Inversion of Atmospheric CO 2 Concentrations
299
TCCON The continuous data experiment, looking at the role of high-frequency (typically hourly) data, as opposed to the monthly mean data used in most
inversions.
TransCom exists primarily to study the impact of model error in CO 2 inversions
(although inversion needs to be regarded more broadly than the Bayesian synthesis
that is the “standard” TransCom case). However, one of the consequences of working on a template negotiated well in advance of the calculations is that as inversions,
TransCom calculations will not be state-of-the-art, but rather closer to a lowestcommon-denominator of what can be achieved by a number of groups.
Some of the features where recent TransCom presentations lag current-bestpractice inversions are
Omission of carbon fl uxes into the atmosphere in forms other than CO
•
2 (see
Section 11.4.2).
Omission of any uncertainty in the fossil component.
•
The calculations have low spatial resolution without any specifi cation of the
•
truncation error discussed above.
Indeed, the fi rst two omissions are specifi c simplifi cations for TransCom since a
fossil fuel uncertainty and a CO fl ux were included in the fi rst published Bayesian
inversions of CO 2 (Enting et al., 1993, 1995), and increasing resolution in inversion
calculations is reducing the discretization problems.
11.3.4 EXPERIMENTAL DESIGN
Inversion studies can also be used for purposes of experimental design. The idea is
that an inversion technique that includes a systematic assessment of uncertainty can
be applied to assess the utility (as measured by reduction in uncertainty) of putative
new data.
This is particularly simple in the case of a linear model with multivariate normal error because in this case the posterior uncertainty does not depend on the
values of the putative data, but only on its uncertainty (as indicated by Equation
11.7b). Examples using Bayesian synthesis inversion of CO 2 were given in the initial paper by Enting et al. (1995), considering aspects such as the utility of 13 C data
and improved precision in CO 2 data. My book described joint reductions in uncertainties of the land-ocean partitioning from various forms of 13 C data (Enting,
2002, Fig. 13.1).
A more extensive application of “experimental design” was the evaluation of
the CO 2 sampling network, initially as optimal location of additional stations and
also as reconfi guration of the entire network (Rayner et al., 1996). The latter case
involved a complex optimization with multiple local minima, and a simulated
annealing technique was used, drawing on applications in seismology (Hardt and
Scherbaum, 1994). A similar approach to network design, using simulated annealing as the optimization technique, has also been applied in oceanography. Further
network design studies were reported by Patra and Maksyutov (2002) and Patra
© 2010 by Taylor and Francis Group, LLC
299
TCCON The continuous data experiment, looking at the role of high-frequency (typically hourly) data, as opposed to the monthly mean data used in most
inversions.
TransCom exists primarily to study the impact of model error in CO 2 inversions
(although inversion needs to be regarded more broadly than the Bayesian synthesis
that is the “standard” TransCom case). However, one of the consequences of working on a template negotiated well in advance of the calculations is that as inversions,
TransCom calculations will not be state-of-the-art, but rather closer to a lowestcommon-denominator of what can be achieved by a number of groups.
Some of the features where recent TransCom presentations lag current-bestpractice inversions are
Omission of carbon fl uxes into the atmosphere in forms other than CO
•
2 (see
Section 11.4.2).
Omission of any uncertainty in the fossil component.
•
The calculations have low spatial resolution without any specifi cation of the
•
truncation error discussed above.
Indeed, the fi rst two omissions are specifi c simplifi cations for TransCom since a
fossil fuel uncertainty and a CO fl ux were included in the fi rst published Bayesian
inversions of CO 2 (Enting et al., 1993, 1995), and increasing resolution in inversion
calculations is reducing the discretization problems.
11.3.4 EXPERIMENTAL DESIGN
Inversion studies can also be used for purposes of experimental design. The idea is
that an inversion technique that includes a systematic assessment of uncertainty can
be applied to assess the utility (as measured by reduction in uncertainty) of putative
new data.
This is particularly simple in the case of a linear model with multivariate normal error because in this case the posterior uncertainty does not depend on the
values of the putative data, but only on its uncertainty (as indicated by Equation
11.7b). Examples using Bayesian synthesis inversion of CO 2 were given in the initial paper by Enting et al. (1995), considering aspects such as the utility of 13 C data
and improved precision in CO 2 data. My book described joint reductions in uncertainties of the land-ocean partitioning from various forms of 13 C data (Enting,
2002, Fig. 13.1).
A more extensive application of “experimental design” was the evaluation of
the CO 2 sampling network, initially as optimal location of additional stations and
also as reconfi guration of the entire network (Rayner et al., 1996). The latter case
involved a complex optimization with multiple local minima, and a simulated
annealing technique was used, drawing on applications in seismology (Hardt and
Scherbaum, 1994). A similar approach to network design, using simulated annealing as the optimization technique, has also been applied in oceanography. Further
network design studies were reported by Patra and Maksyutov (2002) and Patra
© 2010 by Taylor and Francis Group, LLC