Quantitative Analysis of Spatiotemporal Patterns and Factor Contributions of Surface Ozone in the North China Plain

Abstract

Over the past decade, surface ozone has emerged as a significant air pollutant in China, especially in the North China Plain (NCP). For effective ozone management in the NCP, it is crucial to accurately estimate the surface ozone levels and identify the primary influencing factors for ozone pollution in this region. This study utilized ozone precursors such as volatile organic compounds (VOCs) and nitrogen oxides (NO_X), meteorological data, land cover, normalized difference vegetation index (NDVI), terrain, and population data to build an extreme gradient boosting (XGBoost)-based ozone estimation model in the NCP during 2019 to 2021. Four ozone estimation models were developed using different NO₂ and formaldehyde (HCHO) datasets from the Sentinel-5 TROPOMI observations and Copernicus Atmosphere Monitoring Service (CAMS) reanalysis data. Site-based validation results of these four models showed high accuracy with R² values above 0.86. Among these four models, two models with higher accuracy and higher spatial coverage ratio were selected, and their results were averaged to produce the final ozone estimation products. The results indicated that VOCs and NO_X were the two main pollutants causing ozone pollution in the NCP, and their relative contributions accounted for more than 23.34% and 10.23%, respectively, while HCHO also played a significant role, contributing over 5.64%. Additionally, meteorological factors also had a notable impact, contributing 28.63% to ozone pollution, with each individual factor contributing more than 2.38%. The spatial distribution of ozone pollution identified the Hebei–Shandong–Henan junction as a pollution hotspot, with the peak occurring in summer, particularly in June. Therefore, for this hotspot region in the NCP, promoting the reduction in VOCs and NOx can play an important role in the mitigation of O₃ pollution and the improvement in air quality in this region.

1. Introduction

Ozone (O₃) has significant negative effects on the atmospheric environment and human health [1,2,3]. In addition, it also affects the Earth’s ecosystem and climate system [4,5,6,7]. First, tropospheric ozone, especially near-surface ozone pollution, poses a serious threat to human health [1,2,3]. Surface ozone is particularly harmful to the respiratory system [2], especially for children, the elderly, and people with respiratory diseases. Second, ozone has a significant impact on the water cycle by prompting plant stomata to close [4]. Ozone has a negative impact on the growth of crops and vegetation, interfering with crop yields and ecosystem stability, threatening plant photosynthesis, reducing crop output, and weakening the carbon absorption capacity of ecosystems [5,6]. Ozone pollution also poses a threat to biodiversity, and increases in ozone have led to significant reductions in bird populations [7]. In addition, as the third largest greenhouse gas, ozone may also disrupt the Earth’s radiation balance and affect the global climate system [5].

The continued deterioration of surface ozone pollution in China has been confirmed by many previous studies [8,9,10,11,12,13,14]. Among them, the NCP (including Beijing, Tianjin, Hebei, Shandong, Henan, and Shanxi Provinces) is the most economically developed and most populous region in China, one of the regions with the highest population density and urbanization level as well as being the region with the most serious air pollution, especially ozone pollution, in China [8,15].

Previous studies have shown that ozone pollution in the NCP is mainly caused by vehicle exhaust [16,17,18], industrial emissions [19,20,21,22,23,24], and domestic pollution sources [25,26,27,28,29]. In addition, topography, meteorological conditions, and seasonal changes also have an impact on ozone pollution. In terms of ozone precursors, NO_X and VOCs are the two main kinds of precursors of ozone pollution. With the improvement in urbanization level in the NCP, the continued increase in the number of vehicles has led to an increase in exhaust emissions. NO_X and VOCs in automobile exhaust are an important emission source of ozone precursors. In addition to automobile exhaust, industrial emissions are another important emission source of ozone precursors [30]. Due to the high population density in the NCP, domestic pollution sources emitted by human activities including household coal burning, emissions from the catering industry, and domestic waste disposal are also important sources of ozone precursors. In addition to ozone precursors, ozone pollution in the NCP is also significantly affected by meteorological conditions. Among the meteorological factors, temperature, radiation, wind, humidity, boundary layer height, precipitation, and cloud cover can directly or indirectly affect the diffusion and transmission of the ozone precursors (such as NO_X and VOCs) as well as the photochemical reaction processes of ozone. Furthermore, the Taihang Mountains and Yanshan Mountains in the northwest of the BTH and the Taishan Mountains in central Shandong affect the diffusion of ozone precursors in the region, further affecting ozone pollution in this region.

In existing studies, surface ozone is generally estimated using three main methods: deterministic models, empirical models, and machine learning models. Deterministic models are based on principles of atmospheric chemistry and aerodynamics, utilizing mathematical models to simulate the transport, diffusion, and chemical reactions of atmospheric gases to estimate surface ozone concentrations. This method, exemplified by models such as WRF-Chem, CMAQ, and GEOS-Chem, demands extensive meteorological and chemical data and a profound comprehension of the atmospheric chemical processes, enabling it to yield more precise results. Empirical models, based on statistical principles, involve analyzing historical observational data and environmental conditions to establish mathematical models that forecast surface ozone concentrations. Models such as land use regression (LUR), Bayesian maximum entropy (BME), generalized additive model (GAM), and geographically weighted regression (GWR) are all statistical models. Statistical models are relatively simple to implement, but these models are often relatively low in accuracy and are limited by data quality and reliability. Artificial intelligence including machine learning and deep learning is popular for its powerful data mining capabilities. It facilitates the extraction of valuable insights from vast datasets, leading to more accurate parameter estimation through various techniques such as neural networks, random forests (RF), and extreme gradient boosting (XGBoost). Machine learning and deep learning models are simpler than the implementation of deterministic models, and the accuracy of parameter estimation is higher than that of statistical models.

Compared with previous studies, this paper utilized high spatial resolution NO₂ and HCHO observations derived from S5P satellite. These S5P-NO₂ and S5P-HCHO datasets have been used sparingly in previous studies and have higher integrity than other satellite observations. In addition, the spatial resolution of the ozone estimation results in most studies was 0.1° × 0.1° or coarser, and only a few studies have estimated surface ozone products with a resolution of 0.05° × 0.05°, therefore, it is necessary to estimate O₃ concentration products with a higher resolution of 0.05° × 0.05°. In recent years, as machine learning models have seen rapid development, the XGBoost model, recognized for its efficiency as an integrated learning algorithm, has been demonstrated to possess robust predictive capabilities and has been projected to have extensive applications in the fields of data mining and predictive modeling. The NCP, which is characterized by a developed economy, dense population, and high degree of urbanization, is also considered to be one of the most polluted areas in China, especially in terms of ozone pollution, so it was selected as the study area for the study. Consequently, an XGBoost-based model was developed to estimate the concentration of ozone in the NCP. Once the model was constructed, its precision was validated. Subsequently, an assessment was conducted to determine the relative significance and the extent of the contribution of various factors affecting near-surface ozone pollution. Subsequently, it was utilized to calculate the gridded daily maximum daily 8-h average (MDA8) concentration data at a resolution of 0.05° × 0.05° for the NCP over the period from 2019 to 2021. Utilizing the MDA8 estimation outcomes, insights into the spatial and temporal patterns of O₃ pollution were uncovered. Additionally, the study provided a further analysis of the spatiotemporal distribution of ozone pollution in the region, considering the frequency of ozone pollution exceedances. These studies have been instrumental in enhancing our comprehension of ground-level O₃ pollution and in the development of targeted environmental protection policies.

2. Data Sources and Methods

2.1. Data Sources

Hourly ground-based ozone concentration observations from 235 ground stations of China’s National Ground-Level Air Quality Observation Network were used to construct a surface ozone concentration estimation model in the NCP. The data from the National Environmental Monitoring Center of China covers the period 2019–2021. To ensure the quality of the data, the 8-h moving average ozone concentration was calculated only when at least six valid measurements had been recorded within the 8-h time period. In addition, the maximum value of the daily 8-h average surface ozone concentration was extracted as the MDA8 value. Strict quality control measures were implemented to exclude records with insufficient samples. In order to safeguard the precision of the surface ozone estimations from being compromised by variations in the data sample sizes, this study eliminated hourly records that lacked surface ozone data for a continuous ten days period within any given month and those with a sample loss exceeding 30% during the entire research timeframe. Moreover, to reduce interference from proximate monitoring stations and enhance the establishment of an accurate ozone estimation model, a threshold distance of 0.05° was implemented to consolidate the observational data from 235 sites in the NCP into a set of 133 sites. Figure 1 presents a visualization of the geographical arrangement of these 133 ozone monitoring locations within the NCP.

2.2. Ozone Precursors

The formation of ozone is predominantly initiated by photochemical processes that involve the oxidation of VOCs and carbon monoxide (CO). These reactions are catalyzed by NO_X and are further stimulated by UV radiation. Thus, the inclusion of these two key ozone precursors into surface ozone estimation models was deemed of paramount importance for elucidating the mechanisms of ozone production and for the prediction of air quality. This research utilized daily nitrogen dioxide (NO₂) and HCHO data collected by the S5P satellite, along with a reanalysis dataset comprising NO₂ profiles at 1000 hPa, HCHO total column densities (with a resolution of 0.1° × 0.1°), and NO_X emissions from the CAMS inventory spanning the years 2019 to 2021. These datasets were integrated to develop an enhanced model for estimating the surface ozone concentrations in the NCP. Furthermore, six VOC datasets (propane, cyclopentene, methane, hydrogen peroxide, hydroxyl radical, and peroxyacetyl nitrate) from the 2019–2021 CAMS emission inventory (0.1° × 0.1°) were also employed to enhance the optimization of the surface ozone estimation model. For the purpose of maintaining spatial uniformity, the datasets of ozone precursors were resampled to conform to a grid size of 0.05° × 0.05°. Concurrently, the daily mean values were computed for the datasets of six VOCs, CAMS-HCHO, and CAMS-NO₂ to align with the MDA8 standards. Notably, as the CAMS-NO_X dataset represents monthly emission estimates, the inventory data for a given month were applied uniformly across all days of that month for consistency.

2.3. Meteorological Factors

The temporal and spatial variation of surface ozone is affected not only by anthropogenic emissions, especially ozone precursors, but also by meteorological conditions. In the construction of the surface ozone estimation model, a comprehensive consideration was given to other factors impacting surface ozone variations, namely meteorological factors. The meteorological variables considered in this study encompassed temperature at a 2-m height above the surface (T), relative humidity (RH), surface pressure (SP), the height of the atmospheric boundary layer (BLH), wind speed (WS), wind direction (WD), the extent of cloud cover (TCC), total precipitation (TP), and exposure to ultraviolet (UV) radiation. The data utilized in this analysis, characterized by a spatial resolution of 0.1° × 0.1° over the period from 2019 to 2021, were sourced from the ERA5 reanalysis dataset. To enhance the accuracy of the surface ozone estimation and align with the findings of prior research, the model incorporated daily aggregated UV radiation and TP data, along with the daily mean values for RH, SP, BLH, and TCC. Additionally, the model considered the daily peak T and the hourly WS and WD measurements recorded at 14:00 PM, a timeframe commonly associated with peak MDA8 and temperature readings. All meteorological factor datasets were resampled into a uniform 0.05° × 0.05° grid in the modeling process of surface ozone estimation. The detailed descriptions and performance metrics of these extensively utilized datasets have been thoroughly documented in prior studies [31].

2.4. Auxiliary Data

Except for the consideration of ozone precursors and meteorological elements, a multitude of datasets were employed in the development of the surface ozone estimation model tailored for the NCP. The compilation of datasets for the study comprised digital elevation model (DEM) information, NDVI derived from satellite data, land use and cover (LULC) classifications, demographic data reflecting population density (POPU), and the CAMS ozone reanalysis product (CAMS-O₃). DEM data with a 30-m spatial resolution were obtained from the China Resources and Environmental Science Data Center (CRESDC). Concurrently, POPU data with a 1 km × 1 km spatial resolution from 2019 to 2021, also derived from the CRESDC, were corrected based on the statistical yearbooks of each city for more accurate surface ozone estimation. The LULC data comprised nine ground-level attributes (LULC1~LULC9) from 2019 to 2021, with a spatial resolution of 30 m × 30 m. Furthermore, NDVI data provided by the 16-day MOD13C1 product generated by the Moderate Resolution Imaging Spectroradiometer (MODIS) on the Terra satellite were employed, featuring a spatial resolution of 0.05° × 0.05° and covering the period from 2019 to 2021 for model construction. To more accurately estimate the surface ozone in the NCP, the 1000 hPa O₃ data from the CAMS reanalysis product were selected as influential factors as they contain surface ozone information. In a treatment similar to CAMS-HCHO, the daily average concentration of CAMS-O₃ was calculated to be consistent with the daily MDA8. For the subsequent analysis of land surface ozone, all gathered explanatory variables, which included data on ozone precursors, meteorological conditions, DEM, POPU, LULC, and CAMS-O₃, were reprocessed to fit a 0.05° × 0.05° grid. Along the temporal dimension, given that the population and LULC data were gathered on an annual basis, the same set of population and LULC data was applied on a daily basis for every year to correspond with the daily MDA8 measurements. Similarly, a uniform dataset of NDVI values was used consistently for each calendar day over a 16-day period. Monitoring stations that lacked adequate sample sizes were omitted, and subsequently, the daily MDA8 time series data were aligned with the daily datasets of relevant explanatory variables to ensure a synchronized analysis.

The dataset encompassed information on ozone precursors, meteorological conditions, DEM, NDVI, LULC, and POPU statistics across 133 separate monitoring sites. For each site, the corresponding data pairs were matched based on their geographical location and temporal correlation.

3. Method

3.1. Statistical Modeling Methods

XGBoost, a gradient boosting tree model, was designed to construct a robust prediction framework by amalgamating various weak learners, typically decision trees. The fundamental concept underlying XGBoost involves the application of ensemble learning to enhance the overall model’s performance through the amalgamation of predictions from multiple models. “Gradient boosting” refers to an iterative process in which each iteration is designed to correct for the residuals of the previous model, leading the model closer to the actual predicted value. The term “gradient boosting” refers to the iterative process where each iteration aims to rectify the residuals from the preceding model, thereby guiding the model incrementally closer to the actual predicted values. XGBoost employs decision trees as its base learner, with each iteration training a new tree model based on the residuals of the prior model, subsequently aggregating the predictions of this new tree with those of the previous model. To mitigate overfitting, regularization terms were integrated into XGBoost including L1 and L2 regularization of leaf node weights. Additionally, XGBoost incorporates the notion of a learning rate, which governs the weight of the new model in each iteration, thus modulating the rate of model update. By judiciously choosing the learning rate and number of iterations, XGBoost is able to significantly enhance the model’s stability and generalization capabilities. The model boasts efficient implementation, leveraging optimized data structures and algorithms, which facilitates faster training on extensive datasets.

Before using XGBoost to build a surface ozone estimation model, it is necessary to collect relevant data including data on various explanatory variables of ozone and monitoring data on surface ozone concentration sites. The construction of a surface ozone estimation model employing XGBoost commenced with the collection of pertinent data, which included the procurement of surface ozone concentration readings and a variety of explanatory variables. The dataset was then divided into a training set, a test set, and a validation set, and the XGBoost regression model was subsequently initialized and trained. By iteratively optimizing the decision tree through gradient boosting, XGBoost gradually revealed and expressed the complex interaction between the surface ozone and influencing factors. After the model was trained, the test set and the validation set were used to evaluate the accuracy of the model, and the accuracy of the model was evaluated by metrics such as the mean square error. Post the training phase, the test set was utilized to evaluate the model, with accuracy assessed using indicators such as the mean square error. By integrating NO₂ and HCHO data obtained from satellite-based monitoring (S5P-TROPOMI) and re-analysis sources (CAMS), four distinct XGBoost models for estimating the surface ozone levels in the NCP were developed, utilizing varied datasets of NO₂ and HCHO, as outlined below:

surface-O₃~f (S5P-NO₂/CAMS-NO₂, S5P-HCHO/CAMS-HCHO, CAMS-O₃, VOCs, CAMS-NO_X, BLH,
RH, T, TCC, TP, SP, UV, WS, WD, DEM, NDVI, POPU, LULC)

(1)

Among them, surface-O₃ denotes the ground-based MDA8 observational data, while S5P-NO₂ and S5P-HCHO correspond to the column densities of tropospheric NO₂ and total HCHO, respectively. The terms BLH, RH, T, TCC, TP, SP, UV, WS, and WD are indicative of the boundary layer height, 1000 hPa relative humidity, surface 2-m temperature, total cloud cover, total precipitation, surface pressure, ultraviolet radiation, wind speed, and wind direction. In the construction of the surface ozone estimation model, a collection of six individual datasets from the CAMS VOC inventory was integrated. This inventory includes C₃H₈ (propene), C₅H₈ (cyclopentadiene), CH₄ (methane), H₂O₂ (hydrogen peroxide), OH (hydroxyl radical), and PAN (peroxyacetyl nitrate), all identified as distinct precursors that influence the formation of surface ozone. The model was subsequently formulated with these components. These six VOCs were integrated into the VOC category in Equation (1), while DEM and NDVI were also employed to denote the altitude and normalized vegetation index for the corresponding location and date, respectively. In a parallel manner, POPU and LULC were designated to denote the population density and the categorization of land use types, respectively. These combined considerations can help to establish a more comprehensive and accurate estimation model of surface ozone. All of the datasets used in this study are summarized in Table S1.

As the training target dataset for the surface ozone estimation model, we clustered the MDA8 observation data from 235 stations in the NCP and integrated them into 133 stations. The distance threshold for sites clustering was set to 0.05°. In the subsequent model training process, we initiated this by isolating 5% of the total 133 locations to form the location-centric validation dataset. Subsequently, we extracted 80% of the samples from the remaining 95% of these sites to constitute the training dataset. Finally, a subset of 20% of the samples from the 95% of locations was employed to formulate a validation dataset that was based on the previously determined training samples. The selection of the input data samples for this validation set was performed randomly, maintaining an impartial methodology for both the training and validation phases of the model development process. Then, we used the XGBoost model to accurately and quantitatively evaluate the contribution and relative importance of each influencing factor to the surface ozone estimation model, quantified the relative importance of each influencing factor in the XGBoost model to the surface ozone estimation, and made full use of the XGBoost model to accurately evaluate each factor importance. To enhance the precision of the MDA8 estimation model, we conducted a series of 50 training iterations, each producing an estimation outcome. We then computed the mean of these 50 model outputs, creating a composite dataset that represents the average MDA8 estimation. This aggregated dataset was subsequently employed as the definitive daily gridded (0.05° × 0.05°) estimation of MDA8 concentrations within the NCP for the years 2019 and 2021. This approach ensured a robust representation of the MDA8 levels by mitigating the impact of random variability inherent in individual model runs.

3.2. Statistical Indicators for Model Accuracy Evaluation

After the XGBoost algorithm model had been utilized to construct a surface ozone estimation model, the accuracy of the estimation model was evaluated and combined with relevant statistical indicators to ascertain the performance of the model. To thoroughly evaluate the accuracy of the model’s surface ozone estimations, a suite of performance indicators was applied. This set included the root mean square error (RMSE) to measure the deviation of predictions, the mean prediction error (MPE) to gauge the overall bias, the relative percentage error (RPE) to assess errors relative to the observed values, and the coefficient of determination (R²) to quantify the model’s explanatory power. Among them, RMSE is a standard indicator for measuring the prediction error of the model and is more sensitive to large errors. The smaller the RMSE value, the better the prediction performance of the model. MPE is a measure of the average percentage error of the predicted value relative to the actual value, which can reflect the direction of the prediction deviation. RPE is used to measure the relative error of the predicted value relative to the actual value and is used for the error calculation of a single sample. R² is an indicator to measure the degree of variation of the actual value explained by the model. The closer the value is to 1, the better the model fits. These metrics provided a multifaceted assessment of the model’s predictive capabilities concerning ozone levels. These evaluation metrics served not only to quantify the model’s overall efficacy, but to also expose its performance across various dimensions, offering valuable insights for further refinement and enhancement. Through this systematic accuracy evaluation, the feasibility and effectiveness of the XGBoost model for surface ozone estimation could be more reliably adjudged, thus augmenting the precision and practical utility of such estimations. The evaluation of XGBoost models typically encompasses a multitude of indicators, with some of the common evaluation indicators being mathematically represented as follows:

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

(2)

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

(3)

$$MPE=100\%\times \frac{1}{n}\sum _{i=1}^n\frac{(y_i-{\widehat{y}}_i)}{y_i}$$

(4)

$$RPE=100\%\times \frac{1}{n}{\sum }_{i=1}^n\frac{\left|y_i-{\widehat{y}}_i\right|}{y_i}$$

(5)

Among them, $y_i$ is the observed value, ${\widehat{y}}_i$ is the predicted value, $\overline{y}$ is the mean of the observed value, and n is the sample number. In the evaluation and comparison of the predictive accuracy of the four combinations models, using different combinations of NO₂ and HCHO from the CAMS reanalysis datasets and S5P-TROPOMI observations, the models were selected for their higher prediction precision and higher extensive spatiotemporal coverage in the estimation results of MDA8. This combined model was determined as the final combined model for surface ozone estimation. Subsequently, the identified models for estimating surface ozone were then utilized to calculate the daily gridded MDA8 values. The mean outcomes derived from these models were treated as the definitive daily gridded MDA8 datasets, characterized by a spatial resolution of 0.05° × 0.05°, representing the NCP for the period spanning from 2019 to 2021. Furthermore, the research delved into a detailed examination of the spatial and temporal dynamics of MDA8 within the NCP and scrutinized the distribution and trends of high ozone pollution episodes, specifically instances where MDA8 exceeded 160 μg m⁻³. This study aimed to comprehensively understand the spatiotemporal characteristics of surface ozone and the potential impact of explanatory variables and external factors on ozone pollution.

4. Results

4.1. Evaluation of the Accuracy of the MDA8 Estimation Model

In this study, a quartet of models for estimating surface ozone concentrations was developed, each underpinned by the XGBoost algorithm. These models incorporated different datasets of ozone precursors, specifically focusing on the following pairings: data from S5P NO₂ alongside S5P HCHO, S5P NO₂ with CAMS HCHO, CAMS NO₂ with S5P HCHO, and finally, CAMS NO₂ with CAMS HCHO. This diverse integration of datasets aimed to enhance the accuracy and reliability of the ozone estimations. The accuracy and spatial coverage of the MDA8 estimation models, derived from these four schemes, were then compared. The two methods that yielded the highest model accuracy and the most comprehensive spatial coverage were selected. Finally, the average near-surface ozone estimation of the NCP was determined by averaging the MDA8 grid products estimated by the two methods.

Figure 2 presents the site-based accuracy verification results of four MDA8 estimation methodologies. In the course of model training, initially, 5% (7 out of 133) of the clustered sites within the NCP were randomly selected and constituted the site-based validation dataset, followed by the selection of an additional 20% of the remaining 95% of sites for the sample-based validation dataset. The results indicated that all four XGBoost models for MDA8 demonstrated a high verification accuracy, with a strong correlation existing between the models’ estimated MDA8 concentrations and the ground-based MDA8 observations. The R² values for the four models exceeded 0.856, the root mean square error was maintained below 19.48 µg m⁻³, the MPE (mean prediction error) was kept under 14.38 µg m⁻³, and the RPE (relative prediction error) was limited to less than 20.15%. Among the quartet of MDA8 estimation models, those employing S5P NO₂ and CAMS HCHO as well as CAMS NO₂ and CAMS HCHO exhibited superior prediction accuracy, with both R² values surpassing 0.857. Furthermore, the spatial coverage of the estimated MDA8 results from both models was the most extensive, each exceeding 96%. Consequently, the MDA8 estimation models based on these two input schemes were chosen for the ultimate estimation of the MDA8 product. To strengthen the reliability of the MDA8 estimation model, a total of 50 separate training experiments were executed independently. This approach was implemented to ensure a comprehensive validation of the model’s predictive capabilities and to confirm the consistency of its performance across various datasets. Ultimately, 100 independent training iterations of these two models were employed, and the average of the model estimation outcomes was adopted as the definitive daily MDA8 concentration estimation product for the 0.05° × 0.05° grid in the NCP from 2019 to 2021.

4.2. The Relative Importance of Explanatory Variables to O₃ Pollution

Figure 3 presents the relative significance of the explanatory variables for surface ozone in the different surface ozone estimation models. Figure 3 illustrates the comparative importance of the explanatory variables in predicting the surface ozone concentrations across the various models employed for ozone estimation. This representation provides insight into which variables play a more substantial role in the modeling of surface ozone levels. The culmination of the quantitative estimation for the two surface ozone models revealed that VOCs, which include C₃H₈, C₅H₈, CH₄, H₂O₂, OH, and PAN, were the principal contributors to O₃ pollution in the NCP, with the relative importance value for regional MDA8 concentration surpassing 23.34%. These results distinctly indicated that ozone pollution in the NCP was predominantly governed by VOCs. Among the six VOCs identified, C₃H₈ and H₂O₂ emerged as the most substantial contributors, with their relative importance values exceeding 4.26% and 4.08%, respectively. The remaining four VOCs including C₅H₈, OH, PAN, and CH₄ were found to contribute more than 3.96%, 3.88%, 3.57%, and 3.54%, respectively. Furthermore, the relative importance of NO₂ (encompassing both S5P NO₂ and CAMS NO₂) across the four models was found to exceed 7.35%, while that of CAMS-NO_X was determined to be over 2.82%. This indicates that NO_X, which includes NO₂, is the second most critical factor in controlling surface ozone pollution within the NCP. Except for VOCs and NO_X, HCHO, either sourced from S5P or the CAMS, was recognized as a significant contributor to surface ozone pollution within the region. The relative impact of HCHO was found to exceed 5.64%, highlighting its critical role in the formation of ozone at the ground level.

Regarding the meteorological factors, nine variables were considered including T, WS, WD, RH, BLH, UV, TCC, TP, and SP. Within the nine meteorological factors, WD, WS, RH, and T were the four most critical factors influencing O₃ pollution, and their relative importance was more than 3.35%. High temperatures, low wind speeds, and stable atmospheric conditions as well as downwind regions of polluted air masses were recognized as conducive to the photochemical production of ambient O₃, and this has been confirmed by previous studies [30,31,32,33,34]. Moreover, UV, BLH, and TCC were acknowledged as three principal factors affecting O₃ pollution in the NCP. This is attributed to the fact that intense UV radiation provides favorable conditions for the photochemical generation of ambient O₃, while BLH influences the vertical mixing of polluted air masses [1,35]. The quantity of cloud cover impacts the reflection and scattering of light in the atmosphere, subsequently altering the rate of ozone production [36,37,38]. Additionally, TP and SP were also found to play a significant role in O₃ pollution, aligning with the findings of prior studies [39,40,41].

Except for the established influences of ozone precursors and weather elements, the NDVI has also been acknowledged as playing a significant role in ozone pollution. This effect is thought to result from the increased release of bioisoprene from plants, which subsequently leads to higher ozone concentrations [42,43]. Furthermore, the DEM, LULC, and POPU have been identified as influential factors in surface ozone pollution [42,44,45].

Generally, VOCs and NO_X emerged as the predominant factors influencing O₃ pollution in the NCP. The implementation of strategies to reduce VOCs and NO_X emissions was instrumental in markedly decreasing O₃ pollution levels, thus aiding in the control of ozone pollution. The results obtained have proven exceedingly beneficial, serving as a critical reference for regional strategies aimed at mitigating O₃ pollution. Furthermore, they have played a significant role in the improvement of environmental air quality within the area under study.

4.3. Spatiotemporal Distribution of the Estimated MDA8 in the NCP

Figure 4 indicates the results of the spatial distributions of the seasonal averaged and annual averaged MDA8 in the NCP. Among these four seasons for the averaged MDA8, as shown in Figure 4a–d, O₃ pollution was highest in summer, followed by spring, then autumn, and the O₃ pollution was the lowest in winter. This is mainly due to the fact that during summer, the ultraviolet radiation is the strongest and the temperature is the highest, which is suitable for the photochemical reaction for the generation of ozone [46,47,48]. From the perspective of the spatial pattern of O₃ pollution, the most serious O₃ pollution area is in the border area of Hebei, Shandong, and Henan Provinces, that is, the southern part of Hebei, the northern part of Henan, and the western part of Shandong. This is due to the dense population of these areas [15,49,50], the high level of human activities [51,52], and the resulting high anthropogenic emissions, while the strong ultraviolet radiation and high temperature in summer [46,47,48] greatly enhances the photochemical reaction level of O₃, and the ozone pollution level is also the most serious. Furthermore, it was observed that the concentrations of O₃ pollution were comparatively reduced during the spring season when analyzed against those prevalent in the summer period. This is because the temperature in spring is lower than that in summer, while the intensity of ultraviolet radiation in spring is weaker than that in summer. The levels of O₃ pollution in winter and autumn were the lowest among the four seasons due to the lowest temperatures and UV radiation intensity in winter and autumn, which affected the photochemical reaction of ozone. For the annual averaged MDA8, Figure 4e–g shows the spatial distribution patterns and temporal variations in the annual averaged MDA8. The serious O₃ pollution is mainly concentrated in the border area of the NCP, that is, the southeast of Hebei, the north of Henan, and the northeast and west of Shandong, which is consistent with the results shown in Figure 4b. In terms of the temporal variations of annual average MDA8, O₃ pollution in the NCP showed a downward trend from 2019 to 2021, which may be related to the reduction in emissions caused by the impact of COVID-19 in 2020 [53,54].

4.4. The Spatiotemporal Patterns of the Exceedances of the Estimated MDA8

Figure 5 illustrates the spatiotemporal distribution of the MDA8 exceedances as well as the month in which the highest number of MDA8 exceedances occurred. Figure 5a–c revealed the annual cumulative distribution of MDA8 exceedances (MDA8 > 160 µg m⁻³) in the NCP from 2019 to 2021. The annual analysis of ozone pollution reflected in the cumulative count of MDA8 exceedances showed that the overall ozone pollution levels in the region have declined year by year from 2019 to 2021. Notably, from 2019 to 2020, the severity of ozone pollution and its spatial coverage decreased significantly. This reduction may be attributed to the lockdown measures implemented in response to the COVID-19 pandemic during the first half of 2020, which led to a decrease in the intensity of human activities including transportation and industrial operations in the region [53,54]. Consequently, there was a reduction in the emissions of substances that contribute to the formation of ozone pollution. This decrease in precursor emissions played a pivotal role in mitigating the intensity of ozone pollution within the NCP. Moreover, the severity and spatial coverage of ozone pollution in Beijing, Tianjin, and Hebei were found to have been further reduced in 2021 compared to 2020. This could be linked to the resurgence of the pandemic in 2021 [55] as well as the persistent efforts by the government to advance collaborative management and emission reduction policies for PM_2.5 and O₃ pollution. These policies include strengthening emission controls in industrial and transportation sectors and promoting the use of clean energy [30]. During the period spanning from 2019 to 2020, the Xiongan New Area emerged as the locality within the NCP that was predominantly affected by the highest levels of ozone pollution. This high level of pollution may be associated with the intensive urban construction and human activities that took place in the area during that period [56,57,58,59,60].

Figure 5d–f reveals the spatial distribution of the MDA8 exceedances for the months from 2019 to 2021 in the NCP. It was observed that from 2019 to 2021, the majority of months with the highest frequency of ozone pollution in the region were concentrated in June. Therefore, the month of June was predominantly identified as the period with the greatest number of ozone content exceedances in the NCP, primarily associated with the prevalence of high temperatures, intense ultraviolet radiation, and scant rainfall in most areas of Beijing, Tianjin, and Hebei during this month [46,47,48,61,62]. Additionally, the peak levels of ozone pollution along the coastal regions of Hebei, Tianjin, and Shandong in 2019 and 2021 were noted to occur in May. This may be correlated with the climatic conditions specific to these coastal areas. In contrast, the highest levels of ozone pollution in these coastal regions in 2020 appeared in July, potentially due to reduced precipitation, elevated temperatures, and robust ultraviolet radiation in the region during that month [62].

In particular, the month characterized by excessive ozone pollution in the southern region of Tianjin during 2019 and 2021 predominantly occurred in July. This occurrence may have been precipitated by reduced rainfall during this month [62] (Figure S1). In both 2019 and 2021, the most severe ozone pollution zones in Hebei Province and northern Beijing were observed in May. This could also be linked to the levels of ultraviolet radiation and precipitation in these areas [46,62]. Overall, the prevalence of excessive ozone pollution in the NCP was primarily noted during the summer months, aligning with findings from previous research. This pattern may be attributed to the region’s distinct spatiotemporal variations in temperature, ultraviolet radiation, and precipitation [30,40,51,52,63]. The summer season, marked by increased temperatures, ample sunlight, and atmospheric stability, fosters the emission of VOCs and enhances the likelihood of photochemical reactions involving pollutants such as NO_X, leading to the formation of ozone. Consequently, these trends indicate that future pollution control efforts should concentrate on summertime emission reductions to mitigate the incidence and severity of ozone pollution.

Figure 6 presents the annual exceedance rate of ozone pollution for each month in the NCP from 2019 to 2021 as well as the annual cumulative number of ozone exceedances in each province. Figure 6a reveals the monthly exceedance rate of ozone pollution in the region over the three years period and indicates notable interannual variations and seasonal disparities in the ozone pollution levels within the NCP. Based on the monthly patterns of ozone pollution, June was identified as the month with the most pronounced ozone pollution, with an average exceedance rate of 42.65% from 2019 to June 2021. Apart from June, May was the second most polluted month, with an average exceeding rate of 16.90% from 2019 to 2021. Furthermore, September was recognized as the third most polluted month, with a three years average exceedance rate of 15.56%. In addition to these three months, April and July also experienced significant ozone pollution. In general, ozone pollution in China is predominantly observed during the summer months, followed by spring and then autumn. Regarding the annual trends in ozone pollution, a downward trajectory was noted for most months from 2019 to 2021 in the NCP. This decline may be correlated with the COVID-19 pandemic [53,54] and is also attributed to the government’s ongoing implementation of collaborative emission reduction policies for PM_2.5 and O₃ pollution [30,64,65,66]. Among all months, June exhibited a trend of initial increase followed by a decrease in ozone pollution from 2019 to 2021, potentially due to the elevated temperatures in June 2021, which contributed to an upsurge in the net photochemical production of ozone [47,48,61,62]. The main reasons for the ozone peak in June are high temperature and strong solar radiation. These conditions promote the photochemical generation of ozone, resulting in a significant increase in ozone concentration. In order to solve this problem, the following measures can be taken. First, strengthening the control of NO_X and VOCs emissions, implementing stricter industrial emission regulations, and promoting low-emission technologies can significantly reduce the emission of ozone precursors. Second, improving traffic management by optimizing traffic flow, promoting public transportation, and encouraging the use of electric vehicles, which can effectively reduce the emission of ozone precursors from traffic sources. Finally, promoting the use of clean energy over fossil fuels, as shifting to renewable energy sources such as wind, solar, and hydropower can reduce the overall ozone precursor emissions. Through these measures, the ozone peak in June can be effectively alleviated, and the air quality improved. Figure 6b presents the aggregate data on the frequency of ozone pollution exceedances across individual provinces within the NCP as well as in adjacent provinces. O₃ pollution was predominantly concentrated in Hebei, Shandong, and Henan Provinces, aligning with the findings depicted in Figure 4b and Figure 5a–c.

5. Discussion

For O₃ estimation, a variety of methods have been proposed and implemented, but certain limitations persist, primarily concerning spatial resolution, the comprehensiveness of factors, and the precision of estimations. The majority of studies have reported results with a resolution of 0.1° × 0.1° or coarser [67,68], while only a few have produced surface ozone products with a finer resolution of 0.05° × 0.05° [69,70]. Concurrently, the high spatial resolution observations of NO₂ and HCHO from S5P-TROPOMI are seldom utilized in current surface ozone estimation research [69,70]. Therefore, it is essential to utilize the high spatial resolution precursor data (S5P-TROPOMI NO₂ and HCHO) to estimate the surface ozone over large regions, thereby achieving higher accuracy, certainty, and improved spatial resolution in the estimation results. Furthermore, studies employing deterministic models to estimate surface ozone at regional and national scales often produce results for a relatively short period [71], and there is a need to improve the estimation accuracy [72,73]. There is also significant variability in the accuracy of estimations across different regions [74]. This highlights the necessity for more sophisticated models and methods that can address these challenges and enhance the reliability and precision of ozone estimations. In this study, we have addressed these limitations by using high spatial resolution S5P-TROPOMI NO₂ and HCHO data, combined with the XGBoost model, to estimate surface ozone at a finer resolution of 0.05° × 0.05°. This approach not only improves the spatial resolution, but also enhances the accuracy (R²: 0.856) and comprehensiveness of the estimations, distinguishing our work from previous studies. By comparing our results with those obtained from other methods, we demonstrate the advantages of our approach in terms of precision and spatial detail, thereby contributing new insights and advancements to the field of surface ozone estimation.

6. Conclusions

In the scope of this research, ground-based surface ozone measurements from the National Air Quality Monitoring Network of China, recorded between 2019 and 2021 in the NCP, were employed. The observational data were enhanced with multiple data sources that included 7 km × 3.5 km resolution NO₂ and HCHO data from the Sentinel-5 TROPOMI satellite, corresponding data from the CAMS, CAMS VOCs dataset, NO_X emissions data from CAMS, meteorological readings, and data derived from the MODIS, NDVI, DEM, POPU, and LULC details. The integration of these diverse datasets facilitated the development of an extensive model for estimating the surface ozone levels in the NCP. In the NCP, a set of four models for estimating the surface ozone concentrations was crafted, merging data from the Sentinel-5 TROPOMI satellite monitoring system with reanalysis data from the CAMS. This integration aimed to enhance the accuracy of the XGBoost-based models. The R² values, which indicate the site-based verification results for the four MDA8 models, were consistently observed to be above 0.856. Concurrently, the RMSE values for these models were reliably kept under 19.48 µg m⁻³, demonstrating a high level of accuracy in the model predictions. The MPE value was kept under 14.38 µg m⁻³, and the RPE value was limited to less than 20.15%. Given that the coverage of CAMS reanalysis data is more extensive than that of the S5P-TROPOMI satellite monitoring data, the estimation coverage using the combination of CAMS-NO₂ and CAMS-HCHO was the most extensive, followed by the combination of S5P-NO₂ and CAMS-HCHO. The outcomes of the MDA8 estimations, derived from the pair of models in question, were synthesized by taking their average values. This synthesis process led to the creation of the ultimate surface ozone estimation product. Utilizing this method, the daily gridded MDA8 data at a resolution of 0.05° × 0.05° for the NCP, covering the years 2019 to 2021, were deduced accordingly. This comprehensive approach ensured a robust and accurate representation of the spatiotemporal distribution of surface ozone in the region.

Utilizing the surface ozone XGBoost estimation model, a quantitative assessment of the principal factors influencing ozone was conducted in this study. The findings disclosed that VOCs and NO₂, recognized as critical ozone precursors, were responsible for the most substantial contributions to ozone pollution, with their impacts amounting to 23.34% and 7.35%, respectively. Furthermore, the role of HCHO was significant, with its contribution surpassing 5.64%. Meteorological elements such as wind direction, wind speed, temperature, relative humidity, and ultraviolet radiation were also identified as having a notable effect on ozone pollution, each with a contribution greater than 3.31%. The spatiotemporal distribution of the estimated MDA8 concentrations indicated that the convergence area of Hebei, Shandong, and Henan Provinces is the primary region grappling with severe O₃ pollution in the NCP, with summer being pinpointed as the peak season for ozone pollution, particularly in June. Further analysis of the spatiotemporal distribution patterns revealed that severe ozone pollution in the NCP predominantly occurs at the tri-province junction of Hebei, Shandong, and Henan, with June and May being the months with the highest frequency of ozone pollution exceedances. This research has not only furnished essential foundational data for a comprehensive comprehension of ozone pollution, but has also offered robust technical backing for the development of effective strategies aimed at ozone pollution prevention and control.