioxide due to air movement.

Remote Sens. 2023,15, 3878 4 of 24
Remote Sens. 2023, 15, 3878 4 of 26
Figure 1. Location and distribution of ground-based air quality testing stations in BTH. (a) shows
the geographical location of the Beijing-Tianjin-Hebei region of China. (b) shows the population
distribution in the study area. (c) shows the location of air quality monitoring stations and eleva-
tion data.)
2.2. Data and Pre-Processing
2.2.1. Ground-Level NO2 Data
The daily surface NO2 concentration data selected for the Beijing–Tianjin–Hebei re-
gion in this article were obtained from the China National Environmental Monitoring
Center and cover the period from 1 January 2019 to 31 December 2022. A total of 80 air-
quality ground monitoring sites were included (Figure 1). The ground NO2 concentra-
tions were measured via a molybdenum converter method and calibrated according to
the Chinese Ambient Air Quality Standard [31]. Due to reasons such as site maintenance,
a small number of missing values during the study period were filled by averaging the
hourly NO2 measurement results of the site to the daily scale to complete the temporal
sequence.
2.2.2. TROPOMI NO2 Data
TROPOMI (TROPOspheric Monitoring Instrument) is a multispectral instrument on
board the Copernicus Sentinel-5 Precursor (S5P) satellite launched in 2017, which can
monitor the concentration of gases and particles in the Earthʹs atmosphere with high ac-
curacy and resolution [32]. Compared to the Ozone Monitoring Instrument (OMI) sensor
data, which has a similar mission, TROPOMI has higher spatial resolution and accuracy,
providing more detailed and accurate air quality data. Its high spatial resolution is 7 × 3.5
km (improved to 5.5 × 3.5 km after 6 August 2019) and has been proven to have a broader
range of applications in environmental monitoring and climate research [33]. The NO2
retrieval algorithm and product information of TROPOMI have been detailed by van Ge-
ffen et al. and Wu et al. [34,35]. The tropospheric NO2 vertical column density (VCD) data
Figure 1.
Location and distribution of ground-based air quality testing stations in BTH. (
a
) shows the
geographical location of the Beijing-Tianjin-Hebei region of China. (
b
) shows the population distribu-
tion in the study area. (c) shows the location of air quality monitoring stations and elevation data.
2.2. Data and Pre-Processing
2.2.1. Ground-Level NO2 Data
The daily surface NO2 concentration data selected for the Beijing–Tianjin–Hebei region
in this article were obtained from the China National Environmental Monitoring Center
and cover the period from 1 January 2019 to 31 December 2022. A total of 80 air-quality
ground monitoring sites were included (Figure 1). The ground NO2 concentrations were
measured via a molybdenum converter method and calibrated according to the Chinese
Ambient Air Quality Standard [
31
]. Due to reasons such as site maintenance, a small
number of missing values during the study period were filled by averaging the hourly
NO2 measurement results of the site to the daily scale to complete the temporal sequence.
2.2.2. TROPOMI NO2 Data
TROPOMI (TROPOspheric Monitoring Instrument) is a multispectral instrument on
board the Copernicus Sentinel-5 Precursor (S5P) satellite launched in 2017, which can
monitor the concentration of gases and particles in the Earth’s atmosphere with high
accuracy and resolution [
32
]. Compared to the Ozone Monitoring Instrument (OMI)
sensor data, which has a similar mission, TROPOMI has higher spatial resolution and
accuracy, providing more detailed and accurate air quality data. Its high spatial resolution is
7
×
3.5 km (improved to 5.5
×
3.5 km after 6 August 2019) and has been proven to have a
broader range of applications in environmental monitoring and climate research [
33
]. The
NO2 retrieval algorithm and product information of TROPOMI have been detailed by van
Geffen et al. and Wu et al. [
34
,
35
]. The tropospheric NO2 vertical column density (VCD)
data in the daily TROPOMI Level-2 data covering the entire Beijing–Tianjin–Hebei region

Remote Sens. 2023,15, 3878 5 of 24
from 1 January 2019 to 31 December 2022, were obtained from GESDISC [
36
]. Tropospheric
NO2 data with QA_value > 0.75 were selected for modeling, and cloud, inversion errors,
and problematic retrievals were excluded [
37
]. To fill the temporal and spatial gaps in
satellite tropospheric NO2 data, the inverse distance weighting (IDW) method and time-
linear interpolation were used in this study.
2.2.3. Spatiotemporal Ancillary Data
To improve the modeling of NO2 gas pollutants, this study selected several spa-
tiotemporal auxiliary variables in the Beijing–Tianjin–Hebei region that are related to NO2
emission types and anthropogenic emissions, which can affect the transmission mode of
ground-level NO2 [
38
]. The variables include a 16-day normalized difference vegetation
index (NDVI) data (500 m) produced by MODIS data to improve spatial resolution. To
meet the research requirements, it was time-linearly interpolated to expand it to daily
data [
39
]. Digital elevation model (DEM) data (90 m) obtained from the Shuttle Radar
Topography Mission (SRTM) were also used [
9
]. Meteorological data related to potential
emission sources and diffusion were obtained from eight meteorological estimates of ERA-5
atmospheric reanalysis products, including wind direction (WD) and wind speed (WS)
at 10 m height, temperature (TEMP) at a height of 2 m, total precipitation (TP), surface
pressure (SP), boundary layer height (BLH), evaporation (ET), and relative humidity (RH),
with a resolution of 10 km [
40
]. Road network (RN) data in the study area were collected
from OSM and their density was calculated using kernel density estimation [
41
]. Yearly
population density data (PD, 1 km) from NASA were also used [
42
]. In addition to this, the
study also obtained data from the China Statistical Yearbook on the number of industrial
enterprise units (IEU), GDP, and the number of resident residents (RR) at the county level
in Beijing, Tianjin, and Hebei for 2019–2022 to refine the modeling [
43
]. Table 1shows the
relevant parameters of the spatiotemporal auxiliary variables used in this study.
Table 1. Basic information sheet for Remote sensing data and Spatiotemporal ancillary data.
Materials Temporal
Resolution
Spatial
Resolution Time Frame Data Type Data Volume Elemental Values
TROPOMI Day 7×3.5 km
/5.5 ×3.5 km
1 January 2019 to
31 December 2022
TIFF 2170 MB Key inversion elements
MODIS_NDVI
16 days 0.5 ×0.5 km 1 January 2019 to
31 December 2022
TIFF 1986 MB Improved spatial resolution
SRTM_DEM / 0.09 ×0.09 km / TIFF 276 MB
Associated with potential
emission sources
and dispersion
ERA-5_WD
Day 10 ×10 km 1 January 2019 to
31 December 2022
TIFF
28.9 MB
ERA-5_WS 28.6 MB
ERA-5_TEMP 28.8 MB
ERA-5_TP 28.6 MB
ERA-5_SP 28.6 MB
ERA-5_BLH 28.9 MB
ERA-5_ET 28.4 MB
ERA-5_RH 28.7 MB
OSM_RN / 0.5 ×0.5 km / TIFF 6.42 MB
NASA_PD Year 1 ×1 km 2019 to 2022 TIFF 350.60 MB
GDP Year County 2019 to 2022 XLSX /
IEU Year County 2019 to 2022 XLSX /
RR Year County 2019 to 2022 XLSX /
In order to understand the basic situation of the above data, this study first selected
several influence factors with the highest correlation as well as the measured data from the
ground monitoring stations, and analyzed the data for characteristic distribution (Figure 2)
to obtain the basic statistical characteristics of these data. These statistical indicators provide
information on the concentration trend and discrete degree of the data, which helps to
understand the overall situation, shape, and change trends of the data. At the same time, the

Remote Sens. 2023,15, 3878 6 of 24
feature distribution analysis carried out on the data facilitates the detection of data outliers
and assists in data cleaning in data preprocessing for model performance and accuracy.
Remote Sens. 2023, 15, 3878 6 of 26
the ground monitoring stations, and analyzed the data for characteristic distribution (Fig-
ure 2) to obtain the basic statistical characteristics of these data. These statistical indicators
provide information on the concentration trend and discrete degree of the data, which
helps to understand the overall situation, shape, and change trends of the data. At the
same time, the feature distribution analysis carried out on the data facilitates the detection
of data outliers and assists in data cleaning in data preprocessing for model performance
and accuracy.
Figure 2. Characteristic distribution of the main influencing factors of the inversion of NO2 at
ground-level.
2.3. GTNNWR Model
2.3.1. Model Development
Considering the close relationship between nitrogen dioxide concentrations and spa-
tiotemporal factors, this study developed a geo-temporal neural network weighted regres-
sion (GTNNWR) model for air quality assessment based on a spatiotemporal weight ma-
trix and a spatiotemporal neighborhood depth neural network [44]. The model aims to
analyze the regression relationships between 16 independent variables related to nitrogen
dioxide production and dispersion and the dependent variable (ground-level nitrogen di-
oxide concentration). To address the complex spatiotemporal heterogeneity and autocor-
relation in the atmosphere, the STPNN (Spatial Temporal Distance Fusion Neural Net-
work) was introduced based on the GNNWR model. It combines temporal and spatial
distances while using STWNN (non-stationary weight calculation neural network) to cal-
culate the weights of the factors.
The GTNNWR model consists of an input layer that receives normalized spatial dis-
tance and feature data as input, followed by a series of fully connected layers and an acti-
vation function layer. In this model, the spatial distances are first processed through the
dense and PReLU activation function layers and then fused with the feature data to form
the final output. The process of constructing the GTNNWR model from a mathematical
perspective is described below [45].
Figure 2.
Characteristic distribution of the main influencing factors of the inversion of NO2 at
ground-level.
2.3. GTNNWR Model
2.3.1. Model Development
Considering the close relationship between nitrogen dioxide concentrations and spa-
tiotemporal factors, this study developed a geo-temporal neural network weighted regres-
sion (GTNNWR) model for air quality assessment based on a spatiotemporal weight matrix
and a spatiotemporal neighborhood depth neural network [
44
]. The model aims to analyze
the regression relationships between 16 independent variables related to nitrogen dioxide
production and dispersion and the dependent variable (ground-level nitrogen dioxide
concentration). To address the complex spatiotemporal heterogeneity and autocorrelation
in the atmosphere, the STPNN (Spatial Temporal Distance Fusion Neural Network) was
introduced based on the GNNWR model. It combines temporal and spatial distances while
using STWNN (non-stationary weight calculation neural network) to calculate the weights
of the factors.
The GTNNWR model consists of an input layer that receives normalized spatial
distance and feature data as input, followed by a series of fully connected layers and an
activation function layer. In this model, the spatial distances are first processed through the
dense and PReLU activation function layers and then fused with the feature data to form
the final output. The process of constructing the GTNNWR model from a mathematical
perspective is described below [45].
To describe the spatiotemporal non-stationarity of geographical relationships and
the differences in weights across locations, this study first reviews the ordinary linear
regression model, which is mathematically expressed as
yi=β0+∑p
k=1(βkxik)+εi(1)

Remote Sens. 2023,15, 3878 7 of 24
where y
i
is the dependent variable, x
ik
is the kth variable of the ith sample, p denotes the
number of independent variables,
β0
and
βk
are the regression coefficients, and
εi
is the
error terms [46].
However, the spatiotemporal non-stationarity of geographical relationships allows
for variability in the regression coefficients across spatiotemporal locations [
24
]. There-
fore, we introduce a spatiotemporal, geographically weighted regression model, whose
mathematical expression is
y(si, ti)=β0(si, ti)+∑p
k=1βk(si, ti)xik +εi(2)
where y(s
i
, t
i
) denotes the dependent variable at the spatiotemporal location (s
i
, t
i
). To
further describe the spatiotemporal non-smoothness and weight differences, the study
introduced a spatiotemporal weight matrix and a spatiotemporal neighborhood deep
neural network to establish the GTNNWR model, which is mathematically expressed as
y(si, ti)=ω0(si, ti)∗β0+∑p
k=1ωk(si, ti)βkxik +εi(3)
In this model,
ω0
(s
i
, t
i
) and
ωk
(s
i
, t
i
) are the non-stationary weights at the spatiotempo-
ral point (s
i
, t
i
), and the corresponding regression coefficients are obtained by multiplying
them by ordinary linear regression estimates.
To determine the spatiotemporal weight matrix and the non-stationary weights, a spa-
tiotemporal neighborhood deep neural network is used to obtain a complex spatiotemporal
neighborhood representation vector ST at the spatiotemporal location (s
i
, t
i
). By combining
ST with the weight kernel function structure, we can obtain the spatiotemporal weight
matrix W(si, ti) of the form:
W(si, ti) =




ω0(si, ti)0 0 0
0ω1(si, ti)0 0
0 0 . . . 0
0 0 0 ωp(si, ti)




(4)
Thus, the spatiotemporal non-smoothness and weight differences in the GTNNWR
model are determined via the spatiotemporal weight matrix W(s
i
, t
i
). This relationship
effectively captures the spatiotemporal variability of geographical relationships and differ-
ences in weight distributions, further improving the accuracy and explanatory power of
the regression model.
In summary, the GTNNWR model combines linear regression and deep learning
methods by introducing a spatiotemporal weight matrix and a spatiotemporal neighbor-
hood deep neural network to better describe and predict spatiotemporal non-stationarity
in geographic relationships. Specifically, the regression coefficients are modeled as the
non-stationary weights of spatiotemporal points multiplied by ordinary linear regression
estimates to obtain estimates of regression coefficients with spatiotemporal proximity. The
detailed architecture of the GTNNWR model is given in Figure 3. The model has been
applied in areas such as marine water quality monitoring, and the results of the empiri-
cal analysis show that the GTNNWR model is highly accurate and effective in modeling
and predicting spatiotemporal non-stationarity, but its application in the atmospheric
domain has not yet been carried out. This study will validate and explore the model’s
capability and advantages in dealing with atmospheric spatiotemporal heterogeneity and
autocorrelation problems.

Remote Sens. 2023,15, 3878 8 of 24
Remote Sens. 2023, 15, 3878 8 of 26
predicting spatiotemporal non-stationarity, but its application in the atmospheric domain
has not yet been carried out. This study will validate and explore the model’s capability
and advantages in dealing with atmospheric spatiotemporal heterogeneity and autocor-
relation problems.
Figure 3. GTNNWR model structure ( S
i
d denotes the spatial distance between individual image
elements, while T
i
d denotes the difference between each image element).
2.3.2. Model Evaluation
To validate the accuracy of the GTNNWR model for regression and prediction of
ground-level NO2 concentrations, three methods—geographic neural network weighted
regression (GNNWR), convolutional neural network (CNN), and random forest (RF)—
were selected for validation against the results of the GTNNWR model in this study [47].
Over 100,000 rows of data from 80 air monitoring sites in the study area for the period
of 2019 to 2022 were first integrated into one data set. The integrated data set was divided
into a training set, a validation set, and a test set, and the cross-validation method was
used to divide the data set into multiple subsets, rotating one of the subsets as the valida-
tion set and the rest as the training set. The training set was trained using each of the four
methods: GTNNWR model, GNNWR model, CNN, and RF. For each model, the trained
model was used to predict the validation set, and four sets of predictions were obtained.
These predicted values were compared with the real data corresponding to the ground
sites in the validation set. By comparing the predicted values with the true values, four
sets of performance metrics were calculated, including R2 (coefficient of determination),
MAE (mean absolute error), RMSE (root mean square error), and MAPE (mean absolute
percentage error). These metrics are used to assess the regression and prediction accuracy
of the models [48]. The following equations were used to calculate the four performance
indicators,
)/(12 SSTSSRR −= (5)
where SSR is the sum of squares of residuals, which is the sum of the squares of the errors
between the predicted and true values, and SST is the total sum of squares, which is the
sum of the squares of the errors between the true and mean values.
Figure 3.
GTNNWR model structure (
dS
i
denotes the spatial distance between individual image
elements, while dT
idenotes the difference between each image element).
2.3.2. Model Evaluation
To validate the accuracy of the GTNNWR model for regression and prediction of
ground-level NO2 concentrations, three methods—geographic neural network weighted
regression (GNNWR), convolutional neural network (CNN), and random forest (RF)—were
selected for validation against the results of the GTNNWR model in this study [47].
Over 100,000 rows of data from 80 air monitoring sites in the study area for the
period of 2019 to 2022 were first integrated into one data set. The integrated data set
was divided into a training set, a validation set, and a test set, and the cross-validation
method was used to divide the data set into multiple subsets, rotating one of the subsets
as the validation set and the rest as the training set. The training set was trained using
each of the four methods: GTNNWR model, GNNWR model, CNN, and RF. For each
model, the trained model was used to predict the validation set, and four sets of predictions
were obtained. These predicted values were compared with the real data corresponding
to the ground sites in the validation set. By comparing the predicted values with the
true values, four sets of performance metrics were calculated, including R2 (coefficient of
determination), MAE (mean absolute error), RMSE (root mean square error), and MAPE
(mean absolute percentage error). These metrics are used to assess the regression and
prediction accuracy of the models [
48
]. The following equations were used to calculate the
four performance indicators,
R2=1−(SSR/SST)(5)
where SSR is the sum of squares of residuals, which is the sum of the squares of the errors
between the predicted and true values, and SST is the total sum of squares, which is the
sum of the squares of the errors between the true and mean values.
MAE = (1/n)∗∑|y_pred −y_true|(6)
RMSE =q(1/n)∗∑(y_pred −y_true)2(7)
MAPE = (1/n)∗∑|(y_pred −y_true)/y_pred|∗100% (8)

Remote Sens. 2023,15, 3878 9 of 24
where nis the number of samples,
∑
denotes summing over all samples, y_pred denotes
the predicted value, and y_true denotes the true value.
The flowchart (Figure 4) shows the training process of the GTNNWR model for
reducing and predicting the ground NO2 concentration when processing remote sensing
data, meteorological data and auxiliary data applications. The GTNNWR model obtains
the predicted values of the ground NO2 concentration for the 1 km data set and the 500 m
data set, and validates and evaluates the results with the actual measured data from the
ground station. In the above equations and flowcharts, the predicted values are the results
given by the model and the true values correspond to the actual observations from the
ground station [
49
]. By calculating these performance metrics, the regression and prediction
accuracy of the model can be evaluated. It is important to note that the exact calculation of
these metrics may vary depending on the subject area or research question. Figure 5shows
the results of the comparison of the performance metrics of the four models for predicting
ground-level NO2 concentrations.
Remote Sens. 2023, 15, 3878 9 of 26

−= trueypredynMAE __*)/1( (6)

−=
2
)__(*)/1( trueypredynRMSE
(7)
%100*_/)__(*)/1( −= predytrueypredynMAPE
(8)
where n is the number of samples, ∑ denotes summing over all samples, y_pred denotes
the predicted value, and y_true denotes the true value.
The flowchart (Figure 4) shows the training process of the GTNNWR model for re-
ducing and predicting the ground NO2 concentration when processing remote sensing
data, meteorological data and auxiliary data applications. The GTNNWR model obtains
the predicted values of the ground NO2 concentration for the 1 km data set and the 500 m
data set, and validates and evaluates the results with the actual measured data from the
ground station. In the above equations and flowcharts, the predicted values are the results
given by the model and the true values correspond to the actual observations from the
ground station [49]. By calculating these performance metrics, the regression and predic-
tion accuracy of the model can be evaluated. It is important to note that the exact calcula-
tion of these metrics may vary depending on the subject area or research question. Figure
5 shows the results of the comparison of the performance metrics of the four models for
predicting ground-level NO2 concentrations.
Figure 4. The process of downscaling and prediction of impact factors is based on the GTNNWR
model.
Figure 4.
The process of downscaling and prediction of impact factors is based on the GTNNWR model.
The results show that the R2 (coefficient of determination) of the GTNNWR model
reached 0.89 according to the data displayed in Figure 5, indicating that the model fits
the observed data well. The GTNNWR model was able to fit the correlation between the
data well compared to other methods, indicating that the model has a clear advantage in
dealing with the spatiotemporal heterogeneity (variability across locations and time) and
autocorrelation between the ground-level NO2 concentration and the influencing factors.
This implies that the GTNNWR model is able to more accurately capture the correlations
between ground-level NO2 concentrations and various impact factors, including taking into
account the variability in location and time as well as autocorrelation between data points.
Taken together, the GTNNWR model excels in dealing with the complex relationship
between ground-level NO2 concentrations and the influencing factors, and is able to better
explain and predict this relationship. This gives the GTNNWR model a clear advantage in
research and application in this field.

Remote Sens. 2023,15, 3878 10 of 24
Remote Sens. 2023, 15, 3878 10 of 26
Figure 5. Performance indicators for the four groups of models.
The results show that the R2 (coefficient of determination) of the GTNNWR model
reached 0.89 according to the data displayed in Figure 5, indicating that the model fits the
observed data well. The GTNNWR model was able to fit the correlation between the data
well compared to other methods, indicating that the model has a clear advantage in deal-
ing with the spatiotemporal heterogeneity (variability across locations and time) and au-
tocorrelation between the ground-level NO2 concentration and the influencing factors.
This implies that the GTNNWR model is able to more accurately capture the correlations
between ground-level NO2 concentrations and various impact factors, including taking
into account the variability in location and time as well as autocorrelation between data
points.
Taken together, the GTNNWR model excels in dealing with the complex relationship
between ground-level NO2 concentrations and the influencing factors, and is able to beer
explain and predict this relationship. This gives the GTNNWR model a clear advantage
in research and application in this field.
In order to further investigate the performance metrics of several models, 100 sets of
data were selected here to calculate the Pearson correlation coefficients and RMSEs be-
tween the results of the four models and the measured data, and the box plots with error
lines (Figure 6) were generated to further express the differences in the result performance
of the four models.
Figure 5. Performance indicators for the four groups of models.
In order to further investigate the performance metrics of several models, 100 sets of
data were selected here to calculate the Pearson correlation coefficients and RMSEs between
the results of the four models and the measured data, and the box plots with error lines
(Figure 6) were generated to further express the differences in the result performance of the
four models.
Remote Sens. 2023, 15, 3878 11 of 26
Figure 6. Box diagram with error lines.
It can be seen that the Pearson correlation coefficient of the GTNNWR model is 0.89,
which indicates that there is a strong positive correlation between the predictions of the
model and the actual observations. Meanwhile, the RMSE of the GTNNWR model is 10.50,
indicating that the average difference between its predicted and actual observed values is
small. In contrast, the other three models perform slightly less well. The Pearson correla-
tion coefficient of the GNNWR model is 0.89, which is the same as that of the GTNNWR
model, but its RMSE is slightly higher at 10.65. The Pearson correlation coefficients of the
CNN and the RF are 0.79 and 0.77, which are lower compared to those of the GTNNWR
model and the GNNWR model, and their RMSE is 14.74 and 14.97, which is a poor per-
formance compared to the GTNNWR results.
Taken together, the GTNNWR model performs beer in terms of error control, with
a strong positive correlation between its predicted results and the actual observations, and
a small average difference. This makes the GTNNWR model potentially more reliable and
accurate in the application of this topic. Based on this, the GTNNWR model will be used
in this study to invert and analyze the daily high-resolution spatiotemporal distributions
of ground-level NO2 concentrations in the Beijing–Tianjin–Hebei region from 2019 to
2022.
3. Results
3.1. Daily High-Resolution Surface NO2 Concentrations
The study used the GTNNWR model to model and analyze 16 factors affecting
ground-level NO2 concentrations in the Beijing–Tianjin–Hebei region during the period
of 2019–2022. By constructing a geo-temporal weighted neural network to model these
factors in association with ground-level NO2 concentration data for prediction and anal-
ysis of NO2 concentrations, daily 500 m resolution ground-level NO2 concentration data
for a total of 1460 days from 1 January 2019 to 31 December 2022 were successfully ob-
tained, which will enable relevant studies to capture NO2 concentration variations in ge-
ographic space more accurately and provide valuable information for further environ-
mental monitoring and air pollution management.
Figure 7 shows the time series variation curves of TROPOMI NO2 as well as the pre-
dicted ground-level NO2 concentrations for 1460 days during the study period. The com-
parison provides insight into the relationship between ground-level NO2 concentrations
Figure 6. Box diagram with error lines.
It can be seen that the Pearson correlation coefficient of the GTNNWR model is 0.89,
which indicates that there is a strong positive correlation between the predictions of the

Remote Sens. 2023,15, 3878 11 of 24
model and the actual observations. Meanwhile, the RMSE of the GTNNWR model is 10.50,
indicating that the average difference between its predicted and actual observed values is
small. In contrast, the other three models perform slightly less well. The Pearson correlation
coefficient of the GNNWR model is 0.89, which is the same as that of the GTNNWR model,
but its RMSE is slightly higher at 10.65. The Pearson correlation coefficients of the CNN
and the RF are 0.79 and 0.77, which are lower compared to those of the GTNNWR model
and the GNNWR model, and their RMSE is 14.74 and 14.97, which is a poor performance
compared to the GTNNWR results.
Taken together, the GTNNWR model performs better in terms of error control, with a
strong positive correlation between its predicted results and the actual observations, and a
small average difference. This makes the GTNNWR model potentially more reliable and
accurate in the application of this topic. Based on this, the GTNNWR model will be used in
this study to invert and analyze the daily high-resolution spatiotemporal distributions of
ground-level NO2 concentrations in the Beijing–Tianjin–Hebei region from 2019 to 2022.
3. Results
3.1. Daily High-Resolution Surface NO2 Concentrations
The study used the GTNNWR model to model and analyze 16 factors affecting ground-
level NO2 concentrations in the Beijing–Tianjin–Hebei region during the period of 2019–
2022. By constructing a geo-temporal weighted neural network to model these factors in
association with ground-level NO2 concentration data for prediction and analysis of NO2
concentrations, daily 500 m resolution ground-level NO2 concentration data for a total of
1460 days from 1 January 2019 to 31 December 2022 were successfully obtained, which will
enable relevant studies to capture NO2 concentration variations in geographic space more
accurately and provide valuable information for further environmental monitoring and air
pollution management.
Figure 7shows the time series variation curves of TROPOMI NO2 as well as the
predicted ground-level NO2 concentrations for 1460 days during the study period. The
comparison provides insight into the relationship between ground-level NO2 concentra-
tions and meteorological remote-sensing satellite observations. In the figure, the hori-
zontal axis indicates the time and the vertical axis indicates the TROPOMI or ground
NO2 concentration.
Remote Sens. 2023, 15, 3878 12 of 26
and meteorological remote-sensing satellite observations. In the figure, the horizontal axis
indicates the time and the vertical axis indicates the TROPOMI or ground NO2 concen-
tration.
Figure 7. Ground-level NO2 predictions and TROPOMI NO2 concentrations daily variation.
By observing Figure 7, it is possible to study the dynamic relationship between TRO-
POMI and ground-level NO2 concentrations. In general, the two sets of data present the
same paern of variation, showing that the highest NO2 concentrations are found in the
study area each year in winter and the lowest in summer. At the same time, the ground-
level NO2 concentrations show a decreasing trend from year to year for the combined four
years as a whole. The results of this figure help us to understand the correlation and spa-
tiotemporal relationship between TROPOMI and ground NO2 concentrations. It should
be noted that Figure 7 presents only the temporal variation curves and cannot directly
derive the causal relationship. Further statistical analysis and modeling work can be done
with these data to explore the causal relationships and interaction mechanisms between
these factors and ground NO2 concentrations.
To further demonstrate the changes in ground-level NO2 concentrations from 2019–
2022, We calculated a map of annual average ground-level NO2 concentrations based on
daily data (Figure 8). This comparison can visually demonstrate the average NO2 levels
at different locations and time periods in the Beijing–Tianjin–Hebei region.
Figure 7. Ground-level NO2 predictions and TROPOMI NO2 concentrations daily variation.

Remote Sens. 2023,15, 3878 12 of 24
By observing Figure 7, it is possible to study the dynamic relationship between
TROPOMI and ground-level NO2 concentrations. In general, the two sets of data present
the same pattern of variation, showing that the highest NO2 concentrations are found in
the study area each year in winter and the lowest in summer. At the same time, the ground-
level NO2 concentrations show a decreasing trend from year to year for the combined
four years as a whole. The results of this figure help us to understand the correlation and
spatiotemporal relationship between TROPOMI and ground NO2 concentrations. It should
be noted that Figure 7presents only the temporal variation curves and cannot directly
derive the causal relationship. Further statistical analysis and modeling work can be done
with these data to explore the causal relationships and interaction mechanisms between
these factors and ground NO2 concentrations.
To further demonstrate the changes in ground-level NO2 concentrations from 2019–
2022, We calculated a map of annual average ground-level NO2 concentrations based on
daily data (Figure 8). This comparison can visually demonstrate the average NO2 levels at
different locations and time periods in the Beijing–Tianjin–Hebei region.
Remote Sens. 2023, 15, 3878 13 of 26
Figure 8. Average ground-level NO2 concentration map for 2019–2022.
It can be clearly observed from Figure 8 that the ground-level NO2 concentration in
the Beijing–Tianjin–Hebei region shows a clear decreasing trend during the period of
2019–2022. Especially in the northern non-important industrial areas, the area of the low-
concentration area shows a trend of expansion. At the same time, the annual average NO2
concentrations in important industrial areas or transportation hubs, such as Tangshan,
Tianjin, and Shijiazhuang, have also decreased to some extent. This trend may be related
to factors such as the development of public health events and environmental protection
policies. However, there is a rebound in overall NO2 concentrations in 2022 compared to
the previous year, which to some extent reflects the gradual recovery of industries in the
studied area, such as industry and transportation, which have adapted to the develop-
ment of public health events.
Figure 9 represents the standard deviation of the seasonal changes in near-surface
NO2 concentrations and TROPOMI NO2 concentrations over different quarters. The hor-
izontal coordinates indicate the quarters, from 1 to 4, representing the first through fourth
quarters of the year, respectively. The vertical coordinate represents the standard devia-
tion value, which measures the magnitude of change in the data over each quarter. By
comparing the information on the graphs, it is possible to observe that the seasonal fluc-
tuations of the two sets of data show a relatively well-fiing correspondence paern as
well as the same seasonal differences as in Figure 8.
Figure 8. Average ground-level NO2 concentration map for 2019–2022.
It can be clearly observed from Figure 8that the ground-level NO2 concentration
in the Beijing–Tianjin–Hebei region shows a clear decreasing trend during the period of
2019–2022. Especially in the northern non-important industrial areas, the area of the low-
concentration area shows a trend of expansion. At the same time, the annual average NO2
concentrations in important industrial areas or transportation hubs, such as Tangshan,
Tianjin, and Shijiazhuang, have also decreased to some extent. This trend may be related
to factors such as the development of public health events and environmental protection
policies. However, there is a rebound in overall NO2 concentrations in 2022 compared to
the previous year, which to some extent reflects the gradual recovery of industries in the
studied area, such as industry and transportation, which have adapted to the development
of public health events.
Figure 9represents the standard deviation of the seasonal changes in near-surface NO2
concentrations and TROPOMI NO2 concentrations over different quarters. The horizontal
coordinates indicate the quarters, from 1 to 4, representing the first through fourth quarters
of the year, respectively. The vertical coordinate represents the standard deviation value,
which measures the magnitude of change in the data over each quarter. By comparing the
information on the graphs, it is possible to observe that the seasonal fluctuations of the
two sets of data show a relatively well-fitting correspondence pattern as well as the same
seasonal differences as in Figure 8.

Remote Sens. 2023,15, 3878 13 of 24
Remote Sens. 2023, 15, 3878 14 of 26
Figure 9. Standard deviation of seasonal variation in near-surface NO2 and TROPOMI NO2 con-
centrations.
In order to observe in detail the differences in air quality from quarter to quarter
during the study period, this study also statistically analyzed the daily data and divided
the near-surface NO2 concentration into four threshold ranges, which are less than 20
µg/m
3
, less than 40 µg/m
3
, less than 60 µg/m
3
, and greater than 60 µg/m
3
, to obtain the
stacked histograms of the percentages of near-surface NO2 concentration at different
thresholds for each quarter of the period of 2019 to 2022 (Figure 10).
Figure 10. Percentage stacked histograms of near-surface NO2 concentrations at different thresh-
olds, per quarter, 2019–2022.
Figure 9.
Standard deviation of seasonal variation in near-surface NO2 and TROPOMI NO2 concentrations.
In order to observe in detail the differences in air quality from quarter to quarter
during the study period, this study also statistically analyzed the daily data and di-
vided the near-surface NO2 concentration into four threshold ranges, which are less than
20
µ
g/m
3
, less than 40
µ
g/m
3
, less than 60
µ
g/m
3
, and greater than 60
µ
g/m
3
, to obtain
the stacked histograms of the percentages of near-surface NO2 concentration at different
thresholds for each quarter of the period of 2019 to 2022 (Figure 10).
Remote Sens. 2023, 15, 3878 14 of 26
Figure 9. Standard deviation of seasonal variation in near-surface NO2 and TROPOMI NO2 con-
centrations.
In order to observe in detail the differences in air quality from quarter to quarter
during the study period, this study also statistically analyzed the daily data and divided
the near-surface NO2 concentration into four threshold ranges, which are less than 20
µg/m
3
, less than 40 µg/m
3
, less than 60 µg/m
3
, and greater than 60 µg/m
3
, to obtain the
stacked histograms of the percentages of near-surface NO2 concentration at different
thresholds for each quarter of the period of 2019 to 2022 (Figure 10).
Figure 10. Percentage stacked histograms of near-surface NO2 concentrations at different thresh-
olds, per quarter, 2019–2022.
Figure 10.
Percentage stacked histograms of near-surface NO2 concentrations at different thresholds,
per quarter, 2019–2022.

Remote Sens. 2023,15, 3878 14 of 24
It can be seen that there is an overall downward trend in near-surface NO2 concentra-
tions from 2019–2022, with better air quality in the second and third quarters compared to
the first and fourth quarters, and almost no NO2 concentration thresholds above 60
µ
g/m
3
occurring throughout the quarter, but the seasonal differences presented on the graphs
have yet to be discussed in depth on a daily basis, in order to explore the reasons for the
differences and the factors that influence them.
Based on these findings, this study will further investigate and discuss the daily NO2
concentration changes for specific areas and specific time periods. Additionally, different
modeling approaches will be used and the results will be compared cross-sectionally with
those of the GTNNWR model to further validate the reliability of our research methods. In
addition, attention will be paid to the results of high-resolution, small-scale areas to capture
the geospatial variation in NO2 concentration more accurately.
3.2. Comparison of the Changes in NO2 Concentration in the Study Area in the New Year
The study is important for analyzing the impact of a global outbreak of public health
events on industrial production, transportation, and other areas from the end of 2019 to
2022. An in-depth analysis was performed by selecting the changes in ground-level NO2
concentrations (2019, Figure 11; 2020, Figure 12) for the adjacent 2019 and 2020 New Year
holidays (4–9 February 2019 and 24–29 January 2020) in the study area.
Remote Sens. 2023, 15, 3878 15 of 26
It can be seen that there is an overall downward trend in near-surface NO2 concen-
trations from 2019–2022, with beer air quality in the second and third quarters compared
to the first and fourth quarters, and almost no NO2 concentration thresholds above 60
µg/m
3
occurring throughout the quarter, but the seasonal differences presented on the
graphs have yet to be discussed in depth on a daily basis, in order to explore the reasons
for the differences and the factors that influence them.
Based on these findings, this study will further investigate and discuss the daily NO2
concentration changes for specific areas and specific time periods. Additionally, different
modeling approaches will be used and the results will be compared cross-sectionally with
those of the GTNNWR model to further validate the reliability of our research methods.
In addition, aention will be paid to the results of high-resolution, small-scale areas to
capture the geospatial variation in NO2 concentration more accurately.
3.2. Comparison of the Changes in NO2 Concentration in the Study Area in the New Year
The study is important for analyzing the impact of a global outbreak of public health
events on industrial production, transportation, and other areas from the end of 2019 to
2022. An in-depth analysis was performed by selecting the changes in ground-level NO2
concentrations (2019, Figure 11; 2020, Figure 12) for the adjacent 2019 and 2020 New Year
holidays (4–9 February 2019 and 24–29 January 2020) in the study area.
Figure 11. Spatiotemporal distribution of ground-level NO2 concentration in Beijing–Tianjin–Hebei
region during the 2019 New Year holidays (4–9 February).
Figure 11.
Spatiotemporal distribution of ground-level NO2 concentration in Beijing–Tianjin–Hebei
region during the 2019 New Year holidays (4–9 February).
The following results can be obtained from the comparative study of Figures 11 and 12:
By comparing the changes in the spatiotemporal distribution of ground-level NO2 concen-
trations during the New Year holidays in 2019 and 2020, it was found that the ground-level
NO2 concentrations during the New Year holidays in 2020 were significantly lower than
those in 2019, from which it can be inferred that the public health event had a more sig-
nificant effect on industrial production and transportation activities in the Beijing–Tianjin–
Hebei region. At the same time, the reduction of industrial production and transportation
activities leads to the improvement of air quality, and it can be seen from Figures 11 and 12
that the event itself had a positive effect on the reduction of ground-level NO2 concentra-
tions. Comparing the changes in ground NO2 in different provinces and cities over the
same period, it can be seen that Tianjin and Tangshan, which are heavy industrial and port

Remote Sens. 2023,15, 3878 15 of 24
cities, are less responsive to the impact of public health events, and although the ground
NO2 concentration decreases to some extent, it still remains high compared to other areas
in Beijing, Tianjin, and Hebei, which may be related to the high level of industry such as
manufacturing or port trade in the region.
Remote Sens. 2023, 15, 3878 16 of 26
Figure 12. Spatiotemporal distribution of ground-level NO2 concentration in Beijing–Tianjin–Hebei
region during the 2020 New Year holidays (24–29 January).
The following results can be obtained from the comparative study of Figures 11 and
12: By comparing the changes in the spatiotemporal distribution of ground-level NO2
concentrations during the New Year holidays in 2019 and 2020, it was found that the
ground-level NO2 concentrations during the New Year holidays in 2020 were significantly
lower than those in 2019, from which it can be inferred that the public health event had a
more significant effect on industrial production and transportation activities in the Bei-
jing–Tianjin–Hebei region. At the same time, the reduction of industrial production and
transportation activities leads to the improvement of air quality, and it can be seen from
Figures 11 and 12 that the event itself had a positive effect on the reduction of ground-
level NO2 concentrations. Comparing the changes in ground NO2 in different provinces
and cities over the same period, it can be seen that Tianjin and Tangshan, which are heavy
industrial and port cities, are less responsive to the impact of public health events, and
although the ground NO2 concentration decreases to some extent, it still remains high
compared to other areas in Beijing, Tianjin, and Hebei, which may be related to the high
level of industry such as manufacturing or port trade in the region.
In conclusion, by analyzing the changes in the spatiotemporal distribution of ground-
level NO2 concentrations during the New Year holidays in 2019 and 2020, it is possible to
gain insight into the effects of public health events on industrial production, transporta-
tion, and other areas, and to preliminarily assess their effects on the improvement of air
quality in the study area. In addition, by identifying and assessing the effects of different
factors, the driving mechanisms of NO2 concentration changes can be explored prelimi-
narily. These results have important reference value for environmental management and
air pollution management strategy development.
3.3. Variations in Ground-Level Nitrogen Dioxide Concentrations during Public Health Events
Building upon the aforementioned understanding, the study will focus on analyzing
the variation of ground-level NO2 concentrations during the public health event that oc-
curred at Xinfadi, Fengtai District, Beijing, from 11 June to 22 June, 2020. Three different
models, namely GNNWR, CNN, and RF, will be utilized to generate data for the Fengtai
District, which will be compared with the results obtained from the GTNNWR model.
This comparison aims to further discuss the reliability of the research methods employed.
Figure 12.
Spatiotemporal distribution of ground-level NO2 concentration in Beijing–Tianjin–Hebei
region during the 2020 New Year holidays (24–29 January).
In conclusion, by analyzing the changes in the spatiotemporal distribution of ground-
level NO2 concentrations during the New Year holidays in 2019 and 2020, it is possible to
gain insight into the effects of public health events on industrial production, transportation,
and other areas, and to preliminarily assess their effects on the improvement of air quality
in the study area. In addition, by identifying and assessing the effects of different factors,
the driving mechanisms of NO2 concentration changes can be explored preliminarily. These
results have important reference value for environmental management and air pollution
management strategy development.
3.3. Variations in Ground-Level Nitrogen Dioxide Concentrations during Public Health Events
Building upon the aforementioned understanding, the study will focus on analyzing
the variation of ground-level NO2 concentrations during the public health event that
occurred at Xinfadi, Fengtai District, Beijing, from 11 June to 22 June, 2020. Three different
models, namely GNNWR, CNN, and RF, will be utilized to generate data for the Fengtai
District, which will be compared with the results obtained from the GTNNWR model.
This comparison aims to further discuss the reliability of the research methods employed.
Figure 13 illustrates the temporal variation of ground-level NO2 concentrations in the
Beijing area from 11 June to 22 June, 2020.
Figure 13 presents the variation of ground-level NO2 concentrations in the Beijing area
during the specified period. It can be observed that after a significant decrease, the NO2
concentrations exhibit a slow recovery trend, particularly in the Fengtai District where Xin-
fadi is located, and its surrounding areas. This highlights the importance of high-resolution
daily data, enabling relevant personnel to make timely and accurate assessments of the
event’s development by incorporating policy interventions or other influencing factors.

Remote Sens. 2023,15, 3878 16 of 24
Remote Sens. 2023, 15, 3878 17 of 26
Figure 13 illustrates the temporal variation of ground-level NO2 concentrations in the Bei-
jing area from 11 June to 22 June, 2020.
Figure 13. Spatiotemporal distribution of ground-level NO2 concentrations in Beijing from 11 June
to 22 June 2020.
Figure 13 presents the variation of ground-level NO2 concentrations in the Beijing
area during the specified period. It can be observed that after a significant decrease, the
NO2 concentrations exhibit a slow recovery trend, particularly in the Fengtai District
where Xinfadi is located, and its surrounding areas. This highlights the importance of
high-resolution daily data, enabling relevant personnel to make timely and accurate as-
sessments of the event’s development by incorporating policy interventions or other in-
fluencing factors.
In order to clearly compare the differences and synergies in the predictive results
among the four methods, Figure 14 was chosen to display the comparison of the 500 m
resolution prediction results for the Fengtai District on 15 June 2020, processed with the
four models.
Figure 13.
Spatiotemporal distribution of ground-level NO2 concentrations in Beijing from 11 June to
22 June 2020.
In order to clearly compare the differences and synergies in the predictive results
among the four methods, Figure 14 was chosen to display the comparison of the 500 m
resolution prediction results for the Fengtai District on 15 June 2020, processed with the
four models.
Remote Sens. 2023, 15, 3878 18 of 26
Figure 14. Four models to deal with the predicted spatiotemporal distribution of ground-level NO2
concentration at 500 m resolution in Fengtai District on 15 June 2020.
Through the comparative analysis presented in Figure 14, it is evident that the
GTNNWR model, in comparison to the other models, effectively captures the develop-
mental trends and details of ground-level NO2 concentrations, while successfully avoid-
ing the “striping” effect caused by noisy data. Particularly, in the vicinity of the Xinfadi
area in the eastern part of Fengtai District, the NO2 concentrations are significantly lower
than in other areas. This observation indirectly reflects the overall effectiveness of policy
interventions in controlling transportation and industrial production during the specified
time period.
This finding suggests that the GTNNWR model exhibits superior accuracy in data
processing and robust noise filtering capabilities compared to the other models, resulting
in more reliable and accurate prediction results. Furthermore, the accuracy and stability
of the GTNNWR model further validate its reliability and effectiveness within the research
methodology.
To further compare the differences in ground-level nitrogen dioxide (NO2) concen-
trations between the provincial capital city and other regions, particularly important in-
dustrial cities and transportation hubs, daily high-resolution, ground-level NO2 concen-
trations from 2 January to 16 January, 2021, in Shijiazhuang, the capital city of Hebei Prov-
ince, were selected for the study (Figure 15).
Figure 14.
Four models to deal with the predicted spatiotemporal distribution of ground-level NO2
concentration at 500 m resolution in Fengtai District on 15 June 2020.

Remote Sens. 2023,15, 3878 17 of 24
Through the comparative analysis presented in Figure 14, it is evident that the
GTNNWR model, in comparison to the other models, effectively captures the developmen-
tal trends and details of ground-level NO2 concentrations, while successfully avoiding the
“striping” effect caused by noisy data. Particularly, in the vicinity of the Xinfadi area in the
eastern part of Fengtai District, the NO2 concentrations are significantly lower than in other
areas. This observation indirectly reflects the overall effectiveness of policy interventions in
controlling transportation and industrial production during the specified time period.
This finding suggests that the GTNNWR model exhibits superior accuracy in data
processing and robust noise filtering capabilities compared to the other models, resulting
in more reliable and accurate prediction results. Furthermore, the accuracy and stabil-
ity of the GTNNWR model further validate its reliability and effectiveness within the
research methodology.
To further compare the differences in ground-level nitrogen dioxide (NO2) concentra-
tions between the provincial capital city and other regions, particularly important industrial
cities and transportation hubs, daily high-resolution, ground-level NO2 concentrations
from 2 January to 16 January, 2021, in Shijiazhuang, the capital city of Hebei Province, were
selected for the study (Figure 15).
Remote Sens. 2023, 15, 3878 19 of 26
Figure 15. Spatiotemporal distribution of daily high-resolution ground-level NO2 concentrations in
Shijiazhuang, the capital of Hebei Province, from January 2 to January 13, 2021.
Figure 15 shows the spatiotemporal distribution of high-resolution ground-level
daily NO2 concentrations from 2 January to 13 January 2021 in Shijiazhuang, the capital
of Hebei Province. According to Figure 15, the NO2 concentration in Shijiazhuang gener-
ally showed a decreasing trend during this period, especially in Shijiazhuang Gaocheng
district and its surrounding areas, which were more affected by public health events dur-
ing this period. The decrease in nitrogen dioxide concentration may be influenced by a
variety of factors.
Firstly, the spatiotemporal distribution of ground-level NO2 concentrations during
the initial phase of the study period could be affected by the collective heating in northern
China during winter and the transportation activities related to the Spring Festival holi-
day. The collective heating in winter may lead to increased emissions of pollutants such
as coal combustion, thus impacting air quality. Additionally, during the Spring Festival
holiday, the substantial population movement and transportation activities could also
contribute to the release of NO2 pollutants, further affecting air quality. Therefore, during
the early phase of the study period, the NO2 concentrations in the urban area of Shijia-
zhuang were relatively high.
Howev