TransCom 3 CO2 inversion intercomparison: 1. Annual mean ... · CO2 INVERSION INTERCOMPARISON 557...
Transcript of TransCom 3 CO2 inversion intercomparison: 1. Annual mean ... · CO2 INVERSION INTERCOMPARISON 557...
Tellus (2003), 55B, 555–579 Copyright C© Blackwell Munksgaard, 2003Printed in UK. All rights reserved TELLUS
ISSN 0280–6509
TransCom 3 CO2 inversion intercomparison: 1. Annualmean control results and sensitivity to transport
and prior flux information
By KEVIN ROBERT GURNEY1∗, RACHEL M. LAW2, A. SCOTT DENNING1, PETER J. RAYNER2,DAVID BAKER3, PHILIPPE BOUSQUET4, LORI BRUHWILER5, YU-HAN CHEN6, PHILIPPE CIAIS4,SONGMIAO FAN7, INEZ Y. FUNG8, MANUEL GLOOR9, MARTIN HEIMANN9, KAZ HIGUCHI10,JASMIN JOHN8, EVA KOWALCZYK2, TAKASHI MAKI11, SHAMIL MAKSYUTOV12, PHILIPPEPEYLIN4, MICHAEL PRATHER13, BERNARD C. PAK13, JORGE SARMIENTO7, SHOICHI TAGUCHI14,TARO TAKAHASHI15 and CHIU-WAI YUEN10, 1Department of Atmospheric Science, Colorado State Uni-versity, Fort Collins, CO 80523, USA; 2CSIRO Atmospheric Research, PMB 1, Aspendale, Victoria 3195,Australia; 3National Center for Atmospheric Research (NCAR), Boulder, CO 80303, USA; 4Laboratoire desSciences du Climat et de l’Environment (LSCE), F-91198 Gif-sur-Yvette Cedex, France; 5National Oceanicand Atmospheric Administration (NOAA), Climate Monitoring and Diagnostics Laboratory, 326 BroadwayR/CG1, Boulder, CO 80303, USA; 6Department of Earth, Atmospheric, and Planetary Science, MassachusettsInstitute of Technology (MIT), Cambridge, MA 0214, USA; 7AOS Program, Princeton University, Sayre Hall,Forrestal Campus PO Box CN710 Princeton, NJ 08544-0710, USA; 8Center for Atmospheric Sciences, Mc-Cone Hall, University of California, Berkeley, CA 94720-4767, USA; 9Max-Planck Institute fur Biogeochemie,D-07701 Jena, Germany; 10Meteorological Service of Canada, Environment Canada, Toronto, Ontario M3H5T4, Canada; 11Quality Assurance Section, Atmospheric Environment Division, Observations Department,Japan Meteorological Agency, 1-3-4 Otemachi, Chiyoda-ku, Tokyo 100-8122 Japan; 12Institute for GlobalChange Research, Frontier Research System for Global Change, Yokohama, 236-0001 Japan; 13Earth SystemScience, University of California, Irvine, CA 92697-3100, USA; 14National Institute of Advanced Industrial Sci-ence and Technology, 16-1 Onogawa Tsukuba, Ibaraki 305-8569, Japan; 15Lamont-Doherty Earth Observatory
of Columbia University, Palisades, NY, USA
(Manuscript received 22 May 2002; in final form 27 November 2002)
ABSTRACT
Spatial and temporal variations of atmospheric CO2 concentrations contain information about surfacesources and sinks, which can be quantitatively interpreted through tracer transport inversion. PreviousCO2 inversion calculations obtained differing results due to different data, methods and transportmodels used. To isolate the sources of uncertainty, we have conducted a set of annual mean inversionexperiments in which 17 different transport models or model variants were used to calculate regionalcarbon sources and sinks from the same data with a standardized method. Simulated transport isa significant source of uncertainty in these calculations, particularly in the response to prescribed“background” fluxes due to fossil fuel combustion, a balanced terrestrial biosphere, and air–sea gasexchange. Individual model-estimated fluxes are often a direct reflection of their response to thesebackground fluxes. Models that generate strong surface maxima near background exchange locationstend to require larger uptake near those locations. Models with weak surface maxima tend to have lessuptake in those same regions but may infer small sources downwind. In some cases, individual modelflux estimates cannot be analyzed through simple relationships to background flux responses but are
∗Corresponding author.e-mail: [email protected]
Tellus 55B (2003), 2
556 K. R. GURNEY ET AL.
likely due to local transport differences or particular responses at individual CO2 observing locations.The response to the background biosphere exchange generates the greatest variation in the estimatedfluxes, particularly over land in the Northern Hemisphere. More observational data in the tropicalregions may help in both lowering the uncertain tropical land flux uncertainties and constraining thenorthern land estimates because of compensation between these two broad regions in the inversion.More optimistically, examination of the model-mean retrieved fluxes indicates a general insensitivityto the prior fluxes and the prior flux uncertainties. Less uptake in the Southern Ocean than implied byoceanographic observations, and an evenly distributed northern land sink, remain in spite of changesin this aspect of the inversion setup.
1. Introduction
A quantitative understanding of the sources andsinks of atmospheric CO2 is of vital importance for re-liably predicting future CO2 levels. The estimation ofCO2 sources and sinks can be approached in a varietyof ways (Schimel et al., 2001). One approach is to useatmospheric CO2 measurements to infer CO2 fluxesthrough tracer transport inversion. This approach ad-justs a series of regionally explicit “trial” fluxes tobest match observed CO2 concentrations to those sim-ulated by an atmospheric transport model. Early inver-sions used two-dimensional transport models to cal-culate the latitudinal distribution of fluxes (e.g. Tanset al., 1990; Enting and Mansbridge, 1989; Ciais et al.,1995). Recent inversions have used three-dimensionalmodels to estimate the longitudinal distribution offluxes (e.g. Enting et al., 1995; Fan et al., 1998;Bousquet et al., 1999a;b; Kaminski et al., 1999). In-terannual variations in fluxes have also been estimated(e.g. Rayner et al., 1999; Law, 1999; Bousquet et al.,2000; Baker, 2001).
A general concern with inversions is the sensitiv-ity of the results to the atmospheric transport modelused. This sensitivity is difficult to assess because theresults to date have employed only one or two trans-port models in a given study. A goal of the TransComseries of experiments has been to assess the influenceof different transport algorithms on the CO2 inversionproblem. The initial phases of TransCom conductedforward simulations of fossil and biospheric emissionsof CO2 (TransCom 1) (Law et al., 1996) and of SF6
(TransCom 2) (Denning et al., 1999) to characterizemodel behaviour. The fossil CO2 experiment indicateddifferences in simulated surface interhemispheric con-centration gradients among the models of up to a factorof two. However, with no ‘fossil-CO2’ observationsto compare against, it was not possible to determinewhich simulated transport, if any, was correct. The SF6
experiment, with relatively well known sources andobservations, was intended to evaluate atmospheric
transport and to identify which processes led to differ-ent model behaviour. Most of the models were reason-ably successful in reproducing the marine boundary-layer observations of SF6, but tended to overestimatethe SF6 at continental locations near sources. Thosemodels that underestimated marine boundary-layervalues due to excessive vertical convective transporttended to give better continental concentrations thanthe models with less convective transport.
Model differences were also large for the annualmean response to seasonal biospheric CO2 exchange.Almost all models simulated elevated CO2 concentra-tions at the surface in the northern middle latitudes dueto the covariance between seasonal exchange and sea-sonal transport, the so-called ‘seasonal rectifier effect’(e.g., Keeling et al., 1989; Denning et al., 1995), butzonal mean surface concentrations varied over 3 ppmamong the models.
While these experiments provided useful insightsinto model behaviour, they did not directly address thesensitivity of inversion results to transport. This is therole of the current TransCom experiment (TransCom3); the aim is to assess the contribution of uncertain-ties in transport to the uncertainties in flux estimatesfor annual mean, seasonal cycle and interannual inver-sions. The experiment can also contribute to our un-derstanding of other sensitivities in the inversion pro-cess (e.g. inversion set-up, observational data choices),since more reliable results are expected by examiningthe sensitivity with a range of transport models thanwith just one or two models.
The first results from TransCom 3 were presentedby Gurney et al. (2002). They showed mean inversionresults for a ‘control’ inversion in which annual meanfluxes were estimated using 1992–1996 data. Here wepresent results from that same control inversion butfor individual models. We also present results from anumber of sensitivity tests related to the specificationof prior flux information. A companion paper (Lawet al., 2003) presents sensitivity tests related to CO2
data issues including network choice, time period, data
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 557
selection and data uncertainty. Seasonal and interan-nual inversions will be presented elsewhere.
2. Method
2.1. Inversion formalism
The control inversion uses a Bayesian synthesismethod (Enting, 2002). The first step is to choose a setof flux patterns, or basis functions ( �Vi ), from whichsolutions will be constructed: �V = ∑
i si �Vi . The in-version procedure reduces to solving for the scalingfactor, si. With a set of concentration observations �Dthe linearity of transport can be used to yield
�D =∑
i
si T( �Vi ) (1)
where T is the atmospheric transport. If we have pre-defined times and places for our concentration obser-vations (obvious in a diagnostic or inverse study) thenwe can repeatedly run the transport model on eachbasis function and sample the output at the observa-tion locations. The concentration field generated fromeach basis function is referred to as its correspondingresponse function. If we sample the response functionat our chosen observation locations, and write the re-sult as a column vector, we produce a matrix M forwhich we can easily show
�D = M�S (2)
The regional fluxes, �S, can now be solved for us-ing conventional least-squares techniques. These tech-niques rely on minimizing the mismatch between mod-elled concentrations, M�S, and observed concentra-tions, �D. Adding prior information about the fluxesacknowledges the under-determinacy of the systemdue to the limitation of few observations to constrainfluxes from all regions. It also allows those aspects ofthe flux distribution that are reasonably well known tobe specified.
With the addition of prior information, the solutionfor �S now involves minimizing the difference betweenmodelled and observed concentrations and betweenpredicted fluxes and their prior estimates. A cost func-tion J can be defined as
J = 1
2
[(M�S − �D)T C( �D)−1(M�S − �D)
+ (�S − �S0)T C(�S0)−1(�S − �S0)]
(3)
where C( ) embodies confidence in the form of anerror or uncertainty covariance. Since C appears in-verted, quantities with large uncertainty confer lessinfluence. Deviations from this prior estimate willpenalize the cost function severely, and hence berestricted. Alternatively, fluxes that are poorly un-derstood can be assigned large uncertainties and,hence, allowed to deviate significantly from the priorestimates.
The measurements are similarly weighted inverselyby the degree to which the predicted concentrationsare required by the inverse process to match the obser-vations. This weighting term, C( �D), has traditionallybeen called the “data uncertainty”. The name is mis-leading, since the term must also account for the in-ability of models with imperfect transport and coarsespatial and temporal resolution to match point obser-vations. Furthermore, the chosen flux shape or “foot-print” assumed for the basis functions may not rep-resent the distribution of the true source field, andhence may add to the inability of the predicted con-centrations to match observations (Kaminski et al.,2001; Engelen et al., 2002). It would be inappropri-ate in the inversion to attempt to fit the data betterthan the sum of all of these errors, yet objectivelyquantifying this overall model–data mismatch remainsdifficult.
Following Tarantola (1987) we minimize eq. (3) toyield
�S = �S0 + M(−1)( �D − M�S0)(4)
where M(−1), the generalized inverse of M, is given by
M(−1) =[MT C( �D)−1M + C(�S0)−1
]−1MT C( �D)−1.
(5)
An estimate of the uncertainty in the predicted fluxescan also be produced. This depends on the uncertaintyin the prior fluxes, the data uncertainty and the re-sponse functions embodied in M. The relation is:
C(�S) =[C(�S0)−1 + MT C( �D)−1M
]−1. (6)
This always represents a reduction in uncertaintyon the prior estimate, with the amount of reductiondepending on the data uncertainty and how well the
Tellus 55B (2003), 2
558 K. R. GURNEY ET AL.
available observing locations sample the chosen fluxpatterns.
The mean of the individual model flux uncertaintiescan be computed as:
C(�S) =√√√√Nmodels∑
n=1[C(�S)]2
/Nmodels
(7)
where the individual model posterior uncertainty esti-mates are taken from the diagonal of the posterior co-variance matrix. We designate this mean uncertaintythe ‘within-model’ uncertainty. The spread of flux es-timates across models is represented by the standarddeviation,
σ (�S) =√√√√Nmodels∑
n=1[C(�S) − C(�S)]2
/Nmodels
(8)
and designated the ‘between-model’ uncertainty.The minimum value of the cost function, Jmin, de-
scribes the degree to which the inversion calculationmatches the data and prior source estimates simulta-neously. This is usually expressed using the reducedχ 2,
χ 2 = Jmin
Nobs=
Nobs∑n=1
(M�S− �D)2
C( �D)2 +Rregions∑
r=1
(�S−�S0)2
C( �S0)2
Nobs.
(9)
Statistical consistency requires χ2 not to be signifi-cantly greater than unity, otherwise the posterior un-certainty is inconsistent with the quality of the fit to thedata. This inconsistency suggests that too much con-fidence has been placed in the ability of the inversionto match the data, so C( �D) is increased accordingly.This will consequently increase C(�S).
2.2. Experimental design
2.2.1. Forward simulations. Forward simulationswere performed with 16 transport models (or modelvariants) (Gurney et al., 2000), the results of whichwere used to perform annual mean inversions forsources and sinks (Gurney et al., 2002). A 17th model(CSIRO) recently submitted simulations. We havenot included results from this model in the calcu-lation of mean results, in order to maintain consis-tency with Gurney et al. (2002), but have included itin presentations of individual model results. Detailsof each transport model are given in Table 1. The
models vary in resolution, advection scheme, drivingwinds and subgrid-scale parameterizations. They in-clude every model used in recently published CO2
inversions. TM2 is the only model that has beenused in precisely the same configuration in all theTransCom experiments, while seven of the currentmodels performed the TransCom 2 SF6 experiment,though some of these have been modified since thatexperiment.
For the annual mean experiment presented here,four-year simulations were performed by each modelfor each of 26 required basis functions. The 26 basisfunctions comprised four “background” global fluxesand 22 regional fluxes. The model output was submit-ted as global monthly mean distributions and as four-hourly time series at 228 specified locations. Samplingof model output was intended to mimic the samplingprotocol for observations: some stations were movedoffshore away from coastlines in the direction of the“clean air sector” identified for flask sampling. Fulldetails of the experimental protocol are presented inGurney et al. (2000). The inversions were performedon the output from each model.
The four background fluxes consisted of two fos-sil fuel emission fields (1990 and 1995 distributions),an annually balanced, seasonal terrestrial biosphereexchange and air–sea gas exchange. These fluxes areincluded in the inversion with a very small prior un-certainty so that their magnitude is effectively fixed,and the 22 regional fluxes estimated by the inversionare deviations from these global fluxes. The 1990 an-nual mean fossil source (1◦ × 1◦) is from Andres et al.(1996) and totals 5.812 Gt C yr−1. The 1995 annualmean fossil source (1◦ × 1◦) is derived from the dataprepared by Brenkert (1998) and totals 6.173 Gt Cyr−1. The seasonal biospheric exchange (1◦ × 1◦) wasderived from the Carnegie Ames Stanford Approach(CASA) model (Randerson et al., 1997), and has anannual total flux of zero at every grid cell. The oceanicexchange (4◦ × 5◦) was produced by Taro Takahashiand colleagues and represents monthly global oceanicexchange derived from measurements of �p(CO2)(Takahashi et al., 1999). The annually integrated flux is−2.19 Gt C yr−1 (into the ocean). The background fos-sil fuel emission fluxes were prescribed without sea-sonality. The annually balanced terrestrial fluxes werepurely seasonal, and the background ocean fluxes wereprescribed with both seasonal variations and annualmean uptake.
The boundaries of the 22 regional basis functionsare shown in Fig. 1. In the forward simulations, each
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 559
Tabl
e1.
Part
icip
atin
gm
odel
desc
ript
ion
and
refe
renc
es
Mod
elM
odel
erW
inds
H.R
es.
#L
evel
sA
dvec
tion
Subg
rid
Ref
eren
ces
CSU
aG
urne
y&
Onl
ine
72lo
n.×
17si
gma
2nd
orde
rC
um.m
ass
flux
(Ara
kaw
a&
Schu
bert
,197
4;D
enni
nget
al.(
1996
)D
enni
ng44
lat.
Ran
dall
&Pa
n,19
93;D
ing
etal
.,19
98),
var.
prog
.PB
L(S
uare
zet
al.,
1983
;R
anda
llet
al.,
1992
)U
CB
fFu
ng&
John
1h
GIS
S72
lon.
×9
sigm
aSl
opes
(Rus
sell
&Pe
netr
ativ
em
ass
flux
Han
sen
etal
.(19
97)
GC
MII
′46
lat.
Ler
ner,
1981
)U
CI
Prat
her,
Pak,
GIS
S72
lon.
×9
sigm
a&
2nd
orde
rPe
netr
ativ
em
ass
flux
w/u
pdra
fts,
Prat
her
etal
.(19
87);
(s,b
)bL
ee,&
GC
MII
′46
lat.
23si
gma
mom
ents
dow
ndra
fts
and
entr
ainm
ent;
Han
sen
etal
.(19
97);
Han
nega
n(P
rath
er,1
986)
diur
nally
vary
ing
PBL
Koc
han
dR
ind
(199
8)JM
AM
aki
6h
JMA
144
lon.
×32
hybr
idH
or:s
emi-
lag.
,V
ertd
iff.
(Mel
lor–
Yam
ada,
leve
l2.0
),Ta
guch
i(19
96)
(199
7)73
lat.
Ver
t:bo
xsc
hem
eC
umul
us(s
impl
edi
ff.)
,PB
L(fi
xed
laye
r)M
AT
CH
:B
ruhw
iler
NC
AR
128
lon.
×28
sigm
aSp
itfire
Cor
rect
ive
tran
spor
t,su
bgri
ddi
ff.,
PBL
Ras
chet
al.(
1997
)C
CM
3C
CM
364
lat.
w/s
tabl
em
ixin
g,cl
oud
pred
ictio
nM
AT
CH
:C
hen
6h
NC
EP
128
lon.
×28
sigm
aSp
itfire
Ver
t.di
ff.w
/dia
gnos
edPB
L;p
enet
rativ
eR
asch
etal
.(19
97)
NC
EP
(199
0)64
lat.
conv
.+lo
calc
onv.
corr
ectio
n(H
ack,
1994
)M
AT
CH
:L
aw6
h64
lon.
×24
hybr
idSe
mi-
lagr
angi
anH
ack
conv
ectio
n(H
ack,
1994
);ve
rt.d
iff.
Ras
chet
al.(
1997
);L
awM
AC
CM
2M
AC
CM
264
lat.
with
PBL
(Hol
tsla
g&
Bov
ille,
1993
)&
Ray
ner
(199
9)N
IES
Mak
syut
ov12
hE
CM
WF
144
lon.
×15
sigm
aSe
mi-
lagr
angi
anV
ert.
diff
.in
trop
osph
ere
(Hac
k,19
93);
Mak
syut
ov&
Inou
e(1
997)
72la
t.pe
netr
ativ
em
ass
flux
(Tie
dke,
1989
);(2
000)
vert
.dif
f.in
PBL
(Sch
uber
teta
l.,19
93)
Tellus 55B (2003), 2
560 K. R. GURNEY ET AL.
Tabl
e1.
(con
t’d.
)
Mod
elM
odel
erW
inds
H.R
es.
#L
evel
sA
dvec
tion
Subg
rid
Ref
eren
ces
NIR
ETa
guch
i6
hE
CM
WF
144
lon.
×15
hybr
idSe
mi-
lagr
angi
anV
ert.
diff
.in
PBL
,red
uced
Tagu
chi(
1996
)C
TM
-96c
(199
5)73
lat.
diff
usio
nat
trop
opau
seR
PN-S
EF
Yue
n&
Onl
ine
128
lon.
×27
sigm
aSe
mi-
lag.
(Ric
hie
&V
ertic
aldi
ffus
ion;
PBL
base
don
turb
ulen
tD
’And
rea
etal
.(19
98)
Hig
uchi
64la
t.B
eaud
oin,
1994
)K
E(B
enoi
teta
l.,19
97)
SKY
HId,
fFa
nO
nlin
e10
0lo
n.×
40hy
brid
2nd
orde
rD
ry/m
oist
conv
ectiv
ead
just
men
tfor
T,q
;M
ahlm
anet
al.(
1994
);60
lat.
non-
loca
lpar
amet
eriz
atio
nfo
rve
rt.
Stra
han
&M
ahlm
an(1
994)
;m
ixin
gin
PBL
Ham
ilton
etal
.(19
95)
TM
2eB
ousq
uet&
12h
EC
MW
F48
lon.
×9
sigm
aSl
opes
(Rus
sell
&M
ass
flux
(Tie
dke,
1989
);st
abili
tyde
pend
ent
Hei
man
n(1
995)
;Pe
ylin
(199
0)25
lat.
Ler
ner,
1981
)ve
rtic
aldi
ffus
ion
(Lou
is,1
979)
Bou
sque
teta
l.(1
999a
)T
M3f
Hei
man
n6
hE
CM
WF
72lo
n.×
19hy
brid
Slop
es(R
usse
ll&
Mas
sflu
x(T
iedk
e,19
89);
stab
ility
depe
nden
tH
eim
ann
(199
5)(1
990)
45la
t.L
erne
r,19
81)
vert
.dif
fusi
on(L
ouis
,197
9,up
date
d)G
CT
Me
Bak
er6
hZ
OD
IAC
256
km2
11si
gma
2nd/
4th
orde
rV
ert.
diff
usio
n,R
inum
ber-
base
dve
rt.m
ixin
gM
ahlm
an&
Mox
im(1
978)
;G
CM
inlo
wes
t3la
yers
tore
pres
entP
BL
.L
evy
etal
.(19
82)
CSI
RO
-K
owal
czyk
&O
nlin
e20
8km
218
sigm
aH
or:s
emi-
lag
Stab
ility
dep.
Ver
t.di
ff.
McG
rego
r&
Dix
(200
1)C
Cg
McG
rego
r(M
cGre
gor,
1996
);(L
ouis
,197
9)+
non-
loca
lsch
eme,
Ver
t:to
talv
arm
ass
flux
conv
.w/d
ownd
raft
s.di
min
ishi
ng
a Part
icip
ated
inT
rans
Com
1an
dT
rans
Com
2bu
tmod
eldi
ffer
ed.
bU
CI
prim
ary
mod
elus
es3-
hour
lyG
ISS
II’
9-la
yer
GC
Mw
inds
and
ispr
imar
ilytr
opos
pher
ic;U
CIb
isa
boun
dary
-lay
erva
rian
twith
PBL
heig
htdi
agno
sed
from
Rin
umbe
r,su
per-
reso
lved
trac
ergr
adie
nts
usin
gse
cond
-ord
erm
omen
ts,a
ndno
n-lo
calc
losu
reon
the
subg
rid
(Har
tke
and
Rin
d,19
97);
UC
Isus
es6-
hour
lyG
ISS
II’
mid
dle
atm
osph
ere
GC
Mw
inds
with
aw
ellr
esol
ved
stra
tosp
here
(Koc
han
dR
ind,
1998
).c E
arlie
rm
odel
vers
ion
part
icip
ated
inT
rans
Com
1,cu
rren
tmod
elve
rsio
nin
Tra
nsC
om2.
dA
non-
loca
lpar
amet
eriz
atio
nfo
rve
rtic
alm
ixin
gin
the
PBL
was
impl
emen
ted
afte
rT
rans
Com
2an
dw
asus
edin
Tra
nsC
om3.
e Part
icip
ated
inT
rans
Com
1an
dT
rans
Com
2w
ithcu
rren
tmod
el.
f Part
icip
ated
inT
rans
Com
2.gM
odel
outp
utfie
lds
wer
ein
terp
olat
edfr
omm
odel
’sco
nfor
mal
-cub
icgr
idto
regu
lar
(2◦
latit
ude,
long
itude
)gr
idbe
fore
subm
issi
on.
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 561
Fig. 1. Basis function regions and the locations of the 76 CO2 observational records used in the inversion. Multiple recordsexist at some locations and are denoted by open circles. The prior flux and prior flux uncertainties are shown for each basisfunction region (Gt C yr−1). The prior constraint on the atmospheric growth rate is 3.274 Gt C yr−1 with a prior uncertaintyof 0.074 Gt C yr−1. The prior global offset concentration is 355 ppm with a prior uncertainty of 100 ppm.
region emits a 1 Gt C yr−1 flux that is scaled inde-pendently during the inversion calculation. The re-gional terrestrial basis fluxes include spatial struc-ture based on the distribution of annual mean net pri-mary productivity (NPP), as given by a CASA modelsteady-state run (Randerson et al., 1997), reflectingan assumption that terrestrial exchanges are locatedwithin regions of higher biological activity. The ter-restrial basis function boundaries were constructedto enclose vegetation of similar seasonal structureand carbon exchange, with some additional bound-ary smoothing. The starting point for the boundaryconstruction was the 1◦ × 1◦ land cover type classifi-cation used by the original Simple Biosphere Model(SiB) (De Fries and Townshend, 1994; Sellers et al.,1996). A uniform spatial distribution was used for theregional oceanic basis functions, with the exceptionof seasonal ice cover in the northern and southern-most regions. The sea-ice cover data were producedby the World Climate Research Programme, Work-ing Group on Numerical Experimentation, Atmo-spheric Model Intercomparison Project (Taylor et al.,1997).
2.2.2. Inversion set-up and observational data.Prior estimates of the fluxes in each of the 22 re-gional basis functions were determined from inde-pendent estimates of terrestrial and oceanic exchange
and are presented in Fig. 1. The land region priorflux estimates incorporate results from recent inven-tory studies (Apps and Kurz, 1994; Kurz and Apps,1999; UNFCCC, 2000; Pacala et al., 2001; Houghton,1999; Dixon et al., 1990; Houghton and Hackler, 1999;Kauppi et al., 1992). Where more than one estimatefor a given region was available, a mid-point of theestimate range was used. The ocean region prior fluxestimates were prescribed as zero.
The prior flux uncertainty is important for keepingthe estimated fluxes within biogeochemically realis-tic bounds. For land regions we chose growing seasonnet flux (GSNF, the sum of monthly mean exchangefor months exhibiting net uptake) as provided by theCASA model of net ecosystem production (Randersonet al., 1997). Since it is unlikely that an annual meanresidual land flux would exceed the GSNF, this pro-vides a reasonable, ecologically relevant upper bound.The prior ocean uncertainties were guided by the ag-gregate estimates given in Takahashi et al. (1999),which suggest a total global oceanic uncertainty ap-proximately 70% of the total oceanic exchange. Be-cause this level of uncertainty tightly constrains someoceanic regions to the prior flux estimate, the aggre-gate uncertainty was increased to 140% of the totalnet oceanic exchange. The uncertainty in individualoceanic basis function regions was proportional to the
Tellus 55B (2003), 2
562 K. R. GURNEY ET AL.
area of the region and the proportion of sampled gridpoints in the region.
We invert 5-yr mean measurements for 1992–1996 at 76 sites taken from the GLOBALVIEW-2000 dataset (GLOBALVIEW-CO2, 2000). GLOB-ALVIEW is a data product that interpolates CO2 mea-surements to a common time interval. Gaps in the dataare filled by extrapolation from marine boundary-layermeasurements. Sites were chosen where the extrapo-lated data accounts for less than 30% of total data forthe 1992–1996 period. Law et al. (2003) varied thisextrapolated data limit to test the sensitivity of the in-version to alternative data sets.
The uncertainty attached to each data value, C(D),was derived from the residual standard deviation(RSD) of individual observations around a smoothedtimeseries as given by GLOBALVIEW-CO2 (2000).This choice was based on the assumption that the dis-tribution of RSD (higher RSD values for northern andcontinental sites and lower RSD values for southernhemisphere oceanic sites) reflects the high-frequencyvariations in transport and regional flux that large-scaletransport models are unable to accurately simulate. Di-rect use of the RSD values for the data uncertaintywithin the inversion results in a reduced χ2 [eq. (9)]that is much smaller than unity. This indicates thatthe predicted concentrations fit the data much betterthan the uncertainty assigned to the data itself, andthat the uncertainty should be reduced. Therefore, theRSD has been scaled as follows: the RSD was dividedby (8P)0.5, where P is the proportion of real data inthe record and the factor 8 accomplishes the χ2 ≈ 1goal. Where this gave values less than 0.25 ppm, theuncertainty was increased to 0.25 ppm. Finally, the un-certainty was increased for those sites that are likelyto occur in the same model grid cell by the square rootof the number of co-located sites. This gave uncer-tainties ranging from 0.25 ppm for remote, “clean air”sites to 2.2 ppm for continental, “noisy” sites and a re-duced χ 2 averaged across the models of 0.97, close tothe desired value of one. Law et al. (2003) demon-strate the sensitivity of the inversion to the choiceof C(D).
Additional prior constraints included in the inversecalculation are the global atmospheric CO2 growth rate(3.274 ± 0.074 Gt C yr−1) and a global CO2 concentra-tion background value (355 ± 100 ppm). The growthrate and uncertainty are the mean and standard de-viation of the 1992–1996 trends at the observationalstations used in this study via a fit that includes a lineartrend and harmonics.
3. Results
3.1. Background simulation results
The forward simulations of the four backgroundfluxes provide a measure of model-to-model transportdifferences, and explain many of the features foundin the inferred regional fluxes to be discussed in Sec-tion 3.2. Figure 2 shows the surface CO2 concentra-tions resulting from the sum of the background fluxesfor each of the models. Maxima associated with thefossil fuel emissions and the annually balanced, sea-sonal biosphere exchange are evident and vary consid-erably from model to model. Maxima associated withthe fossil fuel emissions are centered over the indus-trial source regions of the United States, Europe andthe Asian Pacific. Maxima associated with rectifica-tion of the biosphere exchange are broadly distributedover forested areas in North America and Eurasia.Regional minima in concentration are simulated overthe Southern Ocean and to a lesser extent over theNorth Atlantic Ocean in response to the backgroundocean fluxes. Surface concentration responses rangefrom very weak (TM2 and the UCI variants) to strongbut localized (GCTM) to very strong over large ar-eas (MATCH:NCEP, UCB). As was found previously(Law et al., 1996; Denning et al., 1999), there does notappear to be a systematic difference between mod-els driven by analyzed winds and those using GCMwinds.
Figure 3 shows the zonal mean surface CO2 concen-tration resulting from each of the background fluxesseparately and summed. The maximum fossil fuel con-centrations occur around 50 ◦N and vary by almost 2ppm among the models. Southern Hemisphere concen-trations also vary with CSU, MATCH:MACCM2 andNIRE producing the highest concentrations. The inter-hemispheric difference (IHD) (mean Northern Hemi-sphere surface-layer concentration minus mean South-ern Hemisphere surface-layer concentration) in thesesimulated surface responses provides a convenient in-dex (not always an accurate indicator; see Denninget al., 1999) to summarize both meridional and verti-cal model transport, and is indicated in the figure forthe fossil fuel and total background responses. CSU,TM2 and JMA exhibit the smallest IHD values indi-cating relatively rapid mixing away from source areas(vertical or meridional or both). The ratio of largestfossil fuel IHD to smallest IHD is 1.5 compared to 2.0for the fossil simulation results from TransCom 1, in-dicating a convergence of model simulations for this
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 563
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
CSU Global Mean = 357.0
354.0
354.8
355.6
356.4
357.2
358.0
358.8
359.6
360.4
361.2
362.0
362.8
363.6
364.4
365.2
366.0
366.8
367.6
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
UCB Global Mean = 357.3
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
UCIs Global Mean = 357.2
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
JMA Global Mean = 356.6
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
MATCH:CCM3 Global Mean = 357.1
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
GCTM Global Mean = 357.0
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
UCI Global Mean = 356.9
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
CSIRO Global Mean = 357.0
Fig. 2. Annual mean surface CO2 concentration (ppm) resulting from the combined (relative to a background concentrationof 350 ppm) background fluxes for each of the models. The UCIb model variant is not shown since its surface distribution isvery similar to the UCI standard version. The global mean surface concentration is also computed for each model.
Tellus 55B (2003), 2
564 K. R. GURNEY ET AL.
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
MATCH:MACCM2 Global Mean = 357.3
354.0
354.8
355.6
356.4
357.2
358.0
358.8
359.6
360.4
361.2
362.0
362.8
363.6
364.4
365.2
366.0
366.8
367.6
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
NIES Global Mean = 357.0
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
NIRE Global Mean = 357.5
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
RPN Global Mean = 357.1
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
SKYHI Global Mean = 357.0
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
TM2 Global Mean = 356.7
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
TM3 Global Mean = 357.2
SP
-60
-30
EQ
30
60
NP
180 120 W 60 W 0 60 E 120 E 180
MATCH:NCEP Global Mean = 357.3
Fig. 2. (Contd.)
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 565
CSU.......................
UCB.......................
UCI........................
UCIs.......................
UCIb.......................
JMA........................
MATCH:CCM3.......
MATCH:NCEP.......
MATCH:MACCM2..
NIES.......................
NIRE......................
RPN.......................
SKYHI....................
TM2........................
TM3........................
GCTM....................
3.9
3.6
4.4
5.2
4.5
4.5
4.5
3.6
4.8
4.6
3.4
4.1
4.2
4.7
4.1
4.1
4.2
4.8
4.5
5.5
7.0
5.3
5.5
5.2
4.6
5.9
6.6
5.0
5.9
6.1
5.6
5.8
4.5
5.4
CSIRO....................
CSU.......................
UCB.......................
UCI........................
UCIs.......................
UCIb.......................
JMA........................
MATCH:CCM3.......
MATCH:NCEP.......
MATCH:MACCM2..
NIES.......................
NIRE......................
RPN.......................
SKYHI....................
TM2........................
TM3........................
GCTM....................
CSIRO....................
358
359
360
361
362
363
364
365
-90 -70 -50 -30 -10 10 30 50 70 90
Latitude
CO
2 C
on
cen
trat
ion
(p
pm
)
Background fossil fuel emissions
344
345
346
347
-90 -70 -50 -30 -10 10 30 50 70 90
Latitude
CO
2 C
on
cen
trat
ion
(p
pm
)
Background ocean exchange
349
350
351
352
353
354
-90 -70 -50 -30 -10 10 30 50 70 90
Latitude
CO
2 C
on
cen
trat
ion
(p
pm
)
Background biosphere exchange
352
353
354
355
356
357
358
359
360
361
362
363
364
-90 -70 -50 -30 -10 10 30 50 70 90
Latitude
CO
2 C
on
cen
trat
ion
(p
pm
)
Total background
Fig. 3. Annual mean, zonal mean surface CO2 concentration (ppm) resulting from the individual and combined (relative toa background concentration of 350 ppm) background fluxes for each of the models. The interhemispheric difference (ppm)for the background fossil and combined background CO2 is listed in the key for each model. Note that the scale is differentfor each of the plots.
flux response. The 1990 fossil distribution used in thisstudy is the same as that used in TransCom 1 but thetotal emission is 5.81 Gt C yr−1 compared to the 5.3 GtC yr−1 previously used. Allowing for this difference,eight of the TransCom 1 models had IHDs below thecurrent range (Law et al., 1996) indicating an evolutionto less vigorous interhemispheric transport.
The zonal mean surface concentration for the bio-sphere exchange shows more spread in the northernmid- to high latitudes than for fossil fuel emissions.In TransCom 1 (using a different biosphere exchange)the model results clustered into two groups. Here, thisclustering is less evident, with models spread through-out the 3.5 ppm range of concentrations around 60–70 ◦N. The highest concentrations (strongest rectifica-tion) are produced by MATCH:NCEP, while the weak-est rectification is produced by the TM2 model. TheRPN model produces relatively large equatorial con-centrations, while the NIRE concentrations are greaterthan for other models around 20–30 ◦N. Besides thelarge variation in zonal mean response to these annu-
ally balanced fluxes, the models exhibit large varia-tions in the spatial structure of the rectifier responsewithin latitude zones (Fig. 2). Some strongly rectify-ing models (e.g., SKYHI, MATCH:MACCM2, RPN)limit concentration maxima to the immediate forcingareas over North America and Asia, where there are nostations. Others (NIES, NIRE, UCB, MATCH:NCEP)generate broad concentration maxima over the north-ern high latitudes and consequently have substantialimpact at observing sites.
The zonal mean surface concentrations for the oceanexchange show similar structure for all models. Thetropical source and extratropical uptake results in asmall maximum concentration in the tropics, whichvaries by about 0.5 ppm across the models. Variationsin the concentration minima are larger, particularlyaround 50–60 ◦S. SKYHI produces the lowest concen-trations in both hemispheres, while MATCH:CCM3exhibits low concentrations across all latitudes. RPNexhibits the largest equatorial maximum concentra-tion. The greatest tropical/extratropical gradient is
Tellus 55B (2003), 2
566 K. R. GURNEY ET AL.
exhibited by SKYHI, while UCI and UCIb exhibit thesmallest gradients.
The zonal mean surface concentration for the totalbackground fluxes shows how some models respondsimilarly to the background fluxes while others exhibitoffsetting responses. For example, while respondingstrongly to fossil fuel fluxes over the northern extra-tropical source region, MATCH:CCM3 and RPN ex-hibit only a moderate response to the seasonal bio-sphere exchange. Similarly, MATCH:MACCM2 andNIES, which have a moderate fossil fuel response,show a stronger response to the seasonal biosphereexchange.
The remaining models show some similarity in theirresponse to the background fluxes. In these cases, mod-els that produce large surface gradients for fossil fuelalso tend to produce large gradients for the backgroundocean exchange and exhibit a strong biospheric recti-fier. This is broadly consistent with both vertical trap-ping of CO2 in the lower model layers and with weakeradvective mixing as compared to models exhibitingsmaller gradients.
The spatial structure of surface CO2 simulated byeach model in response to the total of these backgroundfluxes is different from the observed spatial variationof the observations. Because the background fluxesare fixed, the adjustment of the 22 regional fluxes isemployed in the inversion to minimize this residualconcentration response (RCR). The RCR values at allthe observing stations (smoothed) is plotted for eachmodel in Fig. 4. All models show positive RCR valuesin the northern mid-latitudes, implying the need forCO2 uptake there. The greatest differences in RCRvalues between models are over the northern mid-latitudes (Fig. 4a). The standard deviation of theseRCR values across the 17 models is less than 0.4 ppmat most stations, and greater than 1.0 ppm at 6 stations.The greatest variability is at Hungary (σ = 2.4 ppm;47 ◦N, 16.7 ◦E).
Models with smaller background IHD values (CSU,JMA, MATCH:MACCM2) simulate higher RCR val-ues over the southern extratropics relative to those withlarger background IHD values (UCB, MATCH:NCEP,MATCH:CCM3). This suggests a potential relation-ship between interhemispheric transport and the mag-nitude of the inferred fluxes in the southern regions.Models with vigorous interhemispheric transport mayrequire greater uptake over the southern extratropicsrelative to those models that exhibit weaker interhemi-spheric transport. Over the northern middle latitudes,where the model spread is greatest, the consistently
largest RCR values are simulated by MATCH:NCEP,NIRE and UCB, and the smallest are simulated byTM2, CSU, CSIRO and JMA. These “clusters” of re-sponses may produce similarly clustered inferred re-gional flux estimates.
The longitudinal distribution of the RCR valuesover the northern middle latitudes varied considerablyamong the models (Figs. 4b and 4c), particularly atcontinental sites near source regions (ITN, BAL, SCHand HUN). In general, RCR values increased fromwest to east across North America and Europe, whichsuggests the need for terrestrial uptake in these regions.The largest west to east gradients over North Americawere exhibited by NIRE and RPN, while the smallestbelong to UCIb and TM2. UCIb and TM2 also had thesmallest gradients over Europe while MATCH:NCEPand TM3 had the largest.
3.2. Inversion results
3.2.1. Model mean sensitivity. The model mean re-sults for the control inversion were presented by Gur-ney et al. (2002). They found a Southern Ocean sinkconsiderably weaker than the prior estimate of Taka-hashi et al. (1999). Furthermore, this result was notsensitive to the transport model used and was well con-strained by observations. They also found reasonablystrong data constraints over the northern continents,with uptake distributed relatively evenly across thenorthern land regions. Finally, they found a very weakdata constraint in the tropics, with tropical flux esti-mates depending mostly on prior information, massbalance constraint and model transport.
To test the dependence of the model-mean inver-sion on the prior fluxes and their uncertainties, weperformed the series of sensitivity tests described inTable 2. The first represents inversions in which theprior flux uncertainties for both land and ocean basisfunction regions were increased to 2, 5 and 10 Gt Cyr−1, bringing the inversions closer to methods thatdo not use prior information to constrain the flux esti-mates. The results of this test are presented in Fig. 5.For most regions, the mean flux estimates (X symbols)are very insensitive to the prior information, and liewithin the uncertainty range of the control inversion.Exceptions are regions with few (or no) stations: thetropical and South Atlantic (all cases), northern Africa(±5 and ±10 Gt C yr−1 cases) and southern Africa andSouth America (±10 Gt C yr−1 case).
For each model and region, the figure depicts an es-timated flux and two uncertainty measures. The first
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 567
-1.2
-0.2
0.8
1.8
2.8
3.8
-90 -70 -50 -30 -10 10 30 50 70 90
Latitude
Longitude Longitude
resi
du
al c
on
cen
trat
ion
res
po
nse
(p
pm
)
resi
du
al c
on
cen
tati
on
res
po
nse
(p
pm
)
resi
du
al c
on
cen
tati
on
res
po
nse
(p
pm
)
CSU
GCTM
UCB
UCI
UCIs
UCIb
JMA
M:CCM3
M:NCEP
M:MACCM2
NIES
NIRE
RPN
SKYHI
TM2
TM3
CSIRO
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
-140 -130 -120 -110 -100 -90 -80 -70
esp
cba (-163)
pocn30
uta
cmo
car030car040
car050
itn496itn
bmw
bme
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
-10 -8 -6 -4 -2 0 2 4 6 8 10 12
ice (-20)m
hd
stm
prssch
cmn
hun (17)
a)
b) c)
Fig. 4. (a) Smoothed (mean of 10◦ latitudinal bins) residual concentration response (ppm) at the 76 CO2 observational stationsfor each of the models. (b) Smoothed (three station running mean) residual concentration response (ppm) at North Americanstations within the 30 ◦N to 70 ◦N longitudinal band. Longitudes are listed for stations located outside the center of therunning mean. (c) Same as (b) but for Europe.
uncertainty measure is the within-model uncertainty(circles) defined in eq. (7). The magnitude of the de-crease from the prior uncertainty indicates the degreeto which the final flux estimate is constrained by themeasurements. Figure 5 shows that for some regionsthe data constraint dominates. At the other extreme, the
within-model uncertainty for some regions increasessignificantly as the prior flux uncertainty is increased.This indicates that these regions, for example tropicalAmerica, are poorly constrained by the CO2 observa-tions. Note that since the overall growth rate of CO2
in the atmosphere is specified with small uncertainty
Tellus 55B (2003), 2
568 K. R. GURNEY ET AL.
Table 2. Brief description of the sensitivity tests
Test 1 (“Loose Priors”): The prior flux uncertainties for both land and ocean basis function regions are increased to 2, 5and 10 GtC yr−1. These cases bring the inversion closer to those methods that do not use prior informationto constrain the flux estimates.
Test 2 (“Zero Land Priors”): The prior basis function land fluxes are set to zero. Since many tropical regions are notwell observed, it is important to check how sensitive the flux estimates are to the inclusion of land-use changeinformation in the prior flux estimates.
Test 3 (“No Rect”): The background biospheric exchange is set to zero. This tests the sensitivity of the flux estimates tothe rectifier. This case was also shown in Gurney et al. (2002) but here we also show individual model results.
Test 4 (“Adjust Rect”): The background biosphere exchange flux uncertainty is set to ±100%. This tests what rectifiermagnitude the inversion estimates.
Test 5 (“Zero Ocean”): The background ocean exchange is set to zero. This tests whether the inversion is sensitive to thespatial distribution in the background ocean flux.
Fig. 5. Results of ‘Loose Priors’ case (sensitivity test 1). The control inversion is denoted by the leftmost set of symbols[crosses, posterior flux estimate (includes background ocean exchange); circles, within-model flux uncertainty; whiskers,between-model flux uncertainty; outer box, prior flux uncertainty]. The ±2 Gt, ±5 and ±10 Gt C yr−1 prior uncertainty casesare denoted progressively to the right of the control inversion. All uncertainties represent 1σ . Mean does not include CSIRO.
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 569
(0.074 Gt C yr−1), the sum of the fluxes from the poorlyconstrained regions is fixed. As the priors are relaxed,dipoles of unrealistically large sources and sinks ap-pear in these regions, but cancel each other to retainthe observed mass of CO2 in the atmosphere.
The second uncertainty measure is the standard de-viation of the flux estimates across the ensemble ofmodels (error bars) as defined in eq. (8). This measureindicates the degree to which transport model differ-ences contribute to the range of flux estimates. Over-all, the between-model uncertainty increases with thewithin-model uncertainty in Fig. 5. This reflects theincreasing sensitivity of the inversion to differences intransport as the prior flux uncertainty is increased. Asthe priors are “loosened,” we also obtain a somewhatcloser fit to the observations: the mean data mismatch
Flu
x (G
tCyr
-1)
Flu
x (G
tCyr
-1)
Flu
x (G
tCyr
-1)
Fig. 6. Results of sensitivity tests 2–5. The control inversion is denoted by the leftmost set of symbols (as in Fig. 5). The‘Zero Land Priors’ case (sensitivity test 2), ‘No Rect’ case (sensitivity test 3), ‘Adjust Rect’ case (sensitivity test 4), and ‘ZeroOcean’ (sensitivity test 5) are denoted progressively to the right. Mean does not include CSIRO.
drops from 0.89 ppm for the control inversion to 0.77ppm for ±10 Gt C yr−1 case. This modest decrease inthe mean data mismatch further suggests a general lackof sensitivity to the prior flux constraints chosen forthe control inversion. The large between-model uncer-tainty in some regions means that, while the mean esti-mated flux can be relatively stable, individual modelscan give some very large fluxes. Overall, we concludethat in regions with strong data constraint, the inferredfluxes are insensitive to the prior fluxes, whereas in re-gions with few observations the fluxes are sensitive toboth prior information and differences in model trans-port.
Figure 6 shows results from the other four sensitivitytests. When the land prior fluxes are set to zero onlythe estimated tropical Asian flux changes by greater
Tellus 55B (2003), 2
570 K. R. GURNEY ET AL.
than 0.2 Gt C yr−1. Even for this region, the new fluxestimate does not lie outside the control uncertaintyrange, which at 0.74 Gt C yr−1 is not much smallerthan its prior uncertainty of 0.87 Gt C yr−1. This is be-cause none of the 76 sites used in the control inversioncontributes much constraint to this region.
As described in Gurney et al. (2002), removing thebackground biosphere fluxes (third set of estimatesin Fig. 6) has a large impact in some northern regionswhere rectification is pronounced; mean flux estimateschange by up to 1.1 Gt C yr−1 and between-model un-certainty is generally reduced. For example, in almostall models, Boreal Asia changes from a moderate sinkto a moderate source. In Test 4, we allow the inversionto optimize the magnitude of the seasonal rectifier byassociating a large prior uncertainty (±100%) with theannually balanced biospheric flux. This tends to pro-duce fluxes that are mid-way between the control andthe ‘No Rect’ case, consistent with the 50% reduc-tion in the model mean annually balanced biosphericexchange. In the final test, removing the backgroundocean flux had a larger impact in many of the land re-gions compared to oceanic regions, probably due to thegenerally larger prior uncertainties for land regions.However, none of the flux changes is larger than thecontrol within-model uncertainties for the appropriateregion.
3.2.2. Individual model results. Table 3 lists the con-trol inversion flux estimates for the individual modelswhen the land and ocean regions have been aggregatedseparately into the southern extratropics, tropics andnorthern extratropics. Figure 7 shows similar informa-tion, plotted as differences from the model mean fluxfor each region, along with flux differences for the‘No Rect’ (test 3) and ‘Zero Ocean’ (test 5) sensitivitytests. The total land and total ocean flux estimates ex-hibit considerable spread and are anti-correlated sincethe total source is constrained. MATCH:NCEP andCSIRO produce flux estimates that lie outside the rel-atively large within-model uncertainties on the totalland and total ocean regions. Removal of the back-ground ocean flux reduces the extent to which thesetwo models are outliers, with their flux estimates nowinside the within-model uncertainty range. It is notclear why these two models respond in this way, sincethe strength of their responses to the background oceantracer are similar to those of other models.
For the northern, tropical and southern regions, thereis greater model spread for the aggregated land re-gions than for the oceans. This is driven by a com-bination of the larger variability in model response
to annually balanced biospheric versus oceanic back-ground fluxes (Figs. 2 and 3) and the larger prior fluxuncertainties for land versus ocean regions. The in-fluence of the annually balanced biospheric flux re-sponse is indicated by the considerable reduction inmodel spread in the ‘No Rect’ sensitivity test. Fur-ther evidence for this relationship is indicated by astrong correlation (r = −0.74) between the estimatednorthern land flux for each model and the annually bal-anced biosphere IHD values whereas there is no cor-relation with the fossil fuel IHD values (r = −0.04).Finally, the largest northern land changes betweenthe control inversion and the ‘No Rect’ sensitivitytest occur for those models with the largest rectifier:MATCH:NCEP, MATCH:MACCM2, NIES, UCB andTM3.
The tropical land flux estimates are negatively cor-related (r = −0.81) with the northern land estimates.This occurs because the global growth rate is specifiedand with little observational constraint near the tropicalcontinents, these tropical regions act as a repository forthe flux residual remaining after optimization is madeto regions with stronger observational constraints. Notsurprisingly, MATCH:NCEP, with the largest rectifier,exhibits the largest tropical land flux change (2.2 Gt Cyr−1 source to −0.1 Gt C yr−1 sink) when the rectifieris excluded in the ‘No Rect’ test.
The aggregated southern ocean flux estimates showthe least amount of model spread. As suggested bythe correlation (r = 0.55) between the total back-ground flux IHD and the aggregated southern oceanposterior flux estimate, weaker interhemispheric trans-port (large background IHD values) correlates withless aggregated southern ocean uptake relative to thebackground ocean flux (−1.8 Gt C yr−1). This lendssupport to our speculation that weak interhemispherictransport leads to lower southern hemisphere RCR val-ues which, in turn, result in less uptake in the aggre-gated southern ocean relative to the background fluxin this region. However, this relationship does not holdwhen the southern land and southern ocean are com-bined, suggesting the influence of other factors suchas anomalously strong responses at particular stationsor compensating tradeoffs with other regions.
Table 4 lists the individual model flux estimatesand their uncertainties for all 22 land and ocean re-gions. Figure 8 shows similar information plotted asdifferences from the model mean flux for each region,for the control inversion and the ‘No Rect’ sensitiv-ity test. Table 5 lists the change in individual modelflux estimates when the inversion is run without the
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 571
Tabl
e3.
Agg
rega
ted
post
erio
rflu
xes
and
wit
hin-
mod
elun
cert
aint
ies
(1σ,G
tCyr
−1)
for
indi
vidu
alm
odel
s,co
ntro
linv
ersi
ona
Mod
el\R
egio
nN
orth
land
Nor
thoc
ean
Tro
pica
llan
dT
ropi
calo
cean
Sout
hla
ndSo
uth
ocea
nTo
tall
and
Tota
loce
an
CSU
−2.5
±0.
6−1
.7±
0.5
2.8
±1.
40.
9±
0.8
−0.7
±1.
2−1
.6±
0.6
−0.5
±1.
3−2
.4±
1.3
UC
B−2
.9±
0.6
−1.1
±0.
62.
1±
1.3
0.3
±0.
7−0
.4±
1.1
−0.8
±0.
7−1
.1±
1.2
−1.7
±1.
2U
CI
−1.4
±0.
7−1
.6±
0.6
−0.3
±1.
40.
8±
0.7
0.7
±1.
0−1
.0±
0.7
−1.1
±1.
2−1
.7±
1.2
UC
Is−1
.5±
0.7
−1.1
±0.
6−0
.8±
1.3
1.1
±0.
70.
3±
1.2
−0.8
±0.
7−2
.0±
.13
−0.8
±1.
3U
CIb
−1.4
±0.
7−1
.6±
0.6
−0.4
±1.
40.
9±
0.6
0.7
±1.
0−1
.0±
0.7
−1.1
±1.
1−1
.7±
1.1
JMA
−2.0
±0.
7−0
.7±
0.6
0.9
±1.
5−0
.0±
0.7
0.1
±1.
2−1
.1±
0.8
−1.0
±1.
3−1
.8±
1.3
M:C
CM
3−2
.5±
0.6
−0.7
±0.
50.
7±
1.3
0.6
±0.
7−0
.3±
1.1
−0.6
±0.
6−2
.1±
1.2
−0.7
±1.
2M
:NC
EP
−3.6
±0.
5−0
.0±
0.4
2.2
±1.
20.
6±
0.7
−2.0
±1.
00.
1±
0.6
−3.4
±1.
10.
6±
1.1
M:M
AC
CM
2−3
.4±
0.6
−0.3
±0.
51.
8±
1.4
0.8
±0.
8−0
.3±
1.2
−1.4
±0.
7−1
.9±
1.2
−0.9
±1.
2N
IES
−3.2
±0.
6−0
.5±
0.4
2.2
±1.
3−0
.3±
0.6
−0.1
±1.
1−0
.8±
0.7
−1.2
±1.
0−1
.6±
1.0
NIR
E−2
.5±
0.6
−2.1
±0.
41.
2±
1.3
0.5
±0.
50.
4±
1.1
−0.4
±0.
6−0
.8±
0.9
−2.0
±0.
9R
PN−2
.0±
0.6
−1.3
±0.
41.
0±
1.0
−0.4
±0.
7−0
.2±
0.9
0.1
±0.
6−1
.2±
1.1
−1.6
±1.
1SK
YH
I−3
.2±
0.6
−0.9
±0.
32.
9±
1.3
−0.2
±0.
6−0
.5±
1.1
−1.0
±0.
6−0
.7±
1.0
−2.1
±1.
0T
M2
−0.8
±0.
6−1
.6±
0.5
−0.2
±1.
5−0
.0±
0.5
0.6
±1.
2−0
.9±
0.7
−0.3
±1.
2−2
.5±
1.2
TM
3−3
.1±
0.5
0.1
±0.
41.
4±
1.2
0.5
±0.
6−0
.5±
1.1
−1.2
±0.
7−2
.2±
1.1
−0.6
±1.
1G
CT
M−1
.2±
0.5
−1.6
±0.
40.
8±
1.4
0.2
±0.
6−0
.4±
1.2
−0.5
±0.
6−0
.9±
1.0
−1.9
±1.
0C
SIR
O−2
.7±
0.6
−0.1
±0.
41.
3±
1.3
0.5
±0.
8−1
.3±
1.1
−0.5
±0.
7−2
.7±
1.2
−0.1
±1.
2M
ean
−2.
3±
0.6
−1.
1±
0.5
1.1
±1.
30.
4±
0.7
−0.
2±
1.1
−0.
8±
0.7
−1.
3±
1.1
−1.
5±
1.1
a The
back
grou
ndoc
ean
flux
has
been
incl
uded
inth
eoc
eani
cflu
xes.
Reg
iona
lagg
rega
tion
isas
follo
ws:
Nor
thla
nd(B
orea
lNor
thA
mer
ica,
Tem
pera
teN
orth
Am
eric
a,E
urop
e,B
orea
lAsi
a,Te
mpe
rate
Asi
a),N
orth
ocea
n(N
orth
Paci
fic,N
orth
ern
Oce
an,N
orth
Atla
ntic
),T
ropi
call
and
(Nor
ther
nA
fric
a,T
ropi
calA
sia,
Tro
pica
lAm
eric
a),T
ropi
calo
cean
(Wes
tPa
cific
,E
ast
Paci
fic,
Tro
pica
lA
tlant
ic,
Tro
pica
lIn
dian
),So
uth
land
(Sou
ther
nA
fric
a,A
ustr
alia
,So
uth
Am
eric
a),
Sout
hoc
ean
(Sou
thPa
cific
,So
uth
Atla
ntic
,So
uth
Indi
an,S
outh
ern
Oce
an).
The
mea
nso
urce
sdo
noti
nclu
deth
eC
SIR
Om
odel
.Uni
ts:G
tCyr
−1
Tellus 55B (2003), 2
572 K. R. GURNEY ET AL.
Fig. 7. Aggregated posterior flux differences from the model mean for the control inversion (top row in each box), the ‘NoRect’ case (sensitivity test 3, middle row in each box) and the ‘Zero Ocean’ case (sensitivity test 5, bottom row in each box).The box represents the within-model uncertainty (1σ ). Regional aggregation is as follows: North land (Boreal N America,Temperate N America, Europe, Boreal Asia, Temperate Asia), North ocean (N Pacific, Northern Ocean, N Atlantic), Tropicalland (Northern Africa, Tropical Asia, Tropical America), Tropical ocean (W Pacific, E Pacific, Tropical Atlantic, TropicalIndian), South land (Southern Africa, Australia, S America), South ocean (S Pacific, S Atlantic, S Indian, Southern Ocean).
background biospheric exchange. Most model flux es-timates lie within the estimated uncertainty for the re-spective regions in Fig. 8. This demonstrates that thewithin-model uncertainty encompasses most transportdifferences with different models providing the ex-treme or outlying estimates in different regions. Aswas noted in the aggregated flux estimates, some ofthe outlying total zonal estimates can be attributed tothe meridional gradient of the background fluxes. Re-sults for the complete 22 basis function regions ex-hibits some longitudinal tradeoffs as well. For exam-ple, many models with large uptake in extratropicalAsia such as MATCH:MACCM2. MATCH:NCEP andCSIRO show large downwind sources in the North Pa-cific region. Models with small combined uptake in ex-tratropical Asia, such as TM2, show the largest uptakein the North Pacific.
The two models with the greatest combined uptakein North America, RPN (−1.11 Gt C yr−1) and NIRE(−1.06 Gt C yr−1), support the suggestion that largewest to east declines in the RCR values are related toenhanced uptake. However, models with the smallest
North American uptake (UCIb and GCTM) are notthose with the smallest west to east gradient. OverEurope, the models with the smallest west to east gra-dient, UCIb and TM2, are among the three modelswith the smallest European uptake. Among those withthe largest European west to east RCR gradient, bothNIES and TM3 also exhibit among the greatest lev-els of uptake. UCB, however, has an average gradientbut considerable uptake in Europe. In northern extra-tropical Asia, the relationship between the RCR westto east gradient and the estimated flux appears lessconsistent.
Another example of a relationship between thebackground fluxes and the model estimates of region-specific fluxes arises in the Southern Ocean region.Models that estimated the smallest uptake for theSouthern Ocean region, SKYHI and MATCH:NCEP,were also models that responded most strongly to thebackground ocean flux (Figs. 2 and 3), particularly inthe southern extratropics. They were also among themodels with the lowest RCR values in the southernextratropics. Similarly, the responses of the CSIRO
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 573
Fig. 8. Posterior flux differences from the 16 model mean for the control inversion (top row in each box) and the ‘No Rect’case (sensitivity test 3, bottom row in each box). The boxes represent the 16 model mean within-model uncertainty (1σ ).
and SKYHI models to the background ocean flux inthe northern extratropics were the most pronounced(Fig. 3) and they also produced the least uptake in thenorthern ocean.
In contrast to the dependence of the estimated fluxeson the gradients in the background concentrations,there are a variety of instances in which outlyingflux estimates are much more dependent on modelresponses at particular sites. For example, two mod-els produce very large uptake for temperate NorthAmerica, NIRE and RPN, but for different reasons.RPN is one of two models with a pronounced fos-sil fuel response, particularly over temperate NorthAmerica. RPN also exhibits one of the weaker recti-fier responses. As a result, RPN exhibits the largestuptake for the temperate North American region butshows little change in North America when the back-ground biospheric exchange is removed. NIRE haslarger responses than other models at Key Biscaynefor background fossil and biospheric fluxes and forthe temperate North American region. This results insignificant uptake for the NIRE model in the temperateNorth American region, greater than all the models ex-cept RPN. This is confirmed by an inversion in whichCO2 data for Key Biscayne are excluded. The NIREflux difference from the mean is reduced from −0.76to −0.24 Gt C yr−1.
As with the aggregated flux estimates, Table 5 showssignificant changes over the northern land regions
when the background biosphere exchange is removedfrom the inversion. In particular, large changes occurover Boreal Asia where nearly all of the models shiftfrom uptake to emissions. The tropical land regionsalso exhibit shifts when inverting without the annu-ally balanced biosphere. In particular, northern Africabecomes a sink instead of a source in many models (orin the case of NIRE, a much reduced source) when theannually balanced biosphere is removed. This appar-ent compensation between northern rectifier responseand tropical terrestrial fluxes results from the globalmass balance constraint and the relatively small oceanpriors.
Models that exhibit significant surface gradients inresponse to the biospheric exchange are not alwaysthe models showing large shifts when the inversionexcludes this background flux. This is due to the spa-tial pattern of the rectifier in relation to the position ofthe observational stations. For example, SKYHI showsone of the larger rectifiers (Figs. 2 and 3), but showsonly a moderate shift in estimated flux when the back-ground biospheric flux is removed. This is due to thesomewhat limited horizontal transport of backgroundbiospheric flux near the surface over northern and east-ern Asia. Since the four stations nearest to this regionare located towards the eastern edge of the Asian con-tinent, the response at these stations does not reflectthe strength of the SKYHI rectification over centralportions of Asia.
Tellus 55B (2003), 2
574 K. R. GURNEY ET AL.
Table 4. Posterior fluxes and uncertainties (1σ, GtCyr−1) for individual models, control inversiona
MATCH: MATCH:Region\Model CSU UCB UCI UCIs UCIb JMA CCM3 NCEP
Boreal NA 0.21 ± 0.57 0.04 ± 0.23 0.45 ± 0.29 0.00 ± 0.37 0.51 ± 0.27 0.22 ± 0.53 −0.04 ± 0.04 −0.01 ± 0.38Temperate NA −1.02 ± 0.57 −0.16 ± 0.55 −0.64 ± 0.62 −0.34 ± 0.61 −0.55 ± 0.62 −0.93 ± 0.64 −0.40 ± 0.53 −0.77 ± 0.43Trop America 1.28 ± 1.09 1.06 ± 1.06 0.76 ± 1.02 0.83 ± 1.04 0.61 ± 0.99 0.74 ± 1.13 0.72 ± 1.06 0.36 ± 1.09South America −0.13 ± 0.98 −0.18 ± 0.91 −0.24 ± 0.90 0.12 ± 0.96 −0.24 ± 0.90 0.20 ± 1.01 −0.48 ± 0.94 −1.03 ± 0.88Northern Africa 0.30 ± 1.08 −0.10 ± 0.94 −1.16 ± 0.95 −1.36 ± 0.86 −1.16 ± 0.95 −0.41 ± 1.08 −0.85 ± 0.97 0.46 ± 0.91Southern Africa −0.78 ± 0.80 −0.79 ± 0.82 0.37 ± 0.64 −0.31 ± 1.02 0.42 ± 0.66 −0.20 ± 1.13 −0.29v0.78 −1.03 ± 0.74Boreal Asia 0.71 ± 0.52 −0.53 ± 0.55 0.09 ± 0.58 0.00 ± 0.53 −0.18 ± 0.56 −0.89 ± 0.55 −0.63 ± 0.64 −1.41 ± 0.43Temperate Asia −1.62 ± 0.79 −1.21 ± 0.59 −0.99 ± 0.81 −0.64 ± 0.69 −1.03 ± 0.79 0.14 ± 0.79 −0.96 ± 0.60 −0.47 ± 0.54Tropical Asia 1.23 ± 0.73 1.19 ± 0.79 0.08 ± 0.80 −0.22 ± 0.73 0.15 ± 0.80 0.54 ± 0.70 0.83 ± 0.78 1.34 ± 0.76Australia 0.19 ± 0.33 0.59 ± 0.36 0.52 ± 0.34 0.51 ± 0.30 0.53 ± 0.29 0.12 ± 0.26 0.43 ± 0.30 0.05 ± 0.19Europe −0.82 ± 0.50 −1.01 ± 0.39 −0.32 ± 0.52 −0.55 ± 0.43 −0.18 ± 0.54 −0.52 ± 0.54 −0.47 ± 0.51 −0.90 ± 0.30N Pacific −0.46 ± 0.28 −0.20 ± 0.35 −0.68 ± 0.37 −0.22 ± 0.33 −0.58 ± 0.34 −0.07 ± 0.40 0.16 ± 0.31 0.21 ± 0.23W Pacific −0.16 ± 0.30 −0.04 ± 0.35 −0.48 ± 0.35 0.44 ± 0.31 −0.38 ± 0.29 −0.12 ± 0.37 −0.03 ± 0.32 0.00 ± 0.35E Pacific 0.67 ± 0.45 0.70 ± 0.34 1.09 ± 0.33 0.55 ± 0.30 1.10 ± 0.33 0.49 ± 0.36 0.81 ± 0.39 0.53 ± 0.39S Pacific −0.15 ± 0.48 0.09 ± 0.59 0.38 ± 0.56 0.12 ± 0.63 0.37 ± 0.50 0.28 ± 0.61 −0.18 ± 0.52 0.27 ± 0.51Northern Ocean −0.37 ± 0.15 −0.46 ± 0.17 −0.65 ± 0.21 −0.49 ± 0.18 −0.64 ± 0.20 −0.29 ± 0.17 −0.24 ± 0.12 −0.10 ± 0.13N Atlantic −0.87 ± 0.35 −0.48 ± 0.29 −0.24 ± 0.32 −0.38 ± 0.32 −0.37 ± 0.29 −0.36 ± 0.33 −0.59 ± 0.30 −0.12 ± 0.28Tropical Atlantic −0.21 ± 0.31 −0.24 ± 0.34 −0.09 ± 0.36 0.02 ± 0.33 −0.09 ± 0.35 −0.01 ± 0.34 −0.01 ± 0.32 −0.04 ± 0.34S Atlantic −0.05 ± 0.44 −0.12 ± 0.44 −0.04 ± 0.46 0.08 ± 0.43 −0.03 ± 0.46 −0.03 ± 0.43 −0.10 ± 0.45 −0.05 ± 0.43Southern Ocean −0.79 ± 0.29 −0.44 ± 0.23 −0.75 ± 0.36 −0.41 ± 0.26 −0.79 ± 0.35 −0.61 ± 0.27 −0.25 ± 0.22 −0.15 ± 0.19Trop Ind Ocean 0.60 ± 0.50 −0.17 ± 0.32 0.30 ± 0.40 0.06 ± 0.46 0.25 ± 0.33 −0.39 ± 0.43 −0.15 ± 0.38 0.09 ± 0.36S Indian Ocean −0.56 ± 0.37 −0.34 ± 0.36 −0.56 ± 0.41 −0.59 ± 0.39 −0.51 ± 0.39 −0.71 ± 0.38 −0.10 ± 0.33 0.01 ± 0.27Glbl offset (ppm) 352.9 ± 0.2 352.6 ± 0.2 352.6 ± 0.2 352.3 ± 0.2 352.4 ± 0.2 352.9 ± 0.2 352.3 ± 0.2 352.5 ± 0.2
aThe background ocean flux has been included in the oceanic fluxes.
Some ocean regions also show significant shifts be-tween inversions with and without the biosphere ex-change. For example, both the North Pacific and NorthAtlantic exhibit the largest ocean region shifts betweenthe two cases, highlighting the downwind transport ofthe rectifier signal.
4. Discussion and conclusions
The TransCom 3 experiment has afforded the firstthorough investigation of the extent to which trans-port differences among tracer models contribute to theoverall uncertainty in inversion estimates of carbonsources and sinks. Fortunately, most of the models thathave been used to perform carbon cycle inversions inthe last decade participated in the current experiment.In addition to investigating the sensitivity of carbon in-versions to transport, we have been able to compute amodel mean result and test the sensitivity of the modelmean to various aspects of the inversion setup. A com-panion paper (Law et al., 2003) extends these sensitiv-ity tests by exploring the response of the model meanand individual model flux estimates to changes in theobservation network, the observational uncertainties,baseline data selection criteria and time period.
The first and perhaps most important result of thisexperiment is that model transport is as large a contrib-utor to the inversion uncertainty as the error producedby limited CO2 observations. Some of the individualmodel estimated fluxes can be readily attributed to howthey respond to the background flux fields. Modelsthat exhibited large CO2 concentration maxima nearand downwind of large background fluxes estimatelarge uptake in those same regions in order to bestmatch the CO2 observations. Models with small CO2
concentration maxima estimated less uptake near thebackground fluxes but compensated further downwindover ocean regions with weaker sources or small sinks.The model response to background fluxes was not al-ways the best predictor, however. Many of the regionalflux estimates for individual models were the result ofstrong responses at particular stations or subtle trade-offs and compensation among regions.
The response to surface SF6 fluxes was evaluatedin the previous TransCom 2 experiment (Denninget al., 1999). Models exhibiting small surface con-centration maxima like TM2 and an earlier variant ofthe UCB model systematically underestimated the ob-served meridional gradient of SF6 in the remote ma-rine boundary layer, so these models probably also un-derestimate regional fluxes in the present inversions.On the other hand, models exhibiting large surface
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 575
Table 4. (cont’d.)
MATCH:MACCM2 NIES NIRE RPN SKYHI TM2 TM3 GCTM CSIRO
−0.21 ± 0.32 0.62 ± 0.44 0.49 ± 0.44 0.71 ± 0.28 0.20 ± 0.37 0.39 ± 0.51 0.03 ± 0.52 0.57 ± 0.25 0.33 ± 0.41−0.64 ± 0.50 −0.95 ± 0.46 −1.60 ± 0.44 −1.77 ± 0.33 −1.06 ± 0.49 −1.09 ± 0.47 −0.91 ± 0.41 −0.41 ± 0.51 −0.63 ± 0.480.85 ± 1014 1.05 ± 10.2 −0.89 ± 1.07 −0.01 ± 0.95 1.08 ± 0.98 0.40 ± 1.18 1.41 ± 0.92 −0.11 ± 1.17 0.87 ± 1.05−0.39 ± 1.00 0.07 ± 0.94 0.12 ± 0.91 −0.17 ± 0.85 −0.62 ± 0.85 0.77 ± 0.99 −0.33 ± 0.92 −0.12 ± 0.95 −0.85 ± 0.87−0.12 ± 1.13 0.66 ± 0.88 1.44 ± 0.94 −0.15 ± 0.80 0.96 ± 1.06 −1.15 ± 1.08 −0.45 ± 0.95 0.37 ± 1.02 −0.05 ± 1.01−0.58 ± 1.15 −0.33 ± 0.91 0.02 ± 0.93 0.18 ± 0.51 −0.25 ± 1.19 −0.28 ± 1.02 −0.59 ± 1.04 −0.64 ± 1.17 −0.81 ± 1.15−0.99 ± 0.63 −1.34 ± 0.37 −0.94 ± 0.44 −0.18 ± 0.48 −0.52 ± 0.40 −0.08 ± 0.54 −0.63 ± 0.48 −0.76 ± 0.38 −1.70 ± 0.58−1.41 ± 0.71 −0.32 ± 0.62 −0.10 ± 0.68 0.05 ± 0.67 −1.14 ± 0.62 0.04 ± 0.61 −0.61 ± 0.36 0.31 ± 0.59 −0.24 ± 0.701.09 ± 0.79 0.44 ± 0.72 0.67 ± 0.67 1.13 ± 0.66 0.90 ± 0.68 0.57 ± 0.79 0.44 ± 0.72 0.49 ± 0.72 0.44 ± 0.720.65 ± 0.35 0.12 ± 0.24 0.27 ± 0.15 −0.22 ± 0.22 0.37 ± 0.22 0.15 ± 0.14 0.41 ± 0.27 0.36 ± 0.19 0.40 ± 0.36−0.13 ± 0.51 −1.22 ± 0.35 −0.32 ± 0.29 −0.78 ± 0.34 −0.62 ± 0.33 −0.02 ± 0.58 −1.01 ± 0.29 −0.94 ± 0.32 −0.47 ± 0.410.10 ± 0.33 0.06 ± 0.21 −0.66 ± 0.21 −0.44 ± 0.18 −0.63 ± 0.16 −0.92 ± 0.30 0.08 ± 0.26 −0.60 ± 0.21 0.28 ± 0.260.09 ± 0.38 −0.54 ± 0.30 −0.43 ± 0.21 −0.18 ± 0.18 −0.10 ± 0.34 −0.19 ± 0.28 0.08 ± 0.26 0.11 ± 0.26 −0.11 ± 0.360.98 ± 0.33 0.64 ± 0.33 0.69 ± 0.27 0.25 ± 0.32 0.36 ± 0.34 0.47 ± 0.20 0.46 ± 0.32 0.51 ± 0.28 0.76 ± 0.42−0.84 ± 0.56 0.42 ± 0.43 0.44 ± 0.28 0.44 ± 0.43 −0.90 ± 0.52 0.20 ± 0.52 −0.41 ± 0.66 0.13 ± 0.57 0.02 ± 0.60−0.19 ± 0.13 −0.19 ± 0.17 −0.23 ± 0.14 −0.17 ± 0.13 −0.04 ± 0.09 −0.34 ± 0.18 −0.09 ± 0.13 −0.34 ± 0.09 0.09 ± 0.08−0.20 ± 0.30 −0.41 ± 0.29 −1.20 ± 0.26 −0.72 ± 0.26 −0.25 ± 0.20 −0.34 ± 0.27 0.09 ± 0.27 −0.67 ± 0.27 −0.51 ± 0.28−0.04 ± 0.31 0.08 ± 0.34 0.40 ± 0.28 0.11 ± 0.29 −0.14 ± 0.28 −0.13 ± 0.32 −0.06 ± 0.32 −0.14 ± 0.24 0.33 ± 0.32−0.17 ± 0.42 0.02 ± 0.37 0.03 ± 0.39 0.12 ± 0.43 0.15 ± 0.33 −0.08 ± 0.43 −0.13 ± 0.41 −0.17 ± 0.41 0.19 ± 0.37−0.24 ± 0.27 −0.84 ± 0.30 −0.51 ± 0.29 −0.29 ± 0.24 −0.05 ± 0.16 −0.74 ± 0.35 −0.32 ± 0.25 −0.34 ± 0.22 −0.34 ± 0.16−0.24 ± 0.44 −0.47 ± 0.34 −0.18 ± 0.26 −0.53 ± 0.39 −0.31 ± 0.31 −0.15 ± 0.25 0.05 ± 0.33 −0.32 ± 0.29 −0.44 ± 0.43−0.19 ± 0.27 −0.38 ± 0.31 −0.33 ± 0.31 −0.20 ± 0.33 −0.19 ± 0.26 −0.29 ± 0.34 −0.29 ± 0.29 −0.09 ± 0.26 −0.34 ± 0.33352.6 ± 0.2 352.7 ± 0.2 352.3 ± 0.2 352.5 ± 0.2 352.6 ± 0.2 352.7 ± 0.2 352.6 ± 0.2 352.5 ± 0.2 352.6 ± 0.2
concentration maxima like TM3, which performedbetter at marine stations, tended to overestimate conti-nental concentrations of SF6 near source regions, pos-sibly reflecting excessive vertical trapping of tracer.A combination of model resolution, resolved advec-tive transport and subgrid-scale vertical transport de-fines this tradeoff between meridional gradients andregional concentration extrema which produce mostof the variability in model response and therefore influx uncertainty.
When the fluxes in the current experiment are ag-gregated into three zonal bands for the ocean and land,they provide some instructive examples of model be-haviour in estimating zonally integrated fluxes. Thetight ocean constraints of the inversion setup and thegreater variability in model response to the terrestrialbackground fluxes explain the greater model spread inthe estimates for land versus ocean regions. Removalof the background biospheric exchange considerablyreduces the model spread over land. Furthermore, thetropical land exchange is inversely related to the north-ern land uptake, highlighting the influence of the lim-ited observational constraint in the tropical regions andthe required compliance with the overall meridionalgradient in the CO2 observations. The influence of therectifier emphasizes the need to observe and under-stand this phenomenon. Because incorrect spatial andtemporal structure in this background flux cannot be
adjusted by the inversion procedure, errors in the spec-ification of the background flux will be aliased intoerrors in the regional flux estimates. Future TransComwork in which fluxes are adjusted on a monthly ba-sis will eliminate much of the fixed temporal structurerequired by the current annual mean inversion.
The last conclusion that can be drawn from the ag-gregated flux estimates is the relationship between theSouthern Ocean uptake and interhemispheric trans-port; models with large background flux IHDs andhence weak interhemispheric transport tend to esti-mate the greatest reduction in uptake when comparedto the background ocean flux for this region. Like manyof the other broad relationships between the back-ground fluxes and the estimated model fluxes, this re-lationship is not universal to all the models and someexceptions remain.
Examination of the 22 regional flux estimates acrossthe models further indicates some relationships to thebackground flux responses, though generalization ismuch more challenging. For example, models witha strong response to the background ocean flux esti-mate the least uptake in both the Southern Ocean andnorthern ocean regions. Models with large west to eastRCR gradients across North America and Europe areamong the models with the greatest uptake in these re-gions. However, models with small gradients are notnecessarily those with little uptake in these regions.
Tellus 55B (2003), 2
576 K. R. GURNEY ET AL.
Tabl
e5.
Post
erio
rflu
xdi
ffere
nces
inG
tCyr
−1fo
rin
vers
ion
perf
orm
edw
itho
utba
ckgr
ound
bios
pher
eflu
x(c
ontr
olca
sem
inus
‘No
Rec
t’se
nsit
ivit
yte
st)
MA
TC
H:
MA
TC
H:
MA
TC
H:
Reg
ion\
Mod
elC
SUU
CB
UC
IU
CIs
UC
IbJM
AC
CM
3N
CE
PM
AC
CM
2N
IES
NIR
ER
PNSK
YH
IT
M2
TM
3G
CT
MC
SIR
O
Bor
ealN
A0.
48−0
.32
−0.3
9−0
.41
−0.3
1−0
.12
−0.0
5−0
.14
−0.2
90.
03−0
.26
−0.0
40.
070.
050.
090.
250.
11Te
mpe
rate
NA
−0.3
00.
891.
010.
881.
040.
330.
380.
280.
88−0
.01
−0.1
6−0
.40
−0.0
60.
280.
250.
350.
26T
ropi
calA
mer
ica
0.32
−0.0
8−0
.42
−0.8
0−0
.59
−0.1
1−0
.03
0.09
0.31
0.23
−0.1
7−0
.07
−0.5
5−0
.34
−0.0
6−0
.38
−0.2
6So
uth
Am
eric
a0.
020.
290.
140.
270.
160.
06−0
.03
−0.2
5−0
.07
0.22
−0.2
5−0
.10
0.08
0.01
−0.2
0−0
.02
−0.3
3N
orth
ern
Afr
ica
0.68
0.80
−0.2
5−0
.66
−0.0
30.
54−0
.12
1.52
0.67
1.41
1.09
0.12
1.48
0.30
0.75
1.05
1.02
Sout
hern
Afr
ica
0.10
−0.5
80.
410.
490.
410.
170.
05−0
.79
−0.0
4−0
.23
0.36
0.56
0.37
−0.2
8−0
.22
−0.0
70.
01B
orea
lAsi
a−0
.90
−1.1
5−0
.29
−0.8
3−0
.48
−1.5
7−1
.45
−1.7
6−2
.02
−1.6
9−1
.35
−0.3
6−0
.83
−0.7
3−1
.39
−1.2
4−1
.89
Tem
pera
teA
sia
0.03
−0.0
9−0
.37
0.67
−0.3
40.
440.
580.
66−0
.29
0.59
1.20
1.63
0.04
0.74
0.00
1.01
0.90
Tro
pica
lAsi
a0.
110.
58−0
.12
−0.3
1−0
.01
−0.1
90.
300.
640.
26−0
.20
−0.0
6−0
.21
−0.0
6−0
.15
0.17
−0.1
20.
07A
ustr
alia
−0.0
50.
030.
000.
010.
010.
070.
040.
190.
060.
01−0
.05
−0.0
10.
000.
000.
070.
140.
13E
urop
e0.
33−0
.37
0.21
0.16
0.14
0.36
−0.0
2−0
.35
0.53
−0.0
20.
46−0
.18
0.25
0.52
0.08
0.43
0.06
NPa
cific
0.07
0.34
−0.0
30.
35−0
.01
0.47
0.54
0.32
0.47
0.34
−0.4
1−0
.13
−0.4
1−0
.44
0.36
−0.4
00.
32W
Paci
fic−0
.32
−0.1
8−0
.30
0.12
−0.2
6−0
.30
−0.1
9−0
.23
−0.1
6−0
.56
−0.4
2−0
.23
−0.3
5−0
.16
−0.1
8−0
.16
−0.4
1E
Paci
fic0.
210.
240.
310.
020.
33−0
.17
0.12
−0.2
6−0
.03
0.16
0.11
−0.0
3−0
.01
0.03
−0.0
30.
120.
05S
Paci
fic−0
.06
0.15
0.35
0.42
0.33
0.07
0.24
0.07
0.06
0.26
0.43
0.33
0.18
0.31
0.37
0.15
−0.0
4N
orth
ern
Oce
an−0
.10
−0.0
3−0
.10
−0.0
7−0
.08
0.05
0.03
0.13
−0.0
30.
06−0
.05
0.07
−0.1
20.
010.
04−0
.37
0.03
NA
tlant
ic−0
.49
−0.2
4−0
.14
−0.2
6−0
.23
0.07
−0.3
8−0
.05
−0.2
2−0
.08
−0.4
0−0
.38
0.01
−0.0
90.
17−0
.27
−0.1
8T
ropi
calA
tlant
ic−0
.26
−0.1
7−0
.13
0.02
−0.1
2−0
.01
−0.0
7−0
.11
−0.0
30.
010.
17−0
.06
0.14
−0.0
1−0
.09
−0.1
20.
20S
Atla
ntic
−0.0
5−0
.09
−0.0
10.
05−0
.01
−0.0
1−0
.02
−0.1
2−0
.13
−0.1
70.
00−0
.04
−0.1
5−0
.07
−0.1
4−0
.23
0.02
Sout
hern
Oce
an−0
.10
−0.1
3−0
.22
−0.1
6−0
.21
−0.0
9−0
.11
−0.0
2−0
.04
−0.1
1−0
.01
−0.1
5−0
.05
−0.0
2−0
.11
−0.0
8−0
.09
Tro
pIn
dian
Oce
an0.
330.
170.
340.
130.
29−0
.05
0.20
0.11
0.08
−0.2
4−0
.13
−0.3
7−0
.08
0.01
0.08
−0.1
1−0
.07
SIn
dian
Oce
an−0
.06
−0.0
5−0
.02
−0.0
8−0
.05
−0.5
7−0
.01
0.07
0.04
0.00
−0.1
30.
070.
050.
05−0
.02
0.06
0.06
Glb
loff
set(
ppm
)0.
02−0
.20
−0.2
1−0
.26
−0.2
2−0
.09
−0.8
0−0
.15
−0.1
4−0
.03
−0.1
0−0
.24
−0.1
4−0
.10
−0.0
9−0
.02
0.06
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 577
Furthermore, the estimated fluxes in northern extrat-ropical Asia appear to have a limited relationship to thedistribution of background flux response. Asia, how-ever, is where the largest changes occur in individualmodel flux estimates when the background biosphereexchange is removed from the inversion. Most mod-els exhibit large changes in their estimated flux whenthis change is made, often changing from a sink to asource.
In many instances, the model-to-model differencesappear to reflect particular individual model responsesat stations. Some of this is due to the coincidenceof large concentration gradients and station locations,while some may be due to local transport differencesat the surface, such as convective transport or the con-struction of the planetary boundary layer (PBL).
In contrast to the individual model flux estimates,the model mean fluxes outside of the tropical regionsappear to be relatively insensitive to changes in esti-mates of the prior fluxes and prior flux uncertainties.Less uptake in the Southern Ocean than has been im-plied by oceanographic observations and a large andan evenly distributed sink over the northern continentsremain despite dramatically increasing the uncertaintybounds on the prior fluxes, eliminating the prior fluxes,or eliminating the biospheric and oceanic backgroundfluxes. In this sense the model mean estimates can beconsidered robust to these aspects of the inversion set-up. A companion study (Law et al., 2003) finds themodel mean flux estimates to be relatively insensitiveto variations in the incorporation of the observationaldata.
A better understanding of regional carbon budgetsin the middle latitudes depends on improving the sim-ulated transport, whereas confidence in fluxes overthe tropical continents is primarily limited by sparsedata. The tropical land fluxes are determined primar-ily by global mass balance and the fluxes in betterobserved regions such as the northern extratropical re-gions. Combined with the lack of observational con-straint is the tendency for deep mixing in the tropics,which leads to weak responses at the surface in theseregions. The tradeoff between an uncertain rectifierin the northern extratropics and the poor tropical dataconstraint induces compensating fluxes. Models withstrong rectifiers are able to generate large northern up-take and still maintain global mass balance by intro-ducing large tropical sources, and vice versa. Improv-ing the tropical observations might therefore provide abetter constraint on the rectifier effect and therefore onmid-latitude fluxes. Conversely, a better understandingof northern rectifier effects would probably producebetter estimates of tropical carbon budgets.
5. Acknowledgements
This work was made possible through support fromthe National Science Foundation (OCE-9900310), theNational Oceanic and Atmospheric Administration(NA67RJ0152, Amend 30) and the International Geo-sphere Biosphere Program/Global Analysis, Interpre-tation and Modeling Project. S. Fan and J. Sarmientoacknowledge support from NOAA’s Office of GlobalPrograms for the Carbon Modeling Consortium.
REFERENCES
Andres, R. J., Marland, G., Fung, I. and Matthews, E. 1996.Distribution of carbon dioxide emissions from fossil fuelconsumption and cement manufacture, 1950–1990. GlobalBiogeochem. Cycles 10, 419–429.
Apps, M. J. and Kurz, W. A. 1994. The role of Canadianforests in the global carbon balance. In: Carbon Balanceon World’s Forested Ecosystems: Towards a Global As-sessment (ed.M. Kanninen) Academy of Finland, Helsinki,14–39.
Arakawa, A. and Schubert, W. H., 1974. Interaction of a cu-mulus cloud ensemble with the large-scale environment,Part I. J. Atmos. Sci. 31, 674–701.
Baker, D. F. 2001. Sources and Sinks of Atmospheric CO2
Estimated from Batch Least-Squares Inversions of CO2
Concentration Measurements. Ph.D. Dissertation, Prince-ton University.
Benoit, B. M., Desgagne, M., Pellerin, P., Pellerin, S.,Chartier, Y. and Desjardins S. 1997. The Canadian MC2:
A semi-Lagrangian, semi-implicit wideband atmosphericmodel suited for finescale process studies and simulation.Mon. Wea. Rev. 125, 2382–2415.
Bousquet, P., Ciais, P., Peylin, P., Ramonet, M. andMonfray, P. 1999a. Inverse modeling of annual at-mospheric CO2 sources and sinks 1. method andcontrol inversion. J. Geophys. Res. 104, 26161–26178.
Bousquet, P., Peylin, P., Ciais, P., Ramonet, M. and Monfray,P. 1999b. Inverse modeling of annual atmospheric CO2sources and sinks 2. sensitivity study. J. Geophys. Res.104, 26179–26193.
Bousquet, P., Peylin, P., Ciais, P., Le Quere, C., Friedlingstein,P. and Tans, P. 2000. Regional changes in carbon dioxidefluxes of land and oceans since 1980. Science, 290, 1342–1346.
Brenkert, A. L. 1998. Carbon dioxide emission estimatesfrom fossil-fuel burning, hydraulic cement production,
Tellus 55B (2003), 2
578 K. R. GURNEY ET AL.
and gas flaring for 1995 on a one degree grid cell basis.(http://cdiac.esd.ornl.gov/ndps/ndp058a.html).
Ciais, P., Tans, P., White, J. W. C., Trolier, M., Francey, R. andcoauthors. 1995. Partitioning of ocean and land uptake ofCO2 as inferred by δ13C measurements from the NOAAClimate Monitoring and Diagnostics Laboratory GlobalAir Sampling Network. J. Geophys. Res. 100, 5051–5070.
D’Andrea, F. S., Tibaldi, S., Blackburn, M., Boer, G., Deque,M. and coauthors. 1998. Northern Hemisphere atmo-spheric blocking as simulated by 15 atmospheric generalcirculation models in the period 1979–1988. Climate Dy-nam. 14, 385–407.
De Fries, R. S. and Townshend, J. R. G. 1994. NDVI-derivedland cover classifications at a global scale. Int. J. RemoteSens. 15, 3567–3586.
Denning, A. S., Fung, I. Y. and Randall, D. A. 1995. Latitudi-nal gradient of atmospheric CO2 due to seasonal exchangewith land biota. Nature 376, 240–243.
Denning, A. S., Randall, D. A., Collatz, G. J. and Sellers,P. J. 1996. Simulations of terrestrial carbon metabolismand atmospheric CO2 in a general circulation model. Part2: Spatial and temporal variations of atmospheric CO2.Tellus 48B, 543–567.
Denning, A. S., Holzer, M., Gurney, K. R., Heimann, M.,Law, R. M. and coauthors. 1999. Three-dimensional trans-port and concentration of SF6: A model intercomparisonstudy (TransCom 2). Tellus 51B, 266–297.
Ding, P. and Randall, D. A. 1998. A cumulus parameteriza-tion with multiple cloud base levels. J. Geophys. Res. 103,11341–11354.
Dixon, R. K., Brown, S., Houghton, R. A., Solomon, A. M.,Trexler, M. C. and Wisniewski, J. 1990. Carbon pools andflux of global forest ecosystems. Science, 263, 185–190.
Engelen, R. J., Denning, A. S., Gurney, K. R. andTransCom 3 modelers. 2003. On error estimation in atmo-spheric CO2 inversions. J. Geophys. Res. 107(02), 4635,doi:10.1029/2002JD002195.
Enting, I. 2002. Inverse Problems in Atmospheric ConstituentTransport. Cambridge University Press, Cambridge, U. K.
Enting, I. G. and Mansbridge, J. V. 1989. Seasonal sourcesand sinks of atmospheric CO2: Direct inversion of filtereddata. Tellus 41B, 111–126.
Enting, I. G., Trudinger, C. M. and Francey, R. J. 1995. Asynthesis inversion of the concentration and δ13C of atmo-spheric CO2, Tellus 47B, 35–52.
Fan, S., Gloor, M., Mahlman, J., Pacala, S., Sarmiento, J. andcoauthors. 1998. A large terrestrial carbon sink in NorthAmerica implied by atmospheric and oceanic CO2 dataand models. Science 282, 442–446.
GLOBALVIEW-CO2, Cooperative Atmospheric Data In-tegration Project - Carbon Dioxide, CD-ROM, NOAACMDL, Boulder, Colorado, 2000.
Gurney, K., Law, R., Rayner, P. and Denning, A.S. 2000. TransCom 3 Experimental Protocol,Department of Atmospheric Science, ColoradoState University, USA, Paper No. 707. http://transcom.colostate.edu/TransCom 3/transcom 3.html
Gurney, K. R., Law, R. M., Denning, A. S., Rayner, P. J.,
Baker, D. and coauthors. 2002. Towards robust regionalestimates of CO2 sources and sinks using atmospherictransport models. Nature 415, 626–630.
Hack, J. J. 1993. Description of the NCAR community cli-mate model (CCM2). NCAR/TN-382, 108 pp.
Hack, J. J. 1994. Parameterization of moist convection inthe National Center for Atmospheric Research communityclimate model (CCM2). J. Geophys. Res. 99, 5551–5568.
Hamilton, K., Wilson, R. J., Mahlman, J. D. and Um-scheid, L. J. 1995. Climatology of the SKYHI troposphere–startosphere–mesosphere general circulation model. J. At-mos. Sci. 52, 5–43.
Hansen, J., Sato, Mki., Ruedy, R., Lacis, A., Asamoah, K.and coauthors. 1997. Forcings and chaos in interannualto decadal climate change. J. Geophys. Res. 102, 25679–25720.
Hartke, G. J. and Rind, D. 1997. Improved surface andboundary layer models for the Goddard Institute for SpaceStudies general circulation model. J. Geophys. Res. 102,16407–16422.
Heimann, M. 1995. The global atmospheric tracer modelTM2, Technical Report, 10, Deutsches Klimarechenzen-trum, Hamburg, Germany, 51 pp.
Holtslag, A. A. M. and Boville, B. A. 1993. Local versus non-local boundary-layer diffusion in a global climate model.J. Climate 6, 1825–1842.
Houghton, R. A. 1999. The annual net flux of carbon to theatmosphere from changes in land use 1850–1990. Tellus51B, 298–313.
Houghton, R. A. and Hackler J. L. 1999. Emissions of carbonfrom forestry and land-use change in tropical Asia. GlobalChange Biol. 5, 481–492.
Kaminski, T., Heimann, M. and Giering, R. 1999. A coarsegrid three-dimensional global inverse model of the atmo-spheric transport, 2. Inversion of the transport of CO2 inthe 1980s. J. Geophys. Res. 104, 18555–18581.
Kaminski, T., Rayner, P. J., Heimann, M. and Enting, I. G.2001. On aggregation errors in atmospheric transport in-version. J. Geophys. Res. 106, 4703–4715.
Kauppi, P. E., Mielikainen, K. and Kuusela, K. 1992. Biomassand carbon budget of European forests 1971–1990. Science256, 70–74.
Keeling, C. D., Piper, S. C. and Heimann, M. 1989. A three-dimensional model of atmospheric CO2 transport based onobserved winds: 4. Mean annual gradients and interannualvariations. In: Aspects of Climate Variability in the Pacificand the Western Americas, Geophysical Monograph 55.(ed. D. H. Peterson), AGU, Washington, D.C., 305–363.
Koch, D. and Rind, D. 1998. 10Be/7Be as a tracer of strato-spheric transport. J. Geophys. Res. 103, 3907–3917.
Kurz, W. A. and Apps, M. J. 1999. A 70 year retrospectiveanalysis of carbon fluxes in the Canadian forest sector.Ecol. Appl. 9, 526–547.
Law, R. M., Rayner, P. J., Denning, A. S., Erickson, D.,Heimann, M. and coauthors. 1996. Variations in mod-elled atmospheric transport of carbon dioxide and theconsequences for CO2 inversions. Global Biogeochem.Cycles 10, 783–796.
Tellus 55B (2003), 2
CO2 INVERSION INTERCOMPARISON 579
Law, R. M. 1999. CO2 sources from a mass-balance inver-sion: Sensitivity to the surface constraint. Tellus 51B, 254–265.
Law, R. M. and Rayner, P. J. 1999. Impacts of seasonal co-variance on CO2 inversions. Global Biogeochem. Cycles13, 845–856.
Law, R. M., Chen, Y-H., Gurney, K. R., and TransCom 3modelers. 2003. TransCom 3 CO2 inversion intercompar-ison: 2. Sensitivity of annual mean results to data choices.Tellus 55B, this issue.
Levy, H., Mahlman, J. D., and Moxim, W. J. 1982. Tropo-spheric N2O Variability. J. Geophys. Res. 87, C4, 3061–3080.
Louis, J. F. 1979. A parameteric model of vertical eddy fluxesin the atmosphere. Boundary-Layer Meteord. 17, 187–202.
Mahlman, J. D. and Moxim, W. J. 1978. Tracer simulation us-ing a global general circulation model: results from a mid-latitude instantaneous source experiment. J. Atmos. Sci.35, 1340–1374.
Mahlman, J. D., Pinto, J. P. and Umscheid, L. J. 1994.Transport, radiative, and dynamical effects of the Antarc-tic ozone hole: A GFDL “SKYHI” model experiment. J.Atmos. Sci. 51, 489–508.
Maksyutov, S. and Inoue, G. 2000. Vertical profiles of radonand CO2 simulated by the global atmospheric transportmodel. In: CGER report, CGER-I039-2000, CGER, NIES,Japan, v.7, 39–41.
McGregor J. L. 1996. Semi-Lagrangian advection onconformal-cubic grids. Mon. Wea. Rev. 124, 1311–1322.
McGregor J. L. and Dix, M. R. 2001, The CSIRO Conformal-Cubic Atmospheric GCM, IUTAM Symposium on Ad-vances in Mathematical Modeling of Atmosphere andOcean Dynamics. Kluwer Academic Publishers, Dor-drecht, 197–202.
Pacala, S. W., Hurtt, G. C., Baker, D., Peylin, P., Houghton,R. A. and coauthors. 2001. Convergence of land- andatmosphere-based U. S. carbon sink estimates. Science292, 2316–2320.
Prather, M. 1986. Numerical advection by conservation ofsecond-order moments. J. Geophys. Res. 91, 6671–6681.
Prather, M., McElroy, M., Wofsy, S., Russell, G. and Rind, D.1987. Chemistry of the global troposphere: fluorocarbonsas tracers of air motion. J. Geophys. Res. 92, 6579–6613.
Randall, D. A., Shao Q. and Moeng, C.-H. 1992. A second-order bulk boundary-layer model. J. Atmos. Sci. 49, 1903–1923.
Randall, D. A. and Pan, D.-M. 1993. Implementation of theArakawa-Schubert parameterization with a prognostic clo-sure. In: The Representation of Cumulus Convection in Nu-merical Models (eds.K. Emanuel and D. Raymond). Amer-ican Meteorological Society, Boston, MA, 137–144.
Randerson, J. T., Thompson, M. V., Conway, T. J., Fung, I.Y. and Field, C. B. 1997. The contribution of terrestrialsources and sinks to trends in the seasonal cycle of atmo-spheric carbon dioxide. Global Biogeochem. Cycles 11,535–560.
Rasch, P., Mahowald, N. M. and Eaton, B. E. 1997. Represen-tations of transport, convection and the hydrologic cycle inchemical transport models: Implications for the modeling
of short lived and soluble species. J. Geophys. Res. 102,28127–28138.
Rayner, P. J., Enting, I. G., Francey, R. J. and Langenfelds,R. L. 1999. Reconstructing the recent carbon cycle fromatmospheric CO2, δ13C and O2/N2 observations. Tellus51B, 213–232.
Ritchie H. and Beaudoin, C. 1994. Approximations and sensi-tivity experiments with a baroclinic semi-Lagrangian spec-tral model. Mon. Wea. Rev. 122, 2391–2399.
Russell, G. L. and Lerner, J. A. 1981. A new finite-differencing scheme for the tracer transport equation. J.Appl. Meteorol. 20, 1483–1498.
Schimel, D. S., House, J. I., Hibbard, K. A., Bousquet, P.,Ciais, P. and coauthors. 2001. Recent patterns and mecha-nisms of carbon exchange by terrestrial ecosystems. Nature414, 169–172.
Schubert, S., Rood, R. and Pfaendtner, J. 1993. An assimi-lated dataset for Earth science applications. Bull. Am. Me-teorol. Soc. 74, 2331–2342.
Sellers, P. J., Randall, D. A., Collatz, G. J., Berry, J. A.,Field, C. B. and coauthors. 1996. A revised land surfaceparameterization (SiB2) for atmospheric GCMs. Part I:model formulation. J. Climate 9 676–705.
Strahan, S. E. and Mahlman, J. D. 1994. Evaluation ofthe GFDL “SKYHI” general circulation model usingaircraft N2O measurements: 2. Tracer variability anddiabatic meridional circulation. J. Geophys. Res. 99,10319–10332.
Suarez, M. J., Arakawa, A. and Randall, D. A. 1983. Param-eterization of the planetary boundary layer in the UCLAgeneral circulation model: Formulation and results. MonWea. Rev. 111, 2224–2243.
Taguchi, S. 1996. A three-dimensional model of atmosphericCO2 transport based on analyzed winds: Model descriptionand simulation results for TRANSCOM. J. Geophys. Res.101, 15099–15109.
Takahashi, T., Wanninkhof, R. H., Feely, R. A., Weiss, R. F.,Chipman D. W. and coauthors. 1999. Net sea–air CO2 fluxover the global oceans: An improved estimate based on thesea–air pCO2 difference. Proceedings of the 2nd CO2 inOceans Symposium, Tsukuba, Japan.
Tans, P. P., Fung, I. Y. and Takahashi, T. 1990 Observationalconstraints on the global atmospheric CO2 budget. Science247, 1431–1438.
Taylor, K. E., Williamson, D. and Zwiers, F. 1997.AMIP II Sea Surface Temperature and Sea Ice Con-centration Boundary Conditions. (http://www-pcmdi.llnl.gov/amip/AMIP2EXPDSN/BCS/amip2bcs.html).
Tiedke, M. 1989. A comprehensive mass flux scheme for cu-mulus parameterization in large-scale models. Mon. Wea.Rev. 117, 1779–1800.
UNFCCC. 2000. United Nations Framework Convention onClimate Change, National Communications from PartiesIncluded in Annex 1 to the Convention: Greenhouse GasInventory Data from 1990 to 1998 (FCCC/SBI/2000/11).
Zhang, G. J. and McFarlane, N. A. 1995. Sensitivity of cli-mate simulations to the parameterization of cumulus con-vection in the Canadian Climate Centre general circulationmodel. Atmos. Ocean 33, 407–446.
Tellus 55B (2003), 2