Towards robust regional estimates of CO 2 sources and sinks using atmospheric transport models

: Information about regional carbon sources and sinks can be derived from variations in observed atmospheric CO 2 concentrations via inverse modelling with atmospheric tracer transport models. A consensus has not yet been reached regarding the size and distribution of regional carbon fluxes obtained using this approach, partly owing to the use of several different atmospheric transport models. Here we report estimates of surface-atmosphere CO 2 fluxes from an intercomparison of atmospheric CO 2 inversion models (the TransCom 3 project), which includes 16 transport models and model variants. We find an uptake of CO 2 in the southern extratropical ocean less than that estimated from ocean measurements, a result that is not sensitive to transport models or methodological approaches. We also find a northern land carbon sink that is distributed relatively evenly among the continents of the Northern Hemisphere, but these results show some sensitivity to transport differences among models, especially in how they respond to seasonal terrestrial exchange of CO 2 . Overall, carbon fluxes integrated over latitudinal zones are strongly constrained by observations in the middle to high latitudes. Further significant constraints to our understanding of regional carbon fluxes will therefore require improvements in transport models and expansion of the CO 2 observation network within the tropics.

Information about regional carbon sources and sinks can be derived from variations in observed atmospheric CO 2 concentrations via inverse modelling with atmospheric tracer transport models. A consensus has not yet been reached regarding the size and distribution of regional carbon¯uxes obtained using this approach, partly owing to the use of several different atmospheric transport models 1±9 . Here we report estimates of surface± atmosphere CO 2¯u xes from an intercomparison of atmospheric CO 2 inversion models (the TransCom 3 project), which includes 16 transport models and model variants. We ®nd an uptake of CO 2 in the southern extratropical ocean less than that estimated from ocean measurements, a result that is not sensitive to transport models or methodological approaches. We also ®nd a northern land carbon sink that is distributed relatively evenly among the continents of the Northern Hemisphere, but these results show some sensitivity to transport differences among models, especially in how they respond to seasonal terrestrial exchange of CO 2 . Overall, carbon¯uxes integrated over latitudinal zones are strongly constrained by observations in the middle to high latitudes. Further signi®cant constraints to our understanding of regional carbon¯uxes will therefore require improvements in transport models and expansion of the CO 2 observation network within the tropics. We estimate annual average¯uxes for the 1992±96 period using each transport model and a common inversion set-up (see Methods). Methodological choices for this`control' inversion have been selected on the basis of knowledge gained from a wide range of sensitivity tests (to be reported elsewhere). Performing the inversion with multiple transport models gives mean estimated uxes that are relatively insensitive to reasonable variations in the set-upÐand estimated uncertainties that represent a more complete estimate of the true uncertainty. The maximum number of regions in our inversion and the spatial distributions of¯uxes within each region are ®xed, precluding sensitivity tests of these inversion components. Figure 1 shows the mean¯ux estimates (left-hand cross in each box) and two uncertainty measures for the control inversion. The ®rst uncertainty measure is the mean of the individual model¯ux uncertainties (circles) which we designate the`within-model' uncertainty. For any region, this estimated¯ux uncertainty must be smaller than the prior¯ux uncertainty (outer bounds of the boxes). The magnitude of the decrease indicates the degree to which the ®nal¯ux estimate is constrained by the measurements. Figure 1 shows that the northern land regions and Australia are better constrained by the measurements than are the remaining land regions. The Southern Ocean region is well constrained by the atmospheric measurements, in part because it is treated as a single large region. The Atlantic regions are constrained more by their prior¯ux uncertainties, which are relatively small due to better coverage of ocean measurements in these regions.
The second uncertainty measure is the standard deviation of thē ux estimates over the ensemble of models (error bars in Fig. 1). We call this the`between-model' uncertainty. This measure indicates the degree to which transport model differences contribute to the range of¯ux estimates. Large between-model uncertainties are found for northern Africa, tropical America, temperate Asia and boreal Asia (all greater than 0.5 Gt C yr -1 ).
For most regions, the between-model uncertainties are of similar or smaller magnitude than the within-model uncertainties. This suggests that the choice of transport model is not the critical determinant of the inferred¯uxes. Comparing the uncertainties between regions indicates where the inversion would bene®t most from new observations, and where model improvements are most needed. In this particular inversion, new measurements would be most useful over tropical continents and in the South America and South Atlantic regions, while the focus for resolving transport differences would be the northern and tropical land regions.
Regarding the model mean¯ux estimates, two results deserve attention. First, we ®nd consistency between the ocean¯uxes predicted in this study and those based on a global database 10 of CO 2 partial pressure (p CO 2 ), except in the Southern Ocean where the carbon uptake estimated here is roughly half that based on the p CO 2 database. This shift in uptake from south to north is required to match simultaneously large-scale concentration gradients (Fig. 2) and growth rates.
The mismatch between atmospheric and ocean estimates of the Southern Ocean¯uxes had been noted a decade ago 11 . Our sensitivity tests ®nd that the near-uniformity of observed concentration in the Southern Hemisphere and the small uncertainty associated with those measurements make this result robust to the choice of observing network, prior¯ux estimates, global ocean constraint, and transport (see Fig. 2 in Supplementary Information). The discrepancy also cannot be explained by a systematic bias in transport models, as the north±south transport has been investigated in a recent intercomparison 12 where successful simulations of the observed meridional gradient in SF 6 suggested reasonable veracity in gross interhemispheric transport. One possible reconciliation between the p CO 2 database and the inverse result presented here is suggested by recent ocean measurements taken during January and August 2000 in the Indian Antarctic sector of the Southern Ocean 13 . The p CO 2 values south of 508 S showed seasonal variations that require CO 2 uptake in summer and emission in winter. If the seasonality exhibited in this campaign is true for other parts of the Southern Ocean, this would result in a reduction of the Southern Ocean¯ux uptake in the database, which is currently determined predominantly by summer measurements. This seasonality-driven explanation is also consistent with forthcoming results from the second stage of the Trans-Com 3 comparison in which we estimate seasonal cycles (K.R.G. et al., manuscript in preparation). Second, we ®nd carbon uptake over the continents of the Northern Hemisphere to be distributed relatively evenly across North America, Europe and Asia, in contrast to the distribution found in an earlier, widely cited inverse study 2 . We ®nd a temperate North American sink approximately 60% of that found in the earlier study, a small boreal North American source rather than small uptake, and a large sink for Eurasia rather than an approximately neutral¯ux. Estimated uncertainties are moderate (0.4±0.7 Gt C yr -1 ), indicating that regional partitioning remains dif®cult, but the¯ux differences between the two studies lie at the edge of (or outside) the uncertainty ranges.
Although previous studies have challenged the possibility of a large North American sink 3±7 , little systematic exploration has been performed as to how such a result was achieved. The differences are not due to the choice of transport model, because the two models used in the earlier study are included here and lie in the middle of our range. Extensive sensitivity tests (see Tables 3 and 4 in Supplementary Information) indicate that the Eurasian¯ux estimate is very sensitive to the pattern of background¯uxes used in the inversion, especially that representing the seasonal terrestrial biosphere. The difference in North American uptake results from a combination of methodological choices as well as differences in time period and observational stations used.
There are three methodological differences that together appear to be critical. First, recent work 14 suggests that the larger the region size in an inversion, the greater the potential for producing biased ux estimates. Second, the potential bias can be reduced by increasing the data uncertainty for sites in regions with spatially heterogeneous¯uxes. The earlier study 2 inverted for larger regions than used here, and used relatively small (0.6 p.p.m.), spatially invariant uncertainties compared to the generally larger, variable uncertainties used in this study. The third factor is the uncertainty assigned to prior estimates of ocean¯uxes, which were zero in the earlier study. Thus the¯ux adjustment required to match the atmospheric data was applied only to land regions. Together these three factors suggest that the earlier study had greater potential for biased and more sensitive¯ux estimates than the control results presented here.
Although transport uncertainties do not overwhelm our¯ux estimates, one factor appears to be responsible for a signi®cant portion of the model spread; the`recti®er' produced by the covariance between the seasonal biospheric background¯ux and atmospheric transport 15 . The effect of the recti®er can be seen by performing the inversion without the background biospheric¯uxes ( Fig. 1, right-hand symbols within each box). The between-model uncertainty is reduced for almost all regions, and in some regions there are substantial changes to the estimated¯uxes. An increase of 1.1 Gt C yr -1 in boreal Asia changes it from a moderate sink to a moderate source, because recti®cation produces the strongest concentrations downwind of this region in many of the models. Sink strengths increase by 0.35±0.55 Gt C yr -1 for temperate North America, temperate Asia and northern Africa, to maintain the required global source. Measurements indicating the strength of the covariance effect in nature are needed to assess this aspect of model transport.
One way to reduce the large uncertainties in our full calculation is to aggregate our regions after performing the inversion.  Source (Gt C yr -1 ) Figure 3 Mean sources and uncertainties for six aggregated regions and global land and ocean. Symbols as in Fig. 1  globe as a whole. Within-model uncertainties are reduced relative to simply summing from constituent regions, because much of the uncertainty occurs at the scale of the original regions. There is also a reduction in the between-model uncertainty for this aggregation, as some of the model spread involves details of regional transport. At this larger scale, the decrease in the sink in the Southern Ocean and the enhanced sinks over the Northern Hemisphere appear more signi®cant. The uncertainty on the tropical land region is large; lack of atmospheric data in this region means that inversion methods cannot reliably comment on the extent to which sources due to tropical land-use change are balanced by enhanced growth. This ®rst stage of the TransCom 3 intercomparison has explored many aspects of annual mean inversions more comprehensively than previous work. By incorporating a range of transport models, the¯uxes and their uncertainties represent progress towards more robust inverse estimates of regional carbon exchange. Carbon exchange with the ocean is well constrained in this study and, in the case of the Southern Ocean region, is different from¯uxes suggested by p CO 2 measurements. This result is consistent across the 16 transport models used here, and is insensitive to many aspects of the inversion set-up. Flux estimates in the northern extratropical land regions are reasonably robust as zonal means, but are dif®cult to distinguish in the longitudinal direction, and can be biased owing to key methodological aspects of the inversion construction. Seasonal exchange with the terrestrial biosphere is responsible for much of the model spread over these regions. Realistic characterization of this aspect of model transport is essential if this uncertainty is to be reduced in the future.
Flux estimates in the tropical land regions remain very uncertain, owing to few CO 2 observations and the limited in¯uence of extratropical observations on the tropical land¯ux estimates. New observations that can be represented by global-scale transport models are needed in these regions. Future TransCom analysis will focus on the effect of transport model differences on¯ux estimates at seasonal and interannual timescales.

Methods
We use a bayesian synthesis inversion formalism 16 that speci®es prior estimates of both thē uxes and their uncertainty, and optimizes with respect to atmospheric observations that are also uncertain. We estimate¯uxes for 11 land and 11 ocean regions (see Supplementary Information) as differences from`background'¯uxes that are run separately through each transport model and represent fossil-fuel emissions 17,18 , seasonally varying air±sea gas exchange 10 and an annually balanced, seasonally varying¯ux due to terrestrial photosynthesis and respiration 19 . The use of seasonally varying background¯uxes allows the annual mean inversion to include contributions to annual mean concentrations due to the covariance of atmospheric transport and seasonal¯uxes.
We invert 5-year mean measurements for 1992±96 at 76 sites taken from the GLOBALVIEW-2000 data set 20 (see Supplementary Information). GLOBALVIEW is a data product that interpolates CO 2 measurements to a common time interval. Gaps in the data are ®lled by extrapolation from marine boundary layer measurements. We have chosen to use sites where the extrapolated data accounts for less than 30% of the 1992±96 period. The measurements are weighted inversely by the degree to which the predicted concentrations are required by the inverse process to match the observations, which we refer to as 'data uncertainty'. In addition to measurement precision, this uncertainty incorporates the inability of coarse-grid models to adequately represent discrete measurements. The relative uncertainty of one site to another was based on the mean residual standard deviations for 1992±96 from GLOBALVIEW. The absolute magnitudes were chosen to produce a mean square normalized residual out of the inversion of about 1.0, ensuring that the estimated¯uxes were optimized to the data only to an appropriate level commensurate with our ability to model them. A minimum uncertainty was also speci®ed. This gave uncertainties ranging from 0.25 p.p.m. for remote,`clean air' sites to 2.2 p.p.m. for continental,`noisy' sites (Fig. 2).
The 11 land basis region boundaries were constructed to enclose vegetation of similar seasonal structure and carbon exchange based on vegetation classi®cation 21 . Ocean basis regions were chosen to approximate circulation features such as gyres and upwelling regions. Unit emissions of 1 Gt C yr -1 were speci®ed from each region. Subregional-scale variations in emissions rates were prescribed for land regions according to simulated net primary production from the CASA model 21 . This assumes that carbon¯uxes follow the distribution of vegetation productivity. Emissions from ocean regions were prescribed as spatially uniform, except that sea-ice was masked out using seasonally varying fractional ice cover distributions 22 . The inversion requires prior¯ux and uncertainty estimates. Our choices have been guided by ocean and terrestrial¯ux models and observations 10,19 , and are shown in Fig. 1 (also see Table 2 in Supplementary Information). The land region prior¯u x estimates incorporate recent inventory estimates 23±30 . Where more than one estimate for a given region was considered, a mid-point of the estimate spread was used. The prior ux uncertainty was chosen to be large enough to encompass all estimates. Prior¯ux uncertainties re¯ect one standard deviation.
The inversion is run separately for 16 transport models or model variants. The models used ( . The inversion produces estimated¯uxes and their uncertainties for each region individually and for some groups of regions in addition to a background concentration. Here our analysis focuses on¯uxes and uncertainties that are averaged across models. We also specify two measures of uncertainty as described in the main text. We show the`between' model uncertainty as a one standard deviation con®dence interval for meaningful comparison with the within model uncertainty. The obvious alternative, showing a full range, would also produce a con®dence interval that would widen as more models were included. Inspection of the individual¯ux estimates showed them to be close to normally distributed about the mean¯ux for most regions. Aerobic, anoxygenic, phototrophic bacteria containing bacteriochlorophyll a (Bchla) require oxygen for both growth and Bchla synthesis 1±6 . Recent reports suggest that these bacteria are widely distributed in marine plankton, and that they may account for up to 5% of surface ocean photosynthetic electron transport 7 and 11% of the total microbial community 8 . Known planktonic anoxygenic phototrophs belong to only a few restricted groups within the Proteobacteria a-subclass. Here we report genomic analyses of the photosynthetic gene content and operon organization in naturally occurring marine bacteria. These photosynthetic gene clusters included some that most closely resembled those of Proteobacteria from the b-subclass, which have never before been observed in marine environments. Furthermore, these photosynthetic genes were broadly distributed in marine plankton, and actively expressed in neritic bacterioplankton assemblages, indicating that the newly identi®ed phototrophs were photosynthetically competent. Our data demonstrate that planktonic bacterial assemblages are not simply composed of one uniform, widespread class of anoxygenic phototrophs, as previously proposed 8 ; rather, these assemblages contain multiple, distantly related, photosynthetically active bacterial groups, including some unrelated to known and cultivated types.
Most of the genes required for the formation of bacteriochlorophyll-containing photosystems in aerobic, anoxygenic, phototrophic (AAP) bacteria are clustered in a contiguous, 45-kilobase (kb) chromosomal region (superoperon) 6 . These include bch and crt genes coding for the enzymes of the bacteriochlorophyll and carotenoid biosynthetic pathways, and the puf genes coding for the subunits of the light-harvesting complex (pufB and pufA) and the reaction centre complex (pufL and pufM). To better describe the nature and diversity of planktonic, anoxygenic, photosynthetic bacteria, we screened a surface-water bacterial arti®cial chromo-