Unveiling Spatiotemporal Differences and Responsive Mechanisms of Seamless Hourly Ozone in China Using Machine Learning

Abstract

Surface ozone (O₃) is a multifaceted threat that not only deteriorates the environment but also poses risks to human health. Here, we estimated the seamless hourly surface O₃ in China using Extreme Gradient Boosting (XGBoost) with multisource data fusion to investigate spatiotemporal differences in O₃ during multistage COVID-19, and the response of O₃ variation to meteorology and emissions were explored using Shapley Additive Explanations (SHAP) and WRF-Chem. The results indicate that the optimized model demonstrated higher accuracy, with CV-R² of 0.96–0.97 and RMSE of 4.58–5.00 μg/m³. Benefitting from the full coverage of the dataset, the underestimated O₃ was corrected and hotspots of short-term O₃ pollution events were successfully captured. O₃ increased by 16.8% during the lockdown, with high values clustered in the north and west, attributed to the weakened urban NOx titration resulting from reduced emissions. During the control and regulation period, O₃ levels declined year by year. O₃ exhibited significant fluctuations in the Pearl River Delta but remained stable in western China, with both regions demonstrating high sensitivity to meteorological variability. Among these, solar radiation and temperature were the key meteorological factors. The seamless high-resolution O₃ datasets will enable more insightful analyses regarding the spatiotemporal characterization and cause analysis.

1. Introduction

Ozone (O₃) is a potent oxidant that interacts with solar shortwave and outgoing longwave radiation, which can affect climate change [1]. While approximately 90% of O₃ exists in the upper atmosphere or stratosphere, surface O₃ poses significant risks to human health [2], including an increased incidence of respiratory and cardiovascular conditions [3,4]. Since 2013, China has implemented a series of air pollution control measures that have significantly reduced sulfur dioxide (SO₂), nitrogen oxides (NOx), and particulate matter. However, surface O₃ has increased by 20% [5,6]. The daily maximum 8-h average (MDA8) O₃ concentration in urban areas of China during the warm season is accelerating at an alarming rate of 2.4 ppb yr⁻¹ (5%), surpassing the rate of increase in other regions worldwide [7].

The outbreak of the COVID-19 pandemic caused unprecedented social impacts worldwide and led to reductions in air pollutant levels across most regions [8,9,10,11,12]. In order to limit the spread of the virus, China implemented lockdown measures that significantly decreased anthropogenic emissions [13]. During this period, SO₂, NO₂, CO, and PM_2.5 in China decreased by 17.6%, 15.8%, 11.1% and 3.0%, respectively. However, the concentration of surface O₃ increased by 11.4%. The COVID-19 pandemic served as a “natural experiment” occurring during a period of rising O₃ levels, offering a unique opportunity to evaluate O₃ variations and response to emission controls. While several studies have only reported the sharp rise in surface O₃ during the early stage of the pandemic, research on O₃ variations and response mechanisms across the multiple stages of the pandemic remains limited. As a secondary pollutant, O₃ forms through photochemical reactions involving nitrogen oxides (NOx = NO + NO₂) and volatile organic compounds (VOCs) [14,15]. In addition to photochemical reactions, meteorological conditions also alter O₃ levels through various physical and chemical processes [16]. Consequently, the combined effects of complex chemical buffering and varying meteorological conditions may amplify O₃ pollution in certain regions [17,18]. During the multistage COVID-19 pandemic, changing emission patterns and meteorological conditions continuously influenced O₃ [19]. A comprehensive understanding of O₃ variations and response mechanisms during the COVID-19 pandemic can provide a basis for developing effective strategies to mitigate surface O₃ pollution.

Accurate and reliable monitoring data are essential for tackling the persistent surface O₃ pollution. Ground-based observations are the primary method for studying O₃, offering high precision and strong temporal continuity. However, due to the uneven distribution of monitoring stations, remote areas, deserts, and mountainous regions lack sufficient monitoring sites, making them unable to meet the requirements for continuous spatial monitoring [20,21]. To supplement limited ground-based data, satellite remote sensing has emerged as a pivotal tool for O₃ pollution studies, leveraging backscattered radiation from Earth’s atmosphere and surface [22]. Although satellite monitoring provides extensive O₃ distribution, ultraviolet measurements exhibit limited sensitivity to surface O₃, which typically comprises only a minor fraction (a few percent) of the total column [23]. On the other hand, cloud contamination frequently degrades the accuracy of satellite retrievals [24]. These inherent limitations result in notable differences between ground-based and satellite-derived O₃ levels and distribution. Constructing high-resolution, spatiotemporally continuous surface O₃ data remains a major challenge.

To overcome the limitations of single-source observations, previous studies have aimed to establish the connection between satellite column concentrations and surface O₃ through statistical models, such as geographically weighted regression [25] and generalized additive models [26]. Recently, machine learning (ML) has demonstrated key advantages in dealing with complex nonlinear problems. It not only addresses high-dimensional uncertainties but also improves the training effect as sample size increases [27]. Numerous studies have confirmed that algorithms such as decision trees [28], support vector machines [29], random forests [30], and neural networks [31] precisely model the nonlinear relationships between input features and predictions. Among these, the Extreme Gradient Boosting (XGBoost) ensemble algorithm has shown outstanding performance in modeling complex nonlinear relationships among multiple feature variables, exhibiting excellent fitting ability and high training efficiency [32,33]. Numerous studies have demonstrated the predictive capability of existing approaches at daily timescales, but accurate hourly scale O₃ estimation remains insufficiently explored. Moreover, 24-h data are of utmost importance for determining MDA8, analyzing fine spatiotemporal O₃ variations, and the mechanisms of O₃ pollution events.

In this study, we employed XGBoost algorithm to integrate atmospheric composition retrieved from high-resolution satellites, high-temporal reanalysis meteorological data, high-precision ground observation station data, and multisource geographic elements to establish seamless hourly surface O₃ dataset in China during the multistage COVID-19 pandemic from 2019 to 2022. Unlike traditional statistical methods that mostly rely on individual observation data points, our model extends the impact of satellite and meteorological data on O₃ to an hourly scale by performing randomized search each day for hyperparameter tuning, effectively filling in the missing data and enhancing the accuracy of O₃ estimation. Seamless high-resolution O₃ mapping aims to evaluate the spatiotemporal differences in O₃ during the multistage COVID-19 pandemic. Compared to most previous studies that primarily focused on the early phase, this study achieved more comprehensive coverage in both space and time. In addition, the Weather Research and Forecasting-Chemistry (WRF-Chem) model and Shapley Additive Explanations (SHAP) were independently used to quantify the contributions of meteorology and emissions to O₃ formation, offering theoretical support for O₃ pollution control.

2. Materials and Methods

2.1. Study Period

This study is categorized into three periods to analyze the impacts of various emission reduction scenarios on O₃ levels (

Table 1. COVID-19 period name definition.

Definition	Period	Abbreviation
Before COVID-19	20 January 2019–17 February 2019	BC
Lockdown Period	20 January 2020–17 February 2020	LP
Control Period	20 January 2021–17 February 2021	CP
Regulation Period	20 January 2022–17 February 2022	RP

). The first period, from 20 January to 17 February 2020, was identified as the lockdown period (LP), during which China implemented strict lockdown measures, resulting in predominant reductions in anthropogenic emissions. The same period in 2021 was designated as the control period (CP), characterized by relaxed epidemic prevention measures, with a focus on mitigating the rebound of air pollution amid the gradual resumption of economic activities. The corresponding period in 2022 was designated as the regulation period (RP), during which regional control intensity was dynamically adjusted based on the level of COVID-19 risk. A one-month period starting on 20 January was selected as the research timeframe, as it marked the nationwide spread of COVID-19 and the official implementation of national control policies.

2.2. Data Sources

2.2.1. Ground-Based Observations

Surface air pollutant data were available from hourly observations provided by the China National Environmental Monitoring Center (CNEMC). National air quality monitoring commenced in 2013 and expanded to cover 337 prefecture-level cities by 2020. This study used the hourly O₃ and NO₂ from 1619 air quality monitoring stations across various regions of China, with sampling conducted from 2020 to 2022 (Figure 1a). Following the method proposed by Wang et al. [34], we treated observations with values exceeding the upper threshold or negative values as invalid, and excluded data points greater than 700 μg/m³ and less than 0 μg/m³. Second, observations with deviations (observations minus monthly mean value) greater than 100 μg/m³ were also removed. We focused on the spatial distribution of O₃ across the entire territory of China and in typical regions, including the Beijing-Tianjin-Hebei (BTH), Yangtze River Delta (YRD), Pearl River Delta (PRD), Fenwei Plain (FWP), and western China (WC), as illustrated in Figure 1b.

2.2.2. Remote Sensing Measurements

The hyperspectral TROPOspheric Monitoring Instrument (TROPOMI), launched by the European Space Agency (ESA) on 13 October 2017, is the only payload on the Sentinel-5P satellite. The initial imaging resolution was 7 km × 3.5 km [35], but has since been upgraded to the higher resolution of 5.5 km × 3.5 km [36]. TROPOMI features a high signal-to-noise ratio and includes four independent spectrometers, enabling more accurate retrieval of O₃ column concentration at various altitudes in the atmosphere [37,38]. However, extensive cloud cover during winter in high-latitude regions may lead to data gaps in TROPOMI. To address this limitation, we incorporate the OMI-retrieved data as input variables into the ML framework. The two sensors operate on sun-synchronous orbits and have similar local overpass times (approximately 13:30–13:45 local time). ML is a black-box approach capable of integrating information from multiple satellite instruments to improve the spatial representation of O₃ and enable more robust estimations in regions where TROPOMI observations are missing [39,40]. The study used O₃, nitrogen dioxide (NO₂), and formaldehyde (HCHO) column concentrations. To ensure data accuracy, a stringent screening threshold was applied. The extraction boundary of the study area was expanded by 0.3°. The outliers with qa_value < 0.75, zenith_angle < 70° and cloud_fraction_crb > 0.3 were removed, and adjacent swaths were merged [40].

Surface O₃ is produced through nonlinear chemical processes involving reactions of NOx and VOCs. Here, NO₂ is a crucial component of NOx, while HCHO appears as a high-yield intermediate in the photochemical oxidation of VOCs. Therefore, HCHO and NO₂ are essentially considered the tracers of VOCs and NOx [41]. Owing to this, the ratio of satellite-derived column concentrations of HCHO to NO₂ (FNR) can be exploited to identify the formation mechanism of surface O₃ [42,43]. When FNR < 1 represents VOCs-sensitive, conversely, FNR > 2 depicts NOx-sensitive [44]. The FNR value existing between 1 and 2 indicates a transitional state where both VOCs and NOx influence O₃ production [45].

2.2.3. Meteorological and Auxiliary Data

The hourly reanalysis meteorological dataset, comprising 16 O₃-related parameters, was obtained from the European Centre for Medium-Range Weather Forecasts Reanalysis, Version 5 (ERA5). We also gathered auxiliary data with potential impacts on O₃ formation. The elevation data were gathered from the ASTER Global Digital Elevation Model (AGDEM) with a spatial resolution of 30 m. Vegetation and land use influence the air quality; dense vegetation cover typically helps to mitigate pollution, whereas paved surfaces can contribute to warmer temperatures and boost production of secondary pollutants [46]. Land Use and Land Cover Change data (LUCC) were derived from Moderate-resolution Imaging Spectroradiometer (MODIS) satellites, categorizing 17 land types including urban areas, forests, and farmland. The Normalized Difference Vegetation Index (NDVI) was represented by the 1 km MOD13A3 product, which dynamically reflected vegetation distribution and changes across China. Additionally, emissions and economic activities can also influence the pollutant concentrations. The emissions inventory was obtained from the Multiresolution Emission Inventory of China (MEIC). Cloud-free nighttime light data were acquired from the Visible and Infrared Imaging Suite (VIIRS) onboard the Joint Polar-Orbiting Satellite System (JPSS) satellites. Population distribution data developed by the Oak Ridge National Laboratory (ORNL) were also incorporated.

Table 2. Summary of detailed data used in XGBoost model.

Data	Variable	Full Name	Temporal Resolution	Spatial Resolution
Ground-based	O3	Near-surface ozone	hourly	—
Remote sensing	S5P_O3 S5P_NO2 S5P_HCHO	Ozone Nitrogen dioxide Formaldehyde	daily	5.5 km × 3.5 km 0.25°
Meteorology	UVB	Downward UV radiation	hourly	0.25°
	BLH	Planetary boundary layer height	hourly	0.25°
	T2M	2 m temperature	hourly	0.25°
	D2M	2 m dewpoint temperature	hourly	0.25°
	E	Evaporation	hourly	0.25°
	LAI_HV	Leaf area index, high vegetation	hourly	0.25°
	LAI_LV	Leaf area index, low vegetation	hourly	0.25°
	ASN	Snow albedo	hourly	0.25°
	SP	Surface pressure	hourly	0.25°
	SSR	Surface net solar radiation	hourly	0.25°
	STR	Surface net thermal radiation	hourly	0.25°
	ASN	Snow albedo	hourly	0.25°
	TSN	Temperature of snow layer	hourly	0.25°
	TCO3	Total column ozone	hourly	0.25°
	TP	Total precipitation	hourly	0.25°
	U10	10 m eastward wind	hourly	0.25°
	V10	10 m northward wind	hourly	0.25°
Emission	NOx	Nitric oxide	month	0.25°
	VOCs	Volatile organic compounds	month	0.25°
Geography	LON	Longitude	—	—
	LAT	Latitude	—	—
	DEM	Digital elevation model	year	30 m
	LUCC	Land use-cover change	year	500 m
	NDVI	Normalized difference vegetation index	month	1 km
Population economy	POP	Population density	year	1 km
	NL	Nighttime light	year	750 m

illustrates all features used for ML.

2.3. Machine Learning Estimates Surface O₃

2.3.1. XGBoost Model

XGBoost is an ensemble learning algorithm that employs distributed gradient boosting decision trees. Each decision tree was trained on the residual of the preceding tree. New trees were introduced to minimize the loss function, thereby gradually reducing the error. Leaf node predictions and tree complexity were tuned to optimize model performance. The XGBoost model incorporates model complexity along with the traditional loss function to assess the model’s generalization capability and computational efficiency, as follows:

$${Obj}^{\left(t\right)}=\sum _{i=1}^{n}L\left(y_i,\mathrm{ŷ}\right)+\sum _{k=1}^K\mathsf{\Omega }(\left.f_k\right)$$

(1)

where $K$ is the number of trees, $f_k$ is a function in the functional space, and $\mathsf{\Omega }(\left.f_k\right)$ is the complexity of the tree $f_k$. During each iteration, the optimal $\mathrm{ŷ}$ is obtained by minimizing ${Obj}^{\left(t\right)}$. A regularization term is included to prevent overfitting:

$${Obj}^{\left(t\right)}=\sum _{i=1}^m\left[f_t\left(x_i\right)g_i+{{\displaystyle \frac{1}{2}}\left.\left(f_t\right(x_i)\right)}^2{h}_i\right]+\mathsf{\Omega }\left(f_t\right)$$

(2)

$$\mathsf{\Omega }\left(f_t\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda \sum _{j=1}^T{\omega }_j^2$$

(3)

We calculated the score of each leaf node and determined the optimal tree structure. The score accounted for both impurity measures in the decision tree and the model complexity:

$$Obj=\gamma T-{\displaystyle \frac{1}{2}}\sum _{j=1}^T{\displaystyle \frac{G_j^2}{H_j+\lambda }}$$

(4)

2.3.2. Model Development and Validation

The XGBoost model was constructed to estimate surface O₃. Ground-based hourly O₃ observations were used as the target variable, whereas the remaining variables served as input variables. All data were resampled to the same spatial resolution of 0.1° × 0.1° grid using the bilinear interpolation approach [47]. By performing spatiotemporal matching, we addressed spatial mismatch among datasets at varying longitudes and latitudes.

To optimize model performance and eliminate redundant input, recursive feature elimination (RFE) was applied for variable selection (Figure S1). This process reduced redundancy from irrelevant features while preserving variables critical to prediction accuracy. Based on the results of feature screening, the dataset was randomly divided into 70% for model training and 30% for testing. Generally, meteorological variables are the most important, but the ranking of variable importance results is different under different conditions. To account for this variability while maintaining a sufficient sample size, we found that adjusting the model parameters on a daily basis could obtain the best estimate of hourly O₃ on that day. Therefore, we implemented an intensive optimization process that synergistically adjusted multiple hyperparameters. A ten-fold cross-validation (10-CV) combined with randomized grid search was employed to construct a model with the optimal parameter combination each day. These adjusted parameters were used for training the XGBoost model. We did not use a unified model because coarse-resolution modeling approaches may incur information loss when predicting fine temporal resolution data, thereby limiting the accurate characterization of hourly O₃ dynamics.

We divided the dataset into training and validation sets at a nine-to-one ratio for each training iteration. The mapping relationship was validated through the training set, whereas precision of the trained model was evaluated using the validation set [48]. Statistical metrics, including the coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE), are as defined by Equations (6)–(9). $n$ represents the samples. $x_i$ and $y_i$ denote as the model observed values and predicted values, respectively.

$$R^2={\displaystyle \frac{{\left[\sum _{i=1}^n\left(x_i-x̄\right)\left(y_i-ȳ\right)\right]}^2}{\sum _{i=1}^n{\left(x_i-x̄\right)}^2\sum _{i=1}^n{\left(y_i-ȳ\right)}^2}}$$

(5)

$$MAE={\displaystyle \frac{\sum _{i=1}^n\left|x_i-y_i\right|}{n}}$$

(6)

$$MSE={\displaystyle \frac{\sum _{i=1}^n{\left(x_i-y_i\right)}^2}{n}}$$

(7)

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

(8)

2.3.3. Interpretability Analysis of Model

The model decision-making mechanism was investigated using SHAP to ascertain global interpretability, quantifying the contribution of each feature to the target variable. SHAP characterizes the complex relationships between input features and model outputs by computing the positive or negative contributions of each feature, providing insights into the decision-making process of the tree structure model. The formula is as follows:

$${\varphi }_i=\sum _{S\subseteq N\backslash \{\left.i\right\}}{\displaystyle \frac{\left|S\right|!\left(\right|N|-|S|-1)!}{\left|N\right|!}}\left[\nu \left(S\cup \left\{i\right\}\right)-\nu \left(S\right)\right]$$

(9)

where $S\subseteq N/\{\left.i\right\}$ represents an iteration over all possible subsets ($S$) of $N$ that exclude feature $i$ to calculate the prediction difference of feature $i$ after adding subset ($S$); $\nu \left(S\cup \left\{i\right\}\right)$ and $\nu \left(S\right)$ represent the output of the model for different combinations of features when the feature $i$ is present and absent. According to the SHAP value, red indicates feature value has a positive impact on the target variable, whereas blue signifies a negative impact. The proximity of the SHAP value to zero reflects the feature contribution to the prediction: values closer to zero indicate a smaller contribution, while values further from zero denote a greater influence on the prediction target.

2.4. WRF-Chem Simulation

The WRF-Chem ver.3.9.1 is an air quality model developed by the National Oceanic and Atmospheric Administration Forecast System Laboratory, which couples the chemical transport module (Chem) with the meteorological model (WRF) to reflect the real atmospheric conditions [49]. The model is configured with domains covering China with 34 vertical layers up to 50 hPa. For initial boundary meteorological conditions, we utilized the Final Reanalysis Data (FNL) global reanalysis data provided by the National Center for Environmental Prediction (NCEP), which featured a temporal resolution of 6 h and a spatial resolution of 1° × 1°. The model physics configuration included the Lin microphysics scheme [50], the Rapid Radiative Transfer Model (RRTM) radiation scheme [51], the Goddard Shortwave Radiation scheme [52], the Noah Land-Surface scheme [53], the Mellor-Yamada-Janjic TKE planetary boundary layer parameterization [54], and the Xu-Randall cloud effect [55]. The gas-phase chemical scheme integrated the RADM2 chemical mechanism, and the MADE/SORGAM aerosol coupled scheme [56]. Biogenic emissions were estimated from the Model for Emissions of Gases and Aerosols from Nature (MEGAN) [57], including monoterpenes, isoprene, and soil nitrogen. The anthropogenic emissions were based on the Multiresolution Emission Inventory for China (MEIC) at a spatial resolution of 0.25° × 0.25° on a monthly basis [58]. Further details are summarized in Table S1.

Sensitivity simulations were designed to quantify the respective impacts of meteorological conditions and anthropogenic emissions on surface O₃ (

Table 3. Experimental configurations for WRF-Chem modeling system.

Simulation	Meteorology	Emission
Base	BC	BC
Base_LP	LP	LP
EXP_LP	LP	BC
Base_CP	CP	CP
EXP_CP	CP	BC
Base_RP	RP	RP
EXP_RP	RP	BC

). Considering the uncertainty in the emission inventory, we set up the experiment by fixing emissions while changing the meteorological conditions. The meteorology and emissions were fixed in Base from 20 January to 17 February 2019, representing the situation before the epidemic. Similarly, meteorology and emissions for Base_LP, Base_CP, and Base_RP were static for 2020, 2021, and 2022, respectively. Differences between each experiment and the base reflected the combined effects of meteorology and emissions on O₃ formation. By comparing time periods of equal length, the effects of short-term meteorological anomalies can be minimized, thereby reducing potential impacts related to the Spring Festival [59]. For the comparative experiments EXP_LP, EXP_CP, and EXP_RP, emissions were fixed to 2019 levels, whereas the meteorology was synchronized to the corresponding years. The difference between Base_LP and EXP_LP quantified the emission-driven changes in O₃ during the LP. The same approach was applied to CP and RP. Additionally, the difference between each EXP and the base in the corresponding year represented the meteorological impacts on O₃ variations.

3. Results

3.1. Performance Analysis of Model

First, feature variables were selected using RFE. In each step of the RFE, an XGBoost model was built to rank input variables; the least important variable was removed and a new model base was created on this reduced set of input variables. We then determined the optimal set of features that significantly improved model performance. The variables finally used for model training and their importance rankings are shown in Figure 2a–c. The results revealed that downward UV radiation (UVB) and surface net solar radiation (SSR) were the most dominant variables for estimating surface O₃, and scored from 0.08 to 0.12 and 0.08 to 0.16, respectively. Moreover, Satellite NO₂ and O₃ products play crucial roles in estimating surface O₃, while total column ozone (TCO3), 2 m temperature (T2M), 2 m dewpoint temperature (D2M), and surface pressure (SP) also contribute to O₃ formation. Unlike other periods, thermal radiation (STR) and snow layer temperature (TSN) exhibited higher importance scores during the LP period, accounting for 3.81% and 3.60%, respectively. In addition, VOCs and NOx were the key features in the CP and RP.

We first ensured a sufficient number of data samples in constructing the model. For each stage, the number of samples (N) exceeded 90,000, with more than 3000 valid samples. The only exception was 22 February 2020, with N = 2715 after outlier removal. In total, over 640,000 sample instances were included across ten-fold. To better present the results, the outcomes were summarized by period. The overall performance of the model was validated using sample-based 10-CV (Figure 2d–f). The results indicated that surface O₃ estimates derived from the XGBoost model were highly consistent with O₃ observations across China, and were generally concentrated near the 1:1 tangent line. The average CV-R² values during the LP, CP, and RP were 0.96, 0.97, and 0.96, respectively. The mean MAE, MSE, and RMSE ranged from 2.93 to 3.19 μg/m³, 21.00 to 25.02 μg/m³, and 4.58 to 5.00 μg/m³, respectively. The high R² and low error revealed that the model exhibited superior performance in estimating highly nonlinear hourly O₃.

The model performance for the blind test was also evaluated (Figure 2g–i). The blind test results demonstrated high accuracy across the entire domain, with CV-R² ranging from 0.95 to 0.96. The mean MAE, MSE, and RMSE ranged from 3.10–4.03 μg/m³, 20.98–30.10 μg/m³, and 4.58–5.49 μg/m³, respectively. These values were comparable to or slightly lower than those from the 10-CV on the 70% training dataset, indicating a strong generalization capability of the model. This superior efficacy can be attributed to an intensive optimization process, during which multiple hyperparameters were jointly tuned. This coordinated strategy significantly improved the overall accuracy of hourly O₃ estimation. It is noteworthy that the statistical values across the three stages showed minimal differences, further indicating that the model is robust and reliable. Its performance was not affected by spatiotemporal heterogeneity, meeting the requirements for accurate hourly surface O₃ estimation.

Figure 3 illustrates the temporal performance in China and typical regions. All O₃ estimates were calculated using collocated ground monitoring data and validated against corresponding station observations. The results indicate that the O₃ estimates in China show good agreement with observations, with an R² of 0.89, MAE of 3.78 μg/m³, MSE of 21.34 μg/m³, and RMSE of 11.28 μg/m³. The differences between observed and estimated O₃ are generally small. From a regional perspective, the model effectively captured O₃ variations across five typical regions, demonstrating strong predictive powers (R² = 0.85–0.93, RMSE = 4.07–6.61 μg/m³). The model performs best in regions with denser monitoring networks, such as the BTH and YRD, with most stations having high R² above 0.9 and RMSE below 5 μg/m³. By contrast, the model performs relatively less consistently in the PRD, WC, and FWP, with R² values of 0.88, 0.88, and 0.85, respectively. This may be attributed to greater uncertainties due to complex natural and meteorological influences [60]. The model estimates are highly consistent with observed temporal trends and effectively reproduce abnormal fluctuations in O₃ levels, including sudden spikes or significant drops in the urban agglomeration (BTH, YRD and PRD). In relatively stable western regions, the model effectively captures the diurnal variation of O₃, thus the consistently low estimation errors across all stages. Overall, the model exhibited strong performance and successfully captured the dynamic variations of surface O₃, providing reliable data to support the monitoring of spatiotemporal trends.

3.2. Spatiotemporal Differences of XGBoost Model Surface O₃ Estimation

Figure 3 also illustrates the time series of surface O₃ for China and typical regions. The model exhibited strong responsiveness to short-term O₃ pollution events. During the LP, surface O₃ levels across China showed an overall increasing trend, ranging from 47.25 to 73.89 μg/m³. O₃ levels in BTH, YRD, and FWP exhibited fluctuating increases, whereas the PRD experienced exceptionally large variations. In contrast, the WC remained relatively stable. In February 2021, northern China experienced a sharp rise in O₃ concentration, whereas a steep drop of 48.36 μg/m³ (147.8%) occurred in PRD from 1st February to 2nd. Unlike other regions, the PRD exhibited recurrent downward trends in O₃ during the RP. Meanwhile, the WC demonstrated no notable variations, suggesting that emissions played a limited role in the region, while geographic and meteorological factors may have exerted a greater influence.

Based on the hourly surface O₃ dataset generated by fusing multisource data with the XGBoost model, the spatial distribution of O₃ at different stages of the COVID-19 period was mapped. Figure 4 displays the spatial consistency between observed O₃ and estimated O₃. The simulation results demonstrate that areas with high O₃ levels were predominantly located in the north and WC, whereas southern China exhibited relatively lower O₃ levels during the LP. Compared to the LP, the O₃ decreased by 5.41% (3.21 μg/m³) as shown in the CP, primarily in WC such as Xinjiang, Tibet, and Qinghai. In contrast, O₃ levels increased by 10–15 μg/m³ in Guangxi and Guangdong. Under regulatory measures in 2022, O₃ decreased by 4.67% (2.79 μg/m³), primarily in eastern and southern China, with an average concentration of 59.79 μg/m³. However, significant increases in O₃ levels were observed in Tibet and Qinghai. A notable O₃ hotspot was found at the junction of northwest Yunnan and Sichuan, potentially influenced by stratospheric transport [61]. Due to the sparse distribution of monitoring stations, ground observations substantially underestimated O₃ in WC, particularly during the CP. With the complete spatial coverage provided by the model, the estimated results revealed an average increase of 6.82 μg/m³ in O₃ levels in WC. The WRF-Chem simulation results further validated the reliability of the estimated spatial distribution of surface O₃ (Figure S2).

Figure 5 illustrates the fractional changes (DIFF) in surface O₃ estimates across Chinese provinces between different stages of COVID-19 and BC. Compared to BC, surface O₃ increased remarkably by 16.8% during the LP, predominantly in eastern (8.17 μg/m³) and central China (11.41 μg/m³) with DIFF of 22.64% and 23.84%, respectively. The northern region experienced greater changes than the southern region, with increases exceeding 40% in Hebei, Qinghai, Inner Mongolia, and Heilongjiang. These variations are considerably due to differences in NOx and VOCs ratios, changes in meteorology, and interactions with other pollutants [62]. Xinjiang was the only province exhibiting a decrease in O₃ (−3.24%), which is likely attributed to an underestimation resulting from limited ground-based observations. During the CP, O₃ increased across several southern provinces, with the highest rise of 35.17% in Chongqing. Although the national average O₃ level was 4.21 μg/m³ higher than that in 2019, decreasing trends were observed in the FWP, central, and southern China. In contrast, O₃ in WC increased by 14.38%, with Gansu DIFF increased by 43.3%. This increase was mainly attributed to olefin emissions from the petrochemical industry, which significantly contribute to elevated O₃ levels [63].

3.3. Spatial Analysis of the Short-Term Severe O₃ Pollution Event

We also focused on typical cases with daily average O₃ concentration above 65 μg/m³, while extreme values surpassed 100 μg/m³. Figure 6 illustrates the substantial daily variability in O₃ distribution. Short-term O₃ pollution events were mostly concentrated in February and exhibited uneven spatial distribution. High O₃ value areas were distributed in the north and WC, with comparatively lower concentrations in the south. Wintertime coal heating in northern China sustained NOx saturation, whereas anthropogenic emissions in the south remained lower than in the north, even under strict lockdown policies. This regional disparity highlights the impact of heating practices on air pollution patterns [64]. However, abnormal variations may occur during certain periods. For instance, O₃ hotspots formed in northeast China on 12 February 2020, contrasting with a sharp decline in O₃ levels in southern China. Furthermore, daily O₃ in eastern and central China showed minimal variation in February 2022, while the WC exhibited noticeably high O₃ pollution levels (>120 μg/m³). The highest O₃ level in WC was recorded on 12 February 2022.

3.4. Spatial Disparity Caused by Meteorology and Emission

Surface O₃ formation depends on the influence of photochemical processes including precursors and meteorological conditions, and is generally divided into two distinct chemical properties: VOCs-sensitive and NOx-sensitive. We used the satellite-derived indicator FNR to characterize the sensitivity of surface O₃ production. Species with similar lifetimes in the FNR (on the order of hours for NO₂ and HCHO) better represent the interactions between the catalytic cycles that result in O₃ production [65].

TROPOMI satellite observed NO₂ of 0.03 mol m⁻³ during the LP, detailing a 30–40% decrease in contrast to the same period in 2019. More than 90% of regions were NOx-sensitive during this period, and urban areas exhibited VOCs-sensitive and transitional states (Figure 7). The reduction in NOx emissions decreased the NO₂ available for photolysis, thereby decreasing the amount of free oxygen atoms available to form O₃ [43]. Consequently, NO₂ reduced substantially in areas dominated by urban agglomerations (e.g., BTH, YRD, PRD), resulting in a transition from NOx-limited to VOC-limited O₃ generation sensitivity. When NOx reached a saturated state, the lockdown-induced reduction in NOx emissions weakened O₃ depletion. Meanwhile, HCHO levels exhibited a uniform distribution, depicting minimal variations. As productivity approached normal in 2022, VOCs restrictions gradually shifted toward a transition state. Adjusting either a single emission or simultaneously reducing both precursors can effectively reduce O₃ production in urban areas.

Based on the sensitivity simulations performed by WRF-Chem, the contribution of meteorological conditions and anthropogenic emissions to O₃ levels in China is analyzed (Figure 8). Compared to previous model evaluations, the simulated values in this study showed a good correlation with observed values (R² = 0.68–0.84, Table S2) and were suitable for further analysis (Table S3) [15,66,67,68]. During the LP, the combined contribution of meteorology and emissions to surface O₃ increased by 4.06 μg/m³ compared to BC. Emissions from urban agglomerations led to an O₃ increase of 9.06–15.96 μg/m³, which was partially offset by meteorological conditions (0.59–2.69 μg/m³). In contrast, meteorological contributions in FWP and WC contributed 5.70 μg/m³ and 1.57 μg/m³, respectively, which further exacerbated O₃ pollution. It is noteworthy that O₃ levels in the PRD increased by 71.7% during the CP, with contributions of 9.27 μg/m³ and 8.36 μg/m³ from emissions and meteorology, respectively (

Table 4. The impact of changes in O₃ (Total), meteorological conditions (Met), and anthropogenic emissions (Emi) on O₃ levels over China and five representative regions.

Simulation	Lockdown			Control			Regulation
	Total	Emi	Met	Total	Emi	Met	Total	Emi	Met
China	4.06	3.67	0.39	3.68	2.72	0.96	2.55	4.04	−1.49
BTH	13.32	15.96	−2.64	9.43	13.25	−3.80	13.01	15.57	−2.56
YRD	8.47	9.06	−0.59	5.94	5.81	0.13	4.66	8.74	−4.08
PRD	10.27	12.96	−2.69	17.63	9.27	8.36	4.63	13.46	−8.83
FWP	8.12	5.70	2.42	3.98	3.96	0.02	5.44	9.41	−3.97
WC	2.55	1.57	0.98	2.42	2.19	0.23	1.23	1.88	−0.65

). Differently in other periods, both emissions and meteorological conditions contributed to an increase in O₃. At this time, O₃ sensitivity in the Pearl River Delta (PRD) was primarily VOC-limited and transitioning. This suggests that the unexpected rise in O₃ levels in the PRD was not solely due to the disproportionate reduction in NOx and VOCs emissions but was also influenced by unfavorable meteorological conditions. During the RP, O₃ pollution showed improvement across most regions, which can be attributed to reduced NOx emissions under transitional conditions, coupled with relatively stable VOC emission levels. However, anthropogenic emissions contributed significantly to O₃ levels in the BTH (15.57 μg/m³). Based on the spatial distribution of O₃ estimated by XGBoost, we found significant differences in O₃ levels between WC and the South China region during the same period. During the LP and RP, emissions promoted O₃ formation in South China, but favorable meteorological conditions effectively offset the increases caused by emissions. Conversely, meteorological conditions contributed to elevated O₃ levels in WC throughout the COVID-19 pandemic, while emission reductions partially offset the concentration rise.

To understand the decision-making mechanism of the model, an interpretability analysis of variables was performed based on the SHAP algorithm. The SHAP value provides reasonable quantification of input variable contribution to O₃ change, and explains the direction a feature contributes to a specific prediction [69]. During the COVID-19 period, solar radiation (UVB and SSR) ranked first in both the importance ranking and SHAP values, indicating that such features were highly important and had a significant positive impact on O₃ (Figure 9). As solar radiation increases, O₃ also rises accordingly. Despite weaker solar radiation in winter, it still serves as a main driving force for photochemical reactions in the atmosphere [23]. Temperature has a negative impact on O₃, which is attributed to the weak atmospheric instability and temperature inversions in winter that restrict vertical air mixing, leading to the accumulation of O₃ in the lower atmosphere. Furthermore, under the interaction of lower temperatures, high humidity inhibits surface O₃ depletion [70]. NOx received a relatively low importance score in the model but showed a high SHAP value, indicating that although it was not frequently used during model training, it had a significant negative effect on O₃, with contributions of −3.77 and −4.44 during the CP and RP, respectively. During the LP, NOx and TCO3 also exhibited high SHAP values. NOx emissions have dropped significantly under the background of low human activities, which has weakened the titration effect and led to the rapid accumulation of O₃. High O₃ concentrations have enhanced the oxidation of the atmospheric environment and intensified other photochemical reactions [71]. VOCs emerged as an important variable during the RP. O₃ production takes place by photochemical oxidation of VOCs in the presence of NOx, catalyzed by hydroxyl radicals [72]. However, the SHAP values for NOx were notably higher than those for VOCs across all stages, implying that O₃ formation was highly sensitive to NOx.

4. Discussion

4.1. Comparison with Other O₃ Modeling

This study establishes an hourly seamless dataset of surface O₃ in China based on machine learning and multisource data. The results indicate that the model achieves high accuracy and effectively fills data gaps. Compared to satellite and ground-based monitoring products, the hourly surface O₃ maps generated using the XGBoost model exhibit superior spatial coverage. A comparison between O₃ model performance reported in previous studies and that of our model is provided in

Table 5. Comparison with previous similar studies.

Study Case	Modeling	Regions	Resolution	R2
Wei et al. [23]	STET	China	daily	0.87
Kang et al. [73]	XGBoost	East Asia	daily	0.65–0.78
Xiong et al. [74]	LightGBM	China	daily	0.88
Capilla [75]	Neural network	Iberian Peninsula	hourly	0.92–0.95
Yang et al. [76]	MORF	China	hourly	0.94
Chen et al. [77]	Deep learning	China	hourly	0.86–0.94
Gao et al. [78]	LightGBM	China	hourly	0.95
Li et al. [79]	LightGBM	China	hourly	0.94
Xue et al. [80]	RF	BTH	hourly	0.94
This study	XGBoost	China	hourly	0.96–0.97

Note: STET means extremely randomized trees; LightGBM means light gradient boosting machine; MORF means multioutput random forest; RF means random forest.

. Most previous studies have focused on daily O₃ estimation using ML, but reported R² values generally below 0.90. With advances in technology and computational capabilities, recent studies have improved the temporal resolution of O₃ estimation to the hourly scale. Yang et al. constructed a multioutput random forest model and achieved estimation of hourly O₃ in China, with sample-based CV-R2 reaching 0.94. Gao et al. showed hourly ML-derived O₃ with R² values exceeding 0.95 and RMSE below 7.50 μg/m³. Li et al. utilized a geostationary satellite to retrieve hourly O₃, achieving high accuracy (R² = 0.94). In contrast, our model achieves the sample-based CV-R² values of 0.96–0.97 at the hourly scale, which demonstrates that the model has superior performance. Furthermore, the blind test results yielded R² values of 0.95–0.96, further explaining the robustness of the model. Overall, our model demonstrated superior estimation ability than that yielded by previous methods in terms of the various temporal scales.

4.2. Differences in O₃ Levels During Multistage COVID-19

Benefitting from the high-resolution comprehensive coverage of the surface O₃ dataset, this study provides an evaluation of O₃ differences during the multistage COVID-19 pandemic, whereas previous research has mainly focused on the early stage of the outbreak. We found that significant variations in O₃ distribution were observed at different stages, and short-term O₃ pollution events exhibited marked spatiotemporal heterogeneity on a daily scale. No abnormal diurnal patterns or intensification of daily O₃ cycles were detected during pollution events (Figure S3), and thus no further analysis was conducted. Nevertheless, the result can demonstrate that the estimated hourly O₃ is basically consistent with the observed O₃.

In addition, we quantify the contribution of winter meteorology and emissions to O₃. WRF-Chem aims to separate the effects of meteorology and emissions, while SHAP refines the contribution levels of specific factors. WRF-Chem reveals the complexity of O₃ generation in different regions. The emission reductions after the lockdown weakened the urban NOx titration, resulting in increased O₃ levels. Unfavorable meteorological conditions offset the benefits of emission reductions in FWP and WC. The contribution of emissions to O₃ declined during the CP, and more urban areas shifted from NOx restrictions to VOC restriction. The rise of urban O₃ is attributed to increased VOCs and reduced NOx emissions under VOC-limited conditions. Meanwhile, the reduction in NOx and the relative stability of VOCs under the transitional state slowed down the O₃ increases during the RP. Moreover, O₃ shows strong sensitivity to meteorological changes in PRD and WC.

4.3. Uncertainty Analysis

One limitation of this study is that the use of high-resolution modeling and rigorous evaluation methods for estimating hourly O₃ has increased the computational cost. As a result, the surface O₃ dataset was constructed only for the COVID-19 period. Although the study spans three years, the temporal coverage remains relatively limited. In addition, we obtained good prediction results for a limited number of stations in WC, indicating that the model has strong spatial prediction capabilities. However, the accuracy of estimates in regions lacking monitoring stations remains uncertain. To validate the accuracy of O₃ estimates in data-sparse regions with limited monitoring stations, acquiring additional ground-based observations remains crucial for model refinement and evaluation. Future studies could extend the model to longer time periods and incorporate additional observational data from the WC for validation.

5. Conclusions

This study used the XGBoost algorithm to fuse multisource data and establish an hourly full coverage surface O₃ dataset over China. The results indicate that the optimized model demonstrated higher accuracy, with CV-R² of 0.96–0.97 and RMSE of 4.58–5.00 μg/m³. In addition, the blind test yielded CV-R² of 0.95–0.96 and RMSE of 20.98–30.10 μg/m³, further confirming the model’s robustness and reliability. In terms of temporal performance and spatial prediction capability, the results in China and typical regions further highlight the superior accuracy and reliability of our model (R² = 0.85–0.93; RMSE = 4.07–6.61 μg/m³). Moreover, the model effectively fills spatial data gaps and corrects O₃ underestimation. O₃ increased by 16.8% during the LP compared to 2019, with high concentrations mainly distributed in the north and west, while lower levels were concentrated in the south. Additionally, O₃ levels decreased by 5.41% during the CP and declined further under the regulatory measures implemented in 2022. Throughout the COVID-19 period, solar radiation significantly influenced O₃ levels, and NOx also showed a strong association with O₃, particularly during the CP and RP. Meteorological changes can either aggravate or alleviate O₃ levels, but anthropogenic emissions play a crucial role in O₃ formation. It is essential to simultaneously regulate the emission reduction ratios of O₃ precursors while considering the meteorological conditions in PRD and WC.