Global land 1° mapping dataset of XCO2 from satellite observations of GOSAT and OCO-2 from 2009 to 2020

ABSTRACT A global mapping data of atmospheric carbon dioxide (CO2) concentrations can help us to better understand the spatiotemporal variations of CO2 and the driving factors of the variations to support the actions for emissions reduction and control. Greenhouse gases satellites that measure atmospheric CO2, such as the Greenhouse Gases Observing Satellite (GOSAT) and Orbiting Carbon Observatory (OCO-2), have been providing global observations of the column averaged dry-air mole fractions of CO2 (XCO2) since 2009. However, these XCO2 retrievals are irregular in space and time with many gaps. In this paper, we mapped a global spatiotemporally continuous XCO2 dataset (Mapping-XCO2) using the XCO2 retrievals from GOSAT and OCO-2 during the period from April 2009 to December 2020 based on a geostatistical approach that fills those data gaps. The dataset covers a geographic range from 56° S to 65° N and 169° W to 180° E for a 1° grid interval in space and 3-day time interval. The uncertainties of the mapped XCO2 values are generally less than 1.5 parts per million (ppm). The spatiotemporal characteristics of global XCO2 that are revealed by the Mapping-XCO2 are similar to the model data obtained from CarbonTracker. Compared to the ground observations, the overall standard bias is 1.13 ppm. The results indicate that this long-term Mapping-XCO2 dataset can be used to investigate the spatiotemporal variations of global atmospheric XCO2 and can support studies related to the carbon cycle and anthropogenic CO2 emissions. The dataset is available at http://www.doi.org/10.7910/DVN/4WDTD8 and https://www.scidb.cn/en/detail?dataSetId=c2c3111b421043fc8d9b163c39e6f56e.


Introduction
The continuing increases in atmospheric CO 2 concentrations caused by anthropogenic emissions have had a significant impact on global warming (Friedlingstein et al., 2020). In recent years, many countries have put forward strategies to control and reduce anthropogenic carbon emissions and have pursued efforts to achieve the goals of the Paris Agreement (Coulter, Canadell, & Dhakal, 2008). High-quality observations are needed to detect the increases or decreases of anthropogenic CO 2 concentrations to evaluate the effectiveness of emission reduction policies.
Currently, the Japanese Space Agency's Greenhouse Gases Observing Satellite (GOSAT) and NASA's Orbiting Carbon Observatory (OCO-2) have obtained atmospheric column-averaged CO 2 dry air mole fraction (XCO 2 ) data that were derived from the observed hyperspectral data starting from 2009 (Crisp et al., 2004;Yokota et al., 2009). Many studies have shown that these XCO 2 retrievals have the potential to study terrestrial CO 2 sources and sinks and can detect the CO 2 enhancements that are caused by human activities Hakkarainen, Ialongo, Maksyutov, & Crisp, 2019;Hakkarainen, Ialongo, & Tamminen, 2016;Janardanan et al., 2016;Keppel-Aleks, Wennberg, O'Dell, & Wunch, 2013;Kort, Frankenberg, Miller, & Oda, 2012;Lei et al., 2017;Nassar et al., 2017;Schneising et al., 2013;Schwandner et al., 2017;Shim, Han, Henze, & Yoon, 2019;Wang et al., 2018). Globally mapped XCO 2 data, which fill the gaps of XCO 2 retrievals in space and time that are induced mainly by clouds and aerosols, have been developed in several previous studies (Nguyen, Shivadekar, Chukkapalli, & Halem, 2020;Zeng et al., 2014Zeng et al., , 2016. It has been demonstrated that globally mapping of XCO 2 data effectively reveals the spatial and temporal variations of atmospheric XCO 2 and detects the driving factors of the XCO 2 variations by combining the parameters that affect CO 2 releases and uptakes and the anthropogenic emissions inventories He, Lei, Zeng, Sheng, & Welp, 2020a;Sheng, Lei, Zeng, Rao, & Zhang, 2021;Yang, Lei, Zeng, He, & Zhong, 2019). However, the current mapping XCO 2 datasets are still inadequate, such as their irregular distributions (Nguyen et al., 2020), coarse grid resolutions , and short-term periods from 2009 to 2016 (He et al., 2020b;Zeng et al., 2014Zeng et al., , 2016 although there are more XCO 2 retrievals available from GOSAT and OCO-2. Additionally, we also found that the mapping XCO 2 dataset generated from the satellite observations of the SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY (SCIAMACHY) is less efficient in detecting the CO 2 variations that are induced by anthropogenic emissions due to its lower resolution and data accuracy (He et al., 2020b). Therefore, in this paper, we used long-term XCO 2 retrievals combining GOSAT and OCO-2 data from 2009 to 2020 to map a new global XCO 2 data.
We used the newly released XCO 2 retrievals over land areas from 2009 to 2020 from GOSAT (from April 2009 to August 2014) and OCO-2 (from September 2014 to December 2020). The differences between GOSAT and OCO-2 observations, including their sensitivities to CO 2 changes and their resolutions in space and time, are calibrated (Kataoka et al., 2017;Kulawik et al., 2016). By using these integrated XCO 2 data from GOSAT and OCO-2 observations, we generated a global land mapping XCO 2 dataset with a 1° grid and 3-day resolution from 2009 to 2020 based on the spatiotemporal geostatistical method developed by Zeng et al. (2014 and. Users can utilize this dataset to investigate XCO 2 variations at global and regional scales to detect and assess anthropogenic CO 2 emissions and ecological CO 2 fluxes. Example cases include those presented in Bie et al. (2018), He et al. (2018He et al. ( , 2020aHe et al. ( , 2020b, and Sheng et al. (2021).

Methods
We integrated the satellite XCO 2 retrievals from GOSAT and OCO-2 and generated a global land XCO 2 product (Mapping-XCO 2 ) with a 1° by 1° grid and 3-day temporal resolution. Figure 1 shows the flowchart for producing the Mapping-XCO 2 , which includes the data integration and gap-filling processes using satellite XCO 2 data. In Section 2.1, we provide a description of the CO 2 data that were obtained from satellite, model, and ground observations. In Section 2.2, we integrated a long-term XCO 2 dataset by combining GOSAT and OCO-2 retrievals from April 2009 to December 2020. In Section 2.3, we focus on global lands that range from 56° S to 65° N and 169° W to 180° E to fill the large gaps. The spatiotemporal continuous mapping XCO 2 data are generated by applying the spatiotemporal kriging method to the integrated XCO 2 data.

Data source
To produce the global mapping XCO 2 dataset, we collected the XCO 2 retrievals from satellite observations (GOSAT and OCO-2) from April 2009 to December 2020. The parameters of the XCO 2 retrievals from GOSAT and OCO-2 are listed in Table 1. The XCO 2 data from GOSAT observations is the latest Level 2 Lite data product (v9r) derived from the Atmospheric CO 2 Observations from Space (ACOS) retrieval algorithm, which was produced by the ACOS project based on the GOSAT L1B products with calibrated radiances and geolocations (Crisp et al., 2011). For the OCO-2 retrieval, we used the latest version (v10r) of the OCO-2 Level 2 Lite data product's bias-corrected XCO 2 full physics retrievals (Kiel et al., 2019). The retrieval algorithm of the GOSAT product is consistent with what was used to create the OCO-2 data product (O'Dell et al., 2018). Therefore, we can use these two products to generate a longterm dataset without considering the difference of retrieval algorithms. These data products are both obtained from Goddard Earth Science Data Information and Services Center (GES DISC) at the National Aeronautics and Space Administration (NASA) available at https://oco2. gesdisc.eosdis.nasa.gov/data/. They were filtered by cloud-screening and specified criteria that were introduced by releasing product teams to ensure good quality of XCO 2 retrievals. In our studies, only the high gain soundings over land with "good" observations were retained according to the "xco2_quality_flag" and "land_fraction" variables in the dataset files. The posterior errors of the satellite XCO 2 data are less than 2 ppm for GOSAT and 1.2 ppm for OCO-2. The XCO 2 data from GOSAT and OCO-2 are hereafter referred as GOSAT-XCO 2 and OCO2-XCO 2 , respectively.
CarbonTracker simulates global atmospheric CO 2 mole fractions in a spatial resolution of 3° longitude by 2° latitude based on the Transport Model 5 (TM5) offline atmospheric tracer transport model driven by meteorological fields (Jacobson et al., 2020;Peters et al., 2007). The results of version CT2019B provide global CO 2 mole fractions in 25 layers throughout the entire atmospheric column every 3 hours from 2000 to March 2019. These data can be sampled at the observation times of satellite observation to adjust the a priori information of satellite XCO 2 retrievals (Connor et al., 2008;Jacobson et al., 2020;Rodgers, 2004). We used the CT2019B CO 2 profile data from 2009 to 2019 as the reference data to adjust the differences of satellite XCO 2 retrievals from GOSAT and OCO-2 included by the a priori profiles and overpass times. We also collected model XCO 2 data at 13:30 (LST) every day to validate Mapping-XCO 2 . The model data are available in the Earth System Research Laboratories (ESRL) of the National Oceanic and Atmospheric Administration (NOAA) at https://www.esrl.noaa.gov/gmd/ccgg/carbontracker/.
We compared the mapping XCO 2 dataset with the ground-based XCO 2 measurements that were obtained from the Total Carbon Column Observing Network (TCCON) for data validation. TCCON measures high spectral and temporal resolution spectra of the direct solar transmission in the near infrared spectrum by using ground-based Fourier transform spectrometers (FTSs) (Wunch et al., 2011a). From these spectra, accurate and precise column-averaged abundances of CO 2 , CH 4 , N 2 O, HF, CO, H 2 O, and HDO are retrieved, which provide the primary validation dataset for retrievals from space-based instruments. We obtained the XCO 2 dataset spanning from 2004 to 2020 that was retrieved by GGG2014 from the TCCON data archive (https://tccondata.org/) . In this paper, we used 20 TCCON sites listed in Table S1 and calculated the averaged XCO 2 data at 13:30 ± 2 h for every 3 days.

Adjusting the a priori CO 2 profiles of satellite XCO 2 retrievals
In the algorithm of retrieving XCO 2 data from satellite observations, the CO 2 a priori profiles are used as the a priori information, and the column averaging kernel represents the response of satellite sensor to the CO 2 mixing ratio (Crisp et al., 2011;Miller et al., 2007;O'Dell et al., 2018;Osterman et al., 2017). When using long-term XCO 2 data from different satellite observations, it is necessary to consider the differences between the averaging kernel function and the a priori CO 2 profile that are used by the retrieval algorithm. To calibrate the differences between GOSAT and OCO-2 observations that are due to the a priori profile, we used the CO 2 profile data from CarbonTracker as the reference data to adjust the XCO 2 retrievals to a common a priori CO 2 profile. The CO 2 profiles from the model are searched by using 3°×2° grids where the satellite observations are located. These profiles are then interpolated to 20 layers, which are the same levels as those of the satellite XCO 2 retrievals. As shown in Equations (1) and (2), the differences of the vertical CO 2 profiles between the model and satellite data are converted to the form of XCO 2 variations through calculations with an averaging kernel function and pressure weighting function (Connor et al., 2008;Rodgers, 2004).
where XCO2 adj;t is the adjusted XCO 2 value at observation time t obtained by adding the XCO 2 variation (δ1) to the original XCO 2 value (XCO2 ori;t ), h is the pressure-weighted function, A is the column averaging kernel calculated from the retrieval algorithm, I is a unit matrix, X a;t represents the a priori CO 2 profiles of the satellite observations and X M;t represents the model CO 2 profile sampled at observation time t, which has been interpolated into 20 layers.

Adjusting observing time of satellite XCO 2 data
The atmospheric CO 2 concentrations vary with the time in a day which are affected by the diurnal variations of atmospheric emissions and biosphere-atmosphere exchanges (Gurney et al., 2002;Law, Rayner, Steele, & Enting, 2002). As shown in Table 1, the overpass times of GOSAT and OCO-2 observations are 13:00 and 13:36, respectively. The differences of the XCO 2 observations that are caused by different satellite observation times are not trivial and cannot be ignored. We selected the local standard time (LST) at 13:30 as the reference time and calculated the time-shifted coefficients of the model XCO 2 data between the satellite observation times and 13:30 (LST), which are the ratios between the XCO 2 value at different times (He et al., 2020b;Wang et al., 2014). According to Equation (3), the time-shifted coefficients are applied to adjust the satellite XCO 2 observations.
where XCO2 adj;rt is the converted XCO 2 value at reference time (rt), t is the time of the satellite observation, rt is the reference time 13:30 (LST), X M;rt and X M;t are the CO 2 profile data from CT2019B at the corresponding times, and δ2 is the adjusted XCO 2 variation. The XCO 2 variations (e.g., δ1,δ2) in the satellite XCO 2 data that are used in the adjustments of the a priori CO 2 profiles and observation times are shown in Figures S1-S4. Since the CO 2 profile from CarbonTracker was updated only until March 2019, the satellite XCO 2 data after this period were adjusted based on the monthly 1° gridded XCO 2 variations from 2010 to 2015. The equation used is as follows: where adjXCO2 loc;m is the adjusted XCO 2 data by adding the satellite XCO 2 data (XCO2 loc;m ) and XCO 2 variations for the corresponding 1° grid (loc) and month (m).

Unifying spatial and temporal scales of satellite observations
Satellite measurements of GOSAT and OCO-2 have different temporal and spatial resolutions, which are 3 days and 10.5 km for GOSAT and 16 days and 1.29 × 2.25 km for OCO-2, respectively (Kataoka et al., 2017). We unified the spatial and temporal scales of the satellite observations by averaging the XCO 2 values within 10.5 km and 3 days according to Equation (6).
where XCO2 int;rt is the integrated XCO 2 data at 13:30 (LST) and N is the number of satellite observations within 10.5 km and 3 days. The long-term consistent XCO 2 dataset was obtained by integrating the adjusted XCO 2 data from GOSAT and OCO-2, in which the XCO 2 data from April 2009 to August 2014 consist of GOSAT observations and from September 2014 to December 2020 consist of OCO-2 observations. For the overlap period of GOSAT and OCO-2 observations, we only used the XCO 2 data from OCO-2. This is because the adjustments of GOSAT-XCO 2 have a slight increase of 0.05 ppm every year ( Figure S1), which are induced by the secular increases in the differences between the a priori CO 2 profiles of the satellite and model CO 2 profiles (Wunch et al., 2011b). Figure 2 shows the time series of integrated XCO 2 data from April 2009 to December 2020. There are no significant differences in the integrated XCO 2 data between these two periods. From 2009 to 2020, the annual increases in the atmospheric CO 2 concentrations were approximately 2.43 ppm.

Methodology of generating Mapping-XCO 2 dataset
Aiming at resolving the issues of data gaps and irregular distributions of the satellite XCO 2 observations, we applied a spatiotemporal geostatistical method to estimate the XCO 2 values at the centers of 1° grids using the integrated XCO 2 data He et al., 2020a;Zeng et al., 2014Zeng et al., , 2016. Considering the regional characteristics of the XCO 2 variations that are influenced by atmospheric CO 2 biosphere-atmosphere exchange and transport (Bai et al., 2021;Peters et al., 2007;Tang, Zhang, Zhang, Liu, & Bai, 2020;Wunch et al., 2013), we separately generated spatiotemporal continuous XCO 2 data over different continental regions. The regions north of 60° were not processed due to the smaller numbers of satellite observations and large data uncertainties. As shown in Figure 3, the global land area is divided into 7 mapping regions. Africa is divided into two regions because it spans both the Northern and Southern Hemispheres by using 0° as the boundary, in which the seasonal CO 2 variations are significantly different. South America is divided at 15° S due to the different distributions of terrestrial ecosystems from north to south.  According to Zeng et al. (2014), we assume that the XCO 2 data observed by satellites form a spatial and temporal random function Z ¼ Z s; t ð Þjs 2 S; t 2 T f g; s i ; t i ð Þ; i ¼ 1 . . . n. The spatial and temporal variations of XCO 2 can be further modeled by decomposing them into an inherent deterministic trend and stochastic residual component, which are shown in Equation (7). The deterministic XCO 2 component consists of seasonal CO 2 changes and annual CO 2 increases that can be modeled by the curve fitting method. We used the curve fitting function shown in Equation (8), which is a combination of a linear function and harmonic functions (Masarie & Tans, 1995;Wunch et al., 2013;Zeng et al., 2016).
where m s; t ð Þ is the deterministic trend of XCO 2 , R s; t ð Þ is the residual component, t is a time unit of 3 days (122 cycles every year), and ω represents the temporal periods in a year calculated by 2π/122. The period from April 2009 to December 2020 is divided into 1426 time-units. The parameters a 0 ; a 1 ; β i ; γ i are obtained by least squares fitting.
We obtained the deterministic spatiotemporal trends of the XCO 2 values by fitting the integrated XCO 2 data in each 5° latitude band. After removing these trends from the XCO 2 data, the residual component R s; t ð Þ is a stochastic field that satisfies the standard normal distribution. The component is assumed to be homogeneous and locally stationary for the spatiotemporal correlation structures of the integrated XCO 2 data within each mapping region (Schabenberger & Gotway, 2017). We used 3 days and 100 km as the temporal and spatial steps, respectively, to calculate the spatiotemporal variograms in each region based on the sum of the product-sum model and an extra global nugget model (De Iaco, Myers, & Posa, 2001;Zeng et al., 2014Zeng et al., , 2016. The parameters of the variogram models over different mapping regions are listed in Table 2.
Based on these variogram models, a spatiotemporal kriging method with a moving cylinder kriging neighborhood was implemented to estimate the values at the centers of the 1° grids. A detailed description of the spatiotemporal kriging method can be found in Zeng et al. (2014 and. In this study, the initial search ranges of the cylinder are set to 300 km in space and 15 units (45 days) in time. The increment lags are 10 km and 1 timeunit until the numbers of available data exceed than 20 points. As a result, the mapping XCO 2 data are calculated by adding the predicted value and the fitting trend in each grid, which is hereafter referred to as Mapping-XCO 2 . The mapping uncertainties, as quantified by the root kriging variances, are also calculated for each grid and 3-day time unit according to Zeng et al. (2014 and.

Mapping-XCO 2 dataset
The global 1° land mapping XCO 2 dataset (Mapping-XCO 2 ) spans the period from April 2009 to December 2020. The data product is provided in GeoTIFF format and includes two temporal resolutions: 3 days and a month. The 3-day data files include the gridded mapping XCO 2 and mapping uncertainty values, which are named "MappingXCO2_Date.tif" and "MappingUncertainty_Date.tif", respectively. The flag "Date" is defined as the date ID of 1426 time units starting on 20 April 2009. The monthly data files only include the monthly mean XCO 2 data and are named "MappingXCO2_YYYY_MM.tif". The numbers "YYYY" and "MM" represent years and months, respectively. The dataset domain covers global lands that range from 56° S to 65° N and 169° W to 180° E. The spatial reference for the dataset uses Geographic Lat/Lon. The unit of the XCO 2 data is ppm while the no data values were defined to be NaN. The spatiotemporal characteristics of the global XCO 2 data obtained from the Mapping-XCO 2 are shown in Figures 4 and 5. The high XCO 2 values in the Northern Hemisphere are mainly distributed in regions of western United States, East Asia, and Middle East. The XCO 2 values in the Southern Hemisphere are lower than those in the Northern Hemisphere. Both hemispheres exhibit significant long-term increases, with growth rates fluctuating between 1.5 and 3.5 ppm. These increases are influenced by the CO 2 fluxes from terrestrial ecosystems and anthropogenic emissions (Sheng et al., 2021). The seasonal changes are large in the Northern Hemisphere, and the maximum XCO 2 levels occur during the spring months. The regions at northern high and mid-latitudes show significant seasonal cycles with amplitudes of approximately 8 ppm. The XCO 2 data rise steadily without any apparent seasonal changes in the Southern Hemisphere.    Satellite XCO 2 data of GOSAT and OCO-2 are irregularly distributed in space and time due to geophysical factors such as aerosols and clouds, which can be seen in Figures S4 and S5. Influenced by persistent cloudiness and low signal levels, the numbers of satellite observations that are available from GOSAT are very small in regions near the equator and at latitudes above 45° N. The Mapping-XCO 2 data fill the gaps in the distribution of satellite XCO 2 data. The global XCO 2 pattern of the mapping XCO 2 data agrees with that of the original XCO 2 data. As shown in Figure  S7, the Mapping-XCO 2 data exhibit smoother features than GOSAT-XCO 2 and OCO2-XCO 2 , because the extreme XCO 2 values from satellite observations are filtered or averaged in the process of kriging prediction.
The mapping uncertainty largely depends on the quantity of XCO 2 data. Comparing Figure 7 and Figure S5, the spatial distributions of the higher uncertainty areas are consistent with those areas with data gaps or fewer observations. Larger uncertainties of up to 1.5 ppm are mainly distributed at the middle and high latitudes north of 50° N due to the sparse satellite observations. There is higher uncertainty from 2009 to 2014 relative to the period from 2015 to 2020, because the numbers of OCO-2 observations are larger than those observed by GOSAT. Additionally, in the regions north of 50° N and from 20° N to 20° S, the mapping uncertainties have obvious seasonal fluctuations in Figure 8. The changing cycles vary by location and are related to the seasonal variations in the number of available satellite observations. For example, maximum values appear in wintertime in latitudes north of 40° N and at southern low latitudes (from 0° to 30° S), and in summertime in mid-latitude regions.

Comparison with CT2019B-XCO 2
The model XCO 2 data from CarbonTracker (CT2019B-XCO 2 ) are resampled from a 3° × 2° grid to a 1° grid to compare the spatiotemporal characteristics of Mapping-XCO 2 . The long-term global XCO 2 patterns of Mapping-XCO 2 and CT2019B-XCO 2 have similar spatial distributions. Comparing Figure 4 and Figure S6, Mapping-XCO 2 shows more detailed features at regional scales, such as in the Qinghai-Tibet Plateau in China, Western United States, Central Africa, and Middle East regions. The seasonal XCO 2 patterns of Mapping-XCO 2 ( Figure 6) and CT2019B-XCO 2 ( Figure S7) are generally consistent. The main differences are located in tropical Africa during MAM and DJF, and in southern Asia during JJA. From April 2009 to March 2019, the overall bias of the monthly gridded XCO 2 values between Mapping-XCO 2 and CT2019B-XCO 2 is -0.37 ± 1.11 ppm. Their differences show different spatial distributions between the two periods of GOSAT and OCO-2 observations, as shown in Figure 9. The large differences exceeding 1 ppm are mainly distributed in regions north of 50°N and in the 15°S-25°N latitude zone. These regions have fewer number of satellite observations and higher mapping uncertainties, which are shown in Figure S5 and Figure 7. During the GOSAT observation period from 2010 to 2014, the quantities and spatial extents of the differences in East Asia are larger than those in the period from 2015 to 2018.
Further, we extracted the grids with mean absolute differences greater than 1.5 ppm in the regions north of South America, southern Eurasia, and Central Africa. The differences show significant seasonal changes in these three regions. The period containing the maximum values is related to mapping uncertainties and the numbers of satellite observations. Figure 10 shows that the differences in southern Eurasia have a negative maximum value in summer when few data are observed and the mapping uncertainties are large. In contrast, the larger differences in northern South America and Central Africa  appear in other seasons, as there are relatively large numbers of observations during the summer months. In addition, the differences from 2009 to 2014 are larger than those after 2015, especially in southern Eurasia, because the number of OCO-2 observations is much higher than the number of GOSAT observations, as shown in Figure 10(c).

Comparison with TCCON-XCO 2
We computed the 3-day averaged data from GOSAT-XCO 2 , OCO2-XCO 2 and gridded XCO 2 data of Mapping-XCO 2 obtained by ±5° boxes centered on the TCCON sites. Only the averaged XCO 2 data from GOSAT and OCO-2 that satisfy at least 6 observations within 3 days are retained. As shown in Figure 11 and S8, both the Mapping-XCO 2 and satellite XCO 2 data have good relationships with TCCON XCO 2 data (R > 0.9). The fitting slope of the Mapping-XCO 2 versus TCCON-XCO 2 data is closer to 1 than that of the satellite XCO 2 data obtained from GOSAT and OCO-2. The overall bias between the Mapping-XCO 2 and TCCON-XCO 2 data is 0.10 ± 1.13 ppm, and the mean absolute error is 0.88 ppm, which are slightly less than those when the TCCON-XCO 2 data are compared with GOSAT-XCO 2 and OCO2-XCO 2 . These results are probably related to the smoothing effect of the kriging Figure 10. Time series of regional characteristics from 2010 to 2018 in red boxes shown in Figure 9. (a) The mean difference of the monthly gridded XCO 2 between Mapping-XCO 2 and CT2019B-XCO 2 ; (b) regional mapping uncertainty; (c) the number of satellite observations from GOSAT and  algorithm used in the mapping process. These results indicate that the Mapping-XCO 2 data are consistent with the ground observations and that their characteristics are more similar to OCO2-XCO 2 .
We further compared the temporal features of the Mapping-XCO 2 , satellite XCO 2 data, and TCCON data. Figure 12 shows time series of the XCO 2 data at 4 representative sites (e.g., Garmisch, DE, at 47.48°N; Lamont, US at 36.6°N; Park Falls, US, at 45.94°N; Soga, Japan, at 33.24° N; and Wollongong, Australia at 34.41°S), which are in different continents. The Wollongong site is located in the Southern Hemisphere and is less affected by vegetation activities, so there are no obvious seasonal CO 2 changes. For other sites located in the Northern Hemisphere, the CO 2 changes show significant seasonal fluctuations. The temporal CO 2 variations obtained from the Mapping-XCO 2 data are consistent with those observed by TCCON. Compared to the TCCON data, the biases of the Mapping-XCO 2 data at those sites are lower than those of GOSAT-XCO 2 and OCO2-XCO 2 . The results indicate that the mapping data can be used to characterize regional XCO 2 levels by decreasing the influence of large errors in the satellite XCO 2 observations. However, the Mapping-XCO 2 values at Garmisch are 0.76 ppm higher than those of the TCCON data, and the mean absolute difference is 1.02 ppm. The deviations between theMapping-XCO 2 and satellite XCO 2 data are larger than those at other sites, which are related to the small numbers of satellite observations. Therefore, the spatiotemporal scales and uncertainties of the Mapping-XCO 2 data are still not adequate to reveal the instantaneous variations of CO 2 concentrations that are observed by ground stations.

Usage notes
The global 1° land mapping gridded XCO 2 (Mapping-XCO 2 ) data from 2009 to 2020 are produced by satellite XCO 2 retrievals of GOSAT and OCO-2, which fill the data gaps of satellite observations. The comparison of Mapping-XCO 2 with CarbonTracker and TCCON data shows that the mapping XCO 2 data can effectively reveal the temporal and spatial characteristics of global atmospheric CO 2 concentrations. It should be noted that the accuracy of the spatiotemporal kriging method largely depends on the number of valid data points. Due to the small number of satellite observations, there is higher mapping uncertainty and larger differences between the Mapping-XCO 2 and model data in the regions that are located at middle latitudes north of 50° N, north of South America, and Central Africa. The mapping uncertainties during 2009 to 2014 that correspond to GOSAT observations are higher than those for the period of OCO-2 observations, because the data volume of OCO-2 is much greater than that of GOSAT. For regional studies, users can refer to the uncertainty analysis in this study and use the mapping uncertainties to control the data quality.
The mapping XCO 2 dataset is produced by a data-driven method employing satellite observations and uses a geostatistical model to simplify the complicated processes of CO 2 atmospheric transport and CO 2 flux. Although these processes can be modeled based on mechanistic models used in atmospheric physics and ecological simulations, it is still very challenging to track and capture CO 2 footprints by using model simulations. This is because the modeling results are very sensitive to changes in the input parameters that are complicated and variable under different atmospheric conditions. Small disturbances of the input parameters can blur CO 2 footprints and introduce large biases. Satellite observations, as a data source for mapping XCO 2 , can capture the CO 2 footprints with systematic biases. Therefore, the Mapping-XCO 2 data can better describe the pattern of CO 2 variations in space and time and track the CO 2 footprint variations through statistical modeling of satellite observations. Mapping-XCO 2 data provide valuable information on the long-term atmospheric XCO 2 in space and time. Users can utilize this dataset to study the carbon cycle and estimate anthropogenic CO 2 emissions at global and regional scales. For example, previous studies have shown that global mapping of XCO 2 datasets can be used to detect abnormal XCO 2 caused by extreme events , analyze the responses of CO 2 concentrations to anthropogenic emissions and the absorption of terrestrial ecosystems (He et al., 2020a;, and estimate regional anthropogenic emissions (Sheng et al., 2021;Yang et al., 2019).
Zhao-Cheng Zeng is currently an Associate Research Scientist at Caltech. He obtained his PhD in 2016 from the Chinese University of Hong Kong. His research interests relate to the development of theories and algorithms in two fundamental areas of remote sensing: (1) radiative transfer and (2) inverse modeling/retrieval. More information about his research can be found at http://web.gps.caltech.edu/~zcz/.
Weiqiang Rao received the B.S. degrees in University of Electronic Science and Technology of China, Chengdu, China, in 2017. He is currently working toward the Ph.D. degree in Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. His current research interest is hyperspectral image processing, especially anomaly and target detection algorithms based on deep learning.