Hourly surface O3 observations (µg m-3) from 2016 to 2018 were obtained from the China National Environmental Monitoring Center Network (https://air.cnemc.cn:18007/, last access: 24 December 2023) established by the Ministry of Ecology and Environment of China. The MDA8 O3 of each site was calculated with at least 14 valid hourly values from 08:00 to 00:00 LT. A total of 1016 sites were selected after deleting the missing and abnormal data (Fig. 1).

The surface meteorological fields used in this study include sea surface pressure, horizontal wind at 10 m, air temperature at 2 m, downward solar radiation, surface albedo, and total precipitation. The variables selected at 850 and 100 hPa include relative humidity as well as horizontal and vertical velocity. These meteorological variables have been shown by many previous studies to correlate strongly with surface O3 concentrations as discussed above. Hourly reanalysis data for meteorological variables were obtained from the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis dataset (ERA5) with a spatial resolution of 0.25∘×0.25∘ from 1981 to 2019 (https://cds.climate.copernicus.eu/, last access: 24 December 2023). This spatial resolution sets the highest limit of resolution for our hybrid O3 product.

The national land use data with a spatial resolution of 1 km × 1 km for 2013 were obtained from the Resource and Environment Science data center of the Chinese Academy of Sciences (RESDC) (http://www.resdc.cn, last access: 24 December 2023). Six primary types of land use are considered: cultivated land, forestland, grassland, water bodies, construction land, and unused land. Nationwide elevation data were also provided by the RESDC (https://www.resdc.cn/data.aspx?DATAID=123, last access: 24 December 2023), which are resampled based on the latest Shuttle Radar Topography Mission (SRTM) v4.1 data developed in 2000.

The spatial distribution of the harvested areas for four staple crops (wheat, rice, maize, soybean) in China was obtained from the Global Agro-Ecological Zones 2015 dataset (10.7910/DVN/KJFUO1). Crop harvesting dates with a resolution of 0.5∘×0.5∘ were provided by the Center for Sustainability and the Global Environment (Sacks et al., 2010). For crops having more than one growing season in a year, only the primary growing period was considered.

We used the GEOS-Chem global 3-D chemical transport model version 12.2.0 (https://geoschem.github.io/, last access: 24 December 2023), driven by assimilated meteorological data from Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA2) (https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/, last access: 24 December 2023) with a horizontal resolution of 2.0∘ latitude by 2.5∘ longitude and reduced vertical resolution of 47 levels. GEOS-Chem incorporates meteorological conditions, emissions, chemical information, and surface conditions to simulate the formation, transport, mixing, and deposition of ambient O3. It performs fully coupled simulations of O3–NOx–VOC–aerosol chemistry (Bey et al., 2001). Previous studies have demonstrated the ability of GEOS-Chem to reasonably reproduce the magnitudes and seasonal variations of surface O3 in eastern Asia (Wang et al., 2011; He et al., 2012). To provide long-term simulated O3 fields for incorporation into the ML model (see below), we conducted GEOS-Chem simulations at a resolution of 2.0∘×2.5∘; higher resolutions of GEOS-Chem in nested grids are available but computationally prohibitive for multidecadal simulations. The original unit of GEOS-Chem-simulated O3 is parts per billion (ppb), which was converted to micrograms per cubic meter (µg m-3) assuming a constant temperature of 25 ∘C and pressure of 1013.25 hPa (1 µg m-3 is approximately 0.5 ppb) when compared with observations (P. Yin et al., 2017; Gong and Liao, 2019).

Global anthropogenic emissions of CO, NOx, SO2, and VOCs are from the Community Emissions Data System (CEDS), which has coverage over the simulation years of 1950–2014 (Hoesly et al., 2018). Biomass burning emissions are from the Global Fire Emissions Database (GFED4) inventory (van der Werf et al., 2017). Biogenic VOC emissions are computed by the Model of Emissions of Gases and Aerosols from Nature (MEGAN) v2.1 (Guenther et al., 2012), which is embedded in GEOS-Chem. Emissions of biogenic VOC species in each grid cell, including isoprene, monoterpenes, methyl butanol, sesquiterpenes, acetone, and various alkenes, are simulated as a function of canopy-scale emission factors modulated by environmental activity factors to account for changing temperature, light, leaf age, leaf area index (LAI), soil moisture, and CO2 concentrations (Sindelarova et al., 2014).

Dry deposition follows the resistance-in-series scheme of Wesely (1989), which depends on species properties, land cover types, and meteorological conditions, and uses the Olson land cover classes with 76 land types reclassified into 11 land types. Although transpiration is a potential mechanism via which the land cover affects ozone, we do not address it in this study because water vapor concentration in GEOS-Chem is prescribed from assimilated relative humidity (i.e., not computed online from evapotranspiration).

The primary purpose of utilizing ML here was to minimize the biases of model output as compared with observations, whereby the biases could arise from incomplete model physics, input and parameter errors, numerical errors, coding errors, and representation errors (i.e., mismatch in spatial scales between model grid cells and site observations), so that the output of the hybrid model could have the closest values to the observations and enable more accurate impact evaluation. In this study, we used the LightGBM ML algorithm to integrate GEOS-Chem-simulated O3 at a lower resolution with higher-resolution multi-source data to produce higher-resolution hourly O3 and MDA8 O3 fields.

LightGBM is a ML algorithm based on the gradient-boosting decision tree (Chen and Guestrin, 2016), which has a high training efficiency and lower memory footprint and is thus suitable for processing massive high-dimensional data (Zhang et al., 2019). The general steps to build a ML model can be summarized as follows: (1) choose an algorithm that is appropriate for the problem (e.g., regression or classification), (2) clean the data and split them into training and test data, (3) train and tune the model with training data to capture prediction patterns well, (4) evaluate the model performance on test data, and (5) return to step (3) and (4) until an optimal predictive ability is reached. The training and evaluation processes are both performed at the site level in accordance with the observations, whereby the predictor variables and model responses were first sampled at the same locations using the bilinear interpolation approach (Accadia et al., 2003). This approach of handling spatial-scale mismatch between model grid cells and site observations has been commonly used in previous studies (e.g., Li et al., 2021). When predicting the gridded O3 concentrations with the trained model, predictor variables at different spatial resolutions were all regridded to the same resolution of 0.25∘×0.25∘, consistent with the ERA5 meteorological fields. By taking advantage of these higher-resolution datasets, the hybrid approach can not only correct the biases of the GEOS-Chem-simulated O3, but also refine them into a finer resolution. To evaluate if the hybrid approach truly benefits from using higher-resolution meteorological fields, we also repeated the whole training exercise with the input meteorology of GEOS-Chem (MERRA2 at 2.0∘×2.5∘) instead of ERA5.

During the model training process, the model was evaluated with 10-fold cross-validation to ensure the robustness and reliability of the model, whereby the training data were randomly partitioned into 10 subsets of approximately the same size, with 90 % of the data used to train individual models and the ensemble model and the remaining 10 % of data used to examine model performance (Xiao et al., 2018). This process was repeated 10 times so that each data record was left for testing once. The tuning of the hyperparameters was optimized using grid search optimization to improve detection performance and diagnostic accuracy (Wang et al., 2019). Statistical indicators, including the coefficient of determination (R2) and root mean square error (RMSE), were used in a subsequent assessment of model performance for GEOS-Chem alone and for the hybrid approach.

Our analysis revealed that training the model with 1 year or more of data results in only marginal reductions in RMSE and enhancements in R2 (Fig. S1 in the Supplement); thus a timescale of 2 years appears to strike a good balance between computational burden and model accuracy. These results align with the findings of Ivatt and Evans (2020), who suggested that much of the variability in the power spectrum of surface O3 can be captured by timescales of a year or less. Therefore, here we utilized observations from the 2016–2017 period as the training data, which offered a more economical computing cost and improved training time efficiency, and observations in 2018 as the independent test data to evaluate model performance.

Among O3-exposure indices, AOT40 has been widely used during the last 2 decades as it has been found to have a strong relationship with the relative yield of many crop species (Mills et al., 2007) and was thus used to quantify the impacts of surface O3 on crop yields in this study. The flux-based metrics, which require long-term simulations using a process-based stomatal uptake model, were beyond the scope of this study. The AOT40 (ppm-h) is defined as follows: 1AOT40=∑i=1nO3i-0.04, where [O3]i is the hourly mean O3 concentration (ppm) during the 12 h of local daytime (08:00–19:59 LT), and n is the number of hours in the growing season defined as the 90 d prior to the start of the harvesting period according to the crop calendar.

The exposure–yield response functions based on extensive field experimental studies have been established to relate a quantifiable O3-exposure metric to crop yields. It has been suggested that responses of crop yields were found to be greater in Asian experiments than in the American and European counterparts, indicating possibly higher O3 sensitivity of Asian crop varieties (Emberson et al., 2009; Feng et al., 2022). To better understand O3-induced risks to crops in China, the AOT40 exposure–yield functions developed based on field experiments in China are used in this study, which are named AOT40-China. The exposure–yield response functions for soybean are from Zhang et al. (2017), and those of the other three crops are from Feng et al. (2022). The statistical exposure–yield relationships used in this study are summarized in Table S1 in the Supplement.

All-cause mortality, cardiovascular disease mortality, and respiratory disease mortality are selected as the health outcomes of our study due to the high correlation between these endpoints and short-term O3 exposure found in previous studies. A log-linear exposure–response function is widely adopted and recommended by the World Health Organization (WHO) for health impact assessment in areas with severe air pollution. In particular, the log-linear model is the most widely applied exposure–response model at present in China (Lelieveld et al., 2015; H. Yin et al., 2017; Zhang et al., 2022b). The premature mortality is calculated following 2ΔM=δc⋅(RR-1)RR⋅P, where ΔM is the excess mortality attributable to O3 exposure, δc is the baseline mortality rate for a particular health endpoint (P. Yin et al., 2017; Madaniyazi et al., 2016), P is the exposed population, and RR is the relative risk defined as 3RR=exp⁡X-X0⋅β. Here, β is the exposure–response coefficient derived from epidemiological cohort studies (Shang et al., 2013), X represents the model-calculated O3 concentration, and the value of X0 is the threshold concentration below which no additional risk is assumed. Consistent with previous studies (Lelieveld et al., 2015; Liu et al., 2018), we used X0=75.2 µg m-3.

In this study, the mean MDA8 O3 concentrations in warm seasons (May–September) were used to estimate the disease-specific health impacts of short-term exposure to O3. The province-level population and national baseline mortality rate for particular diseases were provided by the National Bureau of Statistics (https://www.stats.gov.cn/english/, last access: 24 December 2023). The spatial differences of baseline mortality in China were not considered without provincial-level data, which means that we assume the baseline mortality is evenly distributed across China (Dedoussi et al., 2020). The exposure–response coefficients were obtained from existing epidemiological studies in China (Table S2). If the corresponding coefficient of a province could not be found in published epidemiological studies, the datum closest to that province would be selected as a substitute. If there were no neighboring provinces, the results of national meta-analysis was used (Zhang et al., 2021).

The final selected features and their importance estimated by the LightGBM algorithm based on 10-fold cross-validation are shown in Fig. 2. GEOS-Chem-simulated O3 is the top predictor of surface O3 concentrations, accounting for 61 % and 58 % of all relative importance in the ML algorithm predicting hourly O3 and daily MD8A-O3, respectively. The result indicates that process-based GEOS-Chem simulations have high utility for O3 predictions under the hybrid approach (Ma et al., 2021). The meteorological variables with a high contribution to both the daily and the hourly models are downward surface solar radiation (SSRD), relative humidity at 1000 hPa (RH_1000hpa), and 10 m horizontal wind (U10 and V10). Other special features, including location (latitude and longitude), elevation, and diurnal and monthly patterns of O3, also contribute to ambient O3 estimations. The spatial distributions of bias-corrected O3 are consistent with observations for both training and test datasets (Fig. S2), indicating that there is no obvious overfitting; i.e., the model is able to generalize from the training set to the test set. The good generalization ability of the model gives us confidence in its ability to make accurate predictions based on new data. In general, the hybrid approach can yield good O3 estimates in the data-intensive regions, including eastern and central China, which are the hotspot areas of O3 pollution.

Figure 3 shows the density scatterplots between O3 measurements and GEOS-Chem simulations, as well as the hybrid approach predictions for 2018. The R2 values of the hybrid approach and GEOS-Chem model are 0.66 and 0.27 at the hourly level and 0.72 and 0.53 at the MDA8 O3 level, respectively. Bias-corrected O3 concentrations have lower RMSEs in comparison with GEOS-Chem-simulated O3 concentrations, reduced from 31.1 to 23.8 µg m-3 for MDA8 O3 predictions and from 38.5 to 26.3 µg m-3 for hourly predictions. The MDA8 O3 model performance is better than that of the hourly model, indicating reduced errors upon temporal averaging. To test if using the higher-resolution meteorological data offers better prediction accuracy compared with the original input meteorology of GEOS-Chem, the MERRA2 dataset driving GEOS-Chem was also used to train the model. We found that the higher-resolution ERA5 dataset performed better in reproducing observed O3 concentrations with moderately smaller RMSEs and larger R2 values (Fig. S3), demonstrating the extent to which a higher-resolution meteorological dataset, despite not being strictly consistent with the input meteorology for the CTM, can help enhance the performance of the hybrid approach and help resolve finer spatial details within the original CTM grid cells. In summary, the result suggests that the CTM-simulated results can be substantially improved by applying ML with multi-source datasets, and the bias-corrected data can improve our understanding of long-term O3 trends and its further implications on crop and human health over China, as discussed in the following sections.

In comparison with previous studies, Liu et al. (2020) used XGBoost to predict O3 in major urban areas of China at a resolution of 0.1∘×0.1∘, and the R2 value and RMSE for MDA8 O3 were 0.74 and 23.8 µg m-3, respectively. Their result indicates that higher-resolution predictions may help enhance model accuracy but represent a trade-off between model accuracy and time efficiency, depending on the purpose. Instead of directly predicting O3 concentrations, Ivatt and Evans (2020) predicted biases in GEOS-Chem-simulated O3 concentrations and then corrected them with XGBoost. They also suggested that the corrected model performs considerably better than the uncorrected model, with RMSE reduced from 32.4 to 15.0 µg m-3 and Pearson's R raised from 0.48 to 0.84. Their greater improvement with larger reduced RMSE than our result is mainly because they selected fewer sites for training, with all the urban and mountain sites (observations made at a pressure < 850 hPa) removed. The removal of these sites can improve the overall apparent performance of the model because O3 formation could have different characteristics in these areas. In general, ML methods have been proven to be a promising tool to improve air pollutant forecasts when a process-level understanding is still incomplete.

Figure 4 demonstrates the spatial patterns of averaged annual and warm-season (May–September) MDA8 O3 from 1981 to 2019. When compared to the high concentrations in the warm season, MDA8 O3 concentrations are relatively lower at an annual level. The annual and warm-season MDA8 O3 concentrations have similar spatial distributions, and both present an increasing trend over the past few decades, with more substantial increases observed between 1981 and 2010. The O3 levels in southern China are lower than those in northern China, but they are still relatively high in the PRD region, which is consistent with findings in previous studies (e.g., Liu and Wang, 2020a). During the first decade of 1981–1990, high-O3-concentration areas are mainly concentrated in the BTHs and northern Shandong. In the next 2 decades, O3 pollution expands extensively to most of eastern and northern China, spreading northward to Jilin and Liaoning; westward to Shanxi and Ningxia; and southward to northern Hunan, Shanxi, and Zhejiang. Moreover, the SCB and PRD regions also experience aggravated O3 pollution during this period. In the last decade of the study period, O3 concentrations remain at high levels in BTHs and SCB without obvious changes. Next we analyze the interannual variability to understand the detailed changes and trends in O3.

Figure 5 shows that the annual averaged MDA8 O3 concentrations increased from 87 µg m-3 in 1981 to 98 µg m-3 in 2019, with a growth rate of +0.26 µg m-3 yr-1, while the warm-season averaged MDA8 O3 concentrations increased from 100 µg m-3 in 1981 to 117 µg m-3 in 2019, having a growth rate of +0.51 µg m-3 yr-1. Moreover, the average annual and warm-season O3 concentrations have a more obvious upward trend before the 2000s, with a growth rate of 0.38 and 0.71 µg m-3 yr-1, compared to that after the 2000s, when O3 concentrations appear to fluctuate within a certain range. GEOS-Chem-simulated O3 has a similar trend to the bias-corrected O3, but it generally overestimates O3 concentrations on a national scale (Fig. S4). The annual and warm-season averaged MDA8 O3 concentrations in BTHs as well as the YRD, SCB, and PRD regions are shown in Figs. S5 and S6. The warm-season increasing trends for BTHs as well as the YRD, SCB, and PRD regions are 0.32, 0.63, 0.84, and 0.81 µg m-3 yr-1 from the years 1981 to 2019.

In recent years, the worsening O3 pollution has fueled numerous studies on ground-level O3 spatial distribution and changes in China, which have been conducted on local, regional, and national scales using different O3 fields from observations, CTMs, and ML estimates. In this study, we mainly focus on the regional and national O3 characteristics, and the O3 trends reported in recent studies are listed in Table 1. By comparing the results of existing works, we find that source-varied O3 fields can induce great uncertainty in the O3 trends. Moreover, the O3 trends are found to be very sensitive to the study period even with the same O3 fields (Wei et al., 2022), which indicates large interannual variability, mostly reflecting the changing anthropogenic emissions and meteorology (Lu et al., 2019; Li et al., 2020). In contrast to the perceptible O3 trends, Liu et al. (2020) suggested that O3 pollution in most parts of China undergoes only modest changes between 2005 and 2017, and their trends were not spatially continuous. T. Wang et al. (2022) also reported that O3 has small positive increase rates for 2013–2021 in many cities, and the O3 increase rates greatly differ from site to site even within the same region.

In comparison, our results indicate no obvious increasing trends in national MDA8 O3 within the same study period. On a regional scale, only BTHs have a perceptible increasing trend in more recent years, while no such trends are found over the YRD, SCB, and PRD regions during the same period. The summertime MDA8 O3 in BTHs has a change rate of +0.81 µg m-3 yr-1, which is much lower than the results using O3 observations (Li et al., 2020). One possible reason is that most observational sites are in urban regions, which usually suffer more serious O3 pollution, while the O3 concentrations from model simulations and ML methods are calculated on the scale of a grid cell with lower domain-averaged values. Moreover, gridded data at a relatively coarse resolution may fail to capture larger site differences, leading to the larger discrepancy between O3 observations and gridded O3 estimates.

Differences in averaged annual and warm-season O3 concentrations indicate that O3 has distinctive seasonal characteristics. Figure 6 shows the seasonal variations in O3 concentrations from 2011–2019, and results for the other past 3 decades are shown in Figs. S7–S9. In winter, pollution is mainly concentrated in the coastal areas of southern China. In spring, O3 pollution primarily occurs in eastern China and the southern part of Yunnan Province. O3 pollution continues to worsen over eastern China in summer, particularly in BTHs, and further extends to SCB. The air quality in eastern and central China is greatly improved in autumn, while southern China experiences the most pollution in this period. In general, the peak and trough values of O3 concentrations appear in summer and winter, respectively. However, O3 concentrations are found to be minimum in summer and maximum in autumn over PRD, which is largely determined by the summer monsoon (Zhou et al., 2013; H. Wang et al., 2018). Figure S10 shows the seasonal averaged MDA8 O3 concentrations in different regions from 1981 to 2019. In winter, O3 concentrations do not experience much change across the four regions over the past few decades, staying mostly between 70–80 μg m-3. Moreover, wintertime O3 concentrations after the 2000s are generally lower than those before the 2000s in BTHs, YRD, and SCB. In contrast, summertime O3 concentrations increase dramatically over the four regions. In spring and autumn, O3 concentrations have an increasing trend in PRD, while it mostly fluctuates within a certain range in the other three regions. The results show that O3 in non-winter seasons has a more pronounced increase during 1981–2019 albeit with regional differences. The regional characteristics of O3 and its influencing factors are further discussed in Sect. 3.4. BTHs as well as the SCB, YRD, and PRD regions have been identified as hotspots of O3 pollution in China. These regions are characterized by high population density (L. Wang et al., 2018) and are also major agricultural areas (Monfreda et al., 2008), which may face greater burdens of crop yield and human health losses with high O3 concentrations. Therefore, here we provide more detailed analysis and investigation of these regions.

Figure 7 shows the bar plots of the seasonal MDA8 O3 concentrations in each region from 1981–2019 for bias-corrected and GEOS-Chem-simulated O3. For the bias-corrected O3, the averaged summertime MDA8 O3 concentrations in BTHs, YRD, and SCB and autumn-time MDA8 O3 concentrations in PRD are 137±8, 119±10, 113±12, and 98±10 µg m-3, with the increasing rate being 0.46, 0.73, 0.98, and 0.69 µg m-3 yr-1 from 1981 to 2019, respectively (Fig. S11). For GEOS-Chem-simulated O3, the averaged summertime MDA8 O3 concentrations in BTHs, YRD, and SCB and autumn-time MDA8 O3 concentrations in PRD are 141±7, 125±11, 120±14, and 100±12 µg m-3, respectively. This shows that O3 concentrations of the four regions have a consistent upward trend in the summer over the past few decades, but there are regional differences in other seasons. Compared to BTHs and YRD, PRD and SCB have more distinctive O3 increases in spring and autumn. Among these four regions, the O3 concentrations have the biggest seasonal differences in BTHs and the smallest seasonal differences in PRD.

The spatiotemporal patterns of O3 in China have been proven to largely depend on both emissions and meteorology. The regional O3 pollution is usually found to be triggered by specific circulation patterns as local meteorological factors are modulated by synoptic-scale circulation patterns. China has a large territory and is affected by different weather systems. The continental high-pressure systems, components of the eastern Asian summer monsoon (EASM) and tropical cyclones, among others, are critical synoptic conditions leading to O3 formation and transport in China (T. Wang et al., 2022; Han et al., 2020). For instance, regional O3 pollution in northern China usually occurs under a typical weather pattern of an anomalous high-pressure system at 500 hPa (Gong and Liao, 2019), which creates favorable meteorological conditions for high O3 levels with high temperature, low relative humidity, anomalous southerlies, and divergence in the lower troposphere. As one of the most important components of EASM, the western Pacific subtropical high (WPSH) strongly influences summertime precipitation and atmospheric conditions in eastern China. A strong WPSH can decrease O3 levels over YRD as enhanced moisture is transported into YRD under prevailing southwesterly winds (Zhao and Wang, 2017). Located on the southern coast of China, PRD features a typical subtropical monsoon climate. There, O3 concentrations are usually the lowest in summer due to the prevailing southerlies with clean air from the ocean and the associated heavy rainfall, while the worst O3 pollution usually happens in autumn, mainly due to the occasional northerly winds during the monsoonal transition, thereby importing precursors from the north, and stable and still relatively warm and sunny weather conditions before the winter starts. Downdrafts in the periphery circulation of a typhoon system can also strongly enhance surface O3 before typhoon landing (Jiang et al., 2015; Lu et al., 2021; Li et al., 2022). On the one hand, the poor ventilation in the peripheral subsidence region of typhoons favors the accumulation of O3 and its precursors. On the other hand, the deep subsidence can transport the O3 in the upper troposphere and lower stratosphere to the surface, causing aggravated O3 pollution. Moreover, smaller-scale circulation patterns, such as land–sea and mountain–valley breezes, also influence O3 in coastal regions (Ding et al., 2004; Zhou et al., 2013; H. Wang et al., 2018).

When compared to the hybrid approach, GEOS-Chem generally has similar O3 distribution and trends over each region, while overestimating O3 concentrations (Table S4). GEOS-Chem particularly overestimates wintertime and autumn-time O3 concentrations in SCB, which are 10±1 and 17±3 µg m-3 higher than those of the hybrid approach, respectively. Previous studies have reported such model overestimates and have proposed a number of explanations involving precursor emissions, dry deposition, and vertical mixing in the planetary boundary layer (PBL) (Lin et al., 2008; Travis et al., 2016; Fiore et al., 2005). Both observational analyses and inter-model comparisons suggested that the summertime dry deposition of O3 calculated by the Wesely scheme in GEOS-Chem could be underestimated, which has been invoked as a cause for model overestimates of O3. The biased emissions in the model, consistent with the biased-high tropospheric NOx columns, result in overestimated O3. Travis et al. (2016) showed that GEOS-Chem with reduced NOx emissions provides an unbiased simulation of O3 observations from aircraft and reproduces the observed O3 production efficiency in the boundary layer. Lin et al. (2008) suggested that the excessive PBL mixing can lead to the biased-high O3 concentrations. The fully mixed O3 throughout the PBL means that the higher O3 concentrations in the upper PBL are brought down to the surface much more efficiently. Moreover, the excessive spatial averaging of emissions at coarser resolutions could also lead to systematic overestimation of regional O3 production (Wild and Prather, 2006). In summary, with a higher prediction accuracy, the hybrid approach lends greater credence to using model simulations to extrapolate historical O3 further back in time, which can furthermore provide us with more accurate estimates of the impacts of O3 on crop production and human health.

Figure 8 shows the relative yield losses (RYLs; RYL = 1 - RY, where RY is the relative yield defined as the ratio of the O3-affected yield to the yield without O3 exposure) calculated with GEOS-Chem and bias-corrected O3 using the AOT40-China metric. For a given crop, the RYLs show generally consistent spatial distribution across the metrics, with BTHs having the most serious crop yield losses due to high O3 concentrations. Compared to the bias-corrected O3, using GEOS-Chem-simulated O3 generally leads to larger yield losses, especially over BTHs and SCB, reflecting overestimated O3 concentrations by GEOS-Chem in cropland areas during the growing seasons (Fig. S12), primarily in spring and summer, which is consistent with the above analysis. GEOS-Chem-simulated O3 leads to slightly underestimated wheat yield loss only over some parts of BTHs, mostly because the primary growing period of wheat there is in winter and spring, and GEOS-Chem has lower O3 estimates than the hybrid approach during this period there (Table S4).

Figure 9 shows the bar plots of the relative yield for each crop using the AOT40-China exposure–yield response relationship. Crop yield losses are generally consistent with the O3 trends as the exposure–yield relationships used here are essentially a set of linear functions. Most crops experience aggravated yield losses over the past 4 decades due to enhanced O3 concentrations, except for wheat, which has the largest yield loss during the period from 1991 to 2000. The reason could be that BTHs have the highest O3 concentrations in spring during the 1990s, which is the primary growing season for wheat (Fig. S13).

The average annual crop RYLs from 1981 to 2019 for wheat, rice, soybean, and maize range from 1.1 % to 13.4 %, 2.7 % to 13.4 %, 6.3 % to 24.8 %, and 0.8 % to 7.4 %, respectively. The differences in yield losses across crops reflect the dependence on crop-specific phenology and ecophysiology. The estimated annual RYLs using bias-corrected O3 for wheat, rice, soybean, and maize from 1981 to 2019 range from 17.5 %–25.5 %, 10.7 %–19.1 %, 7.3 %–17.9 %, and 7.1 %–12.7 %, with a growth rate of 0.03 % yr-1, 0.04 % yr-1, 0.27 % yr-1, and 0.13 % yr-1. Wheat is the most sensitive crop to the O3 concentrations, whereas maize is the least sensitive. Using GEOS-Chem-simulated O3, the estimated annual RYLs for wheat, rice, soybean, and maize from 1981 to 2019 are 18.7 %–28.7 %, 14.0 %–22.0 %, 12.4 %–23.1 %, and 7.9 %–13.2 %, having a growth rate of 0.08 % yr-1, 0.14 % yr-1, 0.23 % yr-1, and 0.11 % yr-1. There are noticeable differences in crop yield estimates using the bias-corrected and GEOS-Chem O3, again indicating the importance of the bias-corrected high-resolution O3 data in related crop issues.

In existing studies evaluating the O3-induced crop losses in China, which also use exposure–yield relationships derived from the experiments conducted in Asia, Zhang et al. (2017) reported that the ambient O3 concentrations in northeastern China cause substantial annual yield loss of soybean ranging from 23.4 % to 30.2 % during 2013 and 2014, depending on the O3 metric used (including AOT40, W126, SUM06, and a flux-based metric). Feng et al. (2022), using AOT40, indicated that the annual average RYLs of wheat, rice, and maize from 2017 to 2019 are 33 %, 23 %, and 9 %, respectively. Our correspondingly estimated RYLs for rice (18.0 %) and maize (10.0 %) are generally consistent with their results, while the RYLs for soybean (16.4 %) and wheat (23.4 %) are much lower than their estimates. Since we used the same exposure–yield response relationships as in their studies, the discrepancies are primarily attributed to the differences in the metrics used (only for soybean), O3 fields, and the sensitivity of the crop to the changes in O3 concentrations (Mukherjee et al., 2021; Feng et al., 2022; Mills et al., 2018). In Zhang et al. (2017), the O3 measurements are obtained from the experimental field (45∘73′ N, 126∘61′ E), and in Feng et al. (2022), the measured O3 concentrations are from over 3000 monitoring sites across eastern Asia. The results of the comparison are consistent with the previous analysis of O3 trends and variability from different sources, where the domain-averaged values of O3 observations are larger than gridded O3 from model simulations (Sect. 3.2) and thus lead to larger estimates of RYLs. On the one hand, it indicates that O3 fields should be considered a great source of uncertainty when comparing the results of previous studies using source-varied O3 fields. Moreover, different degrees of importance should be given for specific crops; for example, the changes in O3 concentrations have a larger impact on wheat crop. On the other hand, it again highlights the necessity and importance of bias correction for model-simulated O3 when studying O3-induced crop reduction.

The estimated annual all-cause premature deaths induced by O3 increased from 55 876 in 1981 to 162 370 in 2019 with an increasing trend of +2979 deaths per year. The annual premature deaths related to respiratory and cardiovascular diseases were 34 155 and 40 323 in 1998 and 26 471 and 79 021 in 2019, having a rate of change of -546 and +1773 deaths per year during 1998–2019, respectively (Fig. 10a). Among three types of health outcomes, only respiratory diseases experienced a decreasing trend in premature mortality, and the premature mortality is constantly below 40 000. The decreasing trend of the respiration-related mortality primarily results from the decreased annual baseline mortality rate over the past few decades (Fig. S14). As the total respiratory-related deaths decreased over the past few decades, respiratory O3-related deaths are decreasing even under aggravated O3 pollution. Based on GEOS-Chem-simulated O3, the corresponding estimated change rate for all-cause disease is +3516 deaths per year from 50 384 in 1981 to 176 741 in 2019. Premature mortality induced by respiratory disease decreased from 37 822 in 1998 to 29 079 in 2019 with a change rate of -584 deaths per year, while cardiovascular disease increased from 44 516 in 1998 to 85 980 in 2019 with a change rate of +1977 deaths per year (Fig. S15). The result shows that using GEOS-Chem-simulated O3 generally gives higher estimates of mortality than using the bias-corrected data. Figure 10b shows the provincial annual average premature mortality of different health endpoints. The five provinces with the highest all-cause mortality are Jiangsu (14 510; 95 % CI: 9022–19 935), Shandong (12 684; 95 % CI: 4258–20 990), Henan (12 290; 95 % CI: 4125–20 343), Guangdong (9268; 95 % CI: 7224–11 416), and Hebei (8276; 95 % CI: 2776–13 706), which are generally consistent with previous studies for China (Zhang et al., 2021, 2022a). Similar distribution can be found for respiratory and cardiovascular diseases but with a different ranking order. Generally, the provinces in densely populated areas (Fig. 10c) with higher O3 concentrations, such as BTHs, YRD, and PRD, have higher health burdens. In contrast, northeastern and southern China (excluding Guangdong) suffer the least life losses induced by O3 exposure (Fig. S16).

When compared with estimates from previous studies, our estimates are generally quite consistent with those given by Maji and Namdeo (2021), who reported that the short-term all-cause, cardiovascular, and respiratory premature mortalities attributed to ambient O3 exposure were 156 000, 73 500, and 28 600 in 2019, respectively. Based on O3 observations in 334 Chinese cities, Zhang et al. (2021) suggested that the national all-cause, respiratory, and cardiovascular mortalities attributable to O3 were 270 000 to 390 000, 49 000 to 63 000, and 150 000 to 220 000 across 2015–2018, respectively, which are much higher than most existing results. Since the methodological approaches are largely similar and since we use the log-linear exposure–response function, we attribute the very high estimated mortalities mainly to the concentration–response threshold X0 assumed to be 0 in their study. A lower X0 means that O3 can cause more adverse impacts on human health even at low concentrations, thus leading to higher mortalities.