Transcript Compatibility of surface and aircraft station networks for

Compatibility of surface and
aircraft station networks for
inferring carbon fluxes
TransCom Meeting, 2005
Nir Krakauer
California Institute of Technology
[email protected]
Zhonghua Yang, Jim Randerson, Paul Wennberg
Motivation
• The effect of vertical
transport on CO2
concentrations near the
surface is a major
uncertainty in estimating
net regional carbon
fluxes
Gurney et al 2004
• More
• The
aerial and column CO2 data is becoming available
effect of using aircraft and column observations in
inversions provides a measure of the influence of error in model
vertical transport on flux estimates from inversions
Diagnosing model vertical transport with
measurements from aircraft: a schematic
CO2 upper troposphere
latitudinal profile
Aircraft data would
imply a northern
source
Modeled profile
Model vertical stratification is too strong –
slow mixing away from source
Spurious northern
sink inferred from
surface data
S
Modeled profile
CO2 latitudinal profile
In the boundary layer
N
(fossil) CO2 emissions
Note: Mixing rates here can be constant.
Additional biases can be introduced by the model representation of variability
in mixing (e.g. seasonal and diurnal rectifier effects).
Inversion set-up
• GLOBALVIEW-2004 stations
• Use mean CO2
concentrations for 20002003, when more aircraft
data collected
• Only stations with 60%
actual data for period
• Transport operators from
TransCom model annualmean output
• Data uncertainty assumed
proportional to station
residual standard deviation
• TransCom regions and
priors
NOAA CMDL
2000-3 flux for all northern regions
(land + ocean), Pg C/ y
Northern flux
Intermodel SD
Priors
-2.06±3.46
Land
Ocean
-0.74±3.17
-1.32±1.40
Surface data
-1.78
0.44
Aircraft data
-1.53
0.25
Combined data
-1.69
0.37
Distribution of
residuals
Models that are too vertically stratified
tend to imply a larger northern sink
Interim conclusions
• Comparing surface with aircraft data sets suggests that
much of the TransCom intermodel variability in northern
sink estimates is due to variation in model vertical mixing
strength
• Most TransCom models may have too little vertical
mixing, so that surface observations imply an overly
large northern sink
• Detailed comparison of modeled vs. observed vertical
CO2 distributions is probably required to diagnose just
where this mixing bias arises (convection? isentropic
transport? diurnal cycle?) and how to reduce it
• Next: for an imperfect model, how do we best combine
surface with other (e.g. aircraft) observations?
Strategies for choosing weights
• 1) Generalized cross
validation
– Through minimizing a
GCV objective function,
parameters such as the
weighting of prior
information (λ) and the
differential weighting of
high- and low-variability
stations (τ) can be set to
optimize the model’s
prediction of left-out
observations
– See my GRL paper
Parameter values
determined with GCV
Krakauer et al 2004
TransCom parameter
values
• 2) Here we try a maximum-likelihood Bayesian
approach (Koch and Kusche 2002)
– Divide inversion data into independent groups (e.g.
surface vs. aircraft observations vs. prior flux
information)
– Scale the initial error covariance matrix of each group
so that the residual size is equal to the group degrees
of freedom (essentially postulating that χ2 ≈ 1 for each
group)
– Do the inversion again with the new scaling, until
convergence
Weights for data & priors
Group
Size Initial mean
variance
Surface
data
Aircraft
data
108 1.01 ppm
Maximum-likelihood
mean variance
(intermodel range)
0.66 (0.61-0.74)
25
0.20 (0.08-0.50)
Prior region 22
fluxes
0.50 ppm
0.80 Pg C/ y 0.58 (0.23-1.18)
2000-3 northern flux with optimized
weights
(land + ocean), Pg C/ y
Northern flux
Intermodel SD
Priors
Land
Ocean
-2.06±3.46
-0.74±3.17
-1.32±1.40
Surface data
Aircraft data
-1.78
-1.53
0.44
0.25
Combined data
-1.69
0.37
-1.60
0.33
(original weights)
Combined data
(optimized weights)
2000-3 results by region (optimized
weights)
Source
Sink
Conclusions
• Observations of the CO2 concentration
vertical structure valuably complement
surface data
• Relative weights for surface and aircraft
observations can be assigned using
statistical methods
• Future work: look at the effect of including
(ground-based or satellite) column CO2
measurements, other data (isotopic, CO,
etc.) in inversions
Acknowledgments
• Funding from the Moore Foundation (NK),
NASA and NOAA (JR)
• The TransCom modelers
Papers cited
Gurney, K. R., R. M. Law, A. S. Denning, et al. (2004), Transcom 3
inversion intercomparison: Model mean results for the estimation
of seasonal carbon sources and sinks, Global Biogeochem.
Cycles, 18(1).
Koch, K. R. and J. Kusche (2002), Regularization of geopotential
determination from satellite data by variance components, J.
Geodesy, 76(5), 259-268.
Krakauer, N. Y., T. Schneider, J. T. Randerson, et al. (2004), Using
generalized cross-validation to select parameters in inversions for
regional carbon fluxes, Geophys. Res. Lett., 31(19).