Synoptic differences between transport models are likely correlated with surface fluxes, since they are influenced by common drivers such as temperature, precipitation and insolation. Therefore, it is important to use realistic fluxes to generate the true scenario. We produce the true surface fluxes by assimilating CO2 data from the National Oceanic and Atmospheric Administration's (NOAA) Global Greenhouse Gas Reference Network (GGGRN) and the TCCON in a TM5 4DVAR atmospheric inversion (described later in Sect. ). The inversion spanned 1 June 2014 to 1 April 2016. This ensured that the true fluxes had realistic land and ocean sinks consistent with the observed global CO2 growth rate. At the end of the optimization, TM5 4DVAR wrote out global 1∘ × 1∘ 3-hourly total CO2 fluxes for transport models to ingest in the next step.

We ran a suite of transport models between 1 June 2014 and 1 April 2016 with the true fluxes produced earlier, starting from the same initial CO2 mole fraction field as the inversion used to produce the true fluxes. The suite consisted of TM5, LMDZ, ACTM, PCTM and GEOS-Chem. Details of the individual models can be found in the respective references in Table . It is important to note here that this suite of models spans the range of transport models currently being used by various members of the OCO-2 Science Team to assimilate OCO-2 XCO2 retrievals. Moreover, these models are driven by four different meteorological reanalysis products: ECMWF ERA-Interim (TM5, LMDZ), MERRA (PCTM), GEOS FP (GEOS-Chem) and JMA-55 (ACTM). These four products span the gamut of meteorological fields used by most atmospheric inversions today. Therefore, the divergence of flux estimates seen in this study can be taken to be a reasonable measure of the divergence expected in real data inversions with these transport models.

The transport models produced hourly (PCTM) or 3-hourly (TM5, LMDZ, ACTM, GEOS-Chem) CO2 fields at their individual lateral and vertical resolutions, which are listed in Table . Note that the temporal granularity listed is the time step at which the CO2 mole fraction field was written out; the time step of the models for calculating transport is usually smaller. The models also wrote out the geopotential heights and atmospheric pressures at the vertical layer edges. As a first check, we verified that global average CO2 mole fractions from the different models, calculated from their own pressure and CO2 fields, closely matched the expected time series from the true fluxes diluting into an atmosphere of 5.123 × 1018 kg, the total dry air mass of TM5. Figure  shows these time series. It is evident that, while all the colored lines are very close to the dashed black line, there are small differences that are seasonally coherent. These differences arise from differences in the molar mass of carbon assumed by the models (e.g., 12 g mole-1 vs. 12.01115 g mole-1), small differences in the air mass between different models and the handling of water vapor in the model atmosphere. Rather than standardize the models to remove these small differences, we decided to keep them since they reflect legitimate differences between the models that would express themselves in real data inversions.

The five different modeled dry air mole fraction CO2 fields were sampled with the same code to produce synthetic observations of CO2 from in situ and satellite platforms. Table  gives the number of samples per year from each data stream, and the generation of pseudo-data are described below.

TM5 4DVAR is a state-of-the-art variational inversion system that has been used to estimate surface fluxes of CO2 , CO , CH4 and N2O . Given a set of prior fluxes xa with their error covariance Sa, a set of measurements y with their error covariance Sϵ and a transport model K connecting fluxes to measurements, a Bayesian flux estimation system tries to minimize the cost function J: J=12(Kx-y)TSϵ-1(Kx-y)+12(x-xa)TSa-1(x-xa). The posterior estimate of x, usually denoted x^, is given by x^=xa+SaKTKSaKT+Sϵ-1y-Kxa=xa+Gy-Kxa, where G=SaKTKSaKT+Sϵ-1 is called the Kalman gain matrix and determines the weighting between prior information and observations. Details about TM5 4DVAR have been documented by . In this work we use the ability of TM5 4DVAR to assimilate in situ and total column CO2 measurements as documented by . We run the TM5 transport model (K in the equation above) at global 3∘ × 2∘ × 25-layer resolution and solve for ocean and land fluxes at 3∘ × 2∘ globally. We have already described our method for constructing the synthetic observations y. Below we describe the remaining elements of this inversion, namely Sa, Sϵ and xa.

OCO-2 has a local overpass time of 13:30, and most surface measurements assimilated in flux inversions – except for mountaintop sites – are from the afternoon once a fully mixed PBL has formed. Therefore, the mid-afternoon CO2 mole fraction difference between models, both in the PBL and in the total column, would contribute to flux differences in our experiment. The zonal average of those differences between 1 Dec 2014 and 1 Mar 2016 are plotted in Fig. , where the lowest 150 hPa is an approximation for the mid-afternoon PBL depth. Maps of these differences for summer, winter and the annual average are shown in Figs.  and  in the Appendix. For each grid cell, the median CO2 mole fraction of all five models was subtracted from each model to highlight model differences instead of large-scale features common to all models. All modeled CO2 fields were mapped to a global 1∘ × 1∘ grid while conserving mass. Since the models had varying resolutions and grid registrations, this resulted in unavoidable checkered patterns in the differences in Fig. . That, however, did not impact the large-scale model-to-model differences shown.

In Fig. , the agreement across models is generally better over the Southern Hemisphere (SH) than over the north. This is primarily driven by larger ocean masses in the south than in the north, since, as Figs.  and  show, the agreement across models is generally higher over oceans than over land. This is expected because (a) vertical transport, one of the major axes of variability across models, is stronger over land than over oceans, and (b) surface flux variability is also higher over land than over oceans, amplifying the difference between transport models when viewed in the CO2 concentration space. Models driven by the same parent meteorology do not necessarily show the same features in the modeled CO2 field. In the Northern Hemisphere (NH) summer, LMDZ shows faster exchange between the continental PBL and the free troposphere (FT) than TM5, evidenced by higher CO2 mole fractions in the continental PBL in Fig. . By similar logic, PCTM shows much slower PBL–FT exchange than GEOS-Chem. In the NH winter, contrary to summertime, at the northern temperate latitudes PCTM and TM5 exhibit faster PBL–FT exchange than GEOS-Chem and LMDZ, respectively. The two models driven by GEOS-derived winds (GEOS-Chem and PCTM) are significantly different in the PBL over North and South America, East Asia and tropical Africa throughout the year. The corresponding difference between the two models driven by ERA-Interim winds (LMDZ and TM5) are smaller. ACTM has an overall low bias of ∼ 0.5 ppm in the PBL, which shows up to a lesser extent in the total column (Fig. ) and the total atmospheric CO2 mass (Fig. ). However, such an overall bias should not affect fluxes estimated from ACTM pseudo-observations (henceforth “pseudo-obs”). ACTM also appears to trap more (compared to the model median) of the wintertime respiration signal from boreal Eurasia in the PBL (Fig. ), which should have implications for boreal flux estimates.

In the total column, GEOS-Chem and PCTM look very different in the NH summer, with PCTM trapping more of the NH summertime uptake and SH wintertime respiration signals in the respective hemispheres. In the NH winter, GEOS-Chem displays the tropical Asian biomass burning signal more strongly in the total column than PCTM, while the East Asian fossil fuel enhancement is higher in the GEOS-Chem XCO2 throughout the year. In the NH summer, LMDZ appears to transport more of the temperate and boreal uptake signal to the south than TM5, leading to slightly higher XCO2 values in the north. In the NH winter, conversely, TM5 appears to transport more of the northern respiration signal to the south.

All five models were run from the same initial CO2 field with the same surface fluxes. The resulting global burden of CO2 in the models were close but slightly different, as shown in Fig. . The increase in the global average CO2 mole fraction between 1 January 2015 and 1 January 2016 ranged from 2.89 ppm (TM5) to 2.97 ppm (LMDZ). That 0.08 ppm range in the mole fraction, given the dry air mass of TM5, corresponds to a range of 0.16 PgC in the change in the global CO2 burden over 2015. Therefore, even if our pseudo-data inversions nail the global CO2 budget for 2015 exactly, we can expect a variation of up to 0.16 PgC in that budget owing to the small model-to-model differences in Fig. .

The global total CO2 flux in Fig.  shows a spread of ∼1.5 PgC yr-1 for in situ inversions, which is larger than the spread seen in earlier inverse model intercomparisons such as . This is because intercomparisons such as typically report the constraint on the multi-year average global growth rate, while here we are looking at the constraint on a single year's growth rate from in situ samples. compared eight different inverse models of a single year using in situ data and found a spread of 1.73 PgC yr-1 across models for the annual growth rate, with a standard deviation of 0.5 PgC yr-1. The inversions in were less controlled than our setup, since they used different flux and measurement covariances as well as different transport models. Therefore, in our more controlled experiment, a spread of 1.5 PgC yr-1 is reasonable among the different in situ data streams. It is noteworthy that the spread in the global total flux in Fig.  for the OCO-2 pseudo-data inversions is ∼ 0.25 PgC yr-1, close to the previously calculated limit of 0.16 PgC yr-1. This reduction in the spread from in situ to OCO-2 inversions is primarily due to the more spatially extensive sampling of OCO-2 and not because of OCO-2's sensitivity to the total column (as opposed to the surface layer), evidenced by the ∼ 0.25 PgC yr-1 spread in the global CO2 flux from IS-LNLG and IS-OG inversions in Fig. . This suggests that, compared to the current in situ network, a more spatially extensive sampling strategy, whether total column or PBL, can provide a stricter constraint on the global CO2 budget that is less sensitive to transport model specifics.

The global atmospheric growth rate of CO2 (denoted C below) is determined by the fossil fuel (Fff) emissions and the global sink from the land biosphere (Fbio) and oceans (Foce): dCdt=Fff+Fbio+Foce, where Fbio includes fire emissions. CO2 inversions typically assume a known Fff and estimate Fbio and Foce from atmospheric observations of CO2. Therefore, in a suite of inversions assuming the same Fff, the global total sink Fbio+Foce is constrained to a number whose uncertainty is determined by how well the global CO2 budget is determined by the CO2 observations assimilated. A plot of the estimated Foce vs. Fbio from the suite should therefore be clustered around a straight line with a slope of -1. The same logic applies for any other two-way partitioning of the global sink, such as Northern versus Southern Hemisphere. Figures  and  show four different two-way partitionings of the global total CO2 sink from our ensemble of inversions of synthetic data. The straight line with slope -1 corresponds to the global total sink of -3.64 PgC yr-1 in our true fluxes used to generate the observations. For each inversion estimate, the distance from that straight line is a measure of how much the estimated global budget deviates from the true global budget for 2015, while the position along the line is an indication of how the inversion splits the global budget into the two partitions. Table  contains summary statistics from Figs.  and . For each data stream (e.g., MBL) and partitioning (e.g., Equator, which corresponds to the partitioning between the Northern and Southern Hemisphere), the table contains the spread in the sum and difference of fluxes between the two partitions. The spread in the sum is a measure of the uncertainty in the global budget as constrained by that data stream, while the spread in the difference is indicative of the uncertainty in the partitioning.

The global budget for a single year is constrained poorly by inversions with IS and MBL pseudo-data, evidenced by the large spread of the global sum in Table  and the scatter of the IS and MBL points around the -3.64 PgC yr-1 straight line in Figs.  and . This is consistent with the larger spread in the global sink estimate of inversions with IS and MBL data in Fig. . Among the models, PCTM pseudo-obs seem to demand a higher CO2 flux consistently, while ACTM and GEOS-Chem pseudo-obs demand slightly lower CO2 fluxes. Since growth in the atmospheric CO2 burden was the same for all the models in 2015 (Fig. ), these differences are due to large-scale transport differences sampled by the in situ network.

Since the OCO-2 pseudo-obs in this OSSE are bias free, differences in the partitioning from different sounding modes (LN, LG, OG and land or LNLG) are purely due to sampling differences. This includes the obvious difference of sampling the atmosphere over land and ocean surfaces, and also a more subtle difference in the timing of the samples, coming from the fact that during the early part of the OCO-2 record up to July 2015 the satellite operated continuously for 16 days in nadir (glint) mode before switching to glint (nadir). As a result, land nadir and land glint samples over the same location could be separated by up to 16 days. Since CO2 fluxes can change significantly over 16 days, this may give rise to differences in LN and LG derived flux estimates. The impact of spatiotemporal differences in sampling are evident in Fig. . Among assimilations of OCO-2 pseudo-obs (LN, LG, LNLG, OG) simulated by a single forward model, there can be a ∼ 0.5 PgC yr-1 spread in the partitioning across a latitude, whether the Equator or one of the tropics, while the land–ocean partitioning is more uncertain, with a spread of up to ∼1.5 PgC yr-1. Interestingly, the land–ocean partitioning seems to be better pinned down by OCO-2 ocean soundings than land soundings, evidenced by the smaller inter-model spread when assimilating OG pseudo-obs than when assimilating LN, LG or LNLG pseudo-obs. The same does not appear to hold for any latitudinal partitioning.

Finally, we contrast the partitioning from LNLG (OG) with that from IS-LNLG (IS-OG) to gauge the impact of transport error on PBL versus total column measurements. The IS-LNLG (IS-OG) network, which has spatially extensive PBL sampling only over land (ocean), has a much larger spread in the land–ocean partitioning than the LNLG (OG) network of column samples. This suggests that, if the goal is to partition land and ocean fluxes, PBL sampling can amplify differences across transport models, which are larger in the PBL than in the total column . Moreover, comparing the spreads of IS-LNLG and IS-OG inversions suggests that these transport differences are larger over land than over ocean. If the goal, however, is to partition the global budget across a latitude (i.e., the other three partitionings in Table ), column sampling does not appear to have an obvious advantage over PBL sampling. This is likely because of the fast zonal mixing of the CO2 flux signal; i.e., the flux signal missed by PBL samples at one location due to incorrectly modeled vertical mixing will be seen by PBL sites downstream within the same zonal band.

The spread in flux estimates across the five forward models, or the transport-driven uncertainty, is very similar in Fig.  and Table  between IS and MBL data streams for most regions. Over some land regions that have seen a significant increase in measurement density since , such as North America and Europe, the additional measurements in IS result in a smaller uncertainty compared to MBL. Over land regions where the coverage of IS and MBL are almost identical, such as Africa and tropical Asia, the uncertainties are (not surprisingly) comparable between IS and MBL. Over ocean regions, the IS and MBL uncertainties are very similar, except over the Pacific, where the increased coverage in IS on the west coast of North America is likely responsible for the reduction in uncertainty. The uncertainty in the global uptake and the global land and ocean fluxes are slightly smaller for the IS network than for the MBL network. However, for most other zonal regions the IS and MBL uncertainties are roughly equal, likely because of the fast zonal mixing in the atmosphere.

The regional annual flux estimates of Fig.  show that the spread among land flux estimates when assimilating OCO-2 pseudo-data over land (LN, LG and LNLG) is often smaller than when assimilating in situ data (IS, MBL). This could be a combination of the total column nature of OCO-2 pseudo-data and its increased spatial homogeneity of coverage. To separate the two effects, we look at IS-LNLG, which has the same coverage as LNLG but only PBL samples instead of total columns. Over certain regions, such as temperate North and South America, and temperate Eurasia, the IS-LNLG spread is larger than the IS spread, which is larger than the LNLG spread. This suggests that over those regions the transport model error – relative to the flux signal – in the total column is smaller than in the PBL, leading to lower transport-drive uncertainty in total column CO2 assimilations than in situ CO2 assimilation. Sampling the PBL more densely over those regions is likely to increase transport-driven uncertainty in fluxes. This is consistent with the hypothesis of . However, over some other regions, such as boreal Eurasia and tropical South America, the IS-LNLG spread is much smaller than the IS spread, suggesting that over those regions, the reduction in uncertainty going from IS to LNLG is primarily due to the more uniform spatial coverage and not due to total column sampling. In fact, over tropical South America the IS-LNLG spread is smaller than the LNLG spread, suggesting that the transport error in the total column is larger than that in the PBL. Finally, over regions such as Europe, the ordering of IS, IS-LNLG and LNLG uncertainties suggests that the reduction in uncertainty in going from IS to LNLG is partly due to the more spatially uniform coverage and partly due to total column sampling.

Over some ocean regions such as the temperate North Pacific and South Atlantic, the IS-OG spread is larger than the OG spread, suggesting that modeling the PBL is more uncertain than modeling the total column over those regions. However, the opposite is true over several other ocean regions, such as the temperate North Atlantic and South Pacific. Thus, the hypothesis of cannot be said to hold over most ocean regions. Finally, one striking features of ocean fluxes in Fig.  is worth pointing out here. The transport-derived uncertainty for IS-LNLG estimates is often the largest among all data streams, which leads to a large uncertainty in the global land–ocean partitioning using the IS-LNLG network. This suggests that increasing the PBL sampling only over land – where the transport models disagree more – is likely to worsen ocean flux estimates in the presence of imperfect-transport models.

The global uptake and its partitioning between land and ocean, or the Northern and Southern Hemisphere, are less uncertain for XCO2 assimilations than for in situ CO2 assimilations IS and MBL. Looking at the IS-LNLG and IS-OG inversions, we conclude that the improvement in the global budget and its north–south partitioning is likely due to a more uniform spatial coverage, while the improvement in land–ocean partitioning is likely due to the total column nature of the OCO-2 pseudo-data. Partitioning the budget in zonal bands – i.e., northern extra-tropics, tropics and southern extra-tropics – has (roughly) the same uncertainty across all inversions. This is likely due to the fast zonal flow in the free troposphere, which ensures that surface flux signals missed by one set of measurements – perhaps due to imperfect transport – are seen by other measurements in the same zonal band.

Traditionally, inversions of surface CO2 data have had larger uncertainty in tropical flux estimates than in northern temperate regions, stemming from the sparse observational coverage in the tropics . The larger interannual variability of the tropical flux, seen by several inversion studies including and , is also ascribed partly to the higher uncertainty in tropical flux estimates. In contrast, the uncertainty in flux estimates stemming from uncertainties in modeled transport does not have the same correlation with observational coverage. For inversions with in situ data, the relatively well-covered regions of temperate North America and Europe show the same transport-derived uncertainty as the poorly covered regions of temperate South America and tropical Asia (Fig. ). In general, we do not find that the uncertainties in flux estimates due to transport model errors are lower over the northern temperate latitudes than over less measured tropical and southern temperate areas.

One final noteworthy aspect of the flux estimates of Fig.  is that for some regions (such as temperate South America, Atlantic tropics, Southern Ocean, south Indian temperate, tropical oceans, Indian Ocean, the southern extra-tropics and southern extra-tropical land) the range of in situ flux estimates does not overlap with the range of LN, LG or LNLG (and sometimes OG) flux estimates. For some other regions such as the Indian Ocean and the Southern Ocean, there is no overlap between the OCO-2 land (LN, LG, LNLG) and ocean (OG) estimates. Since there are no biases between the IS, OCO-2 land and ocean pseudo-data, these flux differences suggest that spatiotemporal coverage differences between different observation networks and OCO-2 sampling modes can lead to flux differences that are larger than uncertainties due to transport.

Figures  and  show the monthly flux estimates for 2015 from TRANSCOM-like land and ocean regions. As before, only the spread across the pseudo-data generated by the five transport models is shown for visual clarity. The reduced sensitivity of OCO-2 pseudo-data inversions to transport model uncertainty is obvious for most months over both land and ocean regions. As before, this reduced sensitivity is from a combination of two factors: (a) spatially uniform coverage of OCO-2 compared to the in situ network and (b) the assimilation of column average XCO2 as opposed to PBL CO2. The relative importance of the two factors – as gauged by the relative sizes of the bars between the OCO-2 (LN, LG, LNLG, OG), real in situ (IS) and hypothetical in situ (IS-LNLG, IS-OG) data streams in Figs.  and  – varies by region and season. For example, in October in sub-Saharan Africa, going from the sparse IS network to the more uniform IS-LNLG network reduces the flux uncertainty significantly, but going from PBL measurements (IS-LNLG) to the total column (LNLG) does not reduce the uncertainty further. In contrast, over the same region in December, the increased PBL sampling of the IS-LNLG network inflates the flux uncertainty compared to the IS network, while going from PBL sampling (IS-LNLG) to the total column (LNLG) brings that uncertainty down significantly. In general, over most land regions and most months, given OCO-2's spatiotemporal sampling, assimilating total column CO2 (LNLG) results in equal or lower transport-driven uncertainty than assimilating PBL CO2 (IS-LNLG). The same relationship holds between IS-OG and OG inversions over ocean regions with a few exceptions (e.g., south Indian temperate in June and July). However, the relationship between the real (IS) and hypothetical (IS-LNLG, IS-OG) networks is less general and reflects the impact of different sampling.

The transport-derived uncertainty in monthly fluxes has clear seasonality over most land and ocean regions. In general, over temperate and boreal land regions, the uncertainty is higher in the summer than in the winter, likely due to stronger convective transport and higher horizontal wind shear in the summer months. Temperate oceans sometimes display the opposite behavior (e.g., temperate North Atlantic and North Pacific), whereby transport-driven uncertainty is lower in the summer and higher in the winter. This is likely because advective, and not convective, transport uncertainty is the dominant uncertainty over oceans. Over the tropics the distinction is less clear cut, with no clear commonality between tropical Asia and tropical South America. Over the tropical Indian Ocean, the uncertainty is lowest in the last third of the year, whereas in the tropical Pacific the uncertainty is lowest in the middle of the year.

Over certain ocean regions (e.g., Atlantic tropics, east Pacific tropics, south Indian temperate, Southern Ocean), the range of monthly fluxes obtained from synthetic XCO2 over land (LN, LG and LNLG) often does not overlap at all with the range obtained from either the ocean data (OG) or in situ data (IS). Sometimes, the OCO-2 land pseudo-data inversions overlap with the ocean pseudo-data inversions but not with the true fluxes (e.g., temperate North Atlantic and North Pacific). Since there are no coherent biases in OCO-2 pseudo-data in these synthetic data experiments, the differences between land and ocean XCO2 inversions, or between either set and the true fluxes, can only be due to differences in sampling the same CO2 field with different sets of sampling times and locations. These sampling differences can lead to flux differences that are larger than the transport-driven uncertainty in fluxes. As noted earlier, this implies that in real data inversions biases can appear between land and ocean XCO2 inversions, or between OCO-2 and in situ inversions, purely due to an imperfect-transport model sampling the same field according to different sampling patterns. This can, for example, lead to biased flux estimates when ocean fluxes are inferred using OCO-2 land soundings, even when the retrievals are unbiased.

A comparison of the spread of flux estimates from the MBL and IS inversions suggests that the added coverage from mostly continental sites on top of the mostly background network considered by can reduce transport-induced uncertainty over land regions, despite the uncertainty in transport over continents. This is likely due to the added observations averaging out some of the transport variability. The added coverage has minimal or negative benefit in reducing transport-induced uncertainty of ocean flux estimates, and estimates over zonal bands, except for the Pacific ocean and its temperate and tropical subdivisions.

For inversions of in situ data, flux estimates over the tropics have been historically less certain than estimates over the northern temperate regions, owing to lower observational coverage over the former. In previous work, the uncertainty of fluxes purely due to transport was also found to be slightly higher over tropical regions than over extra-tropical regions . However, in this work, we see that that demarcation does not hold for flux uncertainty stemming from transport model uncertainty. For example, the spread among IS inversions over temperate North America or Europe in Fig.  is as large as their spread over tropical Asia or temperate South America, respectively, despite the first two being much better covered with CO2 samples.

hypothesized that inversions of column average CO2 may be less sensitive to vertical transport errors than PBL CO2, since redistribution of CO2 in the vertical does not change the column average. However, the variation of column CO2 due to fluxes is also much smaller than in the PBL. The transport model sensitivity of column CO2 inversions depends on the balance between this smaller flux signal and smaller transport error. In our experiments, we see that over TRANSCOM-scale and larger land regions (except tropical South America) inversions using column CO2 data over land (LNLG) are indeed less sensitive to transport errors than inversions using PBL CO2 at the same locations and times (IS-LNLG). Over TRANSCOM-scale ocean regions, however, the picture is more ambiguous, as several regions (e.g., Atlantic Ocean, South Pacific temperate, North Atlantic temperate, Southern Ocean) display a smaller uncertainty when assimilating PBL CO2 (IS-OG) than column CO2 (OG). This is likely because the uncertainty in convective transport over oceans is smaller than on land. The global budget and the partitioning across zonal bands are constrained equally well by column and PBL CO2 samples, provided they have the same spatiotemporal coverage. The partitioning across land–ocean boundaries is noticeably more uncertain when using PBL samples over land than column samples, likely because vertical transport differences near the surface are larger over land than oceans.

It should be noted here that the low sensitivity of column measurements to PBL CO2 variations is often considered a weakness, since surface flux signals are the largest in the PBL. Efforts are currently underway to construct active remote-sensing instruments that are preferentially sensitive to the lower troposphere . Our OSSEs suggest that, were such an instrument to be deployed, the uncertainty of surface flux estimates derived from that instrument might very well be larger than from an OCO-2-like column CO2 instrument due to transport model uncertainty near the surface. In the long term, significant improvement in transport modeling will be needed to benefit from a remote-sensing instrument preferentially sensitive to near-surface CO2.

In our synthetic data inversions, the difference between the fluxes inferred from the same forward-model run but different sampling strategies is purely due to the interaction between non-ideal transport and data coverage, and not because of biases between the different samples. Despite this lack of bias, there are several regions where the entire spread of flux estimates across the five forward models has no overlap between certain types of data, or with the truth. For example, LN, LG and LNLG annual flux estimates from the Indian Ocean have no overlap with either IS or OG estimate or the truth, while XCO2 estimates of temperate South American fluxes are completely detached from all IS estimates. This effect is even more pronounced for monthly flux estimates. This suggests that, in the presence of imperfect transport and no measurement bias, different coverage and sampling can generate biases in flux estimates that are larger than their uncertainty due to transport. We should therefore avoid inferring, say, oceanic fluxes by using only OCO-2 land soundings.