Earth Observation Data and
Carbon Cycle Modelling
(an incomplete and subjective view…)
Marko Scholze
QUEST, Department of Earth Sciences
University of Bristol
GAIM/AIMES Task Force Meeting, Yokohama, 24-29 Oct. 2004
Overview
• Atmospheric CO2 observations
– TransCom
• Model-Data Synthesis
– Oceanic DIC observations: Inverse Ocean Modelling
Project
– Terrestrial observations: Eddy-flux towers
– Atmospheric observations: Carbon Cycle Data
Assimilation system
TransCom 3
Linear atmospheric transport inversion to calculate CO2 sources and sinks:
• 4 background "basis functions" for land, ocean, fossil fuels 1990 & 1995
• 11 land regions, spatial pattern proportional to terr. NPP
• 11 ocean regions, uniform spatial distribution
Solving for 4 (background) + 22 (regions) * 12 (month) basis functions!
TransCom 3 Seasonal Results
Guerney et al., 2004
(mean over 1992 to 1996)
inversion results:
response to background fluxes:
4
Gt C/yr
15
-35
ppm
-5
TransCom 3 Interannual Results
(1988 - 2003)
red: land
blue: ocean
Gt C/yr
darker bands: within-model
uncertainty
lighter bands: betweenmodel uncertainty
• larger land than ocean variability
• interannual changes more robust than seasonal
... but atmosphere well mixed interannually...
Baker et al. 2004
Model-Data Synthesis:
The Inverse Ocean Modelling Project
Recent ocean carbon survey, ~ 60.000 observations
C* of Gruber, Sarmiento, and Stocker (1996) to estimate anthropogenic DIC.
Innumerable data authors, but represented by Feely, Sabine, Lee, Key.
The Inverse Ocean Modelling Project
Jacobson, TransCom3
Meeting, Jena, 2003
The Inverse Ocean Modelling Project
• southward carbon transport of 0.37 Pg C/yr for pre-industrial times
• present-day transport -0.06 Pg C/yr (northwards)
Gloor et al. 2003
Terrestrial observations: Fluxnet
a global network of eddy covariance measurements
Inversion of terrestrial ecosystem parameter values
against eddy covariance measurements by
Metropolis Monte Carlo sampling
A Posteriori parameter PDF for Loobos site
ga,v: vegetation factor of atmospheric conductance
Evm: activation energy of Vm
Knorr & Kattge, 2004
Carbon sequestration at the Loobos site
during 1997 and 1998
Knorr & Kattge, 2004
CCDAS
Carbon Cycle Data Assimilation System
Misfit 1
Misfit to
observations
Forward Modeling:
Parameters –> Misfit
CO2 station
concentration
Atmospheric Transport
Model: TM2
Fluxes
Biosphere Model:
BETHY
Model parameter
Inverse Modeling:
Parameter optimization
CCDAS set-up
2-stage-assimilation:
1. AVHRR data
(Knorr, 2000)
2. Atm. CO2 data
Background fluxes:
1. Fossil emissions (Marland et al., 2001 und Andres et al., 1996)
2. Ocean CO2 (Takahashi et al., 1999 und Le Quéré et al., 2000)
3. Land-use (Houghton et al., 1990)
Transport Model TM2 (Heimann, 1995)
Methodology
Minimize cost function such as (Bayesian form):


1   T -1  
1  
J ( p )  p  p 0  C p 0 p  p 0   M ( p )  D
2
2


T

 

CD M ( p )  D
-1

where


- M is a model mapping parameters p to observable quantities
- D is a set of observations
- C error covariance matrix
 need of  p J (adjoint of the model)
Uncertainties of parameters
  J
C p    2
 p i, j
2



1
Uncertainties of prognostics X
 
  T
X(p)  X(p)

CX 
 Cp

p
p
Gradient Method
1stderivative (gradient) of

J (p) to model
p:
 parameters

 J (p ) p
yields direction of steepest
descent.
2nd derivative
(Hessian)

of J (p):

2
2
 J (p ) p
yields curvature of J.
Approximates covariance of
parameters.

cost function J (p)

Model parameter space (p)
Figure from Tarantola, 1987
Data Fit
Seasonal Cycle
Barrow
Niwot Ridge
observed seasonal cycle
optimised modeled seasonal cycle
Global Growth Rate
Atmospheric CO2 growth rate
Calculated as:
C GLO B  0.25C SPO  0.75C MLO
observed growth rate
optimised modeled growth rate
Error Reduction in Parameters
Relative Error Reduction
Carbon Balance
Euroflux (1-26) and other
eddy covariance sites*
net carbon flux 1980-2000
gC / (m2 year)
*from Valentini et al. (2000) and others
latitude N
IAV and processes
Major El Niño events
Major La Niña event
Post Pinatubo period
Interannual Variability
Normalized CO2 flux and ENSO
Lag correlation
(low-pass
filtered)
correlation coefficient
Outlook
• Data assimilation: problem better constrained without
• Model-Data-Synthesis:
better by
constrained
without
"artefacts" (e.g. spatialproblem
patterns created
station network)
"artefacts"
(e.g.
spatial
patterns
by station
but: cannot
resolve
processes
thatcreated
are not included
in thenetwork)
model
(look at residuals and learn about the model)
but: cannot resolve processes that are not included in the
• Simultaneous inversion of land and ocean fluxes
model
(look at residuals and learn about the model)
• Isotopes
• More data over
tropical of
lands:
• Simultaneous
inversion
landsatellites
and ocean fluxes
• Further data constraints (e.g. Isotopes, Inventories)
• More data over tropical lands: satellites
Posterior Uncertainty in Net Flux
Uncertainty in net carbon flux 1980-200
gC / (m2 year)
Uncertainty in prior net flux
Uncertainty in net carbon flux from prior values 1980-2000
gC / (m2 year)
Atm. Inversion on Grid-cell
Rödenbeck et al. 2003
• prior and posterior
uncertainties
• sensitivities (colors)
Atm. Inversion on Grid-cell
Not really at model grid of TM3, but aggregated to TM2 grid, 8° x 10°,
Underdetermined problem  correlation matrix (e.g. l=1275 km for NEE)
prior/posterior fluxes
and reduction in
uncertainty
Rödenbeck et al. 2003
CO2 Satellite Measurements
Vertical weighting functions
Sciamachy, OCO
Airs (U)
(=Upper limit)
Airs (L)
(=Lower limit)
Houweling et al. 2003
Pseudo Satellite Data Inversion
posterior/prior uncertainty
Houweling et al. 2003