- Top of page
- 1. Introduction
- 2. Methods
- 3. Results
- 4. Discussion
- 5. Conclusions
- 6. Acknowledgments
We evaluate North American carbon fluxes using a monthly global Bayesian synthesis inversion that includes well-calibrated carbon dioxide concentrations measured at continental flux towers. We employ the NASA Parametrized Chemistry Tracer Model (PCTM) for atmospheric transport and a TransCom-style inversion with subcontinental resolution. We subsample carbon dioxide time series at four North American flux tower sites for mid-day hours to ensure sampling of a deep, well-mixed atmospheric boundary layer. The addition of these flux tower sites to a global network reduces North America mean annual flux uncertainty for 2001–2003 by 20% to 0.4 Pg C yr−1 compared to a network without the tower sites. North American flux is estimated to be a net sink of 1.2 ± 0.4 Pg C yr−1 which is within the uncertainty bounds of the result without the towers. Uncertainty reduction is found to be local to the regions within North America where the flux towers are located, and including the towers reduces covariances between regions within North America. Mid-day carbon dioxide observations from flux towers provide a viable means of increasing continental observation density and reducing the uncertainty of regional carbon flux estimates in atmospheric inversions.
- Top of page
- 1. Introduction
- 2. Methods
- 3. Results
- 4. Discussion
- 5. Conclusions
- 6. Acknowledgments
About half of the anthropogenic carbon emitted into the atmosphere remains in the atmosphere each year. The remainder is taken up by the ocean and terrestrial ecosystems through the processes responsible for the natural exchange of carbon between the atmosphere and terrestrial vegetation and the surface ocean (Denman et al., 2007; Forster et al., 2007). Numerous studies (Myneni et al., 2001; Nemani et al., 2003; Potter et al., 2003) show that climate cycles, local weather and ecosystem conditions all affect the interannual variability of this uptake of carbon. Our understanding of the mechanisms governing the dynamics of the carbon cycle has been hampered by a limited ability to locate and quantify these exchanges at sufficiently fine temporal and spatial resolution (Bousquet et al., 2000; Gurney et al., 2002; Ciais et al., 2005; Baker et al., 2006; Peters et al., 2007). Accurate and precise quantification of sources and sinks at regional and continental scales is likely to be increasingly important for evaluation and monitoring of carbon management policies.
Global atmospheric inversions have been used to infer sources and sinks of carbon (both natural and anthropogenic) at continental and ocean basin scale from atmospheric measurements of carbon dioxide using tracer transport models. Model intercomparison projects, including the TransCom (Atmospheric Tracer Transport Model Intercomparison Project) series (Gurney et al., 2002; Gurney et al., 2004; Baker et al. 2006), have been designed to attribute the uncertainties in the continental and ocean basin fluxes estimated by this method. These studies show that transport model differences and the uneven and sparse global distribution of atmospheric carbon dioxide measurements contribute to the uncertainty of the inverse flux estimates. While transport models are improving and the global measurement network for carbon dioxide is expanding, there are still fundamental representation and aggregation errors (Kaminski et al., 2001; Engelen et al., 2002) inherent in the global atmospheric inversion method. There is a mismatch in space and time resolution between the transport models (grid boxes and minutes), the observations (points in space and time) and the inversion solution (continents or subcontinents and months or weeks). Observations are subject to local atmospheric variations. These subgrid scale mesoscale variations are not explicitly accounted for in the transport models; however, we use these local observations to constrain continental and ocean basin results. In addition, global inversions typically require strong and uncertain assumptions about the correlation of fluxes and observations in space and time. If the inversion solution is constructed at the continental scale, for example, it is not possible to evaluate changes in fluxes from subregions within the continent. These strong assumptions about correlations of fluxes in space and time are a weakness; the strengths of such a continental scale global inversion method are the minimum number of unknowns and the global coverage. Some assumptions about coherence in space and time are essential; atmospheric observations will always be uneven and sparse at some level of resolution.
A compelling approach is to invert on the grid and time scale of the transport model (Kaminski et al., 2001; Engelen et al., 2002). Global atmospheric inversions at the spatial resolution of the transport model (e.g. Rödenbeck et al., 2003a; Gourdji et al., 2008; Mueller et al., 2008), aim for the finest resolution possible to minimize representation assumptions, at the expense of larger posterior covariances. Subsequent aggregation into coarser regional and temporal resolution is then used to lessen the posterior error. Regional atmospheric inversions target a geographically limited domain with finer spatial and temporal resolution (Gerbig et al., 2003; Peylin et al., 2005; Lauvaux et al., 2008; Schuh et al., 2009). Both of these approaches involve many more unknowns, which cannot be resolved independently given the current observation density. The underlying assumptions may be minimized, but at the expense of building prior covariance matrices and the increased computational costs required by the finer resolution.
In this experiment we take a pragmatic, middle-ground approach to the continental-scale global inversion by choosing a number of regions roughly matched to the observation density currently available. If we have chosen observation sites that are representative of the regions and sensitive to the surface exchanges in these regions, then we expect that posterior uncertainties and spatial correlations will be reduced and that the problem will be computationally tractable using simple inversion methods. Inversion results can be aggregated to the larger TransCom continental regions for comparison with published results. We can also test the ability of the expanded network to constrain the smaller regions with this method.
Typically global atmospheric measurement network sites have been chosen to facilitate sampling background concentrations of trace gases including carbon dioxide. These background measurement networks have yielded important understanding of interhemispheric gradients in carbon dioxide mixing ratios (Tans et al., 1990; Denning et al., 1995; Keeling et al., 1996) and of the mean annual cycles of carbon emissions and uptake (e.g. Keeling et al., 1995). These data, however, provide limited understanding of the continental carbon cycle. We cannot diagnose continental or regional scale fluxes and determine the factors influencing terrestrial fluxes without observing sites over the continents. Continental carbon dioxide measurements are characterized by strong diurnal and seasonal cycles that reflect a combination of biological fluxes and atmospheric boundary layer dynamics (Bakwin et al., 1998; Yi et al., 2001; Davis et al., 2003). Continental data also contain strong gradients driven by weather (e.g. Hurwitz et al., 2004; Wang et al., 2007; Parazoo et al., 2008). These strong, rapidly varying gradients in the observations may be difficult to simulate in the transport, but the continental data contain information needed to resolve regional sources and sinks of carbon with increasing spatial and temporal resolution.
In this paper, we use the Bayesian synthesis inversion method to demonstrate the impact of including more continental measurement sites in the global measurement network. The added sites are long-running eddy covariance flux towers with high precision carbon dioxide measurements calibrated to global standards. Carbon dioxide measurements at flux towers do not need to be calibrated to global standards for the calculation of net ecosystem exchange of carbon dioxide using the eddy covariance method. The sites used in this study, however, are part of a growing network where the calibration is done with the intent of providing data suitable for application to atmospheric inversion studies. The five towers used in this study have data available during the 2000–2004 time period. Increasing numbers of flux towers are incorporating the calibration processes into their routine processing; this offers opportunities for extending this research in the future.
We focus here on the effect on the North American carbon balance, recognizing the danger that, in an ill-conditioned problem such as this, increasing the density of observations in North America may introduce new challenges. For example, global inversions typically exhibit dipole behaviour or ‘pair-sum’ relationships (Rödenbeck et al., 2003a) where the flux as a sum for two regions can be constrained, while the individual regions cannot. Within North America we may discover dipoles between the subregions that were not apparent in the continental-scale inversion. With the exception of Boreal Asia, the Northern Hemisphere is well represented in the networks tested (Fig. 2). Concentrating observation sites in North America may highlight dipole relationships between Boreal Asia and other regions of the Northern Hemisphere. Some recent inversions, for example, find a larger terrestrial carbon sink in Europe (Baker et al., 2006; Mueller et al., 2008) and others in Asia (Rödenbeck et al., 2003a; Peters et al., 2007). Published inversion results also frequently disagree with estimates of carbon fluxes from biogeochemical models (Janssens et al., 2003; Peters et al., 2007; Potter et al., 2007). We will examine our results in this light, but concentrate on the uncertainty improvement of the added measurement sites in this paper.
Figure 2. Solid lines define the spatial scale of the inversion solution (36 land regions and 11 ocean regions). Symbols mark locations of the observation sites in the three networks tested: Base network (blue); Enhanced network (blue and cyan); Continental Extension network (blue, cyan, red). Symbol shapes indicate the type of observation: quasi-continuous (circle) and discrete (triangle).
Download figure to PowerPoint
In Section 2, we describe the estimation method. In Section 3, we present global and North American results for two typical global measurement networks and a third network including five additional continental sites. The results are discussed in Section 4. We conclude in Section 5 with recommendations for applicability of the method to future experiments.