Sea–air CO 2 fluxes in the Southern Ocean for the period 1990–2009

The Southern Ocean (44° S-75° S) plays a critical role in the global carbon cycle, yet remains one of the most poorly sampled ocean regions. Different approaches have been used to estimate sea-air CO2 fluxes in this region: synthesis of surface ocean observations, ocean biogeochemical models, and atmospheric and ocean inversions. As part of the RECCAP (REgional Carbon Cycle Assessment and Processes) project, we combine these different approaches to quantify and assess the magnitude and variability in Southern Ocean sea-air CO2 fluxes between 1990-2009. Using all models and inversions (26), the integrated median annual sea-air CO2 flux of -0.42 ± 0.07 Pg C yr-1 for the 44° S-75° S region is consistent with the -0.27 ± 0.13 Pg C yr-1 calculated using surface observations. The circumpolar region south of 58° S has a small net annual flux (model and inversion median: -0.04 ± 0.07 Pg C yr-1 and observations: +0.04 ± 0.02 Pg C yr-1), with most of the net annual flux located in the 44° S to 58° S circumpolar band (model and inversion median: -0.36 ± 0.09 Pg C yr-1 and observations: -0.35 ± 0.09 Pg C yr-1). Seasonally, in the 44° S-58° S region, the median of 5 ocean biogeochemical models captures the observed sea-air CO2 flux seasonal cycle, while the median of 11 atmospheric inversions shows little seasonal change in the net flux. South of 58° S, neither atmospheric inversions nor ocean biogeochemical models reproduce the phase and amplitude of the observed seasonal sea-air CO2 flux, particularly in the Austral Winter. Importantly, no individual atmospheric inversion or ocean biogeochemical model is capable of reproducing both the observed annual mean uptake and the observed seasonal cycle. This raises concerns about projecting future changes in Southern Ocean CO2 fluxes. The median interannual variability from atmospheric inversions and ocean biogeochemical models is substantial in the Southern Ocean; up to 25% of the annual mean flux with 25% of this inter-annual variability attributed to the region south of 58° S. Trends in the net CO2 flux from the inversions and models are not statistically different from the expected increase of -0.05 Pg C yr-1 decade-1 due to increasing atmospheric CO2 concentrations. However, resolving long term trends is difficult due to the large interannual variability and short time frame (1990-2009) of this study.

of the RECCAP (REgional Carbon Cycle Assessment and Processes) project, we combine these different approaches to quantify and assess the magnitude and variability in Southern Ocean sea-air CO 2 fluxes between 1990-2009. Using all models and inversions (26), the integrated median annual sea-air CO 2 flux of −0.42 ± 0.07 Pg C yr Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | carbon exchange and are a major influence in setting atmospheric CO 2 levels (Sallee et al., 2012). The Southern Ocean is comprised of a series of key zones in terms of carbon dynamics (Fig. 1). These zones from north to south are: (i) the SubAntarctic Zone (SAZ) between the Subtropical Front (STF) and the Subantarctic Front (SAF), nominally 40 • S- The available observations indicate that the seasonal cycle is the dominant mode of variability in the surface partial pressure of CO 2 (pCO 2 ) and the sea-air exchange 10 for the Southern Ocean (Lenton et al., 2006;Thomalla et al., 2011). Oceanic uptake of CO 2 from the atmosphere corresponds to a negative net sea-air CO 2 flux, and an oceanic source of CO 2 to the atmosphere has a positive net sea-air flux. Increased biological production in summer tends to reduce the pCO 2 and increase the net ocean uptake of CO 2 . Lower biological production and deeper mixing that entrains more CO 2 - 15 rich water into the mixed layer has the opposite effect in winter. Seasonal temperature changes offset the biological and mixing related effects on surface pCO 2 and act to mute the variability in seasonal sea-air fluxes, e.g. Takahashi et al. (2002) and Takahashi et al. (2009). While the biological and physical mechanisms driving this seasonal variability are relatively well-known, their magnitude and phase remain poorly 20 constrained (Metzl et al., 2006;Lenton et al., 2006).
In the Austral Summer the PFZ and SAZ act as a strong sinks of atmospheric CO 2 due to enhanced biological production (Metzl et al., 1999;Takahashi et al., 2009Takahashi et al., , 2012. In the Austral Winter, the net uptake of CO 2 in the PFZ and SAZ is reduced relative to summer, and in some areas a net outgassing occurs as a result of deep winter mixing 25 entraining carbon-rich waters from the ocean interior into the surface mixed layer. When integrated annually, these regions act as a strong net sink of atmospheric CO 2 , with the largest uptake occurring in the SAZ and decreasing southward (Metzl et al., 1999(Metzl et al., , 2006McNeil et al., 2007;Takahashi et al., 2012). The AZ to the south of the PFZ contains a permanently ice free region in the north that transitions poleward to waters covered seasonally by sea-ice with some permanently ice-covered waters at high latitudes. The seasonal ice covered region in the south portion of the AZ acts as a sink of atmospheric CO 2 in summer as sea-ice retreats and increased stratification of the upper water column and greater light avail-5 ability leads to an increase in the biological draw-down of carbon (Ishii et al., 1998;Bakker et al., 2008). Sea-ice cover inhibits sea-air gas exchange in winter, although there is considerable uncertainty as to how the ice-cover impacts sea-air fluxes at high latitudes in winter months (e.g. Rysgaard et al., 2011;Loose and Schlosser, 2011). In the permanently ice free region of the AZ, pCO 2 tends to be lower than the atmo-10 sphere in summer owing to net biological production (Ishii et al., 2002;Metzl et al., 2006;Sokolov, 2008) and higher than the atmosphere in winter due to deep winter mixing. Integrated annually, the observations suggest the AZ acts as a neutral or weak source of atmospheric CO 2 (Takahashi et al., 2009).
In recent decades, a strengthening of the Southern Annular Mode (SAM), driven 15 by increasing greenhouse gases and stratospheric ozone depletion, has resulted in a southward shift and intensification of the zonal winds, and increases in heat and freshwater fluxes over the Southern Ocean (Thompson and Solomon, 2002). During this period, oceanic pCO 2 growth rates based on limited observations (Metzl, 2009;Takahashi et al., 2009) and some atmospheric inversions and ocean biogeochemical 20 models (Le Quéré et al., 2007) have suggested a decrease in the efficiency of the Southern Ocean CO 2 sink in response to these physical changes. The strengthening SAM is believed to increase the upwelling of carbon-rich deep water that results in a decrease in the net ocean carbon uptake due to a change in the sea-air gradient in pCO 2 (Le Quéré et al., 2007;Lovenduski et al., 2007;Lenton and Matear, 2007;Lenton et al., 25 2009). The corresponding changes in primary productivity are small relative to the response of ocean dynamics (Lenton et al., 2009). Furthermore, this change may have enhanced the outgassing of natural CO 2 , and had only a small effect on the uptake of anthropogenic CO 2 (Lovenduski et al., 2008;Matear and Lenton, 2008). Evidence for 290 a reduced efficiency of the Southern Ocean sink is not supported by all atmospheric inversions (e.g. Law et al., 2008). Analysis of long-running of atmospheric CO 2 time series in the Southern Hemisphere and model simulations suggests that it may not possible to robustly detect a slowdown in the Southern Ocean sink from atmospheric CO 2 measurements at present (Stephens et al., 2012). Furthermore, ocean observa-5 tions also suggest the changes in the sink efficiency may not be zonally uniform for the Southern Ocean . Different approaches have been used to estimate the total sea-air CO 2 fluxes (i.e. the anthropogenic + natural (or pre-industrial) in the Southern Ocean), including: (i) synthesis of surface ocean pCO 2 observations with empirical gas exchange parame-10 terizations, and using interpolation schemes applied to ocean pCO 2 observations to address sparse data coverage; (ii) prognostic ocean biogeochemical models based on physical ocean general circulation models coupled with biogeochemical models; (iii) atmospheric inversion studies, based on the inversion of atmospheric CO 2 data; and (iv) ocean inversion models, based on ocean interior measurements and ocean model 15 simulations. Historically, these modelling approaches have produced significantly different annual mean estimates of Southern Ocean sea-air fluxes of carbon, e.g. Roy et al. (2003). A more recent study suggests there is growing agreement in the magnitude of the annual flux for the region (Gruber et al., 2009), although substantial differences remain with regard to the meridional distribution. 20 In this study we compare the total sea-air CO 2 fluxes in the Southern Ocean derived from observations, atmospheric and ocean inverse calculations, and ocean biogeochemical models for the period 1990-2009. The oceanic exchange of CO 2 with the atmosphere through sea-air gas flux is driven by the difference in ∆pCO 2 between the ocean and the overlying atmosphere and is a function of wind-speed. A negative Introduction magnitude of inter-annual variability in the Southern Ocean and investigate the longterm trends of ocean carbon uptake. This work is part of the REgional Carbon Cycle Assessment and Processes project (RECCAP; Canadell et al., 2011) led by the Global Carbon Project, and this paper is part of a special volume of papers assessing the variability in the global carbon cycle over the period 1990-2009.

Datasets
The regional sea-air CO 2 fluxes described below use RECCAP global "Tier 1" global CO 2 flux products (Canadell et al., 2011). These products include datasets of CO 2 fluxes from observations, ocean biogeochemical models, atmospheric and ocean in-10 versions.

Observations
The Southern Ocean remains one of the most poorly sampled ocean regions with respect to carbon (Fig. 1), with some regions in the Eastern South Pacific yet to be sampled. To account for this paucity of sampling, Takahashi et al. (2009) compiled more 15 than 3 million measurements of oceanic pCO 2 globally, and corrected these to the reference year 2000. These values were averaged onto a global grid, and two dimensional advection-diffusion equations were used to interpolate spatially for each month (Takahashi et al., 1997). Wanninkhof et al. (2012) used the Takahashi gridded data and Calibrated Multi-Platform Winds (CCMP; Atlas et al., 2011) to generate a monthly 20 1 • × 1 • climatology of net sea-air CO 2 fluxes for RECCAP. In our subsequent analysis and comparison with different models and inversions, we define observations as this climatology.
The steep meridional gradients in physical, chemical and biological properties of the Southern Ocean, as well as the patchiness of biological activities, means that the observed pCO 2 values vary over a wide range in space and time. This combined with errors associated with sparse data coverage, wind speed measurements and gas transfer coefficients (see Wanninkhof et al., 2012, andSweeney et al., 2007, for more discussion) mean that the observationally derived fluxes contain large uncertainties. Therefore, we conservatively estimated an uncertainty on all sea-air CO 2 flux observations of ±50 %, consistent with Gruber et al. (2009) andSchuster et al. (2012). At present there is insufficient observational data to assess the longer-term (interannual to decadal) variability for the entire Southern Ocean, and comparisons of model simulations and observed fluxes are limited in the following sections to the annual and seasonal variability of the total (anthropogenic and preindustrial) net sea-air CO 2 fluxes.

Ocean biogeochemical models
The simulated sea-air CO 2 fluxes come from five ocean general circulation models coupled to ocean biogeochemical models (Table 1). These models represent the physical, chemical and biological processes governing the marine carbon cycle and the exchange of CO 2 with the atmosphere. All of these models are coarse resolution and 15 do not resolve or permit mesoscale eddies. The simulations were driven with observed reanalysis products for atmospheric boundary conditions over the period 1990-2009. All of these models have been integrated from the pre-industrial to present day with the same atmospheric CO 2 history. The physical models vary in many aspects such as the details of physical forcing, sub-grid scale parameterizations, and experimental config-20 urations that are detailed in the reference for each model in Table 1. In addition, the models incorporate different biogeochemical modules that can substantially influence the simulated fields of surface CO 2 .

Atmospheric inversions
Inversions of atmospheric CO 2 measurements use an atmospheric transport model 25 and an optimization technique to determine carbon fluxes that best fit the atmospheric measurements. In most cases, a first guess or a priori flux based on independent data sets or models is also used to further constrain the flux estimates. Here we use the same set of atmospheric CO 2 inversions as described in Peylin et al. (2012) The inversions (Table 2) vary in many aspects such as, the atmospheric transport model, sea-air CO 2 flux resolution, solution method, a priori estimates and the atmospheric 5 CO 2 datasets. In this study, we consider inversions that solve for carbon fluxes at a variety of different spatial resolutions, dividing the Southern ocean region into between 1 and 14 regions or, in some cases, solving at the resolution of the transport model used in the inversion. In cases where inversions solved for regions that cross the 44 • S boundary 10 used in this analysis, a proportion of the estimated flux is used based on any assumptions about the flux distribution within a region. For the inversions solving at grid-cell resolution, those methods use a correlation length scale to ensure relatively smoothly varying fluxes. Figure 1  were used by only about half the inversions (grey circles). There are three sites that are only used by one or two inversions. Many inversions use monthly mean atmospheric concentrations derived from the pseudo-weekly filtered data from the GLOBALVIEW-20 CO 2 data product (e.g. GLOBALVIEW-CO 2 , 2011 and earlier annual releases). Four inversions use "raw" data, i.e. the inversions include flask observations at the appropriate sampling time, or hourly observations if these are available (though generally with some selection based on time of day). All but three inversions account for interannual variations in atmospheric transport. 25 An important consideration when interpreting the inverse estimates is to understand the influence of any prior information used in the inversion. For example, in almost all the inversions considered here, ocean CO 2 fluxes derived from pCO 2 measurements, e.g. Takahashi et al. (1999) or Takahashi et al. (2002)  the sea-air fluxes, so that the inversion is actually estimating sea-air fluxes that are deviations from the prior Takahashi-based estimates. Depending on the spatial resolution of the inversion and the additional constraints that have been incorporated into the inversion, the basin-scale flux information may be strongly dependent on the prior information rather than new independent information derived from the local atmospheric 5 CO 2 measurements.

Ocean inversions
In this paper we use ten ocean inverse model simulations presented in Gruber et al. (2009) and the mean value across these simulations calculated from three periods: 1995, 2000 and 2005. As this technique only solves for an annual mean state, i.e. does not resolve seasonality or inter-annual variability, these simulations are only used to assess the annual uptake. The specific details on methods and models used in these ocean inversions are detailed in Mikaloff Fletcher et al. (2006 and Gruber et al. (2009). Gruber et al. (2009) used a suite of ten different ocean biogeochemical models to 15 estimate the uncertainty of the flux estimates due to model transport, and weighted these models using a skill score based on each model's ability to accurately represent pre-bomb radiocarbon and CFC's when calculating the mean and standard deviations of the models. However, here we have chosen to give all of the models equal weight in order to be consistent with our analysis of ocean models and atmospheric inversions, 20 where no weighting scheme was used. The skill score weighting has a relatively minor influence on fluxes over most of the ocean. However, in the Southern Ocean, the skill score weighting leads to a smaller net uptake of CO 2 by the Southern Ocean and a different distribution of the uptake between regions. This is because some models used in the ocean inversion tend to 25 over-estimate CFC concentrations in the Southern Ocean relative to observations and also estimate substantially higher anthropogenic CO 2 uptake in the inversion compared with other contributing models (Mikaloff Fletcher et al., 2006 scheme reduces the impact of these models on the weighted mean and therefore leads to a smaller estimated sink.

Study region
Following RECCAP protocols and its regional definitions, we use the latitudinal boundaries of 44 • S-58 • S, and 58 • S-75 • S, to define broad Southern Ocean subdomains 5 (Mikaloff Fletcher et al., 2006). The 44 • S-58 • S circumpolar band includes a large part of the SAZ and the PFZ, while the southern region includes the AZ. The region 44 • S-58 • S is further split into the three major ocean basins: Indian, Pacific and Atlantic ( Fig. 1; Table 3). All together we consider 6 regions in the Southern Ocean (Table 3): the five outlined above and the total Southern Ocean south of 44 • S.

Calculation and assessment of sea-air CO 2 fluxes
The sea-air CO 2 fluxes for ocean models and inversions were calculated as a median and the variability as a median absolute deviation (MAD; Gauss, 1816), consistent with Schuster et al. (2012). The MAD is the value where one half of all values are closer to the median than the MAD, and is a useful statistic for excluding outliers in data sets. 15 The calculation of the annual uptake and seasonal variability of sea-air CO 2 fluxes from atmospheric inversions and ocean biogeochemical models used data from all of the models and inversions listed in Tables 1 and 2. The seasonality in the sea-air CO 2 flux calculated from the individual models was compared to net flux estimates from the surface ocean CO 2 climatology for the year 2000 using 2-quadrant Taylor diagrams 20 (Taylor, 2001). This allows both the phase and magnitude of the seasonal cycle for each model to be assessed individually along with the annual mean value. Finally, in the calculation of trends we only used model simulations in which 10 or more years of output was available and assumed the trends were linear following Le Quéré et al. (2007). The sea-air CO 2 flux into the ocean is defined as negative, consistent with 25 RECCAP protocols.

Results and discussion
The modeled and observational based sea-air fluxes of CO 2 for the Southern Ocean are evaluated at three scales of variability: (i) annual, (ii) seasonal, and, (iii) inter-annual for the period 1990-2009.

Total Southern Ocean 44 • S-75 • S
The median annual sea-air CO 2 fluxes calculated from observations, models and inversions varies between −0.27 and −0.43 Pg C yr −1 for the entire Southern Ocean region (Table 3 and Fig. 3). The flux estimates from ocean inversions and ocean biogeochemical models, although not significantly different from the other estimates, tend to indicate 15 a stronger uptake. This is consistent with the results of Gruber et al. (2009) who used a subset of the ocean and atmospheric inversions considered here. The agreement in the sea-air flux estimates based on the different approaches is a significant improvement in recent years over previously published studies, e.g. Roy et al. (2003). The northern boundary of the Southern Ocean is set at 44 • S to conform to REC-20 CAP protocols. This excludes some of the SAZ region, which extends to the Subtropical Front (STF) at about 40 • S in some basins (Fig. 1) The largest median absolute deviation (MAD) in the annual sea-air flux is for ocean biogeochemical models (Table 3, Fig. 3). However, the small number of ocean biogeochemical models (5) compared to atmospheric inversions (11) can result in greater 5 MAD values for the ocean models.
The spatial plots of the annual mean uptake for 1990-2009 (Fig. 4) show that all the ocean models simulate a negative sea-air flux or uptake equatorward of 50 • S. The CCSM-BEC, CCSM-ETH and CSIRO models have similar patterns of sea-air fluxes as the values derived from observations with areas of CO 2 flux to the atmosphere at high 10 latitudes (poleward of 50 • S) and areas of CO 2 flux into the ocean to the North of the Subantarctic Front at about 50 • S. However, the fluxes to the atmosphere from these models tend to be greater than the values from observations. The CSIRO model also simulates a region of uptake in the eastern Pacific that is not apparent in the observations, although this region has few measurements to constrain the observational based The CSIRO model also simulates regions of net flux to the atmosphere at high latitudes (Fig. 4), but these are largely offset by other zones of uptake in the same latitude range, resulting in only a slight maximum in the cumulative, zonally integrated flux for 5 this model near 50 • S. A band of high CO 2 uptake in the Subantarctic and Subtropical waters produces a cumulative uptake to 44 • S of −0.42 Pg C yr −1 for the CSIRO model.
The cumulative uptake for both NEMO models increases to 44 • S to −0.4 Pg C yr −1 (NEMO-Plankton5) and −0.8 Pg C yr −1 (NEMO-PISCES). The large deviation in the annual uptake for the ocean biogeochemical models appears to be in part due to how 10 the sea-air CO 2 fluxes are simulated at latitudes poleward of 50 • S.
Atmospheric CO 2 inversions are usually constrained not only by the atmospheric CO 2 data, but also by a first guess or a priori flux estimate. The estimated Southern Ocean flux was not very sensitive to this prior information; across inversions prior annual fluxes were clustered around −0.35 Pg C yr −1 or −1.0 Pg C yr −1 but this clus- 15 tering is not maintained in the estimated fluxes from the inversions. This suggests that the observing network for this region may be sufficient to constrain the flux estimates. However other differences between the inversions, such as the modeling of atmospheric transport, contribute to the range in atmospheric results. For example, a transport model with vigorous mixing of higher CO 2 concentration air from the north 20 would require a larger Southern Ocean sink to maintain the north-south gradient of CO 2 concentration, than a transport model with slower mixing of air from the north.
The ocean inversions do not use a priori estimates, but they may be sensitive to biases in the data based techniques used to estimate the components of the observed dissolved inorganic carbon in the ocean due to anthropogenic carbon uptake and sea-25 air gas exchange (e.g. Matsumoto and Gruber, 2005;Gerber et al., 2009Gerber et al., , 2010. Like atmospheric inversions, the ocean inversion is likely to be sensitive to biases in model transport, particularly in the Southern Ocean, but the use of a suite of different models has been employed to help quantify this uncertainty. Although ocean CO 2 inversions have been less widely used than atmospheric CO 2 inversions, this approach thus far seems to be relatively insensitive to the choice of inverse methodology (Gerber et al., 2009(Gerber et al., , 2010.

Southern Ocean, 44 • S-58 • S
The median annual sea-air flux values for the different approaches varies between ing the majority of the net uptake occurs in this latitude band, consistent with previous studies (Metzl et al., 2006;McNeil et al., 2007;Takahashi et al., 2009Takahashi et al., , 2012. The contributions of the Atlantic, Indian, and Pacific Ocean sectors to the annual 10 uptake in the 44 • S-58 • S band are similar when all 26 models are grouped (−0.13 ± 0.04, −0.11 ± 0.03 and −0.11 ± 0.04, respectively; Table 3). The ocean biogeochemical models and observations do suggest that greater uptake occurs in the Pacific sector, which is not apparent in the atmospheric or ocean inversion results.
It is important to note that the flux values for all approaches are associated with 15 a large range, particularly ocean biogeochemical models (see Sect. 3.1.1) and atmospheric inversions. This is most evident in the Indian and Southeast Pacific Oceans where ocean biogeochemical models differ in the sign of the flux (Fig. 4). For atmospheric inversions, the basin split is dependent on the spatial resolution of the inversion; inversions that solve for only 1 or 2 regions rely on their prior information to deter-20 mine the basin split, while inversions that solve for many regions may be susceptible to "over-fitting" the atmospheric measurements and thereby biasing the flux estimates. This may be a particular problem for the Southern Ocean as atmospheric CO 2 variations between observing stations are very small, placing high demands on data quality and a transport model's ability to represent that data. This is consistent with previous studies based on ocean inversions, which have shown that this approach cannot robustly separate the Indian and Pacific Ocean basins based on available data (e.g. Mikaloff Fletcher et al., 2006. 300

Southern Ocean 58 • S-75 • S
The influence of this high latitude band (40 % of the total Southern Ocean surface area) on the annual sea-air CO 2 flux is small relative to the 44 • S-58 • S region, accounting for only about 10 % of the total Southern Ocean sea-air flux (Table 3). Ocean biogeochemical models and ocean inversions indicate that this region is a small net 5 sink of atmospheric CO 2 annually (−0.04 ± 0.09 and −0.07 ± 0.01 Pg C yr −1 ), with atmospheric inversions and observations suggesting a small net flux to the atmosphere (+0.03 ± 0.03 and +0.04 ± 0.02 Pg C yr −1 ). Ocean biogeochemical models again show the largest range in fluxes (Table 3 and Fig. 3). This variability may be due to: (i) the representation of the sea-ice and asso-10 ciated gas exchange differs across models, e.g. Rysgaard et al. (2011); and (ii) the coarse resolution of the models which precludes the proper representation of potentially important features such as polynas as well as other important coastal ocean processes (Marsland et al., 2004). This part of the Southern Ocean remains one of the most poorly sampled of all ocean regions ( Fig. 2; Monteiro et al., 2010) and the 15 uncertainty in the flux estimates based on observations may be underestimated.
This region was regarded as a strong annual sink of atmospheric CO 2 based on summer data (Takahashi et al., 2002), while more recent observations (Takahashi et al., 2009) have suggested that the higher latitude Southern Ocean is neutral or a weak source of atmospheric CO 2 over the annual mean. These estimates are based more 20 on open-ocean observations rather than data collected in marginal seas and coastal margins of the Southern Ocean. These areas that have a high variability in biological production (e.g. Sweeney et al., 2000;Hales and Takahashi, 2004) and as a consequence have been suggested to play an important role in the global carbon budget (Arrigo et al., 2008). Amongst the approaches in this study, only atmospheric inver- 25 sions have the potential to capture the integrated coastal, sea-ice and open-ocean responses in this region. That these inversions suggest that this region is not a large net sink of CO 2 suggests either that: (i) outgassing in the more northward portion of this region offsets any strong uptake; or (ii) or the role of the coastal ocean and sea-ice zone is not well captured.

Seasonal sea-air CO 2 fluxes
In this section we consider how the various modeling approaches represent the seasonality in the sea-air CO 2 exchange compared to observations. This provides insights 5 into the ability of ocean biogeochemical models to represent the complex interplay of physical and biological processes that drive sea-air CO 2 exchange. The ability of a model to reproduce the seasonal cycle also provides some indication of the ocean biogeochemical models ability to correctly represent climate sensitive processes that could influence long-term projections of the ocean CO 2 uptake. The multi-model me-10 dian seasonal anomalies of sea-air CO 2 fluxes are shown in Fig. 6, while the individual models and observations are shown in Fig. 7.

Southern Ocean 44 • S-75 • S
Figures 6a shows the median and MAD over the entire Southern Ocean RECCAP region. Some ocean biogeochemical models do simulate the seasonal cycle in the sea- 15 air CO 2 flux estimated from observations (Figs. 6 and 7). In summer, a CO 2 flux into the ocean surface (negative flux) results from the combined effects of increased surface stratification, warming, and increased biologically driven CO 2 uptake. Deeper mixing and lower production, offset to some degree by surface cooling, results in a reduced flux into the ocean and potentially outgassing of CO 2 in winter (Takahashi et al., 2009). 20 All approaches tend to give similar estimates in summer, but the atmospheric inversions do not capture the strong winter response evident in ocean models and estimated from observations. To explore the relationship between the seasonal cycle and the annual mean uptake, we used a 2-dimensional (2-D) Taylor Diagram (Fig. 8a). The annual mean uptake 25 and timing of the seasonal flux changes relative to observationally based estimates 302 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | were compared for each ocean biogeochemical model and atmospheric inversion. The analysis provides a synthesis of the ability of all models to capture the seasonal and annual sea-air fluxes. In this analysis, we do not account for the uncertainty associated with the observations. The x-axis in each Taylor diagram is the normalized standard deviation of the sea- 5 sonal cycle (σ model /σ obs ); the closer the value is to 1 (denoted as a red arc in Fig. 8) the better it reproduces the magnitude of the seasonal cycle from the observations. The correlation of the model seasonal cycle with observations is shown on the arc; models or inversions with correlations of +1 (x-axis), would simulate the seasonality in the sea-air flux from observations. Models that lay in the right quadrant have positive correlation, while those in the left quadrant have negative, or anti-phase correlations. The colors of individual symbols represent the difference between the annual sea-air flux based on observations and the value for each model (i.e. observations -model/inversions). Ocean biogeochemical models are represented as circles, while triangles represent atmospheric inversions.

15
A diverse set of responses is evident in Fig. 8a. Many atmospheric inverse and ocean biogeochemical models underestimate the magnitude of the seasonal cycle. Several models and inversions appear to capture the magnitude and the phase of the seasonal cycle, however these models tend to strongly underestimate the magnitude of annual mean uptake (blue symbols with large negative values for the observations mi-20 nus model difference). Some models and inversions do a poor job at capturing the phase of the seasonal cycle, but show better agreement in the annual mean uptake. These results are particularly worrisome, given that some models are unable to simulate even the seasonality, which is the largest scale of variability in the Southern Ocean.
The reasons behind the poor seasonality in many atmospheric inversions remain un- in determining the seasonal cycle in atmospheric CO 2 at stations that observe the Southern Ocean, errors in the fluxes estimated from other regions or in the modeled transport of those fluxes to the high latitude Southern Hemisphere could lead to biases in the seasonal cycle. Secondly, some of the inversions that solve for smaller ocean regions (e.g. C13 CCAM law) show very different seasonality between basins, sug-5 gesting some caution should be applied to those results. Possibly the greater flexibility in those inversions to fit the atmospheric data makes them vulnerable to any data quality or representativeness issues as well as transport model errors. When fewer regions are solved for, the inversion compromises the fit across sites effectively ignoring any poorly calibrated data. Conversely, inversions that solve for only a few regions could be 10 missing important spatial variability in the seasonal cycle, which can lead to biases in the inverse estimates (Kaminski et al., 2001). The representation of the seasonal cycle varies widely across the ocean biogeochemical models (Fig. 7). If the summer warming is too strong in the upper ocean, the solubility response can dominate over the biological productivity leading to a peak in 15 the sea-air flux that is several months out of phase with observations (Fig. 8a). The net primary productivity from the models (not shown) has a similar magnitude over summer. This suggests that changes in the seasonal temperature and mixed layer depth are likely to be more important than the differences between biological models in causing the varied responses in the ocean models. Interestingly, capturing the seasonal 20 cycle of the Southern Ocean is not a prerequisite to reproducing the annual mean seaair flux calculated from observations. However, the inability of the models to simulate the observation-based seasonality in the sea-air CO 2 flux does bring into question the ability of these models to realistically project the response of the Southern Ocean CO 2 flux to climate change.

Southern Ocean 44 • S-58 • S
As with the entire Southern Ocean, the median seasonality for ocean biogeochemical models and observations agree in terms of the phase and magnitude of the seasonal 304 Introduction  (Fig. 6b). The atmospheric inverse models simulate relatively weak seasonal amplitude. The similarity in the magnitude and phase with the total response (44 • S-75 • S) is consistent with this region dominating the seasonality in the sea-air flux for the Southern Ocean (Metzl et al., 2006;Takahashi et al., 2009Takahashi et al., , 2012. The 2-D Taylor diagram for individual models (Fig. 8b) shows a diverse range of 5 responses. Some models represent the observation-based phase and magnitude of the seasonal cycle, but do a poor job in capturing the annual uptake. Some models that do represent approximate the annual uptake can display poorer magnitude and phase of the seasonal cycle. While individual models may not represent both the seasonal cycle and annual sea-air CO 2 flux, taking the median of multiple models can improve 10 the situation.

Southern Ocean 58 • S-75 • S
At high latitudes, the magnitude of the seasonal cycle in the sea-air CO 2 flux is larger for fluxes derived from observations than for the ocean models and atmospheric inversions (Fig. 6c). While the annual median values (Table 3) are quite low in this region, 15 observations indicate a relatively large summer flux into the ocean and a flux to the atmosphere in winter. Neither ocean biogeochemical models nor atmospheric inversions show a well-defined seasonal cycle. The atmospheric inversions show no evidence of a large summer CO 2 uptake (negative flux) associated with the coastal ocean and marginal seas as suggested by Arrigo et al. (2008). 20 The 2-D Taylor diagram showing the behavior of the individual models for this region is shown in Fig. 8c. The biogeochemical models underestimate the seasonal cycle relative to observations, but most capture the observed phase and the models tend to agree on the magnitude of the seasonal cycle, explaining the small range of the ocean biogeochemical models in Fig. 6c. The smaller magnitude of the seasonal cycle in The atmospheric inversions show a diverse set of responses in this region. Consistent with the median, we see that the majority of these models underestimate the seasonal cycle relative to observations. Clearly most of the models do a good job capturing the phase of the seasonal cycle but do poorly at representing the annual mean uptake. However some models, while capturing the annual mean uptake well show very 5 little or poor seasonality. Some inversions produce a semi-annual cycle. These results again highlight that a well-represented seasonal cycle is not a prerequisite for capturing the annual mean uptake in atmospheric inverse models well. This is also confirmed by the negligible a posteriori correlations between seasonal anomalies and the mean in atmospheric inversions, e.g. Rödenbeck (2005).

Longer-term variability
Understanding and quantifying inter-annual variability in the Southern Ocean is key to projecting the future response of Southern Ocean sea-air fluxes. The detection of a long-term trend is difficult as observations tend to be most common in the Austral Summer and studies have shown that changes in sea surface temperature and net 15 production can cause considerable variability in sea-air fluxes at regional scales during the Summer months (Jabaud-Jan et al., 2004;Brévière et al., 2006;Borges et al., 2008;Brix et al., 2012). While a few observational studies attempt to describe the longer-term change of the carbon system, they focus on oceanic pCO 2 rather than changes in sea-air CO 2 fluxes (Inoue and Ishii, 2005;Lenton et al., 2012;Metzl, 2009;20 Midorikawa et al., 2012;Takahashi et al., 2009) or are station studies that may have local influences (Currie et al., 2009). Consequently as we are focusing on sea-air CO 2 fluxes we only use atmospheric inversion and ocean biogeochemical models over the period 1990-2009, rather than observations.

Southern Ocean 44 • S-75 • S
The simulated median inter-annual variability and associated uncertainty in sea-air CO 2 fluxes from ocean biogeochemical models and atmospheric inverse models are shown in Fig. 9. The inter-annual variability in the period 1990-2009 from ocean biogeochemical models and atmospheric inversions are of similar maximum value (+0.10 5 and +0.11 Pg C yr −1 , respectively). This represents 20 % of the total median sea-air flux from ocean biogeochemical models and 35 % of the median flux from atmospheric inversions. The positive and negative flux anomalies are of similar magnitude. The region 44 • S-58 • S can explain about 75 % of the inter-annual variability in the Southern Ocean sea-air CO 2 flux in the atmospheric inversions (+0.07 Pg C yr −1 ) and 10 ocean biogeochemical models (+0.08 Pg C yr −1 ; Fig. 9b). The remaining 25 % of the inter-annual variability is attributable to the region 58 • S-75 • S (Fig. 9c). These results suggest that south of 58 • S the median inter-annual variability from ocean biogeochemical models (+0.03 Pg C yr −1 ) and atmospheric inversions (+0.07 Pg C yr −1 ) can be as large as the net annual sea-air CO 2 flux (Table 3). In this region the range in sea-air 15 flux for ocean biogeochemical models is lower than in the atmospheric inverse models (consistent with Sect. 3.1.3). The large inter-annual variability poleward of 58 • S may be due to biogeochemical responses to changes in heat and freshwater fluxes, wind stress and sea-ice cover. This could indicate that sea-air CO 2 fluxes in this region are more sensitive to climate than previously believed. The sea-air CO 2 fluxes in the 44 • S- at Jubany (58 • W, 62 • S), which has periods in 2003-2004 when the CO 2 concentration is 0.7-1.0 ppm lower than at nearby Palmer Station (64 • W, 65 • S). Consequently those inversions that include this data give larger fluxes in 2003 than those that do not. It is interesting that this sensitivity occurs for inversions that span the range of flux resolution that the inversions solve for (i.e. one or many Southern Ocean regions).

5
The positive and negative anomalies in 1997-1998 and 2000 are harder to attribute. Similar, though larger, anomalies are also seen for Southern Hemisphere land regions (Peylin et al., 2012, Fig. 6). This suggests that the atmospheric data may be insufficient to clearly differentiate land and ocean flux anomalies due to the usual practice of selecting atmospheric measurements to be representative of well-mixed air masses and 10 thus removing data that has had recent contact with land. The median of the ocean biogeochemical models shows a larger sea-air CO 2 flux in 2009 relative to the start of the study period in 1990 (Fig. 9). This is expected as the CO 2 gradient between the atmosphere and ocean has increased in response to continuing atmospheric emissions of CO 2 . The median of the atmospheric inversions 15 shows no increase in CO 2 uptake over the study period. However, our confidence in these long-term trends is low, given the magnitude of the inter-annual variability and the relatively short period of simulations. Figure 10a, b depicts linear decadal trends computed as a function of time using a 10-yr sliding window centered on the reported year, following Lovenduski et 20 al. (2008). The ocean biogeochemical models and atmospheric inversions produce mostly positive trends in the 1990's and negative trends in the 2000's. Calculating the linear trend in Southern Ocean fluxes over the maximum period available from any given model (excluding those with less than 10 yr of simulation output) yields trends that range from −0.3 to +0.3 Pg C yr −1 decade −1 . This suggests that linear trends in 25 model output over periods less than 20 yr is unlikely to provide a statistically meaningful statement about the changing rate of Southern Ocean CO 2 uptake. This is supported by the results of McKinley et al. (2011) for the North Atlantic.
308 synthesis of surface ocean observations; (ii) atmospheric inverse models; (iii) ocean inversions; and (iv) ocean biogeochemical models. The goal of this study is to combine these different approaches to quantify and assess how well the models represent the mean and variability of sea-air CO 2 fluxes in the Southern Ocean in comparison to flux estimates derived from observations. We used the recalculated sea-air CO 2 flux climatology of Wanninkhof et al. (2012) as our observational product; five different ocean biogeochemical models driven with observed atmospheric CO 2 concentrations; eleven atmospheric inverse models using atmospheric records collected around the Southern Ocean; and ten ocean inverse models.
Our results show that the median annual sea-air flux from all four approaches The choice of the RECCAP boundary at 44 • S was problematic in comparing our results with published values from other studies (e.g. Metzl et al., 1999;Boutin et al., 2008;McNeil et al., 2007;Barbero et al., 2011;Takahashi et al., 2012). Therefore while we focused on the comparison between the four different techniques presented here, a more comprehensive analysis of individual regions needs to be undertaken in future 5 studies. Such studies would be timely, particularly in light of recent work highlighting the heterogeneity of the anthropogenic carbon transport out of the Southern Ocean (Sallee et al., 2012).
At seasonal times scales, the fluxes estimated from observations and the median of the ocean biogeochemical models capture a well-defined seasonal cycle in the seaair CO 2 flux. Atmospheric inversions showed only very weak or little seasonality in all regions of the Southern Ocean. All approaches tend to show enhanced flux into the ocean in the biologically productive summer period. The largest seasonality was found in the region 44 • S-58 • S for both ocean biogeochemical models and based on observations. South of 58 • S, neither ocean biogeochemical models nor atmospheric 15 inversions were able to capture the magnitude of the observed seasonal cycle. These differences between models and observational estimates may reflect the model formulation and a poor understanding of the high latitude carbon cycle. None of the models were capable of simulating the magnitude and phase of seasonality and the annual mean sea-air flux at the same time in any of the regions. This raises serious concerns 20 about projecting the future changes in Southern Ocean CO 2 uptake. Inter-annually, ocean models and atmospheric inversions show that the variability in the Southern Ocean sea-air CO 2 flux can be as large as 25 % of the annual mean. Atmospheric inversions tend to produce a larger spread in the inter-annual variability of the sea-air flux than ocean biogeochemical models. Both modelling approaches sug- 25 gest that about 25 % of the total inter-annual variability can be explained by the region south of 58 • S. This implies that this variability can be as large as the net annual mean sea-air CO 2 flux in this region. Decadal trends over the Southern Ocean in the period 310 Boutin, J., Merlivat, L., Henocq, C., Martin, N., and Sallee, J. B.: Air-sea CO 2 flux variability in frontal regions of the Southern Ocean from CARbon Interface OCean Atmosphere drifters, Limnol Oceanogr, 53, 2062-2079, 2008 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Piao, S., Fang, J., Ciais, P., Peylin, P., Huang, Y., Sitch, S., and Wang, Table 3. The annual CO 2 sea-air CO 2 flux (negative into the ocean) from observations (with assumed 50 % uncertainty) and multi-model median sea-air CO 2 flux (negative into the ocean) and median absolute deviation (MAD) (MAD) from ocean biogeochemical models, atmospheric and ocean inversions, and all of the models. A, I, P refer to the Atlantic, Indian and Pacific Sectors of the Southern Ocean. All units are Pg C yr −1 .