A comprehensive evaluation of the spatiotemporal variation of CO2 and its driving forces over China

With the improved accuracy and high spatiotemporal resolution, satellite remote sensing has provided an alternative way for monitoring the variations of CO2 in remote areas where field observations are inadequately sampled but the emissions of CO2 are increasing rapidly. Based on CO2 estimates from satellite remote sensing and the atmospheric tracer transport model, this study assessed the spatiotemporal patterns of atmospheric CO2 and its driving forces across China. Results show a consistent increase in CO2 at all levels of the troposphere, with the growth rate exceeding 2.1 ppm/year. Among them, the near surface witnessed obvious spatial heterogeneity with the highest concentrations of CO2 occurring in East China and the lowest in Northwest China. This strong spatial differentiation disappeared with increase in altitude and is replaced by a distinct south–north gradient difference at the upper troposphere. With regard to vertical variations, the concentration and growth rates of CO2 at the lower troposphere are generally higher than those at the upper troposphere. The driving mechanism analysis indicates that the variation of CO2 at the near surface is primarily caused by anthropogenic and biogenic activities, while air motion dominates the distribution of CO2 at the upper troposphere. The findings of the present study could provide a valuable reference for understanding regional carbon cycles and formulating carbon emission reduction strategies on a national scale.


Introduction
Carbon dioxide (CO 2 ) is the main greenhouse gas causing global warming. It contributes more than 60% of the total radiative forcing and has a lifespan of more than 120 years in the atmosphere (Stocker et al., 2013). Due to massive fossil fuel combustion and dramatic land use changes, the level of global average atmospheric CO 2 has increased by 140% relative to pre-industrial levels and reached 414.72 ppm in 2021 (Peters et al., 2011;WMO, 2012). The increasing concentrations of CO 2 have led to a positive energy imbalance of 0.53 ± 0.11 W/ m 2 from 2003 to 2018, causing an increase in atmospheric temperature and sea level (Kramer et al., 2021), which in turn led to a series of meteorological disasters such as the melting of polar glaciers and frequent drought and flood events (IPCC, 2014). The continuous increase in the concentration of atmospheric CO 2 has attracted a significant attention from the OPEN ACCESS EDITED BY Haipeng Yu, Northwest Institute of Eco-Environment and Resources (CAS), China international community and organizations. Accurately and comprehensively assessing the spatial and temporal distribution of atmospheric CO 2 can provide a foundation for understanding the global carbon cycle and will aid in the formulation of policies aimed at reducing carbon emissions (Umezawa et al., 2018).
For this purpose, the long-term measurement of atmospheric CO 2 concentrations in different regions of the world has been established and gradually evolved into a global network of CO 2 observation. Currently, there are more than 300 sites worldwide, where greenhouse gas levels are measured by the World Meteorological Organization/Global Atmosphere Watch (WMO/ GAW) (WMO, 2012;Fang et al., 2014). The integrated carbon observation system provides a reliable dataset for assessing longterm changes of atmospheric CO 2 at the global and regional scales and plays an important role in the early research on carbon sources and sinks. However, restricted to the level of socioeconomic development and topographic factors, the GAW ground-based observations are sparse and unevenly distributed, with most of them located in developed countries or plain areas (Mustafa et al., 2020). In addition, the discontinuities and inconsistencies in multi-sources of observational data further complicate the availability of data, making the spatiotemporal distribution studies of atmospheric CO 2 uncertain and challenging (Fang et al., 2014;Basu et al., 2013).
The development of satellite remote sensing technology has provided an alternative method for monitoring the spatiotemporal distribution of CO 2 on a global scale. By combining various satellite sensors and high-precision inversion algorithms, CO 2 products are generated based on the characteristics of its absorption spectrum at the thermal and near-infrared bands (Schneising, 2008). To date, three satellite projects dedicated to CO 2 observation have been launched, i.e., the Greenhouse Gases Observing Satellite (GOSAT) (Yokota et al., 2009), the Orbiting Carbon Observatory (OCO) (Crisp, 2015), and the Chinese carbon dioxide observation satellite mission (TanSat) (Liu et al., 2018). Compared with the field observation, satellite CO 2 estimates are not subject to topographical factors and can achieve a stable and continuous observation of atmospheric CO 2 at the regional and global scales with high spatiotemporal resolution (Zeng et al., 2013). With a large spatial coverage from near surface to the troposphere and a long temporal period (2009-present), the GOSAT has been widely applied in research studies of carbon sources and sinks and the transport of atmospheric CO 2 (Basu et al., 2013;Mustafa et al., 2020). In general, the progressive development of space-borne sensors and inversion algorithms has made satellite remote sensing the main method of monitoring atmospheric CO 2 variations and has enhanced our understanding of regional and global carbon cycles.
By integrating ground-based and satellite remote sensing observation data, numerous studies have been conducted to explore the spatiotemporal differentiation of CO 2 and the associated driving forces (Cao et al., 2019;Kong et al., 2019;Yang et al., 2021). However, most of the studies focused on Western Europe and the United States. Northeastern Asia, a densely populated region with high CO 2 emissions, has not been adequately explored. As a matter of fact, China has witnessed rapid economic development over the last three decades, and the rapid increase in fossil fuel carbon emissions has made China the leading contributor of global CO 2 emissions (Le Du et al., 2017). In this regard, China has implemented a number of programs aimed at reducing carbon emissions and conserving energy and has committed to achieving carbon peak by 2030 and carbon neutrality by 2060. Therefore, it is important to evaluate the spatiotemporal variation of CO 2 and explore the driving mechanism, as this could provide valuable information for understanding the carbon cycle and constraining carbon emissions on a national scale (Hammerling et al., 2012;Fang et al., 2014).
To assess the characteristics of CO 2 variations and investigate its driving forces on a national scale, we chose a fast-economic growth and high carbon emission area, i.e., China, as the study region and obtained CO 2 estimates from satellite remote sensing (GOSAT) and an atmospheric tracer transport model (Carbon Tracker). Meanwhile, the data of leaf area index (LAI), fossil fuel carbon emissions, and the 3D wind field were also collected. In particular, we aimed to (1) analyze the spatial and temporal variation of CO 2 at different time scales across China, (2) explore the dominant factors affecting the regional concentrations of CO 2 , and (3) examine the relationship between local CO 2 and LAI, fossil fuel carbon emissions, and regional air motion.

Study area
As one of the largest countries in the world, China covers a vast territory (9.6 × 106 km2) with heterogeneities in climatic conditions, complex topographies, and ecological environments, which have caused strong variations in economic development and population growth. A combination of climatic diversity and human activity has resulted in a unique pattern of local carbon budgets with obvious spatial differences in CO 2 concentrations (Fang et al., 2014;Du et al., 2017). To address this issue, six sub-regions characterized by different climatic conditions and socioeconomic backgrounds were delineated in this study, and the detailed zonal information is illustrated in Figure 1.

Materials
Four types of datasets are used in this study, namely, in situ observations, satellite remote sensing, model simulations, and reanalysis. As part of the validation process for gridded CO 2 products, the in situ observation data from WDCGG were utilized, and the satellite and modeled CO 2 products were used to study the spatiotemporal variation of CO 2 at different altitude. A driving mechanism analysis was conducted using gridded data from the LAI and CDIAC, as well as ERA5 reanalysis data. Supplementary Table S1 provides a brief overview of all datasets used in this study.

WDCGG
The World Data Centre for Greenhouse Gases (WDCGG) has been operated by the Japan Meteorological Agency (JMA) since 1990 as part of the WMO/GAW program. As the only World Data Centre (WDC) specializing in greenhouse gases, it serves to collect, archive, and distribute greenhouse gas data from ground-, ship-, aircraft-, and satellite-based observations, contributed by organizations and individual researchers worldwide (WMO, 2012). The objective of the WDCGG is to support the monitoring of climate change and facilitate policy development, thereby helping reduce the risks associated with environmental degradation. In this study, the in situ observation data from the WDCGG were used to evaluate the performance of GOSAT and CT CO 2 products in China.

GOSAT
GOSAT is the world's first satellite dedicated to monitoring greenhouse gases from space, which was launched on 23 January 2009 and is jointly developed by the Ministry of Environment (MOE), the Japan Aerospace Exploration Agency (JAXA), and the Japan National Institute for Environmental Studies (NIES). GOSAT is a Sun-synchronous orbit at an altitude of 666 km, with an approximate 10.5 km diameter at nadir and a revisit cycle of 3 days (Yokota et al., 2009;Suto et al., 2021). A total of two observation instruments are onboard the satellite, the Thermal and Near-infrared Sensor for carbon Observation Fourier Transform Spectrometer (TANSO-FTS) and the TANSO-Cloud and Aerosol Imager (TANSO-CAI), which are designed to detect the three-dimensional distribution of greenhouse gases as well as clouds and aerosols. Both of them are equipped with four bands, and the three SWIR (0.76, 1.6, and 2.0 um) and the wide TIR (5.5-14.3 um) bands of TANSO-FTS are responsible for retrieving the column concentrations and vertical profiles of CO 2 (Imasu et al., 2008). While the spectral channel of TANSO-CAI is used to capture the cloud cover and aerosol properties (Deng et al., 2016), the cloud-contaminated footprints are screened out according to this information. Based on these observations, the CO 2 retrieval algorithms have been extensively developed, including NIES (Yoshida et al., 2013), ACOS (Kulawik et al., 2019), and UOL-FP , and the operational CO 2 products are widely used in estimating the global and regional CO 2 concentrations and fluxes.
In this study, we used the L4B global CO 2 distribution dataset, which was developed by the Japan National Institute for Environmental Studies using the NIES-FP inversion algorithm. The latest version of this dataset is updated to V02.07 and can be freely accessed through www.gosat.nies.go.jp. It covers the period from June 2009 to October 2019, with a spatial resolution of 2.5 × 2.5 and a temporal resolution of 6 h over 17 vertical levels from the surface to 10 hPa. In order to obtain the atmospheric CO 2 concentrations at different levels and time scales, the original netCDF format dataset was converted to raster images in the R programming environment, and then the daily, monthly, and annual average CO 2 concentrations were aggregated from the hourly observations.

CarbonTracker
CarbonTracker (CT) is a data assimilation system for CO 2 developed by NOAA ESRL. It incorporates a two-way nested offline atmospheric tracer transport model, known as transport model 5 (TM5), to simulate the surface fluxes and the distribution of atmospheric CO 2 Peters et al., 2007). CT separately estimates the surface CO 2 exchange originating from fossil fuel emissions, terrestrial biosphere impacts, biomass burning, and ocean fluxes. In addition to in situ observations from tall towers, flask samples collected by NOAA's cooperative air sampling network, and continuous measurements taken by partners, CarbonTracker assimilates more than 100 datasets around the world. By utilizing the technology of an ensemble Kalman filter, the differences between observations and model forecasts are reduced (Peters et al., 2005;Babenhauserheide et al., 2015). The model provides global 3D CO 2 distribution at 25 levels with 3 × 2 (longitude × latitude) spatial resolution and 3 h temporal resolution. In this study, CarbonTracker data of version CT 2019B

Frontiers in Environmental Science
frontiersin.org were collected for evaluating the impacts of anthropogenic, biogenic, wildfire, and oceanic sources on the concentrations of atmospheric CO 2 .

LAI and CDIAC
It is acknowledged that human activity and the biophysical process of vegetation play an indispensable role in affecting nearsurface CO 2 concentrations (Cao et al., 2019;Yang et al., 2021). Therefore, the leaf area index (LAI) and fossil fuel carbon emissions were employed to investigate the relationship between local CO 2 and anthropogenic and biogenic factors. The LAI is defined as the onesided green leaf area per ground surface and is a useful indicator in reflecting the canopy and functions of the vegetation community (Bréda, 2003). According to the study of Berterretche et al. (2005), the LAI shows a linear correlation with the primary production in terrestrial ecosystems, and the capacity of vegetation in carbon sequestration increases with the growing LAI. In this study, the LAI data were obtained from the Climate Data Record (CDR) developed by the National Centers for Environmental Information, which produced a daily LAI dataset on a 0.05 × 0.05-grid based on data derived from Advanced Very High Resolution Radiometer (AVHRR) sensors from 1981 onward. In order to minimize cloud contamination and atmospheric variability, this study employed a maximum value composite (MVC) procedure to generate monthly LAI.
The Carbon Dioxide Information Analysis Center (CDIAC), operated by the United States Department of Energy, is designed to provide global warming data and analysis to the U.S. government and research community (Andres et al., 2014). The primary mission of the CDIAC is committed to obtaining, evaluating, and distributing data related to greenhouse gas emissions and climate change. On the basis of monthly energy consumption data, the CDIAC divides the global total emissions into different sectors, such as fossil fuel emissions, industrial processes, and land use emissions (Oda et al., 2018). The monthly fossil fuel CO 2 emissions of V2016, with a spatial resolution of 1 × 1 for 2009-2013, were selected in the present study.

ERA5
ERA5 is the latest fifth-generation reanalysis dataset developed by the European Centre for Medium-Range Weather Forecasts (ECMWF) (https://www.ecmwf.int) and contains a series of improvements relative to its predecessor, ERA-Interim. The dataset employs an advanced data assimilation and modeling system with additional historical in situ and satellite observations to provide a more accurate representation of atmospheric conditions (Karl and Michela. 2019;Jiang et al., 2021). The spatial resolution of the data is 0.25°×0.25°on 37 vertical levels from the surface up to 1 hPa. It covers the period from 1959 to the present and is updated daily with a latency of 5 days. In this study, the monthly wind data of the reanalysis from 2009 to 2019 were used to derive the mean states of zonal and vertical movement of the atmosphere.

Methods
Due to the possibility that different datasets may differ in terms of their spatial and temporal references, all datasets were projected to the GCS_WGS_1984 geographic coordinate system and converted to local time of Beijing to maintain consistency. Additionally, to make the datasets comparable and facilitate the following analysis, the products of CT, LAI, and CDIAC were aggregated or resampled to a 2.5 × 2.5 regular grid by using the spatial information of GOSAT as a standard.
Prior to applying the estimated CO 2 in the following analysis, the accuracy of GOSAT and CT in China was evaluated through comparison with WDCGG observations. Stations used for validation were screened based on the following criteria: (1) falling within the study area, (2) not being assimilated in generating the products of GOSAT and CT, and (3) having less than 20% of missing observations. For a gauge station, we first identified the grid cell in which that station was located in the spatial dataset of a satellite product. Then, the values of grid data were directly extracted and the Pearson linear correlation coefficient (R) and the root mean square error (RMSE) were used to measure the strength of the linear association and the magnitude of the deviation between observations and estimates. The formula is as follows: where Pe and Po are the estimated and observed CO 2 , respectively; n is the sample size; δ is the standard deviation; and cov() is the covariance between the two variables.
In addition, the Pearson correlation method was also used to study the impacts of anthropogenic and biogenic activities on surface CO 2 concentrations by calculating the correlation coefficients between LAI, fossil fuel carbon emissions, and local CO 2 concentrations.
The interannual variation of CO 2 was evaluated using linear trend fit as expressed in Eq. 3. The slope and statistical significance of the trends were estimated using the ordinary least squares method and the two-tailed Student's t-test, respectively. In this study, a trend was considered statistically significant when it is at the 95% confidence level. In addition, the coefficient of variation (CV) was used to quantify the seasonal variation of CO 2 , and it was defined as the ratio of the standard deviation to the mean in Eq. 4.
where y is the time series of CO 2 concentrations; a and b are the corresponding trend and the intercept, respectively; t represents the year; and ε is the regression error. s and‾x are the standard deviation and the mean of CO 2 , respectively. As the concentration of CO 2 exhibits a strong seasonal variation, it is thus essential to calculate the seasonal indexes and remove the seasonal factor from the time series when studying the multi-year monthly average CO 2 concentrations and conducting the correlation analysis with fossil fuel carbon emissions and LAI (Dettinger and Ghil, 1998). In this study, the ts and decompose functions in R were used to deseasonalize the interannual variation of CO 2 , and the original time series were divided into three components: the trend component, the seasonal component, and Frontiers in Environmental Science frontiersin.org the random component. After that, the seasonal component was subtracted from the time series and was treated as an input to the subsequent analysis.
ts actual ts trend + ts season + ts random , where ts actual is the actual value of the dataset and ts trend , ts season , and ts random are the trend component, seasonal component, and random component, respectively, of the time series.
3 Results and discussion

Accuracy evaluation
After conducting the screening process, only the stations of LLN and HKO were considered to meet the criteria in China. The scatterplots between the observations and estimates are presented in Figures 2A, B. Generally, the CO 2 estimates from GOSAT and CT agreed well with observations, with averaged correlation coefficients of 0.96 and 0.85 for LLN and HKO stations, respectively. The HKO station exhibited a slightly higher RMSE (6.88 ppm) than that in the LLN station (4.44 ppm). Both of the products had a CO 2 estimation accuracy lower than 2%, meeting the requirements for precision described in Rayner and O'Brien (2001). In addition, the general pattern of intra-annual variations in CO 2 can be wellcaptured by GOSAT and CT, which peaks in winter and reaches its lowest level in summer ( Figures 2C, D). In light of the high accuracy and good stability of GOSAT and CT, the estimated CO 2 products can be used to study the spatiotemporal patterns of CO 2 in China.
In order to study the spatiotemporal variations of CO 2 concentrations at different heights of the atmosphere, three typical layers, namely, the 975 hPa, 500 hPa, and 100 hPa, were selected to represent the mean state of CO 2 at the near surface, the middle, and the upper troposphere. The annual, seasonal, and diurnal variations of CO 2 concentrations were analyzed at these three levels if there is no further specification.

Annual variations
The spatial pattern and magnitude of CO 2 concentrations varied among different heights of the atmosphere. A declining trend was observed with increasing atmospheric height, with a mean CO 2 concentration of 400.25 ppm at near surface decreasing to 398.41 ppm and 393.76 ppm at the middle and upper troposphere, respectively (Supplementary Table S2). Among them, the near surface witnessed a strong spatial heterogeneity with the highest concentrations of CO 2 occurring in East China and the lowest in Northwest China ( Figure 3A). This pattern is consistent with the spatial distribution of China's population and economy, indicating a considerable impact of local carbon emissions on near-surface CO 2 . In contrast, CO 2 concentrations at the middle troposphere showed much less variation, with a standard deviation of only 0.23 ( Figure 3C). The insignificant variations may be associated with horizontal and vertical winds, which transport near-surface CO 2 to the atmosphere and then sufficiently mixed, resulting in a uniform spatial pattern of CO 2 concentrations (Cao et al., 2019;Al-Bayati et al., 2020). In the upper levels of the atmosphere, a distinct gradient of CO 2 concentrations was observed from south to north ( Figure 3E). High values of CO 2 A general increasing trend with a magnitude higher than 2.1 ppm/year was detected for all the three typical layers (Figure 4), and almost all the data points passed the significance test at the 0.05 level ( Figures 3B, D, F). Similar to the variation of CO 2 concentrations, the annual change rate was significant at the near surface (2.38 ppm/year), where the high values of annual CO 2 growth were found in part of East and North China. The middle troposphere witnessed a stable increasing trend at about 2.34 ppm/year. When it comes to the upper troposphere, the annual increase of CO 2 shows large discrepancies across China, with high values in the Mid-South and low values in Northeast China. Based on the decreasing change rates from the lower to upper troposphere, it appears that the intensified anthropogenic activities tend to cause significant increase in CO 2 at near surface.

Seasonal variations
In order to study the intra-annual variation of CO 2 , the linear regression method was used to remove the annual growth rate, and then the monthly CO 2 concentrations were derived by calculating the multi-year averages. At near surface, all the six regions exhibited a unimodal fluctuation pattern, with the peak value of CO 2 concentrations occurring in April, followed by a decline, and reaching its trough in August ( Figure 5A). This may be the result of the combined effect of anthropogenic and biogenic activities (WMO, 2017;Buchwitz et al., 2018). The increased photosynthesis of vegetation in summer is responsible for a higher uptake of CO 2 from the atmosphere FIGURE 3 Spatial patterns of the annual mean (A, C, and E) and the change rates (B, D, and F) of CO 2 . It is to be noted that the grid cells filled with black dots indicate that the change rates are significant at the 95% level.

Frontiers in Environmental Science
frontiersin.org (Yang et al., 2021), while the intensified anthropogenic heating emissions and plant respiration lead to a higher level of CO 2 in early spring (Liu et al., 2012). The amplitude of the seasonal variation is found to be largest in Northeast and lowest in East, and the differences between the peak and the trough are 24.14 and 12.80 ppm, respectively. In Northeast China, there is a strong seasonal difference in anthropogenic heating emissions and vegetation activity, while the anthropogenic CO 2 is almost constant throughout the year in East; combined with the reduced seasonal variation in vegetation, the amplitude of seasonal variation of CO 2 at East is smaller than that in Northeast China (Fang et al., 2014). A similar trend of CO 2 seasonal variation was observed in the middle troposphere, with a high value in April and a low value in August ( Figure 5B). The discrepancies among the six regions, however, are generally reduced, indicating the strong dilution effect of CO 2 by vertical wind. In contrast, the CO 2 variations at the upper troposphere exhibited an opposite trend, with the maximum CO 2 concentrations observed in summer and the minimum CO 2 concentrations in winter ( Figure 5C). Such anomalies indicate that the upper troposphere CO 2 was less affected by anthropogenic and biogenic activities. In addition, it was also found that CO 2 concentrations in North China were generally higher than those in South China throughout the entire year. This The seasonal variations of CO 2 were found largest at the near surface, with a high value in spring and winter and a low value in summer and autumn ( Figure 6). In particular, an apparent east-west difference in CO 2 was observed across China, and the CO 2 concentrations in the coastal provinces of Southeast China were generally higher than those in Northwest China throughout the year. At the middle troposphere, however, such spatial differences almost disappeared, while the temporal variations remained consistent with the lower troposphere. When it comes to the upper troposphere at 100 hPa, both the spatial and temporal variations of CO 2 were the smallest among the three levels. The seasonal variation of the different levels was also reflected in the coefficient of variation, as shown in Supplementary Figure S1, with the largest CV of 1.45 found at 975 hPa and decreased to 0.79 and 0.36 at 500 hPa and 100 hPa, respectively. It is likely that the decreasing trend of CV is associated with the distance to the carbon sources and sinks at near surface. Therefore, we conclude that anthropogenic and vegetation activities are the main factors affecting the vertical distribution of CO 2 .

Diurnal variations
The mean diurnal amplitudes of CO 2 variations at different levels and seasons are shown in Supplementary Table S3. In terms of

Frontiers in Environmental Science
frontiersin.org vertical differences, the amplitude of the diurnal variation decreases from the near surface to the upper troposphere. The largest peak-totrough diurnal amplitudes (8.90 ± 6.98 ppm) were found at the near surface, while the middle and the upper levels of the troposphere exhibited a smaller diurnal variability, with an amplitude of 0.88 ± 0.72 ppm and 0.44 ± 0.32 ppm, respectively. Such differences can be attributed to the transportation and dilution effects of the upper atmosphere, which make the diurnal cycles of CO 2 at the middle and upper levels less affected by local sources and sinks. As for the smallest amplitudes of diurnal variation of CO 2 at the middle and upper troposphere, we are primarily interested in regional differences of daily CO 2 variation at the near surface. Overall, a uniform diurnal cycle of CO 2 was observed across the four seasons, with the trough occurring in the afternoon around 15: 00, then turns to accumulate during nighttime and reaching its maximum at about 06:00 the next morning (Figure 7). It is possible that the obvious diurnal variations in CO 2 throughout the whole year may be caused by active photosynthetic/respiratory fluxes in local vegetation (Li et al., 2004). As compared with spring and winter, summer and autumn witnessed relatively large amplitudes of CO 2 variation. As for the regional differences in diurnal variations of CO 2 , Northwest and Southwest China exhibit an overall smaller amplitude than East China and other subregions, and there is almost no clear diurnal cycle during winter, which could be related to weak human and vegetation activities at high altitudes in western China (Shi et al., 2020;Lin et al., 2021).  China. CO 2 concentrations at the near surface are generally higher than those at the top levels. CO 2 concentrations exhibit a strong spatial heterogeneity at the height of 850 hPa and below, with high values observed in eastern China and lower values in western regions, probably caused by the local emissions of human activity. This east-west spatial difference gradually decreases with height and disappears at 150 hPa, a phenomenon that may be explained by the dilution and mixing effects of vertical wind. At the height of 100 hPa and above, there is a distinct south-north difference in the spatial distribution of CO 2 , with high CO 2 concentrated at low latitudes and vice versa. This discrepancy may be attributed to the large-scale atmospheric circulation at the top troposphere. In order to explore the regional differences of CO 2 at different heights, the vertical profiles of CO 2 for the six regions are shown in Figure 9A. The near surface (levels 1-5) witnessed the greatest discrepancy among the six regions, and then it decreased to 0 at the middle troposphere (levels 6-11). Smaller differences were observed at the upper troposphere (levels 12-17). The dilution effect of vertical wind allows local CO 2 emissions to move upward with an increase in convective PBL, explaining the linear decrease in lower tropospheric CO 2 in East and Mid-South China (Newman et al., 2013). When it comes to the middle troposphere, the mixing effect of zonal and vertical winds is the dominant factor, which makes the CO 2 fully mixed and results in a small variation, whereas the continuous and regular atmospheric circulation contributes most to the smaller regional differences of CO 2 at the upper troposphere (Dargaville et al., 2000;Sohn et al., 2019).

Vertical variations
To better understand the seasonal variation of CO 2 at different heights, the linear regression method was used to remove the annual change rate from the calculation of the monthly mean CO 2 concentration. As shown in Figure 9B, the seasonal fluctuations are gradually diminishing with increasing height. The peak-totrough seasonal amplitude can be reached to 14.36 ppm at the height of 975 hPa, but it is reduced to less than 1 ppm above 50 hPa. In addition, a unimodal pattern of CO 2 variation was observed below 500 hPa, with a peak value occurring in April and a trough in August. Nevertheless, at 400 hPa or higher, the peak and trough of CO 2 occur 1 or 2 months later than at near surface. This further indicates that CO 2 in the middle troposphere is transported by ground emissions.

Driving forces of CO 2 variation at near surface
According to the model results of CT, we evaluated the influence of anthropogenic, biogenic, oceanic, and fire sources in each of the six regions. It is to be noted that the biogenic and oceanic modules act as a carbon sink, a capability that has gradually strengthened over the past decade ( Figures 10A, D). While the anthropogenic and fire activities exert a positive effect on local CO 2 and shows an increasing trend as well ( Figures 10B, C), in particular, the biogenic CO 2 shows

FIGURE 7
Diurnal cycles of CO 2 for the six sub-regions across China.

Frontiers in Environmental Science
frontiersin.org a similar pattern across the six regions, with maximum carbon sequestration occurring in September and the minimum occurring in May. The only difference is that there is another trough of carbon sequestration in January in the Northeast and the North, which is likely due to the inactive photosynthesis of vegetation in the cold season at high latitudes (Li et al., 2004). However, the regional differences in biogenic CO 2 are much smaller than those of anthropogenic CO 2 . Carbon emissions from fossil fuels in the East and Mid-South are higher than those in the Northwest and Southwest throughout the year. The maximum difference of anthropogenic CO 2 can reach 15 ppm in winter. Furthermore, a gentle flat variation of seasonal fossil fuel CO 2 was also observed in western China. Contrary to the variation pattern of anthropogenic and biogenic CO 2 , the fire and oceanic CO 2 exhibit the least regional differences with almost no seasonal variation, indicating that these two tracers of CO 2 in China are transported by wildfire emissions and oceanic absorption in other places.
Because of the large differences of simulated anthropogenic and biogenic CO 2 among the six regions, a completely independent dataset was used to verify the correlation between CO 2 concentrations and the fossil fuel emissions and the vegetation activity and to explore the spatial differentiation of this relationship. According to the results of correlation analysis, the near-surface CO 2 is significantly correlated with fossil fuel emissions on a national scale, with a correlation coefficient of 0.66 ( Figure 11B). As for the grid scale, except for some sparsely populated areas such as desert, Gobi, and high mountains, where the correlation coefficient is negative, other areas show significant positive correlations ( Figure 11A). Especially in parts of the Northwest and Southwest, the correlation coefficient is as high as 0.9, which indicates that fossil fuel emissions are the dominant factor affecting near-surface CO 2 in less developed regions. However, this strong correlation appears to be declining in the East and Mid-South (R ≈ 0.7), which can be partly explained by the intensive land use change and the massive cement production (Gregg et al., 2008;Herzog, 2009).
Although a general positive correlation was observed between near-surface CO 2 and fossil fuel emissions, however, there is no evidence to support that this is the cause of seasonal fluctuation in CO 2 . According to the study of Fung et al. (1997) and Van Der  Velde et al., 2013, the flux of δ 13 C in the process of carbon exchange between the atmosphere and biosphere is evidently greater than that between the atmosphere and ocean. In order to study the influence of terrestrial ecosystems on seasonal variation of near surface CO 2 , the LAI was used to conduct a correlation analysis. As shown in Figures 12A, B, a completely opposite trend was observed between the near-surface CO 2 and LAI, with a negative correlation coefficient being as high as −0.85. Photosynthesis of vegetation removes a relatively small amount of CO 2 before March. From April onward, the LAI increases gradually and leads to a decrease in CO 2 , with the largest photosynthesis CO 2 sequestration observed in August. Then, the monthly mean CO 2 increases with a decrease in the LAI from autumn to early spring of the following year. In terms of spatial distribution, approximately 96% of the grid cells show a negative correlation, with strong correlation mainly distributed in eastern China and weak correlation in part of western provinces ( Figure 12A). This east-west spatial difference may relate to the patterns of land use and land cover in China (Liu et al., 2008;Lin et al., 2021). The seasonal variation of CO 2 is highly dependent on the LAI in areas where forestland, grassland, and cropland are concentrated, while showing a weak or even positive correlation in sparse vegetation and bare land.
3.7 Driving forces of CO 2 variation at the middle and upper levels of the troposphere A similar approach was used to investigate whether fossil fuel carbon emissions and vegetation activity in China may have affected the CO 2 concentrations at the middle and upper troposphere. As shown in Figure 13A, a completely opposite trend was observed between the middle tropospheric CO 2 and LAI, with a negative correlation of −0.57. The variations of CO 2 are lagged by 4 months on average relative to the LAI. Our results are consistent with those reported in the Northern Hemisphere, where the shortest lag phase was observed in the low latitudes and the longest in the region between 30°N and 40°N (Cao et al., 2019). When it comes to the upper levels of troposphere, the variation of CO 2 exhibits a uniform pattern with the LAI (R = 0.92). It appears that vegetation carbon sequestration does not have an evident impact on upper tropospheric CO 2 . However, further research is required to determine why there is a strong correlation between CO 2 at high altitudes and surface vegetation. The strong influence of fossil-fuel CO 2 emissions on near-surface CO 2 tends to weaken at higher levels of the troposphere, with a correlation coefficient decreasing to 0.47 and 0.23, respectively, for the middle and upper troposphere ( Figures 13B, C). The reduced correlation indicates a dissipating

FIGURE 11
Correlation analysis between CO 2 and fossil fuel carbon emissions on the grid (A) and national scale (B) at near surface. It is to be noted that the grid cells filled with black dots indicate that the correlation is significant at the 95% level.

FIGURE 12
Correlation analysis between CO 2 and LAI on the grid (A) and national scale (B) at near surface. It is to be noted that the grid cells filled with black dots indicate that the correlation is significant at the 95% level.

Frontiers in Environmental Science
frontiersin.org effect of local carbon emissions on atmospheric CO 2 with increasing altitude. As a result, we can conclude that the upper levels of atmospheric CO 2 are less affected by local carbon sources and sinks; it is therefore important to take into account the influence of regional atmospheric circulation when conducting the driving force analysis (Sohn et al., 2019;Al-Bayati et al., 2020). Based on the wind data generated by ERA5, this study further explored the effects of atmospheric circulation in modulating the distribution of CO 2 at the upper atmosphere. The wind field was decomposed into zonal and vertical winds, and then their influence on CO 2 concentration at different heights was evaluated. At the height of 500 hPa, satellite observations indicate lower concentrations of CO 2 in the northwest and southwest of China. These areas have low CO 2 emission values and are dominated by the wind from the west and the southwest; the relatively low CO 2 concentrations from upstream countries, such as India, Pakistan, and Central Asia, have less impact on western China ( Figure 14B). In addition, the frequent air motion in the upward and downward directions facilitates the mixing of the upper and lower air ( Figure 14A), which assists in CO 2 dispersal. While the upward airflow was most prevalent in the Mid-south and eastern China, high CO 2 concentrations from the ground were carried to the upper levels, in combination with the westerly wind, resulting in high CO 2 concentrations in eastern China. Accordingly, the spatial distribution of CO 2 at the middle of the troposphere is the result of a combination of near-surface carbon emissions and zonal and vertical air motions (Cao et al., 2019). At the height of 100 hPa, an obviously zonal circulation stratification is observed with a uniform westerly wind, whereas the vertical airflow is weak (Figures 14C, D).
The distinct differentiation of CO 2 from north to south reflects the zonal average distribution of atmospheric circulation at the global scale. Therefore, the spatial patterns of CO 2 at the upper levels of the troposphere may largely be explained by zonal winds (Dargaville et al., 2000).

Comparison of this study with prior studies
Our results indicate an obvious increase in CO 2 at the near surface and the middle troposphere at 2.38 ppm/year and 2.34 ppm/ year, respectively, over the period from 2009 to 2019. The growth rate of CO 2 at both levels is higher than the rate averaged for global areas (2.20 ppm/year and 2.34 ppm/year, respectively, for the near surface and middle troposphere) and Central Asia (Supplementary Table S4). As a result of the rapid development of China's economy, such a trend is in accordance with the study of Peters et al. (2011), who reported a strong increment of atmospheric CO 2 well above the global mean in the 21st century. It should be noted that the rate derived from satellite products is much lower than the rate obtained from in situ measurements at the regional scale (Fang et al., 2014;Cao et al., 2017). This difference may be the results of different sampling points used in the studies. As traditional in situ observation systems are not able to obtain the data in complex terrain regions, where the concentrations of CO 2 are relatively low, using the values derived from satellite remote sensing may lower the regional averages. In addition, the different periods used may also have contributed to the differences. The growth rate of the study

FIGURE 13
Interannual variability of LAI and CO 2 (A) and the correlation analysis between CO 2 and fossil fuel emissions (B, C) at the middle (500 hPa) and upper (100 hPa) levels of the troposphere.

Frontiers in Environmental Science
frontiersin.org period including 2015, for example, is generally higher than that of the study period excluding 2015. Because the year 2015 is a typical El Niño-Southern Oscillation (ENSO) year, the growth rate of atmospheric CO 2 is expected to increase due to the anomalous sea surface warming, and covering the analysis period in that year may have yielded a higher rate of annual growth of CO 2 (Schwalm et al., 2011;Kim et al., 2016).

Conclusion
The improved accuracy of satellite CO 2 products has created new opportunities for studying the spatiotemporal variations of CO 2 in areas where field observations are inadequately sampled. In this study, the annual, seasonal, and diurnal variations of CO 2 at different heights across six sub-regions of China were examined. The results show consistently increasing of CO 2 with a magnitude higher than 2.1 ppm/year for all levels of the troposphere, and the seasonal cycles of CO 2 at the near surface and the middle troposphere are similar, with a high value in the early spring and a low value in summer, which exhibit an opposite trend to the upper troposphere. An obvious spatial heterogeneity was observed at the near surface, with the highest concentration of CO 2 occurring in East China and the lowest in Northwest China. This strong spatial heterogeneity, however, disappeared as the height increased and was replaced by a distinct south-north gradient difference at the upper troposphere. The diurnal variation of CO 2 was found to be the largest in eastern China, whereas the western part exhibits a smaller variation. In terms of vertical variation, the concentrations of CO 2 at the lower troposphere are generally higher than the values at the upper troposphere. Similar trends were also found in both the annual and seasonal variations of CO 2 . According to the driving mechanism analysis, the variation of CO 2 at the near surface is mainly affected by the anthropogenic and biogenic activities, whereas the regional atmospheric circulation dominates the spatial distribution of CO 2 at the upper troposphere.
Continuous monitoring of CO 2 is the foundation for understanding the spatial distribution of carbon sources and sinks and for studying the regional carbon cycle (Dettinger and Ghil, 1998;Hammerling et al., 2012;Peters et al., 2012). This study presented a comprehensive analysis of the spatiotemporal patterns of atmospheric CO 2 . The results show a large discrepancy of CO 2 concentrations and driving mechanisms among the six subregions of China. Thus, it is necessary to take the background concentration and the distinctive driving forces into consideration when formulating strategies for reducing carbon emissions (Zeng et al., 2013;Lin et al., 2021). Although the coarser temporal and spatial resolution of GOSAT may limit the representativeness of CO 2 at fine scale, it contributes to understanding the spatiotemporal pattern and the variability of CO 2 in China. With a more extensive CO 2 observation network established and the continuous improvements in the technology of numerical simulation, future

FIGURE 14
Multi-year average climatology of tropospheric circulation at the middle (500 hPa) and upper (100 hPa) troposphere, with (A, C) for the vertical wind velocity, and (B, D) for the zonal wind and the annual mean CO 2 concentration.

Frontiers in Environmental Science
frontiersin.org studies should integrate multi-source datasets from in situ and remote sensing measurements and model forecast to conduct a further in-depth assessment of atmospheric CO 2 in China.

Author contributions
Conceptualization, XJ and XD; methodology, XD; software, RD and QX; validation, JC; formal analysis, XJ; investigation, YH and YW; resources, SZ; data curation, SZ; writing-original draft preparation, XJ; writing-review and editing, XJ and XD; visualization, XJ; supervision, CJ and XD; project administration, MC; funding acquisition, YH and MC All authors have read and agreed to the published version of the manuscript.