East Asian methane emissions inferred from high-resolution inversions of GOSAT and TROPOMI observations: a comparative and evaluative analysis

. We apply atmospheric methane column retrievals from two different satellite instruments (Green-house gases Observing SATellite – GOSAT; TROPOspheric Monitoring Instrument – TROPOMI) to a regional inversion framework to quantify East Asian methane emissions for 2019 at 0.5 ◦ × 0.625 ◦ horizontal resolution. The goal is to assess if GOSAT (relatively mature but sparse) and TROPOMI (new and dense) observations inform consistent methane emissions from East Asia with identically conﬁgured inversions. Comparison of the results from the two inversions shows similar correction patterns to the prior inventory in central northern China, central southern China, northeastern China, and Bangladesh, with less than 2.6 Tg a − 1 differences in regional posterior emissions. The two inversions, however, disagree over some important regions, particularly in northern India and eastern China. The methane emissions inferred from GOSAT observations are 7.7 Tg a − 1 higher than those from TROPOMI observations over northern India but 6.4 Tg a − 1 lower over eastern China. The discrepancies between the two inversions are robust against varied inversion conﬁgurations (i.e., assimilation window and error speciﬁcations). We ﬁnd that the lower methane emissions from eastern China inferred by the GOSAT inversion are more consistent with independent ground-based in situ and total column (TCCON) observations, indicating that the TROPOMI retrievals may have high XCH 4 biases in this region. We also evaluate inversion

XCH4. TROPOMI retrievals use the RemoTeC full-physics method (Hu et al., 2018). The method is prone to spatially and temporally variable biases owing to scattering artefacts (Hu et al., 2018;Lorente et al., 2021;Sha et al., 2021). These biases in general are not reducible with more observations and, if not corrected, can translate into biases in emission estimates in an inversion. Because of spectrally adjacent CO2 and CH4 absorption in the 1.65 μm band, GOSAT retrievals can alternatively use the CO2 proxy method, in which XCH4 is derived from directly retrieved CH4 to CO2 column ratios and independently 70 specified (simulated or assimilated) CO2 columns (Alexe et al., 2015;Frankenberg et al., 2005;Frankenberg et al., 2006;Parker et al., 2015;Parker et al., 2020). The proxy method usually results in reduced variable biases, as scattering artefacts largely cancel out in retrieving CH4 to CO2 column ratios. It also leads to better data coverage over regions with high aerosol loadings or thin clouds, as the method is less sensitive to these interferences compared to the full-physics approach.

75
A number of studies have applied GOSAT data in inversions on a range of scales (Cressot et al., 2014;Feng et al., 2022;Lu et al., 2021;Maasakkers et al., 2019;Monteil et al., 2013;Pandey et al., 2016;. TROPOMI data have also been applied in several regional inversion studies McNorton et al., 2022;Shen et al., 2021;Shen et al., 2022;Zhang et al., 2020) often with the focus on resolving fine-scale emission hotspots.  performed global inversions of GOSAT and TROPOMI observations at 2° × 2.5° resolution in a comparative 80 analysis, and they showed that methane emissions inferred from the two inversions are generally consistent on the global scale but with significant regional discrepancies including over China.
In this study, we will perform high-resolution (0.5° × 0.625°) regional inversions separately for 2019 GOSAT and TROPOMI observations. We focus on East Asia (including China and northern India), which is one of the world's major 85 methane emitting regions and accounts for more than 20% of global emissions (UNFCCC, 2020). The region has been an important contributor to global increases in methane emissions, but the magnitude of the trend and its sectoral attributions are debated (Ganesan et al., 2017;Gao et al., 2021;Liu et al., 2021;Miller et al., 2019;. Here, we will compare East Asian methane emissions inferred from GOSAT and TROPOMI inversions. In the case of discrepancy, we will evaluate against independent observations and discuss the cause of differences. 90 2 Observation Data

Satellite observations
We used XCH4 observations from GOSAT and TROPOMI for 2019 in regional inversions over East Asia. For GOSAT, we use the University of Leicester Proxy XCH4 v9.0 retrievals (Parker and Boesch, 2020). This product is based on the CO2 proxy method, which, as described above, limits variable biases associated with scattering artefacts but is subject to any 95 biases in specified CO2 columns . We use in our inversion only high-quality GOSAT retrievals flagged as "xch4_quality_flag=0" over both land and ocean (glint mode).
For TROPOMI, we use the science product from Lorente et al. (2021). They derived an empirical correction formula to improve surface reflectance dependent biases identified in TROPOMI full-physics retrievals. The correction significantly 100 improves data quality over scenes with low (e.g. snow cover) and high surface albedo (e.g. deserts) which are challenging for a full-physics algorithm. Large corrections are made in East China, Xinjiang China, Southeast Asia, and Siberia ( Figure   S1). Bias-corrected TROPOMI retrievals flagged with "qa_value = 1" are used for inversion. This version of the TROPOMI product does not provide ocean glint-mode retrievals.
105 Figure 1 shows the spatial distributions of annual average XCH4 on the 0.625° × 0.5° grid for GOSAT and TROPOMI. Both datasets show high XCH4 in eastern China and northern India and low XCH4 over Mongolian and Tibetan plateaus, although TROPOMI provides much better spatial coverage than GOSAT over most regions. There are in total 45,018 observations for GOSAT and 8,860,722 for TROPOMI. We take averages of multiple measurements fall in a 0.625° × 0.5° grid cell on any individual day (this procedure affects primarily dense TROPOMI data), and the resulting gridded daily observations are used 110 in the inversion. The spatial distribution of gridded daily observation numbers is shown in Figure S2.

Independent evaluation data 115
We use a suite of independent high-quality methane observations to evaluate the posterior emissions inferred from satellite observations, including surface in situ observations, ground-based remote sensing observations, and tropospheric in situ measurements from commercial airlines. Table S1 provides a descriptive list of these surface sites and Figure 2 shows the locations of surface sites and a representative flight path. These suborbital observations are of good accuracy and precision compared to satellite observations. 120 https://doi.org/10.5194/acp-2022-508 Preprint. Discussion started: 2 August 2022 c Author(s) 2022. CC BY 4.0 License.
Total methane column observations by ground-based Fourier Transform Spectrometers are available at two TCCON sites 130 located in East China, Hefei, China (HF) and Xianghe, China (XH) Yang et al., 2020), and their observations are sensitive to methane emissions from East China. We note that a previous evaluation of GOSAT and TROPOMI against TCCON did not include data from these two sites, as their data were not available then .
We use only measurements with solar zenith angles < 60° to ensure high data quality.

135
All the above surface sites are located distant from northern India, which is a major methane emitting region in the study domain. The only relevant dataset available to us in this area comes from the Civil Aircraft for the Regular Investigation of the atmosphere Based on an Instrument Container (CARIBIC) project (available via the CH4 GLOBALVIEWplus v4.0 ObsPack (Schuldt et al., 2021)), which includes regular flights in the troposphere over northern India. However, these data are collected in earlier years between 2012 and 2014 before the time of TROPOMI. In the absence of better observation data, 140 we compare these 2012-2014 aircraft observations to a simulation driven by a similarly configured GOSAT inversion for an earlier period (2010-2017) . By doing so, we assume that any systematic bias derived from this comparison should still be representative of the 2019 GOSAT inversion.  (Gelaro et al., 2017). The initial concentration fields on January, 1, 2019 and 3-hourly boundary conditions for the nested domain are taken from a global inversion of TROPOMI data for 2019 . We find that the boundary conditions from this global inversion still 155 have biases over East Asia (more discussion in Section 4.3.3), which may be due to the fact that   While methane sinks are not optimized in our regional inversion, they are explicitly simulated in GEOS-Chem simulations.
We use monthly OH fields from a full-chemistry GEOS-Chem simulation (Wecht et al., 2014) and soil absorption from Murguia-Flores et al. (2018).

Inversion procedure 170
We perform analytical Bayesian inversions to optimize a state vector containing annual methane emissions from 600 clusters and average methane column biases at four model boundaries. We optimize emissions on 600 spatial clusters instead of the native 0.5° × 0.625° grid, which are generated based on a Gaussian Mixed Model (GMM) algorithm proposed by . This strategy significantly reduces the computation of an analytical inversion while accounting for major patterns in the distribution of methane emissions. We also optimize for biases in boundary conditions on four sides of 175 our domain (east, south, west, north). Examination of our prior simulation finds domain-wide biases against either GOSAT or TROPOMI observations that can only be attributed to biased boundary condition.
where is prior estimates for and is the observation vector containing either TROPOMI or GOSAT observations, and is a function of representing the forward model. A and O are respectively prior and observation error covariance matrices. We take A as a diagonal matrix and assume a 50% standard deviation for prior emissions and a 1% standard deviation for boundary conditions. O is also taken as diagonal and is populated following the residual error method (Heald 185 et al., 2004), which finds that observation error standard deviations average 16 ppbv for TROPOMI and 18 ppbv for GOSAT.
is a regularization parameter to balance prior and observation information (Hansen, 1998;Rodgers, 2000) and is introduced to prevent overfitting from omitting error correlations in O . We determine following Lu et al. (2021) and , and find = 0.09 for TROPOMI and = 0.6 for GOSAT ( Figure S3). A smaller for TROPOMI reflects a higher degree of spatial correlation among denser TROPOMI observations. 190 The forward model (GEOS-Chem) can be described by a linear equation: where = ∇ is the Jacobian matrix, which describes the sensitivity of observations to the state vector. The cost function is minimized at ∇ ( ) = 0, which yields the optimal estimate (̂) 195 with the posterior error covariance matrix ̂ ̂= ( −1 + −1 ) −1 (4) and the averaging kernel matrix that describes the sensitivity of the optimal solution to the true value: The trace of is referred to as the degree of freedom for signals (DOFS), which represents the number of independent pieces of information constrained by an observing system.  (Figure 3). While the GOSAT inversion suggests that methane emissions over IND should be increased and those from EC decreased relative to prior estimates, the TROPOMI inversion finds the opposite. As a result, regional total methane emissions inferred by the two inversions differ by 7.7 Tg a −1 (27%; TROPOMI: 24.7±0.6 Tg a −1 , GOSAT: 215 32.4±0.7 Tg a −1 ) (errors reported for regional estimates are 1 standard deviations derived from posterior error covariance matrices) over IND and 7.0 Tg a −1 (32%; TROPOMI: 28.3±0.9 Tg a −1 , GOSAT: 21.3±0.9 Tg a −1 ) over EC (Figure 3c). In addition, the two inversions also disagree over the northwestern part of the domain (NWD including parts of Kazakhstan and northern Xinjiang, China and SXJC including mainly southern Xinjiang, China), where TROPOMI indicates large upward adjustments while GOSAT finds agreement with the prior inventory. 220 canceled out by differences in northern India. For China, we attribute 69.1 Tg a −1 for the TROPOMI inversion and 63 Tg a −1 for the GOSAT inversion to anthropogenic emissions, based on prior sectoral fractions in each spatial cluster. These values are at the high end of previous inversion-based estimates of 43-62 Tg a −1 (Deng et al., 2022;Lu et al., 2021;Miller et al., 230 2019;Saunois et al., 2020;Stavert et al., 2022;Wang et al., 2021;Zhang et al., 2022) and are higher than China's latest submission to the UNFCCC (55 Tg a −1 ) for 2014 (UNFCCC, 2020).

Comparison of methane emissions from TROPOMI and GOSAT inversions
These previous inversions mainly used GOSAT observations but differ greatly in their inversion setups (e.g., time, domain coverage, spatial resolution, transport model), thus resulting in a considerable range of estimates. In contrast, the differences in inversions presented in this work are fully due to satellite observations. Our TROPOMI inversion results are consistent 235 with a recent TROPOMI inversion study by Chen et al. (2022) who reported estimate of China's total, anthropogenic, and natural methane emissions of 70.0 (61.6-79.9), 65.0 (57.7-68.4), and 5.0 (3.9-11.6) Tg a −1 .

Evaluation of inversion results with independent observations
Both TROPOMI and GOSAT posterior simulations can reduce errors against their respective "training" data relative to the prior simulation (Figure 4), which is expected for successful inversions. However, concentration fields from the two 240 simulations show varied degrees of agreement across the domain (Figure 5a). In this section, we use independent highquality observations to evaluate whether GOSAT and TROPOMI inversion results are consistent, and in the case that they are not, which one is more in agreement with independent data. Table 1 summarizes performance metrics against these independent observations. GOSAT and TROPOMI inversions 245 perform similarly at background sites such as PDI, UUM, WLG, and LLN. Both posterior simulations achieve relatively good agreement with in situ observations at PDI, UUM, and WLG (absolute biases < 7 ppbv and 2 between 0.39-0.72). An exception is LLN (a high-mountain background site in the southeast of the domain) where biases grow larger in both posterior simulations (11.0 ppbv for GOSAT and 16.7 ppbv for TROPOMI). This is mainly caused by large seasonal biases in the eastern boundary (Figure 5c) (see Section 4.3.3 for more discussion). In fact, mean biases at LLN decrease from prior 250 to posterior simulations during January to May of the year (Prior: -10.8 ppbv; GOSAT: 1.2 ppbv; TROPOMI: 3.7 ppbv) but increase for June to December (Prior: 7.5 ppbv; GOSAT: 17 ppbv; TROPOMI: 24.7 ppbv).
On the other hand, methane concentrations from the TROPOMI and GOSAT posterior simulations differ by ~10-20 ppbv at sites in methane source regions (i.e., XH and HF within EC and AMY in Korea downwind EC) (Figure 5a). Their 255 differences in concentrations are due mainly to higher methane emissions inferred by the TROPOMI inversion than GOSAT over EC (by 7.0 Tg a −1 ) and Korea (Figure 3). Our evaluation against in situ measurements at AMY and total column measurements at XH and HF shows consistently high biases of ~15-25 ppbv by the TROPOMI posterior simulation and a comparatively better agreement (bias ~8 ppbv) with the GOSAT posterior simulation (Table 1). Smaller mean biases are achieved by the prior simulation at XH and HF (Table 1) caused by biases in prior boundary conditions. Overall, our results at AMY, XH, and HF supports the lower methane emissions from EC inferred by the GOSAT inversion over the TROPOMI inferences and indicates that TROPOMI XCH4 retrievals may have regional high biases over EC (more discussion in 4.3.1).
Methane concentrations from the TROPOMI and GOSAT posterior simulations differ by 5.2 ppbv on average along the 265 CARIBIC flight tracks over the Ganges Plain (Figure 5a). This difference is mainly due to different IND methane emissions between the two inversions ( Figure 5b) with minor contributions from boundary condition bias inferences (Figure 5c). In the absence of concurrent independent observations over IND, we use CARIBIC aircraft observations that are only available from 2012 to 2014 to evaluate the inversions. Since these observations predate TROPOMI, we can only indirectly evaluate by using a simulation driven by methane emissions from a GOSAT inversion for earlier years as an inter-comparison 270 platform. We take inversion results from a previous study , which performed an East Asia inversion also using GOSAT proxy XCH4 retrievals. Their inversion is almost identically configured as this study except that it was for 2010-2017. Consistent with our GOSAT results, the GOSAT inversion from Zhang et al. (2022) also found that IND methane emissions should be adjusted upward.

Figure 4: Differences in XCH4 between simulations and satellite observations from GOSAT (a and c) and TROPOMI (b and d). (a) and (b) show results for the prior simulation, (c) for the posterior simulation driven by the GOSAT inversion, and (d) for the posterior simulation driven by the TROPOM inversion. Root-mean-square errors (RMSE, in ppbv) and mean biases (MB, in ppbv, simulation -observation) are inset.
https://doi.org/10.5194/acp-2022-508 Preprint. Discussion started: 2 August 2022 c Author(s) 2022. CC BY 4.0 License.   (Table 1). On the other hand, the 2019 posterior simulation from the GOSAT inversion is about 5.2 ppbv higher than that from the TROPOMI inversion along flight tracks (Figure 5a). Assuming that our 2019 GOSAT inversion is consistent with the 2010-2017 GOSAT inversion by Zhang et al. (2022) (mean bias 14.9 ppbv), it thus suggests that the TROPOMI inversion likely agrees better with the CARIBIC observations (mean bias 9.7 ppbv) than the GOSAT inversion. Unlike the EC case, we do not find over IND 300 significant differences in TROPOMI and GOSAT XCH4 retrievals ( Figure 6). Our analysis suggests that good data coverage https://doi.org/10.5194/acp-2022-508 Preprint. Discussion started: 2 August 2022 c Author(s) 2022. CC BY 4.0 License.
of TROPOMI over IND is likely responsible for its better performance in quantifying methane emissions (see section 4.3.2 for more discussion).

Regional retrieval bias
To understand the cause of differences in the inferred methane emissions, we first compare coincident TROPOMI and 310 GOSAT XCH4 retrievals. The comparison is done following Zhang et al. (2010) where a CTM simulation is used as an intercomparison platform to account for differences in prior profiles and vertical sensitivity between TROPOMI and GOSAT retrievals. TROPOMI XCH4 are on average higher than GOSAT XCH4 over EC by ~6 ppbv, SXJC by ~10 ppbv, and NWD by ~10 ppbv (Figure 6b), which lead to higher methane emissions inferred by the TROPOMI inversion over these regions ( Figure 3). These differences persist throughout the year in EC and SXJC but appear to be highly seasonal in NWD. The Independent ground-based observations are more consistent with the GOSAT inversion and do not support high emissions 320 from EC inferred by the TROPOMI inversion, which indicates that TROPOMI retrievals have systematic regional high biases over EC. In addition, even with enhanced methane emissions in EC, SXJC, and NWD from the TROPOMI inversion, the posterior simulation cannot fully capture these high XCH4 concentrations (Figure 4d). This is also a hint of retrieval biases, as it indicates that the inversion finds it difficult to reconcile these high XCH4 patterns with known methane sources and wind information, given our specification of error parameters ( , , and ). 325 In addition to EC, large XCH4 differences between GOSAT and TROPOMI are also found in the northwestern part of the 330 domain (SXJC and NWD). Although we do not have independent observations over these regions, we speculate that TROPOMI retrievals have positive biases. SXJC is featured with high surface albedo (desert), while in NWD large TROPOMI and GOSAT differences occur during Dec and Mar when surface albedo is low (snow and/or ice cover) ( Figure   S4). High and low surface albedo scenes are known to be challenging for the full-physics retrieval. We suggest to apply the "blended albedo" filter to TROPOMI observations over these regions before inversion Wunch et al., 335 2011).
In our study, we use the TROPOMI science product from Lorente et al. (2021), who applied a posterior correction for surface albedo dependent biases identified in originally retrieved TROPOMI data. We find that this bias correction scheme does overall improve the agreement between TROPOMI and GOSAT in both their methane column concentrations ( Figure  340 S5) and posterior methane emissions ( Figure S6), however, the agreement is not improved in EC, SXJC, and NWD.

Spatial coverage of observations
Although methane emissions from IND (along the Ganges Plain) inferred by the GOSAT inversion are considerably larger than those inferred by the TROPOMI inversion, we find only small differences in coincident XCH4 retrievals there (except https://doi.org/10.5194/acp-2022-508 Preprint. Discussion started: 2 August 2022 c Author(s) 2022. CC BY 4.0 License.
for November and December) (Figure 6), indicating that retrieval biases are unlikely the dominant cause of discrepancies. 345 We have shown above that indirect comparison with CARIBIC tropospheric aircraft measurements favors lower emissions from IND estimated by the TROPOMI inversion (Table 1). In this section, we explore whether differences in data coverage between TROPOMI and GOSAT may contribute to the discrepancies in inferred emissions.  Figure 7 compares the ability of TROPOMI and GOSAT inversions to constrain the distribution of methane emissions, measured by averaging kernel sensitivities (diagonal elements of the averaging kernel matrix). This measure accounts for spatial and temporal data coverage, measurement and model errors (through ), and error correlations between closely 355 located observations (through ). The sum of averaging kernel sensitivities over a region represents the number of pieces of independent information (also known as degree of freedom for signals, DOFS) constrained by an observation system. Figure   7c shows that the TROPOMI inversion has a larger DOFS value (23) than does the GOSAT inversion (19) in IND indicating that methane emissions from IND are better resolved by TROPOMI observations. More importantly, the GOSAT inversion results in highly uneven spatial patterns in averaging kernel sensitivities with much lower values found in the east Ganges 360 https://doi.org/10.5194/acp-2022-508 Preprint. Discussion started: 2 August 2022 c Author(s) 2022. CC BY 4.0 License.
Plain (corresponding DOFS is 4.5 for GOSAT and 7.4 for TROPOMI) (Figure 7a) because of a small number of GOSAT observations there ( Figure S2) which indicates that the large upward adjustment by the GOSAT inversion over IND ( Figure   3a) is associated with large uncertainties. In contrast, more uniform patterns in averaging kernel sensitivities are achieved by the TROPOMI inversion (Figure 7b).

365
The above analysis demonstrates that the TROPOMI inversion benefits from better data coverage for estimating methane emissions from IND. However, over the entire East Asia domain, TROPOMI and GOSAT achieves almost the same (70) DOFS, with similar spatial patterns of averaging kernel sensitivities (Figure 7). Although the number of TROPOMI observations is much larger, strong error correlations in densely distributed data reduce the efficacy of individual observations, as shown by the difference in the regularization parameter determined for TROPOMI ( = 0.09) and GOSAT 370 ( = 0.6) observations.  found in coarse-resolution (2° × 2.5°) global inversions that GOSAT achieves ~50% more DOFS than TROPOMI. In our high-resolution regional inversion, TROPOMI achieves relatively higher DOFS which reflects a lower level of error correlation on the 0.5° × 0.625° resolution than 2° × 2.5°. It can also be conjectured that TROPOMI observations can provide more information than GOSAT observations in an inversion at a spatial resolution better than 0.5° × 0.625°. 375

Regional boundary conditions
Our evaluation against surface observations shows improved agreement at background sites (i.e., PDI, UUM, and WLG) by both inversions (Table 1). This is achieved through simultaneous optimization for biases in boundary conditions together with emissions. As WLG, UUM, and PDI are respectively sensitive to the west, north, and south boundaries, this result suggests that satellite observations can correct biases along these boundaries, supporting our inversion configuration. 380 Furthermore, we find that a sensitivity inversion not optimizing for boundary condition biases (S0) cannot reduce large prior biases at WLG and PDI and leads to unrealistically high methane emissions over East Asia (222 Tg a −1 ) including China (102 Tg a −1 ). Table 1 is LLN (a high-mountain background site in the southeast of the domain) where biases are increased 385 by both inversions. Although the site AMY is also close to the east boundary, it has little influence from the southeast monsoon ( Figure 5c). The biases show strong seasonality, with the largest occurring in summer consistent with ocean-toland (southeast to northwest) transport by summer monsoon. Our analysis suggests that this increase in biases is caused by large adjustments at the east boundary (GOSAT: 3.7 ppbv; TROPOMI: 24.9 ppbv) rather than changes in methane emissions ( Figure 5). This result indicates that satellite observations that are mainly over land are insufficient to constrain the east 390 boundary which consists mainly of ocean.
We then assess the impact of biases along the east boundary on inferred methane emissions. We perform sensitivity inversions using varied levels of fixed (not optimized by the inversion) east boundary conditions, and find relatively small effects on quantifying annual emissions as expected from prevailing westerlies in midlatitudes. A positive bias of 10 ppbv 395 would result in a reduction of annual methane emissions by 2.9 Tg a −1 (~2%) over the East Asia domain, 1.6 Tg a −1 (~2%) over China, and 0.7 Tg a −1 (~3%) over EC (the most affected region) (Figure 8). Although the inversion has a weak constraint on the east boundary conditions, it does not have a great influence on the posterior emissions. However, if the inversion is performed on the monthly or seasonal basis (as opposed to annually in this study), summer results will be more severely affected, leading to seasonal biases in inferred methane emissions. 400

Conclusions 405
We estimate methane emissions from East Asia for 2019 by applying atmospheric methane column retrievals from two different satellite instruments (GOSAT and TROPOMI) to a high-resolution regional inversion framework, in which methane emissions are optimized on 600 spatial clusters with up to about half degree horizontal resolution.
The two inversions estimate a similar magnitude of methane emissions from East Asia (TROPOMI: 143.5 Tg a −1 ; GOSAT: 410 146.2 Tg a −1 ) as compared to prior estimate (130 Tg a −1 ) but differ by ~10% in China (TROPOMI: 74.9 Tg a −1 ; GOSAT: 68.1 Tg a −1 ). Comparisons at the regional scale show that the GOSAT and TROPOMI inversions find consistent results over Central North China, Central South China, Northeast China, and Bangladesh, where the inferred emissions differ by less than 2.7 Tg a −1 . However, the two inversions show large differences over some of the important regions including northern India and East China. The inferred methane emissions by GOSAT observations are 7.7 Tg a −1 higher than those by 415 TROPOMI over northern India but 7.0 Tg a −1 lower over East China. Large differences in inferred emissions are found in northwestern China and Kazakhstan (SXJC and NWD).
We evaluate the inversion results by comparing GOSAT and TROPOMI posterior simulations with independent observations. We find that independent ground-based in situ observations at AMY and total column observations at XH and 420 HF are more compatible with lower methane emissions from East China inferred by the GOSAT inversion than those by the TROPOMI inversion. We also indirectly evaluate against tropospheric aircraft observations over India during 2012-2014 by using a consistent GOSAT inversion of earlier years as an inter-comparison platform, which favors lower methane emissions from northern India inferred by the TROPOMI inversion over those by the GOSAT inversion.

425
The fact that high East China emissions inferred from TROPOMI are inconsistent with independent observations suggests high regional biases in TROPOMI retrievals over East China. Large retrieval differences between GOSAT and TROPOMI are also found in the northwestern China and Kazakhstan, which also leads to substantially higher methane emissions inferred by the TROPOMI inversion. Unfortunately, we do not have independent observations to evaluate the results in these two regions. However, we note that large TROPOMI XCH4 variations in Kazakhstan and northern Xinjiang are coincident 430 with seasonal changes in surface albedo, suggesting possibly over-correction of surface albedo dependent biases in TROPOMI retrievals at the regional level.
The two inversions show large discrepancies in emissions over northern India along the Ganges Plain, although GOSAT and TROPOMI XCH4 values agree reasonably well. We find that the discrepancy in emissions from norther India is due mainly 435 to differences in data coverage. Analyses of the averaging kernel matrices show that the TROPOMI inversion can better constrain emissions from northern India (especially the eastern part of the Ganges Plain), owing to its good spatial coverage in the region as compared to highly uneven coverage by GOSAT. Over the entire East Asia domain, however, the two inversions show similar ability in constraining the distribution of methane emissions, despite a much larger number of TROPOMI observations. This is due mainly to strong error correlations in dense TROPOMI data at the 0.5° × 0.625° 440 resolution.
Both inversions show improved agreement at background sites supporting our optimization of boundary condition biases. An exception is LLN where both inversions show large positive concentration biases against in situ measurements, which results from over-corrections at the eastern boundary by inversions. However, our simulations demonstrate that methane 445 concentration biases at the eastern boundary have relatively small impacts on annual emission inferences. The newer version of the TROPOMI methane product includes glint-mode ocean observations, which may benefit the optimization of eastern boundary conditions. https://doi.org/10.5194/acp-2022-508 Preprint.

Author Contributions
RL and YZ designed the study. RL performed the inverse modelling with contributions from YZ, JL, WC, PZ, ZQ, and ZC. 460 RL analysed and interpreted results with contributions from YZ, CC, HM, GS, ZQ, MZ, RJP, HB, AL, JDM, and IA. RJP and HB provided the GOSAT methane retrievals. AL, JDM, and IA provided the TROPOMI methane retrievals. MZ and PW provided ground based FTIR methane retrievals at the Xianghe site. RL and YZ wrote the paper with inputs from all authors.
Meteorological and Hydrological Administration, China Meteorological Administration, and NOAA for providing surface measurements through GLOBALVIEWplus CH4 ObsPack and WDCGG. We thank the High-performance Computing 480 Center of Westlake University and National Supercomputing Center at Wuxi for facility support and technical assistance.