Surface ozone concentrations increased in many regions of
China from 2015 to 2019. While the central role of meteorology in modulating
ozone pollution is widely acknowledged, its quantitative contribution
remains highly uncertain. Here, we use a data-driven machine learning
approach to assess the impacts of meteorology on surface ozone variations in
China for the period 2015–2019, considering the months of highest ozone pollution
from April to October. To quantify the importance of various meteorological
driver variables, we apply nonlinear random forest regression (RFR) and
linear ridge regression (RR) to learn about the relationship between
meteorological variability and surface ozone in China, and contrast the
results to those obtained with the widely used multiple linear regression
(MLR) and stepwise MLR. We show that RFR outperforms the three linear
methods when predicting ozone using local meteorological predictor
variables, as evident from its higher coefficients of determination
(
Over the last decade, Chinese policymakers have been successfully
implementing air pollution control policies and strategies, such as The
Clean Air Action Plan in 2013 (Chinese State Council, 2013), to reduce
harmful air pollutants. As a result, annual mean concentrations of fine
particulate matter (PM
Surface ozone is an air pollutant that can induce severe harm to both human
health and ecosystems (Lefohn et al., 2018; Lelieveld et al., 2015). In the
troposphere, it is primarily produced through photochemically induced
reaction chains involving volatile organic compounds (VOCs), NO
In conjunction with the effects of changing ozone precursor emissions, the effect of meteorological conditions on ozone concentrations should always be considered. It is well-known that ozone variations are strongly co-determined by meteorological factors such as incoming solar radiation, temperature, humidity, atmospheric stagnation and precipitation (e.g. Otero et al., 2018; Zhang et al., 2018; Lu et al., 2019a). For example, solar radiation is pivotal to the photochemical production and destruction of ozone (Finlayson-Pitts and Pitts, 2000). Higher surface temperatures, and in general tropospheric temperatures, change the chemical reaction rate of many ozone-relevant chemical reactions and will affect biogenic emissions of VOCs such as isoprene and monoterpenes which are also important ozone precursors (Lu et al., 2019a; Doherty et al., 2013; Guenther et al., 1993; Xie et al., 2008; Archibald et al., 2020). Work by Lu et al. (2019b) further indicated that hotter and drier weather conditions were the main drivers for background ozone increases in 2017 in major city clusters of China. Similarly, Ma et al. (2019) suggested that high biogenic emissions of VOCs and meteorological conditions indicative of heatwaves such as high temperature, low wind speed and no precipitation can elevate ozone pollution in NCP. Furthermore, studies by Wang et al. (2021) and Pu et al. (2017) also found enhanced ozone concentrations during heatwaves in the Pearl River Delta (PRD) and Yangtze River Delta (YRD). Such links between meteorology and ozone pollution provide clear evidence for the necessity to quantify the influence of meteorological factors or even climate change on ozone pollution in China (e.g. Lu et al., 2019a; Meehl et al., 2018). Characterizing the major meteorological drivers behind ozone variations in different regions of China will also be crucial for achieving effective mitigation of ozone pollution now and under future changes in climate.
To quantify the importance of meteorological drivers, previous studies such as Li et al. (2019a) and Han et al. (2020) adopted stepwise multiple linear regression (MLR) to derive linear relationships between meteorological factors and measured surface ozone concentrations across China. Both of these studies demonstrated the significant skill of stepwise MLR in modelling ozone and in quantifying the driver–response relationships. Nevertheless, a key limitation of stepwise MLR or conventional MLR is that these methods are not able to accurately capture nonlinearity, which is a severe constraint given that nonlinear relationships between meteorological factors and ozone, e.g. between temperature and ozone, are to be expected (e.g. Pu et al., 2017; Gu et al., 2020; Archibald et al., 2020). In addition, MLR can suffer from a severe loss in predictive skill and reliability in settings where a large number of (collinear) meteorological factors are considered as predictors (cf. the curse of dimensionality in high-dimensional regression problems; Nowack et al., 2021; Bishop, 2006). Although the stepwise MLR approach adopted by Li et al. (2019a) can overcome collinearity and overfitting to some extent (i.e. because only a few predictors that are particularly strongly influencing ozone concentrations are kept), it is inevitable that many relevant meteorological factors will be excluded from the final MLR predictions using such an approach.
In order to capture any nonlinear relationships between many meteorological factors and ozone and to overcome the potential limitations of considering collinearity and high-dimensional settings in MLR, we will use a machine learning approach as the next logical step to advance such controlling factor analyses of ozone pollution. Specifically, we will adopt random forest regression (RFR) (e.g. Grange et al., 2018; Stirnberg et al., 2021) as a nonlinear approach and contrast the results to a linear statistical learning approach that is also robust in high-dimensional settings in the form of ridge regression (RR) (e.g. Nowack et al., 2018). Both RFR and RR will also be compared with more conventional statistical methods such as MLR and stepwise MLR.
Our paper is structured as follows: in Sect. 2, we describe the data used in this study and the modelling framework of the two machine learning algorithms, namely RFR and RR. In Sect. 3, the performances of RFR and RR will be discussed first and then compared to those achieved with MLR and stepwise MLR. We then summarize the most important meteorological drivers for surface ozone as identified by RFR and RR. Finally, we conduct a trend analysis of recent surface ozone changes in China, and use our method to estimate the contribution of meteorological effects.
The surface air quality measurement data used in this study were obtained
from
To study regional meteorological drivers of ozone, we distinguish four
regions of particularly high population density known as
Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD), Pearl River Delta
(PRD) and Sichuan Basin (Sichuan), with definitions frequently used in
previous studies (e.g. Li et al., 2019a; Han et al., 2020). The boundaries
of these four regions are adjusted to ensure that stations in each region
have similar topography and equivalent elevation. The four regions are also
known as the target areas for air pollution reduction in Chinese government
plans (MEE;
Elevation height (m) and locations of all ground-based stations
and the four megacity cluster regions, BTH (blue box; 114–120
For the meteorological data, we use the gridded ERA5 reanalysis product
(Hersbach et al., 2020) available at
The variables of T2, BLH, SLP, RH, TP, U10 and V10 can also be found as predictors in the controlling factor analyses from the studies of Han et al. (2020) and Li et al. (2019a). The SSRD is included in this study instead of adding a cloud coverage term as done by Han et al. (2020) and Li et al. (2019a). Essentially, we consider SSRD to more directly characterize the local photochemical environment for ozone production and loss than cloud coverage. Zonal and meridional wind at 10 m may be important for the dispersion of ozone's precursors on a local scale. Both zonal and meridional winds at 850 hPa are adopted in this study in order to encompass the effect of transport of more polluted or cleaner air from remote regions. Wind at 850 hPa is less likely to be affected by orography than wind at 10 m altitude, and it is thus better suited for considering the effect of larger-scale transport and dispersion. Additionally, we represent the role of vertical transport of air masses by including vertical velocity at 850 hPa as another factor.
Prior to modelling ozone, we pre-processed the meteorological data by
averaging the raw hourly data over different periods each day and this
process is summarized in Table 1. The averaging periods were not the same
for all meteorological variables. For example, T2, SSRD, SLP, RH and
W850hPa are averaged between local time (UTC
Summary of the meteorological controlling factor variables and the respective times of day considered in their averages. The motivation behind each selected time period is provided in the main text. Note: a positive zonal wind means westerly; positive meridional wind means southerly; positive vertical velocity means downward motion.
To model the relationships between meteorological variables and MDA8 ozone concentrations in China, we use two regression algorithms, a nonlinear approach, RFR and a linear approach, RR. Within our framework, the predictors are the deseasonalized meteorological variables from ERA5 and the dependent variable is the deseasonalized ground-based MDA8 ozone. For RR, both the deseasonalized meteorological variables and the deseasonalized ozone time series are standard-scaled (normalized to zero mean and unit standard deviation) to avoid an imbalance of factors in the regularization part of the RR cost function (Nowack et al., 2018).
Both RFR and RR have been extensively described elsewhere (e.g. Nowack et al., 2018; Grange et al., 2018; Mansfield et al., 2020; Nowack et al., 2021) and it is beyond the scope of this study to provide an in-depth description. Briefly, RFR is based on learning an ensemble of decision trees, where each individual tree splits data into groups until reaching certain pre-set definitions for data “purity” (Breiman, 2001; Grange et al., 2018). The RR is a least-squares linear regression method augmented by L2-regularization aiming to avoid overfitting in high-dimensional regression settings, especially in regression problems with strong collinearity (McDonald, 2009). Both RFR and RR are known to handle collinearity comparatively well (e.g. Dormann et al., 2013), which is key given that many meteorological variables such as temperature and solar radiation are correlated with each other. To assess whether these two machine learning algorithms can improve the accuracy of ozone modelling compared to conventional statistical methods, we will contrast our results to MLR, which may not be highly capable of handling collinearity and overfitting, and stepwise MLR. For MLR, we simply adopt the same modelling framework of RFR and RR; all 11 meteorological variables are ingested into MLR as predictors. For stepwise MLR, we adopted a similar approach as Li et al. (2019a): we start with 11 deseasonalized meteorological variables as predictors in MLR and remove one predictor at a time, based on the smallest significance of the regression coefficient in each new subset of predictors, until there are only three meteorological predictors left. These three predictors are considered to be important predictors and are used in the final model of stepwise MLR for modelling deseasonalized MDA8 ozone.
Supervised machine learning approaches such as the two algorithms we use here require distinct training, validation and testing phases to tune the relevant hyperparameters (explained in detail below) and to validate the skill of the resulting predictive functions on new, unseen data not used in the training and tuning process (e.g. Bishop, 2006). During the training phase, both predictors (i.e. deseasonalized meteorological variables) and dependent variables (i.e. deseasonalized MDA8 ozone) are available and each machine learning regression algorithm is fit to this dataset, assuming different combinations of values for the hyperparameters of each algorithm. The best objective estimate for the combination of hyperparameters is then found in the validation step by predicting ozone values for a validation dataset not used at the training stage (e.g. for a different year in the data record). During the testing phase, the trained and validated algorithm is used operationally to make new predictions for ozone values given a new dataset for the meteorological variables as input to the machine learning function. These test set predictions can then be used to measure the out-of-sample skill of the algorithm in predicting ozone pollution given certain meteorological conditions. In this study, we split the 5 years of data (2015–2019) systematically into training/validation and testing datasets, one at a time and in a rotating fashion. Specifically, 4 of these 5 years are classified as training/validation data, leaving 1 year for testing. To ensure that we are measuring the true predictive performance and relationships, our predictive results and model evaluations are only conducted for the test data, which has not been used at the training and validation stages. This process rotates until ozone data for each year have been assigned once as test data so that all 5 years of data can be predicted by RFR and RR.
Machine learning regressions such as RFR and RR optimize their predictive
performance by tuning certain sets of hyperparameters. To determine the most
suitable set of hyperparameters, we use a statistical cross-validation
method. Specifically, the 4-year training/validation set is further split
into four folds (1 year per fold). We then run a grid search over
pre-defined combinations of hyperparameters by training on three folds and
predicting on the fourth fold in a classic four-fold cross-validation
procedure. We finally select the best-estimated set of hyperparameters on
the basis of the average validation data prediction performance as measured
by the coefficient of determination, i.e. the
The ranges of hyperparameters we search over for both RFR and RR are set as follows: for RFR, the maximum depth for trees growing is iterated in a step of 1 from 8 to 15. The maximum percentage of features and maximum samples (with the bootstrap method) is set from 20 % to 90 % and 30 % to 80 % with 10 % incremental steps, respectively. The total tree number for the forest is set at 200 as a compromise between model complexity and runtime. Optimizations further showed that the minimum samples per leaf are best set to three in our RFRs so that we finally kept this value constant in our grid searches. In terms of RR, the regularization strength is iterated over a range of 1 to 199 with an incremental step of 2, which appeared to encapsulate the best solution in each case. A detailed explanation of these hyperparameters for RFR and RR is provided in Nowack et al. (2021).
Both RFR and RR can enable the identification of the most important meteorological drivers for MDA8 ozone and can help to quantitatively evaluate their relative importance. For RFR, we measure the importance of each meteorological predictor through a metric called Gini importance. A greater Gini importance implies a greater influence of a particular predictor (i.e. the deseasonalized meteorological variable) on the dependent variable (i.e. deseasonalized MDA8 ozone) (e.g. Menze et al., 2009; Zhao et al., 2019; Kuhn-Régnier et al., 2021). Since we train the RFR five times given each possible set of 4-year training/validation data, we average the Gini importance scores for each meteorological predictor across all five runs for our final discussion below. For RR, similar to MLR, the importance of each predictor is evaluated by the magnitude of each predictor's averaged slope (linear regression coefficient) across all 4-year training/validation datasets, which represents the linear effect of each predictor on ozone, given that all predictors are standard-scaled (see Sect. 2.3).
It is important to first assess how well the selected machine learning
algorithms can model ozone by using only meteorological variables as
predictors. Therefore, we adopt the coefficient of determination (
To begin with, the predictors used by RFR and RR are only the local meteorological variables; i.e. each ERA5 grid point's meteorological variables are used as predictors to model the averaged deseasonalized MDA8 ozone for that particular grid location. The average prediction performance of RFR and RR, by comparing predictions across all test years against the deseasonalized measured MDA8 ozone data across China, is illustrated in Fig. 2.
Coefficient of determination (
Coefficient of determination (
Overall, the model performance of RFR generally surpasses the one of RR over
most regions of China, with higher
In order to examine whether RR can improve the model performance by
addressing overfitting, we also applied MLR with all 11 meteorological
predictors and the stepwise MLR approach with the three most important
meteorological factors in the final MLR for comparison. Although most
Weather systems that affect ozone (e.g. high-pressure systems) usually consider large spatial domains, driving regional temperature anomalies and suppressing or accelerating airflow in certain directions. Consequently, it seems intuitive that a meteorological controlling factor framework for ozone might benefit from including additional nonlocal information in the regressions, i.e. if we were to consider surrounding meteorological context information that is not just limited to the predicted ozone grid point in question (Ceppi and Nowack, 2021).
We thus ran a second version of our controlling factor analysis to investigate the spatially wider effect that meteorology may have on a two-dimensional (2D) domain of meteorological variables. This is possible since both RR and RFR better address collinearity and overfitting in high-dimensional regression settings than simple non-regularized MLR approaches, meaning that the additional information included in the regressions might well outweigh the cost of adding more predictors.
In detail, for each ozone target grid point, we include a meteorological
context by adding each meteorological variable within a 7.5
For clarity, Table 2 summarizes the averaged
Averaged
In order to assess the performance of the algorithms in modelling the
regional average ozone, we further compared our regionally averaged machine
learning predictions by RFR, RR and RR–2D against observations for each of
the four selected regions in China (Fig. 4), whereas previously we compared
regional averages based on predictions for individual grid points whose
Using this calculation method, the regional
In summary, all machine learning methods show high skill in modelling meteorologically driven ozone variability. However, similar to results by Han et al. (2020), all linear fits of predicted versus observed ozone values in all regions for both RFR and RR have slopes lower than 1, suggesting a systemic underprediction of ozone for the highest observed ozone levels (higher than the deseasonalized zero mean) and overpredictions of ozone for low ozone pollution regimes (lower than the deseasonalized zero mean). As previously mentioned, such a mismatch may – at least to a degree – arise from non-meteorological factors such as the effect of precursor emissions, which are not taken into account here given the assumption that certain but not all emissions are related to the meteorological factors. Although regionally averaged prediction skill is less affected by local emissions, it will not be completely free from such effects. The increase of the magnitude of the slopes in RR–2D (closer to 1) also suggests that considering nonlocal meteorological variables may help improve the performance of ozone pollution controlling factor analyses, even if nonlinearity is not intrinsically taken into account.
We next aim to quantify how important each local meteorological predictor is for ozone pollution across China. For RR, we use the regression slope as a standard metric to measure how important each of the meteorological predictors are on ozone pollution. A large positive value for the slope (regression coefficient) of a meteorological predictor indicates that the predictor has a strong positive effect on ozone levels and vice versa. Since each set of 4-year training data are learned from independently, we will show averaged results. For RFR, we measure each predictor's importance through Gini importance (see Sect. 2.5). The highest absolute value for both the RR slope or RFR Gini importance is interpreted as the most important meteorological driver variable identified through our data-driven learning procedure. Note that Gini importance only allows measuring relative influences of predictor variables on ozone variability, but not the sign of the influence, i.e. a high value of Gini importance score is not able to determine whether the predictor has a positive or negative effect on ozone.
The Gini importance scores estimated by RFR and the slopes learned by RR for each region are shown in Fig. 5. Both Gini importance scores and slopes are initially estimated for every ERA5 grid location within each region and then averaged across the entire region and across all five learned regression functions.
In general, both RFR and RR show good agreement in terms of identifying the dominant meteorological drivers for each region. The temperature at 2 m is found to be the most important meteorological driver for ozone in BTH, followed by SSRD, albeit the relative difference between these two variables differs more clearly for RFR, which might be caused by nonlinearity in the ozone–temperature relationship (Fig. S5). Temperature was also identified as the most important meteorological variable in BTH by Li et al. (2019a) using MLR. Moreover, a more pronounced positive correlation between daily maximum temperature and MDA8 ozone is found in northern regions of China (Fig. 6a), which is consistent with the findings of these two machine learning algorithms. Since temperature is identified as the key meteorological factor in BTH, more severe ozone pollution with increasing temperature is expected and may be caused by increased rates of chemical kinetics for ozone's production (e.g. Lu et al., 2019a), the contribution of biogenic emissions (e.g. Ma et al., 2019) and anthropogenic emissions such as solvent evaporation which may be intensified in hot weather (e.g. Song et al., 2019; Qi et al., 2017).
For both YRD and Sichuan, surface solar radiation is the most important determinant of ozone variations, with RR slopes again indicating the expected positive relationship between sunny, clear-sky days and high ozone pollution. Furthermore, surface solar radiation is found to be central in BTH and PRD by RFR and RR. Given that Li et al. (2019a, 2020) and Han et al. (2020) did not consider this meteorological variable in their analyses, we recommend that it could be used more generally in the future. High solar radiation stimulates the photochemical environment, which has been suggested as one of the key mechanisms in YRD by Pu et al. (2017). From a large-scale meteorological point of view, such clear-sky conditions in YRD that may enhance severe ozone pollution in this region may be modulated by the western Pacific subtropical high (WPSH) (Shu et al., 2016; Chang et al., 2019; Shu et al., 2020). In the Sichuan, with complex terrain that can complicate atmospheric circulation, ozone pollution is often associated with the occurrence of high-pressure systems associated with clear-sky conditions and high temperatures (Ning et al., 2020), which is also identified by both RFR and RR.
A distinct difference in the weather–ozone coupling relationships is found
for PRD, where RH is the dominant meteorological driver.
Specifically, a negative slope of RH in RR suggests that drier conditions
are strongly favourable for peak ozone concentrations in PRD. As one of many
possible effects of humidity, ozone may be more destroyed through the
photolysis reaction of
Additionally, previous studies (Jiang et al., 2015; Z. Chen et al., 2021; Qu
et al., 2021; Wei et al., 2016) also indicate the importance of vertical
downward transport of ozone in the southern regions of China due to the impact of
typhoons. The effect of such a downward transport may not be well captured
by regressions with only local meteorological predictors as it is a
larger-scale meteorological phenomenon. Therefore, we refer back to our
two-dimensional (2D) approach for RFR in the PRD region first introduced and
described in Sect. 3.2. We show the Gini feature importance scores for this
2D domain approach in Fig. 7a. Since we have multiple values of the
feature importance for each meteorological variable in this setup (i.e.
one for each grid point in the 2D predictor domain), we sum up Gini importance scores
for all grid points within the expanded domain for each meteorological
variable; and this summed value is denoted as the importance for that
particular meteorological variable. As illustrated in Fig. 7a, the
relative feature importance of vertical velocity at 850 hPa (W850hPa)
increases compared to RFR using only local predictors (see Fig. 5b), likely
reflecting the larger-scale influences of downward transport of air masses
in the PRD region. Other key meteorological drivers (RH, surface solar
radiation and meridional wind at 850 hPa) remain in a similar order to what
was identified by purely local regressions. The model performance is
slightly improved by adding the 2D information with an increase of
Trends of MDA8 ozone during April–October from 2015–2019.
Panel
Observational, meteorological and residual trends of regional
averaged MDA8 ozone (ppbv a
Across China, we found that there is a consistency in the identification of the three most important meteorological drivers by RFR and RR: temperature, surface solar radiation and RH (Fig. 8). Overall, there is a distinctive distribution pattern of the three major meteorological drivers in China. The temperature at 2 m is dominant over northeast China, covering BTH and expanding to the northern region of China. Most areas in the mid-latitude regions of China, including east China (e.g. YRD) and Sichuan, show surface solar radiation as the main meteorological driver for ozone, suggesting the necessity of including this variable for analyses. The dominance of surface solar radiation gradually expands northward and southward from this region while being overtaken by temperature in the north and RH in the south. Ozone in southern China is primarily driven by RH. Such a distinctive spatial distribution of meteorological drivers may be related to the characteristics of regional climatology. For instance, as described above, regions in southern China such as PRD are particularly influenced by variations in incoming moist air masses, leading to the importance of humidity surpassing temperature and surface solar radiation. The relatively drier northern regions do not have such changeable humidity conditions, making temperature and surface solar radiation the key meteorological factors driving ozone.
Finally, we explore how our new approach could be used to study the quantitative influence of meteorology on historical ozone variability and trends in China. To facilitate a comparison to previous work, we use a similar method as Li et al. (2020) to establish estimates for observed surface ozone trends in China. We note that our exercise is somewhat limited by the slightly shorter period considered here, i.e. from 2015 to 2019, instead of starting from earlier years. Given this very short period, we are aware that any long-term trend analysis is explorative and has to be interpreted carefully, as will also become evident from low statistical significance in many detected trends. We nevertheless attempt such an analysis to demonstrate how our method can be used in such contexts and to also evaluate if any statistically significant trends are robust after accounting for meteorological influences.
For trend analyses, we first convert MDA8 ozone concentrations from mass
concentrations (
Table 3 summarizes the regionally averaged observed trends from 2015 to
2019, which is estimated by ordinary linear regression in the four regions.
We additionally list our meteorologically estimated trends and the residual
trends. Overall, the three machine learning methods and MLR provide
relatively similar estimates of meteorologically driven trends in BTH, YRD
and Sichuan, while we find indications that the meteorologically
driven trend in PRD may be underestimated by only using local meteorological
factors; using RR–2D we estimate a meteorologically driven trend of 0.84 ppbv a
In terms of the raw observed trends, both BTH and PRD show significant
increases in ozone pollution (
Finally, we aim to calculate trends on a ERA5 grid-by-grid point basis.
Although both RFR and RR–2D show overall better skill in modelling ozone
across China, RR–2D exhibited particularly increased predictive skill in
southern China. Therefore, for assessing meteorologically driven trends of
MDA8 ozone across all ERA5 grid locations in China, we will only be
examining the results for RR–2D. Figure 9 shows trends during April–October
from 2015 to 2019 across China. Overall, the observed average trend across
China is 1.1 ppbv a
Ozone pollution in China can be strongly influenced by meteorological conditions. This study examines the major meteorological drivers for ozone pollution across China during months with particularly high ozone pollution (i.e. April to October, from 2015 to 2019) using a controlling factor framework and two machine learning algorithms, namely RFR and RR.
The results obtained with RFR and RR are also compared with conventional
approaches i.e. MLR and stepwise MLR, using
consistent out-of-sample cross-validation methods. When considering local
meteorological factors only, RFR outperforms the linear approaches RR and
MLR, which in turn perform better than stepwise MLR that uses only the three
local, most significant meteorological factors. The better performance of
RFR is for example evident from the overall increase in predicted versus
observed coefficients of determination (
A key advantage of our approach is that both RFR and RR allow for a straightforward interpretation of the predictive models (explainable machine learning). Reassuringly, we find a good agreement regarding the identification of the dominant local meteorological drivers for each region. In general, ozone pollution in northern China such as in the BTH region is predominantly driven by temperature fluctuations while ozone in southern China like in the PRD region is particularly strongly controlled by humidity, possibly due to the important role of humid weather in preventing significant ozone pollution episodes in this region. Besides, we observe a strong influence in PRD of air exchange with pristine marine regions, leading to a greater influence of large-scale wind directions, e.g. through the transport of clean marine air into the region, or through air stagnation and ozone accumulation under large-scale sinking atmospheric motion. Surface solar radiation plays a major role in general due to its importance in setting the conditions for ozone photochemistry, which is particularly dominant in YRD and Sichuan. Our work thus highlights that surface solar radiation might be a key predictor to consider in future controlling factor analyses. In summary, hot, dry and sunny weather tends to provide more favourable conditions for ozone pollution in China, which is not entirely unexpected but carries important implications for future changes in air pollution under climate change, while simultaneously considering the pivotal role of targeted emission control strategies on ozone precursors.
In terms of ozone trends, we find a linear MDA8 ozone increase of about 1.1 ppbv a
The original air quality data including hourly and 8 h rolling mean of ozone are available at
The supplement related to this article is available online at:
PN, XW and GLF designed the study. XW performed the modelling and analysis of the data, supervised by PN and GLF. XW wrote the paper with input and revision from PN and GLF.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Peer Nowack was supported through an Imperial College Research Fellowship. We thank the Ministry of Ecology and Environment in China for supporting the nationwide observation network and publishing air quality data. We are grateful to Xiaolei Wang for archiving the air quality data. We thank the European Centre for Medium-Range Weather Forecasts (ECMWF) for providing the ERA5 reanalysis product.
This paper was edited by Qiang Zhang and reviewed by Ke Li and one anonymous referee.