Carbon dioxide (CO₂) is the most important anthropogenic greenhouse gas (GHG), with atmospheric concentrations having risen from the pre-industrial level of 280 to 426 ppm in 2025 (Joos and Spahni, 2008; Lan et al., 2025). To achieve the Paris Agreement's goal of limiting global temperature rise to below 1.5 °C above pre-industrial levels (UNFCCC, 2015), effective carbon management is imperative. This entails not only controlling anthropogenic emissions but also improving our understanding of carbon sink mechanisms, as major natural sinks such as terrestrial ecosystems and oceans currently absorb roughly half of global emissions (Friedlingstein et al., 2023). Net Ecosystem Exchange (NEE) represents the net CO₂ exchange between terrestrial ecosystems and the atmosphere and reflects the balance between photosynthetic uptake and ecosystem respiration. It is widely used to quantify the strength of land carbon sinks (Lian et al., 2023; Munassar et al., 2022; Reichstein et al., 2005). In parallel, air–sea CO₂ flux describes the net exchange of CO₂ between the ocean and the atmosphere and constitutes a major component of the global carbon budget. However, significant uncertainties remain regarding the capacity and dynamics of these natural sinks (IPCC, 2023). This problem is particularly acute in East Asia, one of the world's fastest-growing carbon-emitting regions (Gilfillan and Marland, 2021). Despite its critical role, previous studies have struggled to accurately estimate regional carbon fluxes due to the limited number of in situ CO₂ observation sites in Asia compared to Europe or North America (Park and Kim, 2020), which poses a limitation for robust regional carbon flux estimation.

Carbon fluxes are commonly estimated using two main approaches: bottom-up and top-down. Bottom-up methods combine observations with statistical upscaling or process-based models (Jung et al., 2020; Kondo et al., 2020; Sitch et al., 2008, 2015). In contrast, top-down methods infer surface fluxes by applying inverse techniques to atmospheric CO₂ concentration data, a process commonly referred to as atmospheric inverse modeling. Among top-down techniques, atmospheric inversions driven by a chemical transport model (CTM) are widely used (Basu et al., 2018; Nassar et al., 2011; Palmer et al., 2003; Peylin et al., 2013). Building on these approaches, international efforts to quantify regional carbon fluxes have continued. REgional Carbon Cycle Assessment and Processes (RECCAP) is an international initiative aimed at quantifying regional greenhouse gas budgets, including CO₂ (Canadell et al., 2011). Coordinated assessments have also been conducted for East Asia. In particular, Wang et al. (2024) provided a comprehensive evaluation of greenhouse gas budgets over East Asia for the 2000s and 2010s using both top-down and bottom-up approaches. In their framework, the top-down estimates represented integrated net land–atmosphere CO₂ fluxes at the regional scale rather than NEE alone. Bottom-up NEE estimates were also reported, although these were based on the TRENDY v9 dynamic global vegetation model ensemble (Sitch et al., 2015) rather than being newly derived within that study. These estimates are briefly compared with the results of this study in Sect. 5.

While several studies have examined carbon fluxes in East Asia, most have either focused on China or provided only limited quantitative assessments of flux uncertainties. For example, Wang et al. (2020) estimated Chinese carbon fluxes from in situ data, assigning prior uncertainties of 50 % for land and 40 % for ocean, which were prescribed as simple percentage values rather than derived from data variability. Thompson et al. (2016), as part of the RECCAP initiative, used a seven-model inversion ensemble for Asia, but applied inconsistent prior fluxes and uncertainties across models. Jiang et al. (2013) estimated carbon uptake in China using ground observations. In their framework, land prior uncertainties were derived from net primary production, while a uniform prior uncertainty was assumed for the ocean.

Since in situ CO₂ measurements are highly precise (typical observational errors < 0.2 ppm), they have been extensively used in inversion frameworks (Baker et al., 2006; Deng and Chen, 2011; Gurney et al., 2003; Jiang et al., 2013; Monteil et al., 2020; Peylin et al., 2013). Their major limitation is sparse spatial coverage, especially over data-poor regions such as the oceans and much of Africa. Satellite retrievals, by contrast, offer broad spatial coverage. The Greenhouse Gases Observing SATellite (GOSAT), launched in 2009, provides global column-averaged CO₂ (XCO2) observations. GOSAT has a footprint of approximately 10.5 km in diameter with an observation error of about 1 ppm (Kulawik et al., 2019). Whereas Wang et al. (2019) excluded oceanic soundings due to concerns over glint-mode retrievals (Wunch et al., 2017), such exclusions may not be optimal for East Asia, where strong anthropogenic emissions are transported eastward over adjacent oceans, making ocean soundings particularly informative for constraining continental outflow signals. We acknowledge that ocean-glint retrievals can exhibit systematic biases distinct from those over land, which has motivated their exclusion in previous inversion studies. However, the ACOS v9 product applies mode-specific bias correction that reduces global mean biases to below 0.2 ppm, with residual seasonal biases of 0.2–0.6 ppm against OCO-2 v10 and single-sounding scatter (∼ 1 ppm) that is comparable to or smaller than over land (Taylor et al., 2022). These residual biases are modest relative to the XCO2 gradients that drive regional flux inversions. To further assess the impact of potential systematic biases in ocean retrievals, we conducted sensitivity experiments by perturbing ocean XCO2 by +0.2, +0.4, and +0.6 ppm, following the range reported by Taylor et al. (2022). The sensitivity tests suggest that the inferred fluxes are not substantially affected by these perturbations (Fig. S1 in the Supplement). We therefore retain both land and ocean soundings, weighting them through their reported retrieval uncertainties in the observation error covariance matrix.

In contrast to previous global inversion systems, the present study employs a regional nested inversion framework over East Asia, enabling higher-resolution meteorological fields and improved representation of regional transport processes. Such a configuration is particularly important in East Asia, where strong emission gradients and complex circulation patterns can amplify transport representation errors in coarse-resolution global inversions. In addition, we explicitly account for prior uncertainties in both terrestrial and oceanic fluxes using data-informed estimates from multi-model ensembles. Terrestrial uncertainties are derived from the standard deviation of the TRENDY ensemble (Sitch et al., 2015), while ocean flux uncertainties are based on the standard deviation among ocean models contributing to the Global Carbon Project (Friedlingstein et al., 2023), rather than prescribing fixed percentage values. We further incorporate both land and ocean GOSAT soundings as observational constraints through uncertainty-based weighting, thereby maximizing observational coverage while accounting for retrieval-specific errors. These methodological features provide a more regionally consistent and physically constrained estimate of East Asian NEE, strengthening the robustness of the inferred carbon fluxes. Such refinements support evidence-based policymaking and climate-mitigation strategies.

To evaluate the reliability of the inversion results, we compared the simulated CO₂ concentrations using the prior and posterior fluxes with independent observational datasets, namely WDCGG and TCCON, which were not assimilated into the inversion system (Feng et al., 2020; Jiang et al., 2021; Jin et al., 2018; Wang et al., 2019). This approach allows for an objective assessment of the inversion performance. Three statistical metrics were employed for the evaluation: correlation coefficient (R), root mean square error (RMSE), and normalized mean bias (NMB), which quantify the linear relationship, overall error magnitude, and systematic bias between the simulated and observed CO₂ concentrations, respectively.

To ensure that the evaluation reflects large-scale, well-mixed CO₂ variability rather than local influences or large representation errors, sites with model–observation RMSE exceeding 7.0 ppm were excluded. This threshold approximately corresponds to the annual amplitude of the seasonal cycle at Mauna Loa, a globally representative background site (Lan et al., 2025). Errors exceeding this threshold suggest that the station is influenced by sub-grid variability that GEOS-Chem cannot resolve at its native resolution, making such sites unsuitable for model evaluation. Following the approach of Jiang et al. (2022a), which excluded sites with inadequate model performance, we removed three WDCGG stations (KIS, HKG, and HKO), representing Kisai (Japan), Hong Kong Hok Tsui (China), and Hong Kong King's Park (China). All TCCON stations met the performance criterion and were retained.

We analyzed the inversion results at eight WDCGG and three TCCON observation sites (Sect. 2.1). Since WDCGG provides point-based ground-level measurements, we selected the nearest model grid cell to each observation site based on latitude, longitude, and altitude for comparison. Among the WDCGG sites, all except YON showed improvements in all three statistical metrics, correlation coefficient (R), root mean square error (RMSE), and normalized mean bias (NMB), after the inversion (Table 2). The YON site, located at the southernmost edge of the domain, lies on a small island (∼ 28.9 km²), which likely introduced substantial representation errors due to the mismatch with the coarser model resolution. For the TCCON observations, which represent column-averaged CO₂ concentrations, we computed the simulated XCO2 using Eq. (2) to ensure a consistent comparison. All three TCCON sites showed improvements across all evaluation metrics.

The posterior simulation improved the overall model performance, reducing the mean RMSE from 3.08 to 2.94 ppm and the mean NMB from 0.33 % to 0.28 %, while maintaining a high correlation (R= 0.95). Although the overall improvements were moderate, they represent consistent enhancements at 10 of the 11 sites and are statistically significant. A paired t-test across all WDCGG and TCCON sites confirmed significant improvements after the inversion: the correlation coefficient increased (ΔR= +0.005, p= 0.012), the normalized mean bias decreased (ΔNMB = -0.03 %, p= 0.037), and the RMSE decreased by 0.15 ppm on average (p= 0.006). Furthermore, both overestimations (positive NMB at most sites) and underestimations (negative NMB at LLN and TAP) were reduced after optimization, suggesting that the improvement was not coincidental but systematic. A moderate level of improvement, which is commonly reported in CO₂ inversion studies, arises because CO₂ fields are already well constrained by the background state, while the remaining discrepancies are primarily attributed to transport and representation errors. For instance, Kou et al. (2023) reported only marginal improvements (RMSE: 2.65 → 2.63 ppm; R: 0.66 → 0.66; MAE: 2.03 → 2.02 ppm), emphasizing that such modest statistical changes are typical in atmospheric CO₂ inversions.

The spatial distributions of prior and posterior uncertainties are shown in Fig. S2. Posterior uncertainties are generally reduced relative to the prior uncertainties across most regions, although the magnitude of reduction varies spatially depending on observational coverage and regional flux sensitivity. To quantify this improvement, we calculate the uncertainty reduction (UR), a key metric for evaluating inverse-modeling performance that measures the reduction in prior uncertainty (Deng et al., 2007).

The mean UR values for each region during 2010–2019 are summarized in Table 3. The UR in China is relatively high, likely due to its large spatial extent, which allows for the inclusion of more GOSAT XCO2 pixels, thereby indicating stronger observational constraints. In contrast, Taiwan, due to its much smaller spatial extent, includes relatively fewer GOSAT XCO2 pixels, resulting in weaker constraints. The UR of regional carbon flux estimates varies substantially across time and space (Deng et al., 2014; Takagi et al., 2011). Over ocean regions, the UR is lower than over land, primarily due to the limited spatial coverage of GOSAT over the ocean. In addition, ocean fluxes are generally much smaller than land fluxes at the grid scale, resulting in a weaker contribution to XCO2 variability and making them more difficult to constrain using satellite observations. This spatial pattern is consistent with the findings of Deng et al. (2014), who demonstrated that UR is closely related to the spatial coverage of GOSAT XCO2 observations. While seasonal differences in observational coverage are relatively small, observations are denser over land and more limited over the ocean (Fig. S3). Similarly, Jiang et al. (2021) reported that UR over land ranged from 5.9 % to 27.2 %, whereas ocean UR remained relatively low, ranging from 0.12 % to 3.7 %. Such large spatial variations in UR highlight its strong dependence on observational density. These results suggest that dense and spatially extensive observational coverage is essential for achieving tighter constraints on regional carbon fluxes.

This section describes regional changes in prior and posterior estimates of carbon fluxes. The 10-year mean NEE increased from -0.17 ± 0.08 to -0.31 ± 0.06 PgC yr⁻¹ (mean ± interannual standard deviation) (Fig. 2a, b), while oceanic uptake showed a slight increase from -0.20 ± 0.03 to -0.21 ± 0.03 PgC yr⁻¹, although this change lies within the range of prior uncertainty and is therefore not statistically significant (Fig. 2a, b). Most regions exhibited a trend toward enhanced carbon uptake, as shown in the difference map (Fig. 2c).

In particular, Mongolia, characterized by its vast grasslands, initially showed very weak carbon uptake of -0.01 PgC yr⁻¹ in the prior estimate, which increased to -0.05 PgC yr⁻¹ in the posterior. Most regions in China experienced increases in carbon uptake, although the magnitude of enhancement varied across subregions. In contrast, carbon uptake weakened in Southwest China, while Northeast China remained nearly neutral with little change from the prior estimate. On the Korean Peninsula, carbon uptake increased, and Japan exhibited a similar level of enhancement. Taiwan, however, showed little to no change. Oceanic regions showed no substantial change.

To help interpret the inferred carbon flux variability, we examine vegetation activity using the Enhanced Vegetation Index (EVI), a widely used satellite-based proxy for photosynthesis. Terrestrial carbon uptake responds non-linearly to complex environmental drivers such as drought and El Niño events (Yue et al., 2017). As a result, vegetation indices cannot perfectly represent variations in carbon fluxes. Despite these limitations, carbon uptake remains fundamentally linked to photosynthetic activity, and EVI provides one of the most practical and widely used proxies for vegetation activity by reflecting vegetation greenness. Noumonvi and Ferlan (2020) also demonstrated that EVI serves as one of the best satellite-based indicators of NEE, even though it cannot fully capture respiration-related processes or short-term environmental stress.

Previous studies (e.g., Wang et al., 2020; Jiang et al., 2021) have used satellite-derived vegetation indices such as EVI, NDVI, and LAI to estimate carbon fluxes. These analyses were generally conducted at coarse spatial scales, typically at continental or subcontinental levels, without resolving fine-scale regional heterogeneity. Following this approach, our comparison also focuses on the domain-averaged behavior. Figure 3 presents the time series of domain-averaged EVI with seasonal variations removed. This increasing trend in EVI suggests enhanced vegetation activity, supporting our finding of increased carbon uptake across most regions of East Asia. Similarly, Wang et al. (2020) attributed China's substantial carbon uptake to the annual rise in vegetation indices.

We examine regional interannual variability and associated supporting evidence, such as El Niño–Southern Oscillation (ENSO) and EVI, that may help explain observed flux patterns. Notably, 2015–2016 coincided with one of the three strongest Super El Niño events on record (1982–1983, 1997–1998, and 2015–2016; Ren et al., 2017; WMO, 2017). ENSO is known to influence photosynthesis and carbon uptake by altering temperature and precipitation patterns (Cox et al., 2013; Fang et al., 2017; Wang et al., 2013, 2014). Accordingly, we focus on ENSO-related impacts and extend the analysis of EVI by conducting correlation analyses to assess its temporal relationship with fluxes.

Figure 4 presents annual CO₂ fluxes for all regions considered in this study over 2010–2019, allowing for direct comparison of prior and posterior estimates across East Asia. The Korean Peninsula acted as a weak carbon sink with low interannual variability. For all years, posterior estimates consistently showed stronger uptake than prior estimates. Japan exhibited a similar pattern, with posterior values exceeding prior ones, and overall low variability. In Mongolia, prior estimates indicated a weak sink, while posterior estimates showed markedly enhanced uptake. Except for 2017, which showed a shift toward a weak source, all years suggested a sink. In Taiwan, posterior fluxes were comparable to or slightly lower than the prior, and overall fluxes remained relatively stable.

During 2015–2016, reduced carbon uptake was observed in several regions across East Asia, coinciding with the Super El Niño. Bastos et al. (2018) reported that this event substantially reduced terrestrial carbon uptake globally by suppressing ecosystem productivity. Within our study domain, ENSO-related climate anomalies were particularly evident over China, where several studies (Ma et al., 2018; Zhai et al., 2016) consistently reported a characteristic south-flood north-drought pattern.

In northern China (North, Northwest, and Northeast China), precipitation deficits prevailed during the 2015 El Niño peak, especially in North China, where severe summer droughts were reported (Zhai et al., 2016), followed by near-normal or slightly wetter conditions in 2016 (Ma et al., 2018). These anomalies are consistent with our results, which indicate a transition from carbon release in 2015 (0.008 PgC yr⁻¹) to weak carbon uptake in 2016 (-0.005 PgC yr⁻¹; Fig. 4). In Northwest China, by contrast, the residual effects of the 2015–2016 El Niño brought unusually high rainfall during 2016 (Lu et al., 2019), particularly in spring and autumn, when precipitation exceeded 150 % of the climatological mean (Ma et al., 2018). As noted by Liu et al. (2024), vegetation in arid regions tends to respond positively to increased moisture availability, and our posterior flux estimates indeed indicate sustained or even enhanced carbon uptake during this period. Specifically, the mean flux during 2015–2016 (-0.078 PgC yr⁻¹) was more negative than the decadal mean excluding those years (-0.054 PgC yr⁻¹), suggesting strengthened carbon uptake under wetter conditions. In Northeast China, interannual flux variability was large, with strong uptake in 2016, but the statistical correlation with ENSO remained insignificant (p>0.05; Ma et al., 2018). This region encompasses diverse vegetation types and spans arid to humid zones (see Jiang et al., 2022b; Fig. 1b), potentially explaining its high interannual flux variability.

In southern China (East, South Central, and Southwest China), the El Niño–induced precipitation anomalies were generally opposite to those in the northern China. Southwest China represented an exception. While East and South Central China experienced excessive rainfall and flooding, Southwest China underwent persistent drought due to weakened southward moisture transport (Ma et al., 2018). This region suffered from prolonged drought conditions from summer 2015 through spring 2016, leading to vegetation water stress and reduced carbon uptake, with net carbon emissions of 0.011 and 0.023 PgC yr⁻¹ during these two years. This drought-induced water limitation likely explains the reduced carbon uptake observed over Southwest China during 2015–2016. In contrast, the summer 2016 flood in East China was particularly severe. The WMO reported that flooding across the Yangtze River Basin in summer 2016 was the most serious since 1999 (WMO, 2017). This extreme rainfall event coincided with a marked shift toward positive NEE (+0.092; carbon release) in 2016 (Fig. 4). South Central China similarly exhibited enhanced precipitation and frequent flooding during 2015–2016 (Ma et al., 2018), corresponding to nearly neutral and carbon-releasing conditions in those years (-0.001 and +0.011 PgC yr⁻¹).

While numerous studies have addressed the effects of ENSO on temperature, precipitation, and extreme weather events, few have explored its direct influence on regional carbon fluxes. When comparing the temporal patterns of precipitation (ERA5; Hersbach et al., 2020) and carbon flux anomalies, a statistically significant time-series correlation is difficult to identify. This is likely because carbon flux variability is influenced by multiple environmental drivers beyond precipitation alone (e.g., temperature and radiation), as well as ecosystem nonlinearity, potential lag effects, and regional climatic heterogeneity. Nevertheless, during the strong 2015–2016 El Niño event, the precipitation anomalies and corresponding flux responses appear qualitatively consistent, providing additional support for our event-based interpretation (Figs. S4 and S5). Our analysis therefore provides new evidence that ENSO-related climatic variability can influence terrestrial ecosystem carbon uptake across East Asia, helping to bridge this critical research gap.

We also analyzed the correlations between EVI and carbon uptake, defined here as NEP (= -NEE) so that positive values indicate uptake. Overall, the correlations strengthened across most regions (Table 4), particularly in the northern part of the domain, including Northwest China, Korean Peninsula, and Japan. For example, the correlation coefficients increased from 0.60 → 0.75 in Korean Peninsula, 0.55 → 0.69 in Japan, and 0.09 → 0.78 in Northwest China, respectively. In North and Northeast China and Mongolia, the correlations shifted from negative to positive, while East China showed a slight increase.

However, it is unrealistic to expect consistent improvement in vegetation–carbon correlations across all regions. For reference, Jiang et al. (2021) compared the relationships between carbon sinks and two vegetation-related indicators (SDA and LAI; see their Table 5) and reported improvements in correlations in fewer than half of the regions examined. In our study, correlations weakened in South Central and Southwest China, whereas the negative correlation persisted in Taiwan. These southern regions are dominated by evergreen broad-leaved forests (Zhu and Tan, 2024). According to Buchmann and Schulze (1999), broad-leaved forests differ from other ecosystems in that leaf area index (LAI) does not significantly correlate with carbon uptake, due to self-shading and increased ecosystem respiration that offset photosynthetic gains. Although EVI differs from LAI, Potithep et al. (2013) reported a high correlation between the two in broad-leaved forests (r2= 0.96), suggesting a close relationship. This may explain why EVI–carbon uptake correlations did not improve in South and Southwest China and Taiwan, where broad-leaved forest characteristics dominate.

In this study, we examine the characteristics and discrepancies of our posterior carbon flux estimates by comparing them with a suite of established products derived from diverse estimation frameworks. The comparison encompasses FLUXCOM (NEE), GCAS2021 (NEE), TRENDY (NEE), OCO-2 v10 MIP (NEE and ocean), CMS-Flux Ocean v3 (ocean), and the Global Carbon Project ocean ensemble (process-based and observation-based).

The FLUXCOM RS product estimates global terrestrial carbon fluxes by applying multiple machine learning algorithms, including Multivariate Adaptive Regression Splines (MARS), to satellite-based remote sensing inputs (Jung et al., 2020). As the FLUXCOM dataset is available only through 2018, while the other products extend to 2019, the comparison for FLUXCOM is limited to that period. GCAS2021 (Jiang et al., 2022a) provides a NEE product derived from GOSAT XCO2 retrievals using the Global Carbon Assimilation System (GCAS), an inverse modeling framework that shares a satellite-based foundation with this study. The TRENDY (Trends in net land–atmosphere carbon exchange) project is a multi-model ensemble (bottom-up framework) designed to assess long-term trends in global terrestrial carbon fluxes. It integrates multiple Earth system and dynamic global vegetation models driven by common input datasets, including atmospheric CO₂ concentration, meteorological forcing, and land-use changes (Sitch et al., 2008). In this study, we used an ensemble of eight models – CABLE-POP, CARDAMOM, CLASSIC, DLEM, EDv3, IBIS, OCN, and YIBS – all of which simulate the full carbon cycle processes encompassing photosynthesis, respiration, carbon storage, and land-use change.

We further include the OCO-2 v10 Model Intercomparison Project (MIP), specifically the LNLGOGIS inversion configuration, which assimilates OCO-2 XCO2 retrievals from land nadir, land glint, and ocean glint observations together with in situ measurements (Crowell et al., 2019). The OCO-2 MIP ensemble provides an independent set of atmospheric inversion estimates using multiple transport models and prior flux assumptions, thereby serving as an additional benchmark for evaluating both terrestrial and ocean carbon flux estimates. It should be noted that the OCO-2 v10 MIP provides net biosphere exchange (NBE) rather than net ecosystem exchange (NEE). To ensure consistency with our flux definition, we therefore subtracted fire emissions from the NBE estimates using the GFED4 fire inventory, which is also used in our inversion framework.

For the ocean domain, the CMS-Flux Ocean v3 product (Liu and Bowman, 2024) represents a posterior estimate generated under NASA's Carbon Monitoring System (CMS), combining GOSAT and OCO-2 observations with an atmospheric transport model to infer global air–sea CO₂ exchange. In addition, we compare our estimates with ocean flux products from the Global Carbon Budget (GCB; Friedlingstein et al., 2023), which include both process-based Global Ocean Biogeochemistry Models (GOBMs) and observation-based reconstructions derived from surface pCO₂ measurements. The process-based ensemble consists of ten ocean biogeochemical models – ACCESS, CESM, CNRM, FESOM, IPSL, MOM, MPIOM, MRI, NEMO, and NORESM – which simulate large-scale ocean circulation and marine carbon processes governing global air–sea CO₂ exchange. The observation-based ensemble reconstructs air–sea CO₂ fluxes using surface pCO₂ measurements and statistical or machine-learning interpolation approaches. In this study, we use eight observation-based products: CMEMS-LSCE-FFNN, CSIR-ML6, JENA-MLS, LDEO-HPD, NIES-ML3, OceanSODA-ETHZv2, UExP-FNN-U, and VLIZ-SOMFFN.

As shown in Fig. 5a, our posterior estimates consistently indicate enhanced terrestrial carbon uptake relative to the prior and are comparable to other top-down products (FLUXCOM and GCAS) as well as the bottom-up ensemble (TRENDY). The magnitude of the terrestrial carbon sink inferred in this study (-0.31 PgC yr⁻¹) is broadly consistent with the terrestrial sink estimate reported by Wang et al. (2024) for East Asia during 2000–2019 (-0.27 PgC yr⁻¹; Fig. 5).

The posterior results show closer agreement with these datasets than the prior does. However, in 2016, although the posterior estimates remain closer to the other products than the prior, a slight discrepancy persists, likely due to the nearly neutral prior flux that year. Over the ocean, the posterior estimates show a comparable magnitude of carbon uptake to both the prior and alternative products (Fig. 5b). Although the posterior mean suggests a marginally stronger sink, the difference lies within the uncertainty range and is therefore not statistically significant. This result remains consistent with the bottom-up ensemble (Global Carbon Budget Ocean).

The ocean remains a region of limited observational coverage, where variability in data availability and input types can lead to differences among products. The Northwest Pacific, our primary ocean focus region, is particularly characterized by complex coastal geometries and sparse surface pCO₂ observations, thereby contributing to elevated uncertainties and product-level discrepancies (Wu et al., 2025). Further contributing factors include differences in observational datasets and model configurations. The CMS-Flux Ocean v3 product assimilates satellite observations from GOSAT v7.3 and OCO-2 within an atmospheric inversion framework. GCAS2021 also uses the GOSAT v9 retrievals employed in this study, but differs by adopting CT2019B as the prior flux and MOZART-4 as the transport model instead of GEOS-Chem. The OCO-2 v10 MIP estimates used here are based on the LNLGOGIS atmospheric inversion system, which integrates OCO-2 XCO2 observations from multiple viewing geometries along with in situ CO₂ measurements. In addition, the Global Carbon Budget ocean products include both process-based ocean biogeochemical models and observation-based pCO₂ reconstructions, which rely on different observational constraints and modeling approaches. These methodological differences likely contribute to the discrepancies observed among the flux estimates.

The carbon budget of East Asia for 2010–2019, incorporating the sink estimated in this study, is summarized as follows (Fig. 6; see Appendix A for details of the calculation method). Fossil fuel and biomass burning emissions are derived from ODIAC and GFED4, respectively. Fossil fuel emissions amount to 3.86 PgC yr⁻¹. Compared with the global total fossil fuel emissions of 9.6 PgC yr⁻¹ (Friedlingstein et al., 2020), East Asia accounts for about 40 % of the global fossil carbon release.

Biomass burning contributes 0.11 PgC yr⁻¹, while the regional NEE and ocean uptake are -0.31 PgC yr⁻¹ and -0.21 PgC yr⁻¹, respectively. These yield a combined sink of -0.52 PgC yr⁻¹, offsetting only 13.6 % of fossil fuel emissions. Consequently, the residual carbon that is not compensated by natural sinks accumulates in the atmosphere, leading to an increase in atmospheric CO₂ concentrations. This imbalance between emissions and sinks explains the persistently high atmospheric CO₂ levels observed over East Asia (Yeh et al., 2023). The atmospheric carbon stock over East Asia is estimated at 38.97 PgC, representing the amount of CO₂ currently retained within the regional atmosphere.

In our East Asia domain, the net surface flux (fossil fuel + biomass burning + NEE + ocean uptake, with NEE and ocean uptake typically negative) is +3.45 PgC yr⁻¹ for 2010–2019, indicating a strong net source to the atmosphere. Over the same period, the vertically integrated atmospheric carbon mass within the domain increases at a mean rate of ∼ 0.24 PgC yr⁻¹, implying that only about 7 % of the emitted carbon remains stored locally in the atmospheric column. The remaining ∼ 3.21 PgC yr⁻¹, ∼ 93 % of the net source, is exported out of the domain by large-scale transport. Most of the carbon emitted from East Asia is transported beyond the regional boundaries. Therefore, East Asian emissions are not confined to a local issue but are linked to downstream transport influencing other regions.

Despite gradual increases in NEE and ocean uptake due to fertilization effects and enhanced solubility associated with pCO₂ gradients, East Asia remains dominated by large fossil fuel emissions. Given this limited natural sink capacity, achieving carbon neutrality will require substantial reductions in fossil fuel use and the enhancement of anthropogenic removals, such as carbon capture and storage (CCS).

This study provides a top-down estimate of regional carbon fluxes across East Asia (18.5–54° N, 73–146° E) for the period 2010–2019, using a Bayesian inversion framework constrained by GOSAT ACOS v9 XCO2 retrievals. By applying the GEOS-Chem chemical transport model and incorporating region-specific prior uncertainties based on the standard deviation of terrestrial and ocean carbon fluxes, we optimized both terrestrial and oceanic fluxes. The posterior estimates indicate enhanced carbon uptake compared to the prior, with mean terrestrial NEE ranging from -0.17 to -0.31 PgC yr⁻¹ while oceanic uptake changed slightly from -0.20 to -0.21 PgC yr⁻¹, showing no statistically significant difference.

Evaluation against independent surface-based CO₂ observations (WDCGG and TCCON) showed consistent improvements across most stations in terms of correlation, RMSE, and bias, supporting the robustness of the inversion framework. Uncertainty reduction (UR) was generally more substantial over continental regions such as China, whereas smaller or oceanic regions showed limited improvements due to observational constraints.

At the regional scale, most regions acted as persistent carbon sinks throughout the decade, with interannual variability influenced by climate events. Notably, the 2015–2016 Super El Niño was associated with temporary flux reversals, primarily over several regions in China. These reversals were largely driven by ENSO-induced floods and droughts, which suppressed vegetation photosynthetic activity and, in some regions, led to near-neutral or even positive NEE values, indicating temporary carbon release. This suggests that terrestrial carbon sinks can be substantially weakened not only by natural climatic variability such as ENSO, but also by extreme weather events intensified under climate change. An increasing trend in the Enhanced Vegetation Index (EVI), along with improved correlations between EVI and posterior carbon uptake, further supports the credibility of the flux estimates. However, regions dominated by broadleaf forests exhibited persistent negative correlations, likely due to self-shading effects of dense canopies.

Comparison with other top-down and bottom-up flux products showed general agreement in both trend and magnitude. Nonetheless, discrepancies remain, largely due to differences in observational inputs, modeling frameworks, and prior flux assumptions. In particular, oceanic uptake estimates tend to diverge more than terrestrial ones, as ocean regions are more sparsely observed and often include complex coastal zones (Wu et al., 2025). In addition, fossil fuel emissions were prescribed and not optimized in this study. Thompson et al. (2016) estimated that uncertainty in the growth rate of these emissions accounted for about 32 % of the uncertainty in the inferred East Asian land sink. Given the large magnitude of anthropogenic emissions in East Asia, differences among fossil fuel emission inventories may influence inversion-based estimates of terrestrial carbon fluxes and should therefore be considered when interpreting our results.

Although the optimized posterior fluxes indicate enhanced carbon uptake compared to the prior, the East Asian domain remains highly fossil-fuel-dominant. Approximately 7 % of the residual carbon accumulates within the regional atmosphere, while the remaining 93 % is transported out of the domain by large-scale circulation. Considering the limited capacity of natural carbon sinks, new strategies will be required to mitigate both the persistently high atmospheric CO₂ concentrations over East Asia and the downstream transport of these emissions to other regions.

Overall, this study estimates carbon sinks over East Asia by incorporating region-specific uncertainties and demonstrates the effective use of satellite constraints and a chemical transport model in inverse modeling. The results were evaluated against independent observations and compared with other flux products, while the interannual variability was interpreted through ENSO and vegetation indices. However, the relatively limited observational coverage over ocean regions resulted in smaller uncertainty reductions, highlighting the need for denser and more continuous oceanic CO₂ observations to further constrain regional flux estimates. Despite this limitation, this study provides valuable insights into the East Asian carbon cycle, which is critical for carbon management, and can support policy strategies aimed at mitigating climate change.