Evaluation of Multi-Source Satellite XCO₂ Products over China Using the Three-Cornered Hat Method and Multi-Reference Comprehensive Comparisons

Highlights

What are the main findings?

• OCO-series XCO₂ products (OCO-2/3) consistently outperform GOSAT and GOSAT-2 across China.
• The Three-Cornered Hat (TCH) proves highly robust for multi-satellite XCO₂ uncertainty analysis.

What are the implications of the main findings?

• Clarifying the data-priority direction lays a superior data foundation for accurate CO₂ monitoring across China.
• Breaking through the limitations of traditional assessment provides a powerful tool for gaining deeper insights into the reliability of satellite XCO₂ products.

Abstract

As one of the most important greenhouse gases, carbon dioxide (CO₂) exhibits spatiotemporal variations that directly affect the accuracy of global carbon inventories. In recent years, multiple satellites have successively been deployed for observing the column-averaged CO₂ dry-air mole fraction (XCO₂). However, these satellites perform quite differently, so it is crucial to evaluate their XCO₂ products systematically for both scientific and practical reasons. Most existing studies rely on ground-based observations or the CarbonTracker (CT) model data as reference benchmarks. Nevertheless, because ground-based stations are sparsely distributed and model data are subject to prior errors, biases may be introduced into the evaluation results. In contrast, the Three-Cornered Hat (TCH) method can estimate the relative errors of multi-source data without true values. Based on this, the current study systematically evaluates the XCO₂ products of the four following satellites—Greenhouse Gases Observing Satellite (GOSAT), GOSAT-2, Orbiting Carbon Observatory 2 (OCO-2), and OCO-3—over China by integrating the TCH method, ground-based observations and CarbonTracker model data. The results show that the monthly coverage of the four satellite XCO₂ products in China is limited. In terms of overall performance, the OCO-series outperforms the GOSAT-series, with OCO-3 showing the relatively best performance. Additionally, the TCH method proves to be applicable and reliable for uncertainty analysis of XCO₂ data. This study provides a new perspective for the quality grading and fusion application of multi-source satellite XCO₂ data, and is of great significance for carbon assimilation models.

1. Introduction

The continuing increase in greenhouse gas concentrations in the global atmosphere caused by human activities such as the burning of fossil fuels has significantly enhanced the greenhouse effect, prompting widespread concern from the international community [1]. The rising concentration of carbon dioxide (CO₂) represents the principal driver of global warming. Since the onset of Industrial Revolution, anthropogenic activities have emitted substantial quantities of CO₂ into the atmosphere, elevating its global concentration from approximately 280 ppm in the mid-19th century to more than 420 ppm in 2025. This increase has been accompanied by the intensification of extreme heat events, drought, heavy precipitation and glacial retreat [2,3,4]. To mitigate the continuous rise in atmospheric CO₂ concentrations, a series of international emission reduction and control initiatives have been implemented, exemplified by the 2016 Paris Agreement [5]. Consequently, accurately monitoring and quantifying the spatiotemporal dynamics of greenhouse gas emissions is not only an essential prerequisite for the effective implementation of the Paris Agreement, but also a strategic cornerstone for strengthening global climate governance and promoting the green transformation of industrial systems.

Early measurements of atmospheric CO₂ concentrations relied primarily on globally distributed surface stations and their networked systems [6,7], such as the Total Carbon Column Observing Network (TCCON). These stations provide authoritative true-value references for XCO₂ through high-precision and traceable measurements [8], serving as the benchmark datasets for remote sensing products and model data. However, surface observations are inherently constrained by sparse station coverage, limited representativeness and sensitivity to local sources, sinks and complex terrain, making large-scale continuous monitoring impossible [9]. In contrast, satellite-based remote sensing, characterized by wide spatial coverage and high spatiotemporal resolution, has become the cornerstone of top-down emission estimation, enabling the generation of global XCO₂ products [10,11,12].

At present, global XCO₂ remote-sensing missions can be grouped into “two passive, one active” approaches. The first is the thermal infrared hyperspectral sensors, exemplified by the Atmospheric Infrared Sounder (AIRS). AIRS adopts infrared grating array spectroscopy technology, with spectral channels covering the 650–2700 cm⁻¹ region, and is widely used in large-scale observation of global atmospheric XCO₂. Study have shown that the CO₂ measurement channel of AIRS is at 12.6 μm, with a measurement uncertainty of approximately 3–7 ppm [13]. The second category is short-wave near-infrared hyperspectral sensors. It utilizes the radiation spectra of CO₂ absorption bands (mainly including the weak absorption band at 1.6 μm and the strong absorption band at 2.0 μm) detected by hyperspectral sensors. After eliminating interference from various influencing factors, it extracts CO₂ content information. Furthermore, it combines the atmospheric molecular number concentration estimated from the oxygen A-band (O₂-A) absorption band (0.76 μm) to further calculate the XCO₂ [14]. The third is the active integrated path differential absorption (IPDA) lidar. It detects XCO₂ based on a dual-pulse differential absorption mechanism, enabling day-and-night operation. It features excellent vertical detection capability and low susceptibility to aerosol and cloud scattering. By conducting atmospheric vertical detection along the flight track, it can acquire fine-scale and high-precision greenhouse gas distribution information, with the accuracy of XCO₂ better than 1 ppm [15].

Currently, near-infrared hyperspectral technology serves as the primary approach for spaceborne greenhouse gas detection [16]. Satellites such as Greenhouse Gases Observing Satellite (GOSAT) [17,18,19], GOSAT-2 [20,21,22], Orbiting Carbon Observatory 2(OCO-2) [23,24,25], OCO-3 [26,27,28], and Tansat [29,30,31], equipped with mature retrieval algorithms and comprehensive validation systems, have become indispensable core data sources in the field of carbon emission research. Despite the wide variety of satellite XCO₂ products, their accuracy still requires systematic verification. The evaluation results not only provide a basis for users to select data but also feed back directions for improvement to retrieval algorithms. Existing studies can be categorized into three types: (1) independent accuracy validation for single-satellite products (e.g., GOSAT, OCO-2) [32,33,34,35,36]; (2) cross-comparison of multi-satellite products [24,37,38]; and (3) comparison between satellite observations and model data [18,19,25]. The aforementioned research efforts have collectively provided critical support for the continuous optimization and wide application of CO₂ remote sensing data. However, existing methods still exhibit notable limitations. Single-satellite validation is constrained by the sparsity and uneven distribution of ground stations. Although multi-satellite cross-comparison can reflect differences among products, they often fail to isolate common-mode systematic biases, making it difficult to diagnose specific error sources. Furthermore, the comparative analyses between satellite observations and model data are easily affected by model-inherent uncertainties, thereby reducing the reliability of the validation results. However, the inherent limitations of the above methods make it difficult to quantify the uncertainty of satellite XCO₂ data comprehensively and objectively. Therefore, uncertainty analysis is essential for a comprehensive assessment of the reliability and robustness of satellite XCO₂ data, and helps to select the most suitable XCO₂ products for carbon flux inversion, emission inventory verification, and climate policy evaluation. The Three-Cornered Hat (TCH) method neatly fills these gaps: it quantifies the relative errors of multiple datasets without requiring ground truth, is insensitive to common-mode systematic biases, and can simultaneously compute the relative uncertainties of three or more inter-related datasets, thereby avoiding the need for true-reference data in single-satellite validation, the interference of common-mode errors in cross-comparisons, and the intrinsic uncertainties of model-based validation. TCH method has been widely applied in fields such as precipitation [39], total water storage [40], evapotranspiration [41], and soil moisture [42]. Therefore, this study introduces the TCH method to assess, for the first time, the relative uncertainties of multi-source satellite XCO₂ products, providing a more objective uncertainty metric for the comprehensive evaluation of regional satellite observations.

This study performs a comprehensive evaluation of four mainstream satellite XCO₂ products (GOSAT, GOSAT-2, OCO-2 and OCO-3) over China. Ground-based observations and CarbonTracker (CT) model data are jointly used as dual references to assess the accuracy of the satellite products, while the spatial patterns of XCO₂ and the capability of each sensor to capture seasonal variations are systematically examined. In addition, the TCH method is employed to quantify the relative uncertainty of each XCO₂ datasets, providing an objective measure of their reliability and applicability over China.

2. Materials and Methods

2.1. Study Area

The study area encompasses mainland China, which is primarily divided into ten major river basins. Its terrain slopes from west to east in a three-step staircase pattern, descending progressively from west to east. Plains are scarce and mountains are widespread, with a vast elevation difference across the land. Mountains, plateaus and hills account for roughly 67% of the land area, while basins and plains make up 33%. Moreover, mountains and plateaus are mainly concentrated in the western part of the country. Precipitation varies greatly across China, generally decreasing from the southeastern coast to the northwestern interior. The annual precipitation along the southeastern coast mostly exceeds 1600 mm, whereas most areas in the northwest receive less than 50 mm [43]. Population distribution is extremely uneven. Taking the “Heihe-Tengchong Line” as the boundary, the southeastern half accounts for 42.7% of the land area but concentrates about 94.1% of the population, with a population density generally exceeding 300 persons/km²; the northwestern half occupies 57.3% of the land area, with only about 5.9% of the population, and a population density mostly below 50 persons/km² [44]. The location of the study area is shown in Figure 1.

2.2. Datasets

2.2.1. XCO₂ Satellite Remote Sensing Products

This study selects four long-term and widely recognized XCO₂ products including GOSAT, GOSAT-2, OCO-2 and OCO-3 (details in

Table 1. Detailed Information of Four Satellite Remote Sensing Products [37].

Satellite	GOSAT	GOSAT2	OCO-2	OCO-3
Launch time	January 2009	October 2018	July 2014	May 2019
Sensor (km)	TANSO-CAI	TANSO-CAI2	OCO-2	OCO-3
Spatial resolution	10.5 km × 10.5 km	9.7 km × 9.7 km	1.29 km × 2.25 km	1.29 km × 2.25 km
Revisit cycle	3 days	6 days	16 days	-
Data source	NASA/ACOS	NASA/ACOS	NASA/OCO-2	NASA/OCO-3
Data version	ACOS_L2_Lite_FP_9r	ACOS_L2_Lite_FP_9r	ACOS_L2_Lite_FP_10r	ACOS_L2_Lite_FP_10r
URL	https://data2.gosat.nies.go.jp, accessed on 10 March 2025	https://data2.gosat.nies.go.jp, accessed on 10 March 2025	https://disc.gsfc.nasa.gov/, accessed on 17 August 2025	https://disc.gsfc.nasa.gov/, accessed on 17 August 2025

, the technical specifications of the satellites are summarized in Table S1) to systematically evaluate their accuracy, uncertainty and spatiotemporal variability over China. To ensure consistency and comparability among the datasets, the common temporal period from 2020 to 2021 is uniformly extracted, and all products are analyzed at monthly temporal resolution (the data processing workflow and the spatiotemporal matching strategy are thoroughly documented in Sections S1 and S2 of the Supplementary Material).

GOSAT was the world’s first satellite dedicated to greenhouse gas monitoring. Its Thermal And Near-infrared Sensor for carbon Observation-Fourier Transform Spectrometer (TANSO-FTS) instrument has three channels in the near-infrared that are used to retrieve XCO₂, while the Thermal And Near-infrared Sensor for carbon Observation Cloud and Aerosol Imager (TANSO-CAI) sensor is employed to detect clouds or elevated aerosols within the FTS field of view, thereby enabling the screening of spectra valid spectra for reliable retrievals [17]. The TANSO-FTS-2 and TANSO-CAI-2 instruments onboard GOSAT-2 are fully upgraded versions of their predecessors. The FTS-2 incorporates a 2.3 μm weak-CO absorption band and offers a higher signal-to-noise ratios, while CAI-2 provides seven ultraviolet-to-shortwave-infrared bands through fore- and aft-viewing that cover the principal cloud and aerosol detection channels [45]. Its Short-Wavelength InfraRed (SWIR) spatial resolution has been enhanced to 920 m, providing substantially improved cloud and aerosol screening capability compared with the original CAI instrument [16]. However, OCO-2 is not equipped with a dedicated cloud and aerosol detector. Instead, it utilizes three high-resolution grating spectrometers and filters out data contaminated by thick clouds or high-concentration aerosols through observations in the O₂-A band. In addition, OCO-2 operates in three observation modes, namely nadir, glint and target, which are specifically designed for high-precision carbon emission and carbon cycle monitoring, thereby significantly enhancing its capability for quantifying anthropogenic greenhouse gas emissions [46]. Compared with the observation modes of OCO-2, OCO-3 adds an additional Snapshot Area Maps (SAMs) observation mode. Similar to the target mode, this mode features two-dimensional scanning capability, enabling the detection of a 100 km × 100 km area within the target observation region [47].

2.2.2. TCCON Ground Station Observation XCO₂

The Total Carbon Column Observing Network (TCCON) consists of more than 30 ground-based Fourier transform spectrometer stations worldwide. Since 2004, it has inverted the column concentrations of greenhouse gases by measuring direct solar spectra in the near-infrared band. These stations are far from local emissions and can represent true values at the regional scale. With an accuracy of approximately 0.2 ppm for atmospheric XCO₂, TCCON has been widely used to calibrate and validate various carbon satellite products, serving as an authoritative ground-based benchmark for tracking the global carbon cycle and evaluating emission reduction effectiveness [48]. The GGG2020 version of the data was used in this study. GGG2020 builds upon HITRAN 2016 and incorporates TCCON-2020 line-parameter patches to optimize absorption-line parameters, performs more accurate instrument-line-shape calibrations for each site’s spectrometer, and updates the atmospheric prior profiles. These improvements markedly reduce systematic errors, making this version the most reliable data source to date.

2.2.3. CarbonTracker 2022 Model Data

CarbonTracker 2022 (CT2022) is a high-precision atmospheric carbon assimilation dataset developed by the National Oceanic and Atmospheric Administration (NOAA) of the United States, primarily serving global and regional carbon cycle research and carbon emission assessment [49]. Relying on data assimilation technology, it integrates multi-source CO₂ observation data and related models through algorithms such as the Ensemble Kalman Filter, and optimizes the inversion of surface carbon fluxes and atmospheric CO₂ concentrations to reduce uncertainties. Covering the global scope from 2000 to 2020 with a resolution of 3° × 2°, the dataset can be applied to the validation research of satellite XCO₂ products.

2.3. Method

2.3.1. Interpolation Method

Geostatistical interpolation techniques mainly comprise Kriging, Inverse-Distance Weighting, Multiple Regression, and Thiessen Polygons. Among them, Ordinary Kriging (OA) is the most frequently used for spatial interpolation. Therefore, this study employs OA interpolation method, which provides unbiased, optimal estimates of regionalized variables within a limited area based on semi-variogram theory and structural analysis [50]. Its formula is as follows:

$$Z\left(s_0\right)={\sum }_{i=1}^n{\mathit{\lambda }}_{i}Z\left(s_i\right),$$

(1)

where Z($s_0$) denotes the estimated value at the point to be estimated $s_0$; n represents the number of observation points used for estimation; ${\mathit{\lambda }}_i$ is the weight coefficient at the observation point $s_i$; and Z($s_i$) stands for the measured value at the observation point $s_i$.

2.3.2. Uncertainty Quantification and Analysis Method

As a robust method for evaluating the relative uncertainties of multi-source data without requiring prior knowledge, the Three-Cornered Hat (TCH) method has been widely applied in the error assessment of remote sensing products [39,40,41,42]. Each satellite-observed XCO₂ dataset can be expressed as {X_i}_{i = 1,2,…,N}, and the TCH method decomposes the data into two components:

$$X_i=T+{\mathit{\epsilon }}_i{\forall }_i=1,\dots ,N,$$

(2)

In the formula, T is defined as the true time series of XCO₂ data, ${\mathit{\epsilon }}_i$ is the error term, and N is the total number of satellite products (N = 4 in this study). One set of observation data is selected as the reference, and the differences between the remaining observation data and this reference sequence are calculated:

$$Y_{iN} =X_i-X_N={\mathit{\epsilon }}_i-{\mathit{\epsilon }}_N {\forall }_i=1,\dots ,N-1,$$

(3)

where $Y_{iN}$ denotes the relative error, and $X_N$ represents the reference sequence. $Y$ is a matrix containing N-1difference sequences, and its covariance matrix is denoted as S:

$$S=\mathit{cov}\left(Y\right),$$

(4)

where cov() denotes the covariance operator. By introducing an N × N covariance matrix R for individual noise and an auxiliary matrix J, the relationship between matrix S and matrix R is defined as follows [51]:

$$S=J·R·J^T.$$

(5)

In the formula, matrix J is expressed as follows:

$$J_{N-1, N}=\left[\begin{array}{ccc}\begin{array}{c}\begin{array}{cc}1& 0\\ 0& 1\end{array}\end{array}& \dots & \begin{array}{c}\begin{array}{cc}0& -1\\ 0& -1\end{array}\end{array}\\ \begin{array}{cc}⋮& ⋮\end{array}& \ddots & \begin{array}{cc}⋮& ⋮\end{array}\\ \begin{array}{cc}0& 0\end{array}& 0& \begin{array}{cc}\dots & -1\end{array}\end{array}\right].$$

(6)

Matrix R is expressed as follows:

$$R =\left[\begin{array}{ccc}\begin{array}{cc}\begin{array}{c}{r}_{11}\\ r_{12}\end{array}& \begin{array}{c}{r}_{12}\\ r_{22}\end{array}\end{array}& \begin{array}{c}\dots \\ \dots \end{array}& \begin{array}{c}\begin{array}{c}{r}_{1N}\\ r_{2N}\end{array}\end{array}\\ \begin{array}{cc}⋮& ⋮\end{array}& \ddots & ⋮\\ \begin{array}{cc}{r}_{1N}& r_{2N}\end{array}& \dots & r_{\mathit{NN}}\end{array}\right].$$

(7)

By solving the above matrix, the elements in matrix R are as follows:

$$r_{ij}=s_{ij}-r_{NN}+r_{iN}+r_{jN}.$$

(8)

However, since the number of unknown parameters is greater than that of equations, the unknowns cannot be solved based on the equations. To address this issue, Galindo and Palacio proposed an objective constraint based on the Kuhn-Tucker theory [52]. The objective function is defined as follows:

$$T_1(r_{1N},\dots ,r_{\mathit{NN}}) = {\displaystyle \frac{1}{K^2}}{\sum }_{i < j}^N{r}_{\mathit{ij}}^2,$$

(9)

where K $= \sqrt[N-1]{\left|S\right|},$ Its constraint function is defined as follows:

$$H_{(q_{1N}\dots q_{NN})}=-{\displaystyle \frac{\left|Q\right|}{\left|S\right|K}}<0.$$

(10)

To ensure the initial values fall within the constraint conditions, the initial values for iterative calculation are set as follows:

$$r_{iN}^0=0,i<N,$$

(11)

$$r_{\mathit{NN}}^0 ={\displaystyle \frac{1}{2S^{*}}}, S^{*}= [1,\dots 1]S^{-1}[1,\dots 1],$$

(12)

In Equation (9), Q is a diagonal matrix composed of r₁₁, r₂₂, …, r_NN. These parameters can be obtained by iteratively minimizing the initial conditions. The square roots of the diagonal elements of R (r₁₁, r₂₂, …, r_NN) represent the relative uncertainties of each XCO₂ dataset.

2.3.3. Data Quality Assessment Indicators

In this study, the Root Mean Square Error (RMSE) and the Correlation Coefficient (R), which quantifies how closely the estimated values linearly relate to the true values, are used as assessment indicators. The satellite XCO₂ observation data accuracy is quantitatively evaluated by taking the TCCON station observation data and the CT2022 XCO₂ data as the true values, respectively. Their expressions are as follows:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{m}}\sum _{i = 1}^m{(x_i-y_i)}^2},$$

(13)

$$\mathrm{R}={\displaystyle \frac{\sum (x_i-\stackrel{-}{x}\left)\right(y_i-\stackrel{-}{y})}{\sqrt{\sum {(x_i-\stackrel{-}{x})}^2\sum {(y_i-\stackrel{-}{y})}^2}}} ,$$

(14)

where m is the number of samples; x_i is the ith satellite observation; yi is the ith true value.

3. Results

3.1. Spatial Coverage Analysis of Four XCO₂ Satellites Remote Sensing Products

Figure 2 shows the spatial coverage of XCO₂ observation data from four satellites (GOSAT, GOSAT-2, OCO-2, and OCO-3) between January 2020 and December 2021. The spatial coverage is defined as the ratio of the number of grids with observation points to the total number of grids in this study. Overall, the coverage of three satellites (GOSAT, GOSAT-2, and OCO-2) exhibited obvious seasonal fluctuation characteristics: it was relatively high in winter and spring, and low in summer and autumn. The seasonal variation in the effectiveness of satellite observations is primarily due to significant periodic changes in atmospheric interference factors such as cloud cover, water vapor, and aerosols. In contrast, OCO-3 showed a more prominent pattern where coverage was low in odd months and high in even months. This is because the OCO-3 mission employs an odd-even month observing strategy. The odd months are dedicated to nadir observations to obtain uniform background CO₂ data, while the even months are dedicated to SAMs intensive sampling of key regions. Specifically, the peak coverage of GOSAT was approximately 1%, and it dropped to a minimum of about 0.2% in summer. The peak coverage of GOSAT-2 was also close to 1%; its coverage curve overlapped with that of GOSAT in some periods, but its performance in summer was better than that of GOSAT, with the minimum value maintained at around 0.5%. The peak coverage of OCO-2 was slightly higher, reaching up to 1.7%, while its valley value in summer was also close to 0.5%. By comparison, OCO-3 performed the best: its coverage exceeded 1% in half of the months, with the peak even surpassing 3%, and the valley value remained above 0.5%. In summary, among the four satellites, OCO-3 has the highest spatial coverage, followed by OCO-2, while GOSAT and GOSAT-2 have relatively the lowest coverage.

Figure 3 shows the annual spatial distribution of XCO₂ observation data from four satellites (GOSAT, GOSAT-2, OCO-2, and OCO-3) in 2020 and 2021, providing an intuitive view of the coverage differences among the four sensors over China. Despite disparities in total data volume and sampling strategies, the spatial patterns exhibited strong consistency: observations clustered mainly between 30°N and 50°N, with the highest densities over North and Northwest China, whereas Central, South and Southwest China contained markedly fewer soundings. In addition, GOSAT and GOSAT-2 samples were relatively concentrated and spatially limited; in particular, some GOSAT footprints were sampled up to 124 times, with most others near 50, while GOSAT-2 footprints typically numbered around 20. OCO-2 delivered a larger number of observations over a wider area, outperforming the GOSAT series in coverage. OCO-3 showed the greatest sampling density, exhibiting the most extensive spatial distribution and significantly higher coverage in key regions than the other three satellites. Notably, all four sensors collected dense observations in the vicinity of the Xianghe station.

The number of observation points of OCO-series satellites is several times that of GOSAT-series satellites, mainly due to the difference in their observation strategies: OCO-series adopts a continuous push-broom observation mode, acquiring high-density, strip-shaped observation data along the satellite orbit, with a single detection element corresponding to only about 3 km². In contrast, GOSAT-series uses a point observation mode, where each observation point covers an area of approximately 85 km², and is limited by the pointing system, resulting in sparse data acquisition [53].

Notably, all four datasets are concentrated in North and Northwest China, where data density is highest, whereas Central, South and Southwest China are poorly sampled. This pattern reflects the climatic prerequisites of near-infrared measurements. Northern regions experience long, dry, cloud-free periods, stable winter snow cover and low aerosol loading, which are conditions conducive to high-quality retrievals. Conversely, South and Southwest China suffer year-round cloud and rain, high aerosol burdens and persistent mountain fog, leading to aggressive cloud masking and far fewer valid observations. Within the GOSAT-series, GOSAT-2 significantly outperforms its predecessor in Northeast China. TANSO-FTS-2 offers a cross-track field of regard of ±40°, twice that of the original FTS, and incorporates a fully programmable, intelligent pointing system. The satellite can adjust its viewing angle in real time, actively avoid clouds and target clear-sky scenes, roughly doubling the number of usable observations in cloudy regions [45].

3.2. Accuracy Analysis of Four Satellites XCO₂ Data

Further, this study took long-term ground-based observations as the benchmark to evaluate the accuracy of interpolated XCO₂ data from multiple satellites. However, only the TCCON Xianghe station has continuous records from 2020 to 2021; meanwhile, the CT2022 product (up to 2020) was selected as a supplementary reference field. Considering data availability, the experimental period was uniformly fixed to the entire year of 2020. As shown in Figure 4, CT2022 itself exhibits the highest accuracy (R = 0.92, RMSE = 1.78 ppm) with minimal systematic and random errors. Among the four satellites, OCO-3 performs best overall (R = 0.87, RMSE = 1.71 ppm, slope = 0.70) and shows the smallest systematic and random deviations, followed by OCO-2 (R = 0.82, RMSE = 2.07 ppm). Although GOSAT-2 has the highest correlation coefficient (R = 0.90), its slope is only 0.60, and the observation data generally show a significant overestimation, with an RMSE as high as 4.02 ppm; GOSAT has the worst consistency (R = 0.58, RMSE = 2.37 ppm), with large data dispersion and the lowest accuracy.

Additionally, we compared the 2020 raw satellite XCO₂ observations with the CT2022 model data. As shown in Figure 5, GOSAT-2 (N = 18,798) achieves the highest correlation (R = 0.92) and the lowest RMSE (1.73 ppm), yielding the best overall agreement with CT2022, although a subset of XCO₂ values exhibits noticeable bias. OCO-2 delivers roughly 34 times more data than GOSAT-2 while maintaining R = 0.82 and RMSE = 1.67 ppm, giving the most balanced quality. OCO-3 further expands the data size by a factor of 46; its R remains identical to that of OCO-2, while the RMSE increases only slightly to 1.86 ppm, demonstrating robust performance. GOSAT provides only 6514 observations, less than one third of the GOSAT-2 observations, with R dropping to 0.66 and RMSE rising to 3.72 ppm, and thus shows the poorest accuracy. In summary, OCO-3 demonstrates superior data quality, followed by OCO-2; GOSAT-2 shows a marked improvement over its predecessor, whereas GOSAT remains the least accurate.

3.3. Annual Mean Spatial Distribution and Seasonal XCO₂ Increments of Satellite XCO₂

Figure 6 presents the annual-mean XCO₂ distributions for 2020 derived from the five XCO₂ data. CT2022 (Figure 6e) exhibits a pronounced southeast-high to northwest-low gradient: values exceed 414 ppm over Central, East and South China, peak above 414.5 ppm, fall to ~413 ppm across the Northwest and Northeast, and drop below 412.5 ppm over Xizang. All four satellite XCO₂ data reproduce this broad pattern, yet differ in magnitude and regional detail. GOSAT (Figure 6a) is systematically low by 0.5–1 ppm and the high-value band over Central, East and South China is markedly narrower. GOSAT-2 (Figure 6b) captures the spatial extent of the CT2022 maxima, but overestimates XCO₂ by ~1.5 ppm across the Northwest and Northeast; the largest overestimation (~7 ppm) occurs in Xinjiang, likely due to excessive sensitivity to surface albedo and aerosol assumptions. OCO-2 (Figure 6c) and OCO-3 (Figure 6d) resemble CT2022 most closely, both showing a ~1 ppm underestimation in the Northeast. Over the Tibetan Plateau, OCO-3 is high by ~0.5 ppm whereas OCO-2 is low by ~0.5 ppm. Within the high-value band, the spatial extent is contracted in both OCO XCO₂ data and local underestimates of ~1 ppm appear; the deviation is slightly smaller in OCO-2, indicating superior data quality.

Additionally, to verify the ability of the four satellites to capture the dynamic changes in XCO₂, we calculated the seasonal XCO₂ increment (ΔXCO₂) for 2020. Seasons are defined for the Northern Hemisphere: winter (January–February), spring (March–May), summer (June–August), and autumn (September–November). As shown in Figure 7a, with vegetation regrowth and anthropogenic emissions rebounding rapidly after the heating season (winter–spring period), CT2022 shows a general positive ΔXCO₂ of 0.5–2 ppm across the country. However, the four satellites exhibit significant differences compared with CT2022: GOSAT is generally 1–2 ppm lower; GOSAT-2 is substantially higher, with ΔXCO₂ generally ranging from 3–5 ppm nationwide and even reaching as high as 10 ppm locally in Xinjiang. The spatial distribution patterns of OCO-2 and OCO-3 are the closest to CT2022, but they still show “one high, one low” deviations: OCO-2 is generally higher, with local values in Northeast China being 2–3 ppm higher; OCO-3 is generally lower, with values in Northeast China ranging from −3–0 ppm lower. In the high-altitude regions of Xinjiang-Xizang, both satellites show higher values: OCO-2 is 1–2 ppm higher, while OCO-3 is locally 0.5–1.5 ppm higher. Notably, the four satellites show consistent performance in the central region, with ΔXCO₂ all falling within the range of −3–0 ppm, consistently lower than that of CT2022.

Entering the peak growing season (spring–summer period, Figure 7b), XCO₂ generally shows negative growth over China. The ΔXCO₂ of CT2022 ranges from −5 to −1 ppm, and in Northeast China in particular, the decline reaches approximately −8 ppm. In contrast, XCO₂ in Xizang shows a slight increase, with ΔXCO₂ of only 0–1 ppm. Through comparison, it is found that all four satellites can capture the dynamic changes in XCO₂ well. However, in Northeast China, the ΔXCO₂ of GOSAT and GOSAT-2 is significantly lower, generally around −10 ppm, with some local values dropping to as low as −15 ppm. In addition, the ΔXCO₂ of these two satellites is also lower in the Xinjiang and Xizang regions. OCO-2 and OCO-3 have the closest spatial distribution to CT2022, with a better variation range than the GOSAT-series, but OCO-2 slightly overestimates in Northeast China.

Furthermore, entering late summer and early autumn, vegetation photosynthesis weakens and the heating season begins (summer–autumn period, Figure 7c). CT2022 exhibits a spatial pattern of “increasing in the north and decreasing in the south”: in Northeast and North China, ΔXCO₂ reaches 3–5 ppm due to the combination of post-harvest farmland respiration and residential heating; in southern China, photosynthesis decreases but respiration remains vigorous, leading to an overall negative growth of approximately −1 to −3 ppm; in the northwest deserts and Xizang, affected by a sharp temperature drop and weakened local carbon sources, a negative growth of −1–−2 ppm also occurs. Overall, the four satellites have a consistent spatial distribution with CT2022, but significant differences exist in local areas. Specifically, the ΔXCO₂ of GOSAT is approximately 3–5 ppm higher in Northeast, Central, and Northwest China (e.g., Xinjiang); GOSAT-2 and OCO-2 also show an overestimation of around 5 ppm in the high-value band of Northeast China; in contrast, the spatial distribution of OCO-3 is the most consistent with CT2022, showing the best overall consistency. Based on the above analysis, OCO-3 has the best performance in terms of XCO₂ spatial distribution pattern and the ability to capture seasonal dynamic changes in XCO₂; OCO-2 ranks second in consistency, with overall controllable errors; GOSAT and GOSAT-2 perform relatively poorly.

3.4. Uncertainty Analysis of Satellite XCO₂ Data in Ten Major River Basins of China

Figure 8 presents the uncertainty distributions of the four XCO₂ datasets across the ten major river basins from 2020 to 2021. For intercomparison, all datasets were resampled to a uniform spatial resolution of 0.1° × 0.1°. On average, OCO-2 exhibits the lowest uncertainty (1.22 ppm/month), followed by OCO-3 (1.79 ppm/month), while GOSAT-2 (3.06 ppm/month) and GOSAT (3.11 ppm/month) show relatively high uncertainty. Specifically, OCO-2 maintains low uncertainty in almost all basins. OCO-3 also demonstrates low uncertainty in most regions, except for the Southeast River Basin (SERB), where pronounced peaks occur in March (12.66 ppm/month), June (8.12 ppm/month), and August (11.52 ppm/month). In contrast, GOSAT and GOSAT-2 exhibit moderate to high uncertainty in most basins, though with distinct spatial patterns. GOSAT shows elevated uncertainty in the Huai River Basin (HuRB), Liao River Basin (LRB), and Northwest River Basin (NWRB), for example, 8.24 ppm in HuRB in June, 12.16 ppm in HuRB in November, 17.96 ppm in LRB from June to August, and 33.61 ppm in NWRB in March. GOSAT-2’s uncertainty peaks are mainly located in the Hai River Basin (HRB), Songhua River Basin (SRB), Yellow River Basin (YRB), and NWRB, with maximum values of 7.93, 7.84, 9.17, and 9.63 ppm/month, respectively.

Further analysis indicates that uncertainty is jointly influenced by multiple factors such as geographical location and climatic conditions. Specifically, the uncertainty of OCO-series satellites is generally lower than that of GOSAT-series satellites. However, in the SERB, the June–August peak uncertainty of OCO-3 is instead higher than that of GOSAT and GOSAT-2. Even within the observation system of the same satellite, significant differences in uncertainty levels exist among different basins: taking GOSAT as an example, its monthly average uncertainty in the Yangtze River Basin (YZRB) and YRB generally remains at 4–6 ppm/month, while in the LRB, it can surge to 17–34 ppm/month in June–August. Similarly, the uncertainty of OCO-3 in the SERB and HRB during the 3–10 reaches September–December ppm/month, much higher than the 2–5 ppm/month in the surrounding areas of the Qinghai-Xizang Plateau (NWRB, SWRB) during the same period. In addition, the uncertainty in humid regions is slightly higher than that in arid regions, which further highlights the important impact of climatic and environmental factors on the reliability of satellite XCO₂ observation.

4. Discussion

4.1. Analysis of Regional Differences in Satellite XCO₂ Observation Data and the Influencing Mechanisms of Retrieval Accuracy

The amount of satellite XCO₂ observation data in South China is much smaller than that in North China and Northwest China, which is mainly constrained by the coupling physical properties of the atmosphere and underlying surface on near-infrared observations, resulting in significant differences at the level of radiative transfer and physical mechanisms. North China and Northwest China are dominated by temperate continental climate and temperate monsoon climate, featuring long dry and clear-sky periods throughout the year, low aerosol loading, and stable snow cover in winter (with uniform albedo). From the perspective of radiative transfer, the dry and clear-sky environment can minimize the absorption and scattering interference of water vapor and aerosols on near-infrared channels (e.g., the 1.6 μm and 2.0 μm channels of OCO-2). The combination of low water vapor content (minimizing radiation loss), low aerosol loading (reducing signal distortion), and uniform surface albedo (preventing signal fluctuations) collectively ensures a high retention rate of high-quality pixels. Meanwhile, with plains and plateaus being the dominant terrain, this region experiences no significant mountain-valley inversion and enjoys stable vertical transport of water vapor. The physical characteristics of climate, atmospheric composition, and terrain together construct an observation environment with low interference and high stability, further ensuring data density.

In contrast, South China is influenced by the East Asian monsoon, resulting in frequent clouds and rainfall, high aerosol loading, and a predominantly mountainous and hilly terrain. During the radiative transfer process, the strong scattering of clouds causes severe signal distortion; high-concentration aerosols enhance radiation attenuation and reduce the retrieval signal-to-noise ratio; the weak-to-moderate overlap between water vapor absorption lines and CO₂ absorption lines in the near-infrared 1.6 μm band masks the weak CO₂ absorption signal, leaving residual errors even after algorithmic correction. From the perspective of physical mechanisms, frequent clouds and rainfall coupled with high humidity exacerbate the spatiotemporal variability of water vapor; mountain-valley inversion induces persistent clouds and fog; high-concentration aerosols also distort the radiation propagation path. The lack of accurate Aerosol Optical Depth (AOD) data further leads to deviations in atmospheric optical depth calculations. The superposition of these factors results in the elimination of a large number of pixels by cloud masking and quality control algorithms, ultimately leading to a significant shortage of observation data in South China.

However, the difference in regional data volume is merely a superficial phenomenon; the deeper challenge lies in the direct impact of different underlying surface environments on retrieval accuracy. The fact that “OCO-2/OCO-3 outperform GOSAT/GOSAT-2 in accuracy” is well-documented in previous studies [18,19,37,38], which note OCO-series’ superior performance in regions where GOSAT errors are large. This performance gap reflects not just data-level error differences but is fundamentally rooted in the physical basis of their distinct sampling mechanisms. The reason lies in that the approximately 1.3 km pixels of OCO-series satellites can capture the CO₂ signals of a single surface type more accurately, while the approximately 10 km pixels of GOSAT are prone to interference from urban-rural mixing and land–sea boundaries. For example, in the YZRB, where vegetation, water bodies and farmland are intimately mixed, the uncertainty of OCO-series is significantly lower than that of GOSAT-series. OCO-series satellites can focus more accurately on a single surface type (such as small patches of farmland, independent water bodies, or vegetation patches), effectively reducing signal mixing of different surface types within the same pixel and minimizing CO₂ concentration biases caused by underlying surface heterogeneity. In contrast, GOSAT-series satellites struggle to distinguish fine boundaries between urban-rural areas, vegetation, and water bodies in the basin, and are prone to diluting effective CO₂ signals due to large-scale surface type mixing, thereby introducing higher uncertainty. Therefore, the uncertainty of satellites in humid regions is significantly higher than that in arid regions, which is also consistently confirmed by Figure 6c,d. Furthermore, previous study has pointed out that the uncertainty in semi-humid regions is higher than that in arid regions [36], and this result further indirectly supports the conclusions of this study.

4.2. Validation of Multi-Satellite XCO₂ Data Accuracy and Analysis of the Causes of Regional Discrepancies

Comparative analyses of multi-satellite XCO₂ products are essential for understanding the spatial distribution of atmospheric CO₂ across China. Building on previous work, this study extends the comparison to GOSAT, GOSAT-2, OCO-2, and OCO-3.

Validation against TCCON indicates that CT2022 XCO₂ is the closest to ground-based observations (RMSE = 1.78 ppm, R = 0.92), providing a reliable benchmark for comparing the four satellite XCO₂ with CT2022. OCO-2 and OCO-3 exhibit excellent overall performance, whereas GOSAT shows poor accuracy (R = 0.58, RMSE = 2.37 ppm). Numerous studies have indicated that the accuracy of OCO-2 XCO₂ is higher than that of GOSAT on a global scale [18,19,53,54], and in the Chinese region, the accuracy of OCO-2 XCO₂ is also better than that of GOSAT XCO₂. In contrast, GOSAT-2 exhibits significantly lower accuracy (RMSE = 4.02 ppm) and a systematic positive bias relative to TCCON, consistent with the official GOSAT-2 validation report based on global TCCON stations [55,56]. It should be noted that TCCON observations are point-scale measurements of XCO₂, while satellite XCO₂ data are area-scale results estimated by integrating satellite XCO₂ observations in the surrounding area centered on the TCCON station. This scale difference may reduce the consistency between ground-based and satellite XCO₂ data [19]. However, this study further compared the satellite XCO₂ data with CT2022, which effectively compensated for the uncertainty caused by the scale difference. The results show that the accuracy of GOSAT-2 data has been significantly improved, but there are still large regional errors. OCO-2 and OCO-3 still perform well, but the main reason why their overall indicators are slightly inferior to those of GOSAT-2 may be the random errors introduced by excessive data. GOSAT, as always, performs poorly.

In addition, the results in Figure 6 show that among all satellites, GOSAT has the lowest annual mean XCO₂ (411.7626 ppm), while GOSAT-2 has the highest (414.4188 ppm). OCO-2 (412.6308 ppm) and OCO-3 (412.8703 ppm) are the closest to CT2022 (413.3125 ppm), with annual mean differences of −0.6817 ppm and 0.4422 ppm, respectively. The annual mean difference between GOSAT and OCO-2 is −0.8682 ppm. Previous studies have reported similar results [49,57], and some studies have pointed out that the overall bias between the two is −0.21 ± 1.3 ppm [58], indicating that the results of this study are consistent with existing research. It should be noted that in desert areas such as Northwest and North China, the annual mean XCO₂ of GOSAT is relatively low. The previous research result show that the annual XCO₂ in the Inner Mongolia region is systematically underestimated by 1 to 1.5 ppm [34]. In contrast, the annual mean XCO₂ of GOSAT-2 is significantly overestimated, which may be caused by the excessive influence of dust-aerosol scattering in desert regions. Bie et al. [36] also pointed out that the uncertainty of satellite XCO₂ increases with the increase in albedo and aerosol optical depth. By contrast, OCO-2 and OCO-3 achieve accurate observations through comprehensive upgrades in hardware, strategies, and algorithms; their data quality far surpasses that of the GOSAT-series in dust-prone regions such as northwestern and northern China. However, OCO-series performance degrades markedly in the high-value band (Figure 6). The reason may be that the observation points of OCO-series in this region are too sparse (Figure 3), and only local mean values can be obtained during the interpolation process. This leads to excessive smoothing of the spatial field, resulting in underestimation of peaks and ultimately introducing significant biases.

4.3. Applicability of the TCH Method to Satellite XCO₂ Data

The TCH method is a classic tool for assessing the relative uncertainty of multi-source data. It estimates the relative error contributions of each dataset by computing closure residuals, without requiring reference values. This feature renders TCH well suited to evaluating the uncertainty of XCO₂ data from the four satellites (GOSAT, GOSAT-2, OCO-2, OCO-3), especially in regions lacking high-resolution ground references, such as the highly heterogeneous areas of China. The successful application of the TCH method for assessing the relative uncertainty of multi-satellite XCO₂ products hinges on satisfying two core prerequisites: the independence of measurement errors and the linearity of the variable relationships. Regarding error independence, although the four satellites all belong to the passive near-infrared observation system, they exhibit significant heterogeneity in key dimensions such as spatial and temporal resolution, core hardware (e.g., the TANSO-FTS/TANSO-CAI sensors on the GOSAT-series versus the grating spectrometers on the OCO-series), observation modes (point observation versus swath observation), and retrieval methods. This inherent diversity avoids common-mode errors at their source. Furthermore, generational hardware upgrades within each satellite series (e.g., sensor optimizations in GOSAT-2 and the addition of the SAMs mode on OCO-3) further ensure the spectral independence of their error sources. Coupled with an independent data preprocessing pipeline for each satellite, this approach effectively severs potential inter-satellite error correlations, thereby fully satisfying the independence of measurement errors requirement.

Concerning the linearity of variable relationships, the CT2022 product, as an atmospheric CO₂ concentration data assimilation product, is primarily constrained by in-situmeasurements from ground-based observation networks. It does not directly assimilate raw satellite observations, thus maintaining good independence from the satellite data used in this study. Moreover, as validated by ground-based stations (shown in Figure 4), the CO₂ concentrations simulated by CT2022 are highly credible and can serve as an effective proxy for the true atmospheric CO₂ concentration. Based on this premise, our study analyzed and verified the linear relationship between the observations from the four satellites and the corresponding CT2022 data (as shown in Figure 5). Since CT2022 itself exhibits a strong linear correlation with the true atmospheric CO₂ concentration (containing only minimal assimilation error), the transitivity of linear relationships allows us to deduce that the observations from the four satellites also maintain a fundamentally stable linear relationship with the true atmospheric CO₂ concentration. It is important to note, however, that some satellites might exhibit slight nonlinear deviations due to differences in hardware configuration or inherent limitations of their retrieval algorithms. These deviations do not disrupt the overall linear trend and thus do not affect the applicability of the TCH method. The above logical reasoning and empirical results comprehensively satisfy the core prerequisites of the TCH method, providing a solid scientific foundation for the subsequent solution of relative uncertainties and ensuring the reliability and rationality of the assessment results.

4.4. Spatial Representativeness Limitation of Single-Station Validation and Its Impact on Conclusion Robustness

This study uses the Xianghe station of TCCON as the direct validation benchmark for satellite XCO₂ data; however, its spatial representativeness is inherently limited. Located on the North China Plain at 39.75°N, 116.96°E, the station exhibits pronounced local typicality: the underlying surface is dominated by cropland, urban built-up areas and surrounding vegetation, while the climate is temperate monsoon, strongly influenced by East Asian monsoons and human activities such as industrial emissions and agricultural production. These specific geographic and environmental conditions mean the station can only reflect XCO₂ characteristics of mid-latitude plains under a temperate monsoon climate with intensive anthropogenic emissions, and cannot cover China’s complex and diverse regional types such as the cold high-altitude Tibetan Plateau or the deserts and mountains of the arid Northwest. Satellite data performance is closely related to surface type, climate and topography, so validation based on a single station cannot fully characterize accuracy across different regions; this is the core spatial-representativeness limitation of the present study.

Nevertheless, the central conclusion that OCO-2/OCO-3 exhibit overall better observation accuracy than GOSAT/GOSAT-2 in the study area is only weakly affected by this limitation. The dominant factor behind the accuracy difference is the inherent observational mechanism such as the high spatial resolution of the OCO-series versus the coarse resolution of the GOSAT-series, a mechanism-based difference that is stable and unlikely to be reversed by changes in validation region. Comparative validations using spatial distributions relative to CT2022 (Figure 6) and inter-satellite relative uncertainties (Figure 8) further demonstrate that the relative ranking of satellite performance is broadly valid at regional scales, and single-site validation does not distort this core pattern.

5. Conclusions

This study is the first to integrate ground-based validation, model data validation, and TCH uncertainty analysis to systematically evaluate the accuracy, spatial coverage, spatiotemporal patterns, and ability to capture seasonal XCO₂ dynamic changes of four satellite XCO₂ products (GOSAT, GOSAT-2, OCO-2, and OCO-3) over China. Additionally, it further verifies the applicability of the TCH in the evaluation of satellite XCO₂ data. Based on the results and analysis, the following conclusions can be drawn:

(1)

The monthly spatial coverage rate of the four satellite XCO₂ products across China is still insufficient. Among them, OCO-3 has the highest coverage rate, with a peak value of over 3%; OCO-2 ranks second, with a coverage rate of approximately 1%; and the coverage rates of GOSAT-2 and GOSAT are less than 1%.

(2)

In terms of overall accuracy, OCO-3 achieves the highest accuracy (RMSE = 1.71 ppm), followed by OCO-2 (2.03 ppm), while GOSAT-2 and GOSAT exhibit substantially larger errors (4.02 ppm and 2.37 ppm, respectively). In summary, the accuracy of XCO₂ products from OCO-series is comprehensively ahead of that from GOSAT-series.

(3)

The spatiotemporal distribution pattern of OCO-series shows the highest consistency with CT2022: OCO-3 performs the best, followed by OCO-2; GOSAT-2 has a systematic overestimation, while GOSAT shows an overall underestimation. On the seasonal scale, the ability of OCO-series to capture the dynamic changes of XCO₂ is also significantly superior to that of GOSAT-series, among which OCO-3 has the most excellent capture accuracy.

(4)

The TCH demonstrates excellent applicability and reliability in the uncertainty analysis of XCO₂ data. Moreover, its evaluation results are highly consistent with both TCCON and CT2022 reference data.

Overall, the four satellites can provide high-precision XCO₂ products, offering strong support for global carbon monitoring tasks and global carbon inventory; the OCO-series demonstrates more prominent advantages in terms of accuracy and stability. In addition, this study further verifies the applicability of the TCH in quantifying the uncertainty of multi-source satellite XCO₂. Although the TCH method cannot provide absolute accuracy metrics, it serves as an effective tool for assessing the relative uncertainties among different XCO₂ products, and holds important reference value for multi-source data fusion and the construction of CO₂ assimilation models. Furthermore, the TCH method is equally applicable to other greenhouse gases such as CH₄ and NO₂ and can be extended to any region worldwide; this study provides a viable approach and paradigm for their uncertainty assessment. It should be noted that these conclusions are valid only for the 2020–2021 study period; caution is needed when extrapolating to other time frames.