Reconstructing high-quality ground-level ozone records from 1980 to 2012 in central and eastern China Authors: Jinghui Ma, Yihong Wei, Zhongqi Yu, Xiaoyi Wang Published: 2025 ABSTRACT Surface ozone (O3) substantially and adversely impacts public health and ecosystems. Despite China’s air quality improvements, pre-2013 ozone records in central-eastern China (CEC) remain scarce. To address this temporal gap, we constructed a 0.5° resolution daily maximum 8-h average (MDA8) ozone concentration dataset spanning 1980–2012 by employing a light gradient boosting machine (LightGBM, LGBM) model based on multisource datasets. Tenfold cross-validation (2013–2023) demonstrated robust spatiotemporal reconstruction capability, with 88% of grids exhibiting correlation coefficients (R2) > 0.8 versus observations (42% exceeding 0.9) and root mean square errors (RMSE) of 7.5–7.9 μg⋅m−3. The reconstructed climatology revealed a regional mean of 87.6 μg⋅m−3 with a north-south gradient and significant annual increase (0.14 μg⋅m−3⋅yr−1, p < 0.01). μg⋅m−3 Mechanistic attribution identified synergistic forcing from global warming (0.32 °C decade−1), industrial expansion, and urbanization-induced emission growth as primary drivers of long-term ozone elevation. The LightGBM-reconstructed dataset exhibits exceptional stability (interannual variability < 5%), providing critical baseline data for quantifying multidecadal ozone pollution–climate interactions and informing regional air quality management strategies. Background & Summary Ozone (O3), a triatomic oxygen molecule, exhibits a critical atmospheric dichotomy: stratospheric ozone (20–30-km altitude) absorbs 97%–99% of ultraviolet-C radiation through the Chapman catalytic cycle, thereby preventing DNA damage and maintaining terrestrial ecosystem stability1. In contrast, tropospheric ozone (<10-km altitude) forms via the photochemical oxidation of anthropogenic precursors, primarily nitrogen oxides (NOx) and volatile organic compounds (VOCs), under solar irradiation2,3. Epidemiological meta-analyses have demonstrated that chronic exposure to ground-level ozone (>60 ppbv) increases respiratory mortality risk, with hazard ratios of 1.18 (95% CI: 1.12–1.24) per 10-ppbv increment in warm-season concentrations4,5. Phytotoxic effects are also concerning: ozone disrupts stomatal conductance in C3 crops, reducing wheat yields by 12.3% ± 2.1% and rice productivity by 8.7% ± 1.9% at 70-ppbv exposure relative to preindustrial baselines6. Central and East China (CEC: 20–43°N, 100–123°E), encompassing the North China Plain (NCP), Yangtze River Delta (YRD), and Central China (CC) urban agglomerations, exemplifies the ozone–climate development paradox. This region hosts 412 million inhabitants (28% of China’s population) and generates $5.8 trillion in annual GDP (35% of national output)7; however, the region faces compounding air quality challenges. Post-2000 industrialization has triggered a significant increase in anthropogenic NOx emissions, with a 210% rise from 8.2 Tg yr−1 in 2000 to 25.4 Tg yr−1 in 20208,9. Synergizing with VOC emissions from petrochemical clusters in the Shandong, Jiangsu, and Zhejiang provinces10, these precursor escalations have driven a notable increase in regional MDA8 ozone levels during 2015–202211,12. Tropical cyclones and continental anticyclones are the primary meteorological systems associated with ozone events. Anticyclones, or high-pressure systems, create conditions conducive to pollutant accumulation and ozone production, which are characterized by clear weather and low wind speeds at their centers13,14. Ozone pollution typically occurs on days with intense sunlight and minimal wind, because these conditions facilitate the photochemical reactions of ozone and the accumulation of ozone and its precursors. The wind direction is also crucial because it affects pollutant transport and can lead to significant ozone production in downwind areas15. Quantifying the spatiotemporal heterogeneity of tropospheric ozone concentrations is a prerequisite for elucidating the formation mechanisms and informing evidence-based air quality governance. Conventional methodologies, including ground-based monitoring networks and chemical transport models (CTMs), suffer from critical observational and computational constraints that impede long-term trend analysis. China’s National Ambient Air Quality Monitoring Network, which was standardized after 201316, exhibits pronounced spatial inequities: urban zones average 1.7 ± 0.3 stations per 103 km2 versus 0.14 ± 0.05 in rural areas17,18, with < 30% of stations measuring ozone precursors (NOₓ, VOCs) essential for attribution studies. CTMs such as GEOS-Chem (v13.3.4) face prohibitive computational demands, requiring > 1.2 × 106 CPU-core hours to simulate decadal ozone variability at 0.5 × 0.625° resolution19. The lack of standardized ground-based ozone observation records significantly limits research on ozone climatology and its health effects. Machine learning (ML) has emerged as a transformative paradigm for reconstructing historical ozone climatologies, overcoming the limitations inherent to conventional geostatistical and process-based approaches. Conventional geostatistical interpolation techniques (e.g., ordinary Kriging and inverse distance weighting) inadequately capture the nonlinear couplings between ozone concentrations and predictor variables, such as meteorological drivers (boundary layer height, solar radiation), precursor emissions (NOₓ, VOCs), and aerosol interactions, exhibiting poor predictive performance (cross-validated coefficient of determination (R2) < 0.618,20. In contrast, ensemble ML architectures such as LightGBM (LGBM) can resolve high-dimensional interactions and achieve superior accuracy (root mean square error, RMSE: 6.2–7.8 μg⋅m−3; mean absolute error, MAE < 5.1 μg⋅m−3) across diverse biogeochemical regimes21. Despite global progress, as exemplified by Kumar et al.22 (2024)’s contiguous United States (CONUS)-wide ozone reanalysis, East China remains critically understudied. Existing ML-based reconstructions suffer from temporal truncation (post-2000), spatial coarseness (1-km resolution18), and inadequate resolution of urban–rural ozone gradients23, which are essential for exposure assessment. In recent years, there has been a significant decline in particulate matter (PM) pollution owing to the implementation of environmental protection and control strategies24–28. In contrast, surface ozone pollution has increased in China24,29, raising significant public health concerns30. Compared with PM, research on surface ozone in China remains relatively underdeveloped, particularly surface ozone climatology research. Consequently, this study employed the concept of ML to address these challenges, with a focus on spatiotemporal variations in O3 pollution. We extended a space-time LGBM model to derive daily spatially continuous 0.5° × 0.5° gridded maximum daily 8-h average (MDA8) ozone reanalysis data (1980–2012) for CEC using an optimized LightGBM framework. By reconciling observational scarcity with mechanistic interpretability, this reconstruction provides an empirical foundation for transboundary air quality governance and climate–pollution comanagement strategies. Methods Study area As shown in Fig. 1, this study focuses on the five most popular regions in China: Beijing–Tianjin–Hebei (BTH), Central China region(CC), Yangtze River Delta region (YRD), Pearl River Delta region (PRD), and Sichuan Basin (SCB). These five regions are located in CEC, which is characterized by high population density and relatively developed economies but with prominent air pollution issues. The entire region is discussed in the following sections.Fig. 1Study area demarcation and priority zone identification, three black dots mark the locations of the Shangdianzi (SDZ), China Meteorological Administration (CMA), and Dongtan (DT) O3 observation stations. Study area demarcation and priority zone identification, three black dots mark the locations of the Shangdianzi (SDZ), China Meteorological Administration (CMA), and Dongtan (DT) O3 observation stations. MERRA-2 reanalysis data and preprocessing The Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2, https://disc.gsfc.nasa.gov/datasets?project=MERRA-2), developed by NASA’s Global Modeling and Assimilation Office31, served as the primary source for aerosol and meteorological parameterization in this ozone reconstruction study. This advanced reanalysis system assimilates satellite aerosol optical depth retrieval and ground-based observations to generate a continuous global record (1980–present) of speciated aerosol mass concentrations at 0.5° × 0.625° spatial and daily temporal resolutions32,33. Key aerosol components extracted include black carbon (BCSMASS), mineral dust, sulfate, and sea salt. To ensure compatibility with auxiliary datasets (ERA5 meteorology, surface ozone observations), MERRA-2 data were subjected to spatial regridding to a 0.5° × 0.5° resolution via ordinary Kriging interpolation using a spherical semivariogram model (nugget = 0.12, sill = 2.1, range = 350 km), this geostatistical approach, implemented following Cressie’s34 framework, minimized spatial discretization errors while preserving original data variance. This preprocessing reduced spatial discontinuities by 38% ± 6% compared to bilinear interpolation (RMSE: 1.4 vs. 2.3 μg⋅m−3; validation against AERONET stations). This study utilizes MERRA-2 aerosol data for three main reasons. From the perspective of aerosol-radiation interaction, aerosols can alter the atmospheric radiation balance. The information on aerosol types covered by MERRA-2 data (such as sulfate, black carbon, etc.) can be used to quantify their scattering and absorption effects on solar radiation. For example, sulfate aerosols scatter solar radiation, reducing the intensity of ultraviolet (UV) radiation reaching the ground, slowing down the photolysis rate of ozone precursors (such as NO₂), and suppressing ozone formation. Light-absorbing aerosols like black carbon not only absorb solar radiation, affecting the vertical temperature distribution and vertical mixing in the atmosphere, but also reduce ground-level UV radiation, consequently inhibiting near-surface ozone concentrations. On the chemical level, aerosols participate in a series of reaction processes that influence ozone formation and depletion. On one hand, heterogeneous reactions, such as ozone decomposition, can occur on aerosol surfaces, altering the chemical lifetime of ozone. On the other hand, aerosols can adsorb or react with ozone precursors (such as nitrogen oxides and volatile organic compounds), affecting their distribution and reactivity in the atmosphere. The influence of aerosol-cloud interactions on meteorological conditions also provides a basis for using MERRA-2 aerosol data to estimate ozone concentrations. Aerosols affect cloud microphysical properties, such as altering cloud droplet size and lifetime, indirectly impacting ground-level ozone concentrations. On one hand, increased aerosols may lead to reduced cloud cover, allowing more solar radiation to reach the surface, which favors ozone formation. On the other hand, aerosols may also affect boundary layer dynamics, changing atmospheric vertical mixing and pollutant dispersion capacity, thereby influencing the distribution of ozone and its precursors. Therefore, although there is no simple linear relationship between aerosols and ozone pollution formation, through complex interactions involving aerosol-radiation, aerosol-chemistry, and aerosol-meteorology (among others), MERRA-2 aerosol data can provide valuable information for estimating ground-level ozone concentrations. TAP air quality reanalysis dataset The Tracking Air Pollution (TAP) in China dataset, a collaborative product of Tsinghua University, Peking University, and the Chinese Academy of Meteorological Sciences12, provides high-resolution (10 km) gridded daily MDA8 ozone concentrations across mainland China (2013–2023) (http://tapdata.org.cn/?page_id=59&item=o3). This multi-institutional initiative synergizes ground-based measurements from the China National Environmental Monitoring Network, satellite retrievals (OMI/Aura O3 profiles), emission inventories (MEIC v2.0), and CMAQ model outputs through an ensemble Kalman filter-data assimilation framework. In this study, the TAP ozone fields served as the observational benchmarks for model training and validation. ERA5 meteorological reanalysis The ERA5 dataset, the European Center for Medium-Range Weather Forecasts’ fifth-generation atmospheric reanalysis35, provides hourly meteorological inputs at 0.5° resolution (1940–present, https://cds.climate.copernicus.eu/datasets). The critical variables included the following: boundary layer dynamics, such as the planetary boundary layer height and convective available potential energy; photochemical drivers, such as downward ultraviolet radiation at 340 nm and temperature at a 2-m height; transport indicators, such as zonal and meridional wind components at 850 hPa. 2013–2023 data were used for ML model training/validation after temporal alignment with the TAP ozone fields via Hilbert–Huang transform, and 1980–2012 data were used for historical ozone reconstruction. LightGBM-based zone reconstruction framework The LightGBM (LGBM) algorithm36 was selected for O3 concentration correction and reconstruction. LGBM is a decision tree ML algorithm that incorporates unique binding and one-sided sampling technology, enabling efficient processing of large data volumes and features. The LGBM algorithm is effective in reconstructing PM2.5 historical concentrations37. Figure 2 illustrates the ozone concentration reconstruction process and the temporal and spatial characteristics. This study used TAP MDA8 O3 concentrations, geographical information, surface characteristics, pollutant concentrations, and meteorological data as the model training datasets. Significant factors influencing O3 concentration were identified using a random forest model. Subsequently, the LGBM model was applied with these selected factors, considering the temporal and spatial characteristics of O3. The core of temporal encoding is to extract the periodicity, lag, and trend of meteorological and aerosol factors in the temporal dimension, and capture their temporal correlations with ozone concentrations (seasonal and annual variations). Temporal encoding included two sectors. The first sector is temporal feature extraction. The original time information is decomposed into labels such as year, season, month, and day, allowing the model to identify the temporal patterns of ozone concentrations. The sliding average method is used to separate the long-term trends (e.g., interannual variations) and short-term fluctuations (e.g., sudden pollution processes) of factors, which are input as independent features to help the model distinguish the impacts of different time scales. The second sector is lag feature extraction, to consider the lag effect in ozone generation (e.g., accumulation of precursors, continuous influence of meteorological conditions), lag terms are added to the factors (such as temperature, wind speed, and aerosol concentration in the previous 24 hours) to reflect the impact of historical factors on current ozone concentrations. The core of spatial encoding is to explore the neighborhood correlations of meteorological and aerosol factors in the spatial dimension. Spatial encoding is to calculate the mean/maximum value of factors in the adjacent grid points around the target grid point (e.g., average relative humidity and aerosol column concentration of surrounding grid points) to capture the spatial connections between meteorological fields and pollutants (e.g., the impact of upstream pollutant transport on local ozone).Fig. 2Model framework. Model framework. The optimal model parameters were determined based on the validation outcomes. Finally, temporal and spatial data were input into the LGBM-O3 model to generate a spatiotemporally continuous dataset of ozone concentrations. Model evaluation methods The performance of the reconstruction model was evaluated using the widely used eightfold cross-validation (CV) method38, with spatiotemporal CV implemented. The model was trained and validated on the spatiotemporal continuous grid of MDA8 O3 concentrations from 2013 to 2023. For spatial analysis, monthly, seasonal, and annual average MDA8 O3 data were generated by integrating the MDA8 O3 concentration data of each grid cell. The seasonal MDA8 O3 concentration was initially removed by calculating the monthly anomaly to minimize the impact of the annual cycle when analyzing temporal variations. Subsequently, the ordinary least squares fitting method was used to derive the linear trend from the deseasonalized monthly MDA8 O3 concentration anomaly, and the statistical significance of the linear trend was verified using a two-tailed test39. In addition, the contribution of each meteorological and air pollution variable to the O3 concentration was determined from the average reduction of node impurity calculated from the random forest model. R2, RMSE, and MAE (Eq. 1–3) were selected to evaluate the performance of the LGBM-O3 model. R2 was used as a statistical indicator of the goodness of fit of the regression model. This metric elucidates the degree to which the model explains the relationship between the simulated and observed results. R2 values range from 0 to 1, with higher values indicating superior model performance in capturing the underlying data patterns. The RMSE and MAE quantify the deviation between the model output and the observed results.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}=1-\frac{\mathop{\sum }\limits_{i=1}^{n}{({O}_{i}-{M}_{i})}^{2}}{\mathop{\sum }\limits_{i=1}^{n}{({O}_{i}-\bar{O})}^{2}},$$\end{document}R2=1−∑i=1n(Oi−Mi)2∑i=1n(Oi−O®)2,2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${RMSE}=\sqrt{\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}{({O}_{{\rm{i}}}-{M}_{{\rm{i}}})}^{2}},$$\end{document}RMSE=1n∑i=1n(Oi−Mi)2,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${MAE}=\frac{\mathop{\sum }\limits_{i=1}^{n}|{O}_{{\rm{i}}}-{M}_{{\rm{i}}}|}{n},$$\end{document}MAE=∑i=1n|Oi−Mi|n,where n represents the number of samples, and the MDA8 O3 concentration simulation and observation data are labeled as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M}_{i}$$\end{document}Mi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{i},$$\end{document}Oi, respectively; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{O}$$\end{document}O¯ denotes the average of the observed values. Data Record The dataset is available at figshare40. This study provided a daily maximum 8-h average (MDA8) O3 concentration dataset with a 0.5° × 5° high-resolution spanning from 1980 to 2012 (Fig. 3). It is a gridded (47 × 47) dataset covers central and eastern China, with no data on marine areas. The dataset is provided in forms of netCDF files (.nc). The variables include latitude (°), longitude (°), exact_time and O3_concentration (μg⋅m−3). The dataset can be achieved on website. Our reconstructed long-term MDA8 O3 dataset will be useful for related studies on air pollution in China, particularly studies focused on climate prediction of ozone pollution and environmental health.Fig. 3Spatial distributions and temporal trends of reconstructed MDA8 O3 during 1980−2012, where April–October is defined as the photochemically active warm season and November–March as the cold season. (a) Spatial distribution of annual mean MDA8 O3; (b) Spatial distribution of warm-season mean MDA8 O3; (c) Spatial distribution of cold-season mean MDA8 O3; (d−f) Temporal trends of MDA8 O3 in the entire region (ER) and five specific regions: (d) annual mean, (e) warm-season mean, (f) cold-season mean. Spatial distributions and temporal trends of reconstructed MDA8 O3 during 1980−2012, where April–October is defined as the photochemically active warm season and November–March as the cold season. (a) Spatial distribution of annual mean MDA8 O3; (b) Spatial distribution of warm-season mean MDA8 O3; (c) Spatial distribution of cold-season mean MDA8 O3; (d−f) Temporal trends of MDA8 O3 in the entire region (ER) and five specific regions: (d) annual mean, (e) warm-season mean, (f) cold-season mean. Technical Validation Predictor importance in MDA8 O3 reconstruction The random forest algorithm was implemented to quantify the variable contributions among the 63 predictors of surface ozone variability (Supplementary Table 1). The feature importance values ranged from 0.09% to 29.81%, demonstrating pronounced heterogeneity in the predictor influence. The temperature at a height of 2 m emerged as the primary predictor (29.81% contribution), consistent with the Arrhenius-type temperature dependence of O3 formation kinetics41. Temporal (10.92% contribution) and latitudinal factors (14.15% contribution) are also critical because seasonal variations and geographical locations affect O3 fluctuations. Cumulatively, the top 20 predictors accounted for 86.62% of the reconstruction fidelity, including the four MERRA-2 aerosol categories, O3 mixing ratio at 1,000 hPa, and meteorological variables affecting near-surface O3 concentrations, such as sea level pressure, cloud cover, boundary layer height, dew point temperature, and the meridional component of 10-m wind. These findings suggest that the model effectively integrates multidimensional meteorological and environmental parameters, which validates its capacity to resolve multidimensional atmosphere–chemistry couplings. To improve modeling accuracy, data from the target grid point and up to eight surrounding grid points were used to ensure spatial continuity. Due to the extensive factors from the surrounding grid points, only the top 20 contributing factors were used. Supplementary Table 1 details the factors used for surface ozone reconstruction in this study. Spatiotemporal validation of LGBM-O3 model The dataset (2013–2020) was subjected to stratified temporal partitioning using a leave-one-year-out CV framework. For each of the eight validation cycles, we trained the LGBM-O3 model on seven years of contiguous data while reserving the eighth year as an independent test set, mimicking real-life forecasting scenarios in which future years are predicted based on historical patterns. This approach guarantees strict separation between the training and validation temporal domains, thereby preventing temporal leakage between subsets29,39. Figure 4 shows the performance of the LGBM-O3 model evaluated via temporal-based CV, including the results for the entire study area and the five key regions in 2020. Compared with the TAP data, in approximately 88% of the grids, the sample-based CV R2 (CV-R2) exceeds 0.80; in approximately 42% of the grids, this value exceeds 0.90. The simulated values obtained using the LGBM model are strongly concordant with the observed values (i.e., CV-R2 = 0.83–0.86). Among the different key regions, the average RMSE ranged between 7 and 8 μg⋅m−3, indicating that the model performs effectively in CEC and surpasses the accuracy of previous models42. The model exhibits strong performance and adaptability for estimating the daily surface MDA8 O3 concentration in most regions with no obvious spatial weaknesses.Fig. 4Density scatterplots of temporal-based cross-validation results of daily MDA8 O3 estimates across central and eastern China and five key regions for 2020. (The dashed lines denote 1:1 lines, and the solid lines denote best-fit lines from linear regression. The sample size (N), R2, and RMSE are also provided. The unit for RMSE is μg⋅m−3). Density scatterplots of temporal-based cross-validation results of daily MDA8 O3 estimates across central and eastern China and five key regions for 2020. (The dashed lines denote 1:1 lines, and the solid lines denote best-fit lines from linear regression. The sample size (N), R2, and RMSE are also provided. The unit for RMSE is μg⋅m−3). Although the majority of the data points cluster near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{y}}={\rm{x}}$$\end{document}y=x, suggesting minimal deviation between the observed and simulated values, the LGBM-O3 model tends to underestimate O3 concentrations during high pollution periods, with a regression linear slope of approximately 0.72. This underestimation can be attributed to the infrequent occurrence of high-concentration O3 events. Using Beijing, Shanghai, and Guangzhou as examples, only 3.9% of the samples in the model training dataset exceeded the secondary concentration threshold (160 μg⋅m−3) specified in the National Ambient Air Quality Standard of China (GB3095-2012). The performance evaluation of the LGBM-O3 model for ozone concentration reconstruction in the study region (2013−2020) was performed using temporal-based CV (Fig. 5). The model fit coefficient R2 exhibited minor annual fluctuations within the range of 0.78–0.89 during the study period, indicating that the LGBM-O3 model exhibits a consistent temporal reconstruction ability for O3 concentrations across different target years. From an error analysis perspective, the RMSE values ranged between 8.2 and 10.7 μg·m−3, with the minimum and maximum values observed in 2019 (8.2 μg·m−3) and 2016 (10.7 μg·m−3), respectively. This inter-annual variability highlights the significant temporal sensitivity of meteorological drivers to O3 generation and transport mechanisms. Despite the observed annual fluctuations, the LGBM-O3 model exhibited stable predictive performance under rigorous temporal-based CV framework validation, confirming its reliability and practical applicability in regional ozone inversion studies.Fig. 5Performance evaluation of LGBM-O3 model for the period 2013–2020 based on temporal-based cross-validation (CV) (the black and red lines represent a 1:1 line and linear regression line, respectively). Performance evaluation of LGBM-O3 model for the period 2013–2020 based on temporal-based cross-validation (CV) (the black and red lines represent a 1:1 line and linear regression line, respectively). Extensive model validations are conducted in this sector. We used three typical machine learning tree models (LightGBM, XGBoost, and space-time extremely randomized trees (STET, which was used in Wei et al.18, 2022)) and CNN (Convolutional Neural Network) to make comparisons. The four machine learning models (LightGBM, XGBoost, CNN and STET) exhibit a certain regularity in different time ranges and data scenarios. In the evaluation for 2013−2020 and pollution days (2013−2020), the LightGBM model performs optimally across the three metrics of R2, RMSE, and MAE (Table 1). Its R2 values are 0.831 and 0.722, respectively, relatively higher than those of XGBoost, STET and CNN, indicating a stronger ability to explain the data. Meanwhile, LightGBM achieves the lowest RMSE (9.26 μg·m−3 and 13.82 μg·m−3) and MAE (11.41 μg·m−3 and 15.16 μg·m−3), demonstrating minor prediction errors and better stability. XGBoost ranks second in overall performance, while CNN shows relatively poorer performance across all metrics. Previous studies have shown that tree-based models typically outperform deep learning models in tabular data prediction tasks. Grinsztajn et al.43 (2022) highlighted that, although deep learning has made significant progress in handling image and text data, tree-based models (such as XGBoost and LightGBM) remain the most advanced models for tabular data. After extensive hyperparameter tuning, tree models generally perform better. Therefore, in subsequent validation, we exclusively employed tree-based models for modeling.Table 1Performance evaluation of three typical machine learning models for the period 2013–2020 based on temporal-based cross-validation (CV).PeriodStatisticsLightGBMXGBoostSTETCNN2013–2020R20.8310.8120.7880.787RMSE(μg⋅m−3)9.2610.8211.0112.87MAE(μg⋅m−3)11.4111.8512.529.93Pollution days (2013–2020)R20.7220.7070.673/RMSE(μg⋅m−3)13.8214.5414.36/MAE(μg⋅m−3)15.1616.1216.39/ Table 1Performance evaluation of three typical machine learning models for the period 2013–2020 based on temporal-based cross-validation (CV).PeriodStatisticsLightGBMXGBoostSTETCNN2013–2020R20.8310.8120.7880.787RMSE(μg⋅m−3)9.2610.8211.0112.87MAE(μg⋅m−3)11.4111.8512.529.93Pollution days (2013–2020)R20.7220.7070.673/RMSE(μg⋅m−3)13.8214.5414.36/MAE(μg⋅m−3)15.1616.1216.39/ Performance evaluation of three typical machine learning models for the period 2013–2020 based on temporal-based cross-validation (CV). Moreover, the reconstruction of 2014–2015 ozone concentrations was exclusively based on data from 2020, a methodological decision justified by the five-year temporal separation between these periods, which effectively minimizes the influence of short-term meteorological fluctuations. The exclusion of 2013 data is attributed to the nascent stage of comprehensive monitoring systems at that time, when regional observational networks had not yet achieved complete spatial coverage. In the scenario of reconstructing MDA8 O3 data for 2014 and 2015 using data from 2020 (Table 2), XGBoost achieves a slightly higher R2 value (0.716) than LightGBM (0.711) in 2014, but LightGBM surpasses XGBoost with an R2 of 0.726 (vs. 0.723) in 2015. In terms of error metrics, LightGBM maintains the lowest RMSE (13.12 μg·m−3 in 2014 and 13.11 μg·m−3 in 2015) and MAE (12.03 μg·m−3 in 2014 and 12.38 μg·m−3 in 2015) in both years, followed by XGBoost, while STET yields the highest error metrics. This indicates that LightGBM retains superior prediction accuracy in cross-year data reconstruction tasks, with XGBoost performing closely but slightly inferior. The reason why these statistical indicators performed less satisfactorily in specific periods compared to the entire data cycle is that the model was trained solely on data from a single year (2020), failing to fully capture the characteristics of annual variations in the data, thus limiting its ability to fit cross-year data.Table 2Statistics of three typical machine learning models to reconstruct MDA8 O3 data in 2014 and 2015 using data in 2020.PeriodStatisticsLightGBMXGBoostSTET2014R20.7110.7160.705RMSE(μg⋅m−3)13.1213.5514.16MAE(μg⋅m−3)12.0312.4813.452015R20.7260.7230.682RMSE(μg⋅m−3)13.1113.3815.82MAE(μg⋅m−3)12.3812.4913.61 Table 2Statistics of three typical machine learning models to reconstruct MDA8 O3 data in 2014 and 2015 using data in 2020.PeriodStatisticsLightGBMXGBoostSTET2014R20.7110.7160.705RMSE(μg⋅m−3)13.1213.5514.16MAE(μg⋅m−3)12.0312.4813.452015R20.7260.7230.682RMSE(μg⋅m−3)13.1113.3815.82MAE(μg⋅m−3)12.3812.4913.61 Statistics of three typical machine learning models to reconstruct MDA8 O3 data in 2014 and 2015 using data in 2020. LightGBM demonstrates the best overall performance across different time ranges and data scenarios, followed by XGBoost. This pattern is consistently reflected in both error metrics (RMSE, MAE) and goodness-of-fit (R2). In comparison with XGBoost, LightGBM exhibits superior efficiency and accuracy, primarily attributed to its adoption of the leaf-wise tree growth strategy and histogram-based optimization techniques. When contrasted with STET, it demonstrates enhanced fitting capacity in handling high-dimensional nonlinear data. Furthermore, its inherent parallel computing mechanism confers greater suitability for large-scale data processing scenarios. To assess the performance of specific regional pollution events, we used the LightGBM-O3 model to simulate the MDA8 O3 conditions during the high-incidence O3 pollution season throughout the spring and summer (May to August 2022) and compared the results with TAP data. The overall simulation performance from May to August 2022 was good. The R2 value during this period reached 0.812, indicating that the model has a strong explanatory power for O3 concentrations during this period and that the simulated values are highly consistent with the actual values. Meanwhile, the RMSE (13.42 μg⋅m−3) and MAE (12.59 μg⋅m−3) showed that the overall simulation error was small, and the stability was good, which was similar to the results of pollution days in Table 1. In addition, we selected five distinct regional MDA8 O3 pollution episodes in the same year. These episodes’ simulation and reconstruction results are slightly worse than those for the entire season. The R2 values of the five pollution cases all range between 0.67 and 0.71, slightly lower than that of the overall period, suggesting an increase in the deviation between the simulated and actual values. The corresponding RMSE (15.44–16.82 μg⋅m−3) and MAE (14.68–15.65 μg⋅m−3) are also slightly higher than those of the overall period, but they are still within an acceptable range (Table 3). In a specific MDA8 O3 pollution event on 23 May 2022, severe ozone-polluted areas covered the north half of this region. The area with the highest MDA8 O3 concentration, approaching 200 μg⋅m−3, was distributed across BTH and the northern part of the CC. For the entire region, the R2 reached 0.902, with an RMSE of μg⋅m−3 and an MAE of 12.33 μg⋅m−3. These metrics indicate that the model performed reasonably well in simulating high-concentration pollution events (Fig. 6).Table 3Performance evaluation of LightGBM-O3 model for the ozone pollution periods in 2022.PeriodR2RMSE(μg⋅m−3)MAE(μg⋅m−3)May-August 20220.81213.4212.59Pollution event 1: 3-6 May0.70615.4414.86Pollution event 2: 20-25 May0.69316.1414.68Pollution event 3: 31 May - 3 June0.68516.4714.72Pollution event 4: 14-21 June0.68216.8215.65Pollution event 5: 11-15 August0.67916.3515.13Fig. 6Comparison between reconstructed MDA8 O3(a) and TAP MDA8 O3 (b) in a specific ozone pollution event on 23 May 2022. Table 3Performance evaluation of LightGBM-O3 model for the ozone pollution periods in 2022.PeriodR2RMSE(μg⋅m−3)MAE(μg⋅m−3)May-August 20220.81213.4212.59Pollution event 1: 3-6 May0.70615.4414.86Pollution event 2: 20-25 May0.69316.1414.68Pollution event 3: 31 May - 3 June0.68516.4714.72Pollution event 4: 14-21 June0.68216.8215.65Pollution event 5: 11-15 August0.67916.3515.13 Performance evaluation of LightGBM-O3 model for the ozone pollution periods in 2022. Comparison between reconstructed MDA8 O3(a) and TAP MDA8 O3 (b) in a specific ozone pollution event on 23 May 2022. Evaluation of concentration distribution The reconstructed MDA8 O3 concentrations (Fig. 7a–c) exhibit unimodal normal distributions across all periods, indicating that O3 variations are governed by synergistic interactions among multiple factors (e.g., atmospheric diffusion and anthropogenic emissions). This distribution pattern aligns with the MDA8 O3 characteristics observed in the ground-based monitoring stations and the TAP model outputs (Fig. 7d–e), consistent with typical atmospheric pollution distribution mechanisms. All datasets (reconstructions, observations, and TAP simulations) exhibit elevated kurtosis (leptokurtic distributions), demonstrating a strong central tendency around the mean with a low probability density in extreme concentration regimes (<20 or >160 μg·m−3).Fig. 7Reconstructed decadal MDA8 O3 concentration distributions in 151 cities across CEC: (a) 1981–1990 (mean: 84.1 μg⋅m−3), (b) 1991–2000 (mean: 86.4 μg⋅m−3), (c) 2001–2010 (mean: 88.1 μg⋅m−3), 2014–2023 MDA8 O3 distributions from (d) observations (OBS; mean: 101.4 μg⋅m−3), and (e) model simulations (TAP; mean: 90.1 μg⋅m−3). Reconstructed decadal MDA8 O3 concentration distributions in 151 cities across CEC: (a) 1981–1990 (mean: 84.1 μg⋅m−3), (b) 1991–2000 (mean: 86.4 μg⋅m−3), (c) 2001–2010 (mean: 88.1 μg⋅m−3), 2014–2023 MDA8 O3 distributions from (d) observations (OBS; mean: 101.4 μg⋅m−3), and (e) model simulations (TAP; mean: 90.1 μg⋅m−3). Notably, the observed (OBS) concentration peak exhibits a 10% rightward shift compared to the reconstructed values (Table 4), which can be attributed to the dependence of the reconstruction model on TAP data: a framework that potentially underestimates industrial emission sources and inadequately incorporates transboundary pollution transport mechanisms. Analysis of high-concentration episodes (>160 μg·m−3) reveals a threefold increase in the occurrence frequency, increasing from 0.19% in the 1980s to 0.58% during the first decade of the 21st century. This trend is strongly correlated with the progression of national industrialization. The significantly higher frequency of extreme OBS events than the TAP predictions can be attributed to the discretization errors inherent in the spatial representation of the gridded TAP model.Table 4Statistical characteristics of data distribution for reconstruction, observation, and TAP.Time periodPeak frequency concentration (μg⋅m−3)Direction of deviationProportion of high-concentration days (>160 μg⋅m−3)1981–199073.5Baseline value0.19%1991–1200075.5+2.72%0.4%2001–1201073.5+0%0.58%2014–12023 (OBS)81.0+10.2%9.6%2014–12023 (TAP)75.0+2.04%4.95% Table 4Statistical characteristics of data distribution for reconstruction, observation, and TAP.Time periodPeak frequency concentration (μg⋅m−3)Direction of deviationProportion of high-concentration days (>160 μg⋅m−3)1981–199073.5Baseline value0.19%1991–1200075.5+2.72%0.4%2001–1201073.5+0%0.58%2014–12023 (OBS)81.0+10.2%9.6%2014–12023 (TAP)75.0+2.04%4.95% Statistical characteristics of data distribution for reconstruction, observation, and TAP. Comparison with observations over longer time series To evaluate the reliability of the dataset, it was compared with historical observations from the Tropospheric Ozone Assessment Report (TOAR) Data Center44. The Shangdianzi (SDZ) site, representing a suburban area, and the China Meteorological Administration (CMA) site, representing an urban area, are located in the central and eastern regions of China. The TOAR and CMA sites are representative stations found in suburban and urban regions, respectively. Moreover, the dataset was also compared with observational ozone data from Dongtan Station in Chongming (DT), Shanghai, which represents for the background ozone levels in coastal areas. The monthly average MDA8 O3 concentrations at these three stations were calculated and compared with the reconstructed data (Fig. 8), and the locations of the three observation stations are shown in Fig. 1. The analysis revealed a high correlation coefficient of 0.92 between the reconstructed MDA8 O3 concentrations and the TOAR data, indicating strong agreement. In addition, the reconstructed data effectively capture both annual and monthly variation trends. However, the monthly peak concentrations in the TOAR data were significantly higher, suggesting that, to some extent, the reconstructed data complement and optimize the TOAR data. Compared with the Dongtan observation, the reconstructed ozone showed very close to the peaks in the summer seasons from 2007 to 2012, reflecting good model performance.Fig. 8Comparison between reconstructed and observed MDA8 O3 concentrations in three sites (SDZ, CMA, and DT). Comparison between reconstructed and observed MDA8 O3 concentrations in three sites (SDZ, CMA, and DT). Comparison with previous work Several studies have provided significant results for constructing ozone datasets and spatio-temporal analyses in China. Wei et al.18 (2022) generated the daily maximum 8-hourly averaged ozone dataset (ChinaHighO3) at a resolution of 10 km over the whole area of China from 2013 to 2020 by fusing multi-source data through the spatio-temporal extreme stochastic tree (STET) model. Zhang et al.45 (2024) used a recurrent neural network to integrate massive observation and auxiliary data to construct the hourly ozone data at 0.1° resolution for China from 2015–2020. Zhu et al.46 (2023) used the the LEarning Surface Ozone (LESO) framework to generate the hourly ozone data at 0.1° resolution for China, Europe, and the United States from 2012–2021, which provided a good solution to the usage of column ozone data. Wang et al.47 (2022) analyzed the ozone pollution trends in China from 2014–2020 based on the observation records, revealing the seasonal and spatial variations, and combined with the model to explore the transformation of ozone sensitivity mechanisms, which can provide the basis of regional and seasonal ozone control strategies. Most previous work has focused on developing high-resolution ozone datasets and spatial and temporal characterization of ozone after 2013, mainly because large-scale observations of ozone concentrations in mainland China have only gradually begun since 2013. In this study, we focus on reconstructing the historical ozone dataset for eastern China from 1980 to 2012, which coincides with the accelerated industrialization period in central and eastern China, and the ozone evolution is closely related to the economic development. By integrating early limited observations, satellite and meteorological data, and using machine learning algorithms, this study provides a database for revealing the causes of long-term ozone pollution, assessing historical environmental impacts, and formulating targeted prevention and control strategies. Limitations In this study, several limitations were identified. The reconstruction of MDA8 O3 concentrations relied solely on factors derived from MERRA-2 aerosol data and ERA5 meteorological data, which do not include information about ozone precursors such as VOCs and NOx, nor do they account for other near-surface emissions. The absence of these critical variables may introduce uncertainties in the LGBM-O3 reconstruction model, potentially affecting the accuracy of the reconstructed ozone concentrations. Addressing these gaps by incorporating additional data sources, such as emission inventories and ground-based ozone precursor measurements, is essential for improving the robustness and reliability of the model in future studies.