Estimation of Short-Term and Long-Term Ozone Exposure Levels in Beijing–Tianjin–Hebei Region Based on Geographically Weighted Regression Model

Abstract

In recent years, ozone (O₃) concentration has shown a decreasing trend in the Beijing–Tianjin–Hebei (BTH) region in China. However, O₃ pollution remains a prominent problem. Accurate estimation of O₃ exposure levels can provide support for epidemiological studies. A total of 13 variables were combined to estimate short- and long-term O₃ exposure levels using the geographically weighted regression (GWR) model in the BTH region with a spatial resolution of 1 × 1 km from 2017 to 2020. Five variables were left in the GWR model. O₃ concentration was positively correlated with temperature, wind speed, and SO₂, whereas is was negatively correlated with precipitation and NO₂. Results showed that the model performed well. Leave-one-out cross-validation (LOOCV) R² for short- and long-term simulation results were 0.91 and 0.71, and the values for RMSE were 11.14 and 3.49 μg/m³, respectively. The annual maximum 8 h average O₃ concentration was the highest in 2018 and the lowest in 2020. Decreasing concentrations of major precursors of O₃ due to the regional joint prevention and control may be the reason. O₃ concentration was high in the southeast of the BTH region, including in Hengshui, Handan, Xingtai and Cangzhou.

1. Introduction

Ground-level ozone (O₃) is mainly produced by nitrogen oxides (NOx) and volatile organic compounds (VOCs) through complex photochemical reactions [1]. O₃ formation is influenced by the emission of precursors, photochemical transformations and meteorological conditions. O₃ is a typical secondary pollutant, which not only damages the ecological environment and affects the growth of plants and animals, but also poses a certain threat to public health [2,3,4,5,6]. Epidemiological studies had indicated that short- and long-term exposure to severe O₃ pollution may cause many diseases, including asthma, lung function decrements, respiratory infection, heart failure, impaired heart function, and so on [7,8,9,10]. Meanwhile, exposure to O₃ pollution can lead to premature deaths [11,12]. It is estimated that there were 156,173 (95% confidence interval (CI): 79,562–303,843), 104,051 (95% CI: 35,824–200,055), and 33,456 (95% CI: 0–70,548) cases of all-cause, cardiovascular as well as respiratory death led by long-term exposure to O₃ in 331 Chinese cities in 2020 [13].

According to air quality report released by Ministry of Environmental Protection for the Beijing–Tianjin–Hebei (BTH) region and surrounding areas, the number of polluted days with O₃ as a dominating pollutant accounted for 41%, 46%, 48.2%, and 46.6% of the total polluted days from 2017 to 2020, respectively. Meanwhile, the daily maximum 8 h average O₃ concentrations were 193, 199, 196, and 180 μg/m³ during the four years, respectively [14,15,16,17]. None of them meet Class II standard (160 μg/m³) in Ambient Air Quality Standards (GB 3095-2012). Accurate estimation of O₃ exposure level and reasonable analysis of the spatiotemporal distribution characteristics are significant for preventing and controlling air pollution and assessing the impact on human health [18].

At present, multiple data and methods were used to estimate the O₃ concentration. Ground measurement can provide accurate O₃ concentration values at monitoring sites [1,19]. However, it cannot meet the large-scale analysis of spatio-temporal distribution of O₃, and the site maintenance requires a lot of human effort and material resources. Remote sensing techniques are applied to simulate O₃ concentration at large spatial scales. Satellite remote sensing data can provide large-scale spatial and temporal distribution information of O₃, but it is limited by atmospheric climate conditions such as clouds [20]. Chemical transport models (CTMs) can also be used to predict O₃ concentration, such as Weather Research and Forecasting Community Multiscale Air Quality (WRF-CMAQ) model. Such models provide complete spatio-temporal coverage by taking into account emission patterns, meteorological conditions as well as atmospheric chemical reactions. However, there are some drawbacks including high computational costs and incorrect specification of physical processes and initial/boundary state [18,21,22]. In addition, many statistical models have been applied to pollutant concentration estimation. Kriging interpolation estimates pollutant concentrations based on Tobler’s First Law of Geography [23]. Ordinary kriging (OK) is one of the kriging interpolation techniques, which performs well in predicting air quality levels in areas where only several monitoring sites are available [24]. OK assumes a stationary mean of the variable within the search window [25]. However, the observations of O₃ show high spatial heterogeneity in O₃ concentrations [26]. The land use regression (LUR) model uses pollutant monitoring data and spatial predictor variables to estimate the pollutant concentration, which can reflect small-scale spatial distribution of pollutant concentrations. Multiple linear regression (MLR) establishes the global relationship between several independent variables and a dependent variable. However, it neglects the spatial variability of air pollutants [27]. To cope with deficiencies of the above model, geographically weighted regression (GWR) was proposed [28]. GWR is a local spatial statistical method with spatial variability coefficients. It incorporates data location information to predict air pollutant concentrations on a large scale. The model can provide a quantitative spatially varying relationship between air pollutant concentrations and multiple predictor variables. Moreover, the local parameter estimates obtained by the model can show a high spatial variability, compensating for the weakness of the global regression methods. The GWR model has been used in several recent studies to explore air pollutant exposure levels. By the calibration of GWR, Stowell et al. [29] simulated daily PM_2.5 concentration in Southern California, and the results showed that the model performed well (cross-validation coefficient of determination (R²) = 0.80). Based on the GWR model, Shen et al. [30] simulated concentrations of NO₂, O₃, PM₁₀, and PM_2.5 across Europe from 2000 to 2019, and the mean R² values were 0.67, 0.57, 0.61, and 0.77, respectively.

This study attempts to estimate short- and long-term O₃ exposure levels in the BTH region, China, based on GWR model by fusing air pollutant data, meteorological data and the second Modern-Era Retrospective analysis for Research and Applications (MERRA-2) data. In addition, we evaluated the accuracy of the prediction results.

2. Data and Methods

2.1. Study Domain

The BTH region (36°05′–42°40′ N, 113°27′–119°50′ E) is located on the eastern coast of China and covers an area of 218,000 km². The topography is dominated by plains and plateaus, which is high in the northwest and low in the southeast. It comprises Beijing, Tianjin and Hebei Province. The region is not only one of the most economically developed area, but is also a densely populated region. Figure 1 displays the distributions of monitoring sites, meteorological stations and topography in the BTH region.

2.2. Data Collection and Processing

2.2.1. Air Pollutant Datasets

Air pollutant datasets were obtained from the China National Environmental Monitoring Centre (CNEMC) (http://beijingair.sinaapp.com/, accessed on 24 September 2022), including the hourly values of O₃, PM_2.5, PM₁₀, CO, NO₂, and SO₂ from 2017 to 2020. Calculated daily maximum 8 h average O₃ concentration demands at least 20 maximum 8 h O₃ concentrations during a day, and the calculation of the 24 h average for the other air pollutants demands at least 20 hourly average values during a day. A monitoring site used to calculate annual maximum 8 h average was retained when it was recording for at least 324 daily maximum 8 h average values during a year. Otherwise, this monitoring site was considered invalid. There were 90 O₃ monitoring sites. Based on these criterions, 86 sites with valid daily values and 81 sites with valid annual values were reserved for the O₃ concentration estimation from 2017 to 2020, respectively. The 4-year average annual O₃ concentration was regarded as the dependent variable for GWR model development. The distribution of 90 monitoring sites is shown in Figure 1.

2.2.2. Meteorological Data

The meteorological data were derived from China Meteorological Administration (http://data.cma.cn/, accessed on 24 September 2022), containing daily air pressure, wind speed, temperature, relative humidity as well as precipitation of 6887 meteorological stations from 2017 to 2020. Figure 1 also shows the distribution of meteorological stations.

2.2.3. MERRA-2 Reanalysis Data

MERRA-2 is released by NASA’s Global Modeling and Assimilation Office (GMAO) and provides data beginning in 1980. The spatial resolution is 0.5 × 0.625°, and the temporal resolutions are 1 h, 3 h, and 1 month. It provides data for a total of 72 vertical layers from the ground to 80 km altitude. Aerosol optical depth (AOD), planetary boundary layer height (PBLH), and atmospheric temperature data at 10 m, 1000 hPa, and 975 hPa in the BTH region from 1 January 2017 to 31 December 2020 were used in the study. In this paper, atmospheric temperature data at 10 m from the ground, 1000 hPa, and 975 hPa were used to calculate the strength of the temperature inversion. The air pressure was converted to altitude according to Equation (1).

$$H=44300\times (1-{\left(\frac{P}{P_0}\right)}^{\frac{1}{5.256}})$$

(1)

where H represents altitude, P represents air pressure and P₀ is standard atmospheric pressure (0 °C, 101.325 kPa). Then, 1000 and 975 hPa were converted to altitudes of 110 and 323 m, respectively. When the upper layer temperature is higher than the lower, the inversion temperature occurs. The strength of the temperature inversion refers to the ratio of the temperature difference to thickness of the inverse temperature layer, which is calculated by Equation (2).

$$TI=100\times \frac{T_2-T_1}{H_2-H_1}$$

(2)

where TI refers to the temperature inversion in units of °C/100 m. H₁ < H₂. T₁ and T₂ refer to the temperatures at the height of H₁ and H₂, respectively. TI is calculated when T₁ < T₂.

Since the predictor variables included in this paper have different sources and resolutions, all variables were interpolated to the grids with spatial resolutions of 1 km based on ArcGIS 10.2 using kriging, and then, the interpolated variables were matched with the O₃ monitoring data. Meanwhile, to assess the importance of these predictor variables, data were standardized by z-score before performing GWR according to the following formula:

$$z=\frac{x-\mu }{\sigma }$$

(3)

where z represents the standardized variable value, x is the actual value of the sample data, and μ and σ represent the average value and standard deviation of the sample data, respectively.

2.3. GWR Model Building

By establishing a local point-by-point regression model using ArcGIS 10.2, GWR reveals the relationship between independent and dependent variables. It allows regression coefficients to change in space, which can effectively explore the spatial non-stationary of variables. The calculation equation is as follows, where (u_i,v_i) refers to location information of the ith monitoring site, and a_k(u_i,v_i) represents the realization of the continuous function a_k(u_i,v_i) at the ith monitoring site.

$$y_i=a_0(u_i,v_i)+{\displaystyle \sum }_k{a}_k(u_i,v_i)x_{ik}+{\epsilon }_i$$

(4)

In the GWR model, for each sample, the other samples are given different weights according to their different spatial relationships with that sample, so that the regression coefficients are no longer constant but vary with the samples. The calculation formula is as follows:

$$\widehat{a}\left(u_i,v_i\right)={\left[X^{T}W\left(u_i,v_i\right)X\right]}^{-1}{X}^{T}W\left(u_i,v_i\right)y$$

(5)

where W(u_i,v_i) refers to an N × N weight matrix at sample point i. The weighting functions are mainly divided into adaptive kernel function and fixed kernel function. Adaptive kernel function was chosen due to the uneven distribution of O₃ monitoring sites.

Spatial weighting function is core of the GWR model. There are numerous methods for selecting the spatial weighting function, among which the distance threshold method is simple. However, this method suffers from the problem of discontinuity. As a sample for which parameters are estimated varies in space, the weights could change suddenly as other samples move into or out of the circular buffer around the sample. One obvious way to combat this is the Gaussian method, which uses a continuous monotone decreasing function to represent the relationship between the weight w_ij and the distance d_ij. The Gaussian function is shown in Equation (6).

$$w_{ij}=\exp (-\frac{d_{ij}^2}{{\beta }^2})$$

(6)

where β denotes the bandwidth, which refers to the non-negative decay parameter of the function between weight and distance.

Akaike information criterion (AIC) method was used to determine the bandwidth. The method uses two parameters to estimate the model and is a standard to measure the goodness of fit of the statistical model. The AIC value is expressed as:

$$AIC=2nln\left(\widehat{\sigma }\right)+nln\left(2\pi \right)+n\left[\frac{n+tr\left(S\right)}{n-2-tr\left(S\right)}\right]$$

(7)

where $\widehat{\sigma }$ is the maximum likelihood estimate of standard deviation of the random error, and S represents the hat matrix of O₃.

The stepwise regression approach was used to select the variables to be included in the GWR model. First, the regression equations of each predictor variable on O₃ concentration were established in turn. The variable that has the largest adjusted R² was selected as the first factor to be included in the model. The rest of the variables were then entered. Those with the potential to further improve the model performance (adjusted R² increased by at least 1%) were contained in the model in turn. The pattern lasted until no more variables satisfied the requirements.

Variance inflation factor (VIF) is an important indicator of multicollinearity among multiple variables in regression analysis. It is considered that there is no multicollinearity among the variables included in the model when VIF ≤ 3. Therefore, variables with p value > 0.1 or VIF > 3 were removed in the final step, and the model accuracy was reassessed [18].

2.4. Cross-Validation

The accuracy of the GWR model was assessed by the leave-one-out cross-validation (LOOCV) method. O₃ monitoring data were divided into n portions according to the number of the monitoring sites. O₃ monitoring data from one monitoring site was chosen as a validation set. Then, the model was trained using data from the remaining n − 1 monitoring sites, and the model accuracy was verified using the data from the validation set. The above steps were repeated n times until the O₃ monitoring data from all monitoring sites, which were used as a validation set. The predictive performance of the model was evaluated by the R² and the root mean square error (RMSE). The calculation formulas are as follows:

$$R^2=1-\frac{\sum {\left(y_i-{\widehat{y}}_i\right)}^2}{\sum {\left(y_i-{\overline{y}}_i\right)}^2}$$

(8)

$$RMSE=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{n}}$$

(9)

where y_i represents O₃ monitoring value, ${\overline{y}}_i$ is average value of O₃ concentration, ${\widehat{y}}_i$ refers to simulated value of O₃, and n is the sample number.

3. Result

3.1. Exploratory Data Analysis

The variables in the dataset included O₃, PM_2.5, PM₁₀, CO, NO₂, SO₂, precipitation, air pressure, relative humidity, temperature, wind speed, AOD, PBLH, and TI. The statistics of daily maximum 8 h average O₃ concentration are shown in

Table 1. Statistics of daily maximum 8 h average O₃ concentration.

Year	Point	Valid Value	Min (μg/m3)	Max (μg/m3)	Mean (μg/m3)	Median (μg/m3)
2017	83	28,175	1.46	476.63	93.64	83.38
2018	80	27,927	2.50	314.33	95.98	84.90
2019	80	27,820	1.50	319.5	92.25	82.65
2020	82	28,832	3.08	366.58	90.49	81.63

and

Table 2. Descriptive statistics of daily values for the predictor variables from 2017 to 2020.

	O3 (μg/m3)	PM2.5 (μg/m3)	PM10 (μg/m3)	CO (μg/m3)	NO2 (μg/m3)	SO2 (μg/m3)	Precipitation (mm)
Mean	93.07	53.01	95.63	1.02	40.43	15.66	1.30
Min	1.46	2.62	5.19	0.10	1.54	1.00	0
Max	476.63	644.14	1767.46	10.00	188.29	261.45	158.25
Standard deviation	52.23	44.68	68.73	0.70	21.13	14.78	6.03
	Pressure (hPa)	Relative Humidity (%)	Temperature (°C)	Wind Speed (m/s)	AOD	PBLH (m)	TI (°C/100 m)
Mean	1004.24	55.76	13.76	1.66	0.47	815.48	0.45
Min	869.63	7.24	−21.12	0.05	0.02	63.90	0
Max	1043.97	99.56	34.84	7.35	4.37	3651.20	4.95
Standard deviation	26.21	19.25	11.36	0.77	0.37	464.45	0.60

. The mean values of daily maximum 8 h average O₃ concentrations in the BTH area from 2017 to 2020 were 93.64, 95.98, 92.25, and 90.49 μg/m³, respectively, which increased first, and then decreased. The maximum values were 476.63, 314.33, 319.5, and 366.58 μg/m³, which occurred on 6 July 2017, 6 June 2018, 1 August 2019, and 25 June 2020, respectively. The statistics of daily values for the predictor variables from 2017 to 2020 are shown in Table 2, including the mean, minimum, maximum values and standard deviation for all variables.

Figure 2 shows the variation characteristics of the O₃ monitoring data from 2017 to 2020. Hourly O₃ concentration was kept at a high level from 12:00 to 20:00 and reached its maximum at 16:00 p.m. (110.09 μg/m³). Seasonal maximum 8 h average O₃ concentration peaked in summer, followed by spring, autumn, and winter. Annual maximum 8 h average O₃ concentrations were 92.82, 95.98, 92.11, and 89.06 μg/m³, respectively, and the corresponding standard deviations were 9.78, 6.22, 6.11, and 10.89 μg/m³ in 2017, 2018, 2019, and 2020, respectively. The 4-year average O₃ concentration was 92.42 μg/m³.

3.2. Model Fitting and Cross Validation

Table 3. Four-year average GWR model of O₃ concentration based on 81 monitoring sites.

Variables	β	p	VIF
Intercept	93.036	<0.001	NA
Precipitation	−1.385	0.002	1.572
Temperature	4.235	<0.001	1.812
NO2	−3.951	<0.001	2.574
Wind speed	1.304	<0.001	1.043
SO2	0.986	0.041	1.993

presents the predicted variables for the 4-year average GWR model. The five variables were included in the GWR model, which were precipitation, temperature, NO₂, wind speed, and SO₂. The VIF of all five variables was less than 3 and ensured that there were no redundant independent variables in the model. The regression coefficients of predictors in the GWR model for the simulation of 4-year average O₃ concentration indicated that temperature, wind speed, and SO₂ were positively correlated with the O₃ concentrations, whereas precipitation and NO₂ were negatively correlated with O₃ concentrations.

The daily and 4-year average annual maximum 8 h average values of 81 monitoring sites across the BTH region were matched with the GWR model simulations at the corresponding sites. Figure 3 shows the fitting scatter plots of the GWR model for the short- and long-term cross validation results, respectively. The LOOCV R² for the short- and long-term simulation results were 0.91 and 0.71, and the values for RMSE were 11.14 and 3.49 μg/m³, respectively.

3.3. Spatio-Temporal Distribution of O₃ Concentration

The short-term O₃ exposure levels were estimated from 2017 to 2020. Figure 4 showed the spatio-temporal distribution of the short-term O₃ exposure levels based on the GWR model with a spatial resolution of 1 × 1 km in the BTH region for the first day of each month in 2020. The maximum value of daily maximum 8 h average O₃ concentration was 224.84 μg/m³ on 7 June 2020, and the lowest value was 18.85 μg/m³ on 5 January 2020. Short-term O₃ exposure level was high in southern Hebei, such as Hengshui, Cangzhou and Handan, and low in Chengde and Qinhuangdao.

Figure 5 displayed the seasonal average O₃ concentrations and the percentages of O₃ exceedance days. The variation of O₃ concentrations showed an obvious seasonal characteristic in the BTH region. O₃ concentration was the highest in summer with an average value of 140.36 μg/m³, followed by spring (109.48 μg/m³), autumn (70.24 μg/m³) and winter (51.73 μg/m³). In both spring and summer, O₃ concentrations did not meet the Class I standard of Ambient Air Quality Standards of China (100 μg/m³). From the exceedance rate of each season, it can be revealed that the percentage of O₃ exceedance days ranged from 0% to 90.22%. The percentage of O₃ exceedance days in summer was the highest, which were 87.78%, 87.36%, 85.23%, and 90.22% from 2017 to 2020, respectively, while the lowest was in winter, meeting the standard in all four years.

The estimation results of annual and 4-year average annual O₃ exposure levels from 2017 to 2020 are shown in Figure 6 and Figure 7. The annual maximum 8 h average values in the BTH region from 2017 to 2020 were 93.58, 96.60, 92.66, and 91.21 μg/m³, respectively, which showed a clear trend of increasing from 2017 to 2018 and subsequently decreasing from 2018 to 2020. High O₃ concentration areas were gathered in the southern BTH region in 2018, including Hengshui, Cangzhou, Handan, and Xingtai. The O₃ concentrations in these cities all failed to meet the Class I standard (100 μg/m³). In terms of the variation of pollution areas, the O₃ pollution area gradually expanded from the west and southeast to the entire southern region from 2017 to 2018. However, the pollution area has shrunk to Shijiazhuang and Hengshui in the southern part of the BTH from 2018 to 2020. The 4-year average O₃ concentration estimation results presenting the O₃ concentration in the southeast and western regions of BTH were high, and the concentration in the northeast region was relatively low. The center of high O₃ concentrations appeared in Hengshui (97.62 μg/m³), followed by Handan and Xingtai, while lower O₃ concentrations were shown in Chengde and Qinhuangdao.

4. Discussion

The GWR model constructed in this study has achieved higher model performance than those of other comparable studies. Ma et al. [31] built a high-performance random forest (RF) model in the BTH region and estimated daily maximum 8 h average O₃ concentrations, and the R² value was 0.84. Lyu et al. [32] not only analyzed spatial and temporal variations of air pollutions but also predicted O₃ concentrations based on machine learning (ML) algorithms across the BTH region, and ML algorithms including RF and decision tree regression showed the R² values as 0.83 and 0.73, respectively. The GWR model constructed in this study has better prediction accuracy, with LOOCV R² of 0.91 and 0.71 for the short- and long-term O₃ exposure levels, respectively. In addition, O₃ exposure levels estimated in our study have a relatively high spatial and temporal resolution compared to other similar studies. Hu et al. [33] combined the Weather Research and Forecasting with Chemistry (WRF-Chem) model and extreme gradient boosting (XGBoost) algorithm to predict the near-surface O₃ concentrations with a resolution of 0.1 × 0.1° in the BTH region. Wei et al. [34] integrated multi-source datasets and used space-time extremely randomized trees (STET) to estimate surface O₃ concentrations at the resolution of 10 × 10 km across the BTH region. Xue et al. [35] combined the WRF-Chem and RF models to simulate O₃ concentrations with a horizontal resolution of 9 km across the BTH region. Compared with the above research studies, we attempted to predict O₃ concentrations at the spatial resolution of 1 × 1 km. Masri et al. [36] built an LUR model to simulate O₃ concentration spatial distributions in summer and winter in Tianjin, China, while we predicted O₃ exposure levels at temporal resolutions of 1 day, 1 year and 4 years.

Solar radiation is an essential condition for O₃ generation. Strong solar radiation increases temperature, and high temperature can make the photochemical reactions of O₃ formation active, which increases O₃ concentration. In addition, summer temperature is higher than winter temperature, which contributes to the high O₃ concentration in summer [37]. Wind speed was positively correlated with O₃ in the BTH region in this study, which is in line with the result of Wang et al. [38]. In general, high wind speed favors the diffusion of pollutants when there are only local sources of pollution, resulting in the reduction of O₃ concentration. The higher the wind speed, the higher the O₃ concentration, indicating that the influence of external pollution sources cannot be ignored [39]. Wang et al. [40] built the regional transport matrix of O₃ in the BTH region. The result suggested that the transport source was the major contributor for O₃, which provided the supporting evidence. Rainfall had a negative correlation with O₃ concentration. It can accelerate wet deposition by scavenging O₃ and its soluble precursors, e.g., NO₂, resulting in the reduction of O₃ concentration [41,42]. In addition, there may be a negative correlation between NO₂ and O₃ at fine scales, because NO emitted from motor vehicles near the road reacts rapidly to generate NO₂. At the same time, O₃ is reduced and other compounds are formed. This process is referred to as NOx titration.

During these four years, annual O₃ concentrations in the BTH region peaked in 2018 and then began to decrease. Regional joint prevention and control and the decreasing concentrations of major precursors of O₃ may be the reasons. Although a series of stringent air pollution control strategies has been issued over the past few years, for example, Air Pollution Prevention and Control Action Plan, O₃ concentrations were able to decrease until the integrated management of the O₃ precursor (VOCs) was first proposed in the important policy of governance for controlling O₃ in the winter of 2018. The policy had achieved initial improvements in 2019 [43]. O₃ concentration in the BTH region had a seasonal variation. The concentrations were highest in summer, followed by spring, autumn, and winter. In general, high temperature and strong solar radiation levels in summer can stimulate active photochemical reactions, which lead to high O₃ concentrations [44]. Although the temperature in spring was lower than summer, dry and rainless meteorological conditions also contribute to O₃ production [45]. In autumn and winter, O₃ concentrations decreased due to weakened solar radiation and lower temperatures. In terms of spatial variation, O₃ concentrations were higher in the southeastern region, including in Hengshui, Handan, Xingtai and Cangzhou, and were lower in Chengde. The southeastern part of BTH region was dominated by heavy industries, and these polluting industries have high precursors emissions, which are more conducive to O₃ generation [46]. In addition, the number of civil vehicles in Beijing, Tianjin, and Hebei province is gradually rising, which increases emissions from vehicle exhaust and thus worsens O₃ pollution [47,48,49].

There are some limitations in our study. The variables used in the GWR model including air pollutants and meteorology, as well as more variables such as traffic, population, and pollutant emission, should be considered in future studies to make more accurate estimations. In addition, there may not be enough samples for the regions lacking monitoring sites to accurately capture O₃ exposure levels. With the continuous development of the national air quality monitoring network, the performance of our model will be continuously improved.

5. Conclusions

Short- and long-term O₃ exposure levels were estimated based on the GWR model at the spatial resolution of 1 × 1 km in the BTH region from 2017 to 2020. The cross-validation results indicated that simulated concentrations were well consistent with observed O₃ concentrations. By estimating the parameters of each location, the GWR model constructed in this study showed a high ability to simulate O₃ concentration at different spatial and temporal scales. The temperature, wind speed, and SO₂ were positively related with O₃ concentrations, whereas precipitation and NO₂ were negatively correlated with O₃ concentrations. We found that O₃ concentration increased with wind speed, and the possible reason was that the contribution of the transport source was greater than the contribution of the local source. Compared with long-term O₃ exposure level, short-term O₃ exposure level obtained a better simulated performance. The estimations based on the GWR model showed spatio-temporal variation of O₃ concentrations across the BTH region. In summary, the model could accurately estimate O₃ exposure levels, which could be helpful for epidemiological studies and health risk assessment in the future.