Estimating Terrestrial CO2 Fluxes from XCO2 Data using an EnKF: Sensitivity to Glint-view Measurements

1
Estimating Terrestrial CO2 Fluxes from XCO2 Data
using an EnKF Sensitivity to Glint-view
Measurements Spatial Resolution of Control
Variables
Liang Feng, Paul Palmer http//www.geos.ed.ac.uk/e
ochem Hartmut Bösch and Sarah Dance
2
Observing System Simulation Experiments

• Overall Aim Determine the potential of
  space-borne XCO2 data to improve 8-day surface
  CO2 flux estimates over tropical continental
  regions of size 12º15º.
• How sensitive are these estimates to changes in
  alternative measurement and model configurations?

3
XCO2 Data
Model XCO2
8-day Flux Forecasts (climatology)
GEOS-Chem
8-day forecast (3-D CO2, T H2O etc)
Prior error
Posteriori error
Obs operator
8-day OCO XCO2
ETKF (Living and Dance, 2008)
4
XCO2 Data
Model XCO2
8-day Flux Forecasts (climatology)
(Perturbations)
Surface CO2 Ensemble
GEOS-Chem
GEOS-Chem
8-day forecast (3-D CO2, T H2O etc)
8-day forecasts (3-D CO2, T H2O etc)
Prior error
Posteriori error
Obs operator
Obs operator
8-day OCO XCO2
ETKF (Living and Dance, 2008)
Model XCO2 Ensemble
5
XCO2 Data
Model XCO2
8-day Flux Forecasts (climatology)
Surface CO2 Ensemble
GEOS-Chem
GEOS-Chem
8-day forecast (3-D CO2, T H2O etc)
8-day forecasts (3-D CO2, T H2O etc)
Prior error
Posteriori error
Obs operator
Obs operator
8-day OCO XCO2
ETKF (Living and Dance, 2008)
Model XCO2 Ensemble
6

• Realistic XCO2 observation operator

1) Sampled along Aqua orbits
GEOS-Chem transport model (4x5 degree
resolution) Biosphere (CASA), Biomass (GFED),
Fossil fuel (NDIAC), Ocean (Takahashi)
1-day
3) Averaging kernels applied
2) Scenes with cloud or AOD gt 0.3 removed
Glint mode
Pressure hPa
Jan
Averaging kernels
7
Ensemble Kalman Filter Approach

• Based on Kalman Filter

Analysis
Forecast
KPfHT(HPfHTR)-1 is the Kalman gain matrix H is
the Jacobian (adjoint) matrix.
EnKF samples the forecast error covariance of the
forecast using an ensemble of forecasts.
Advantages no adjoint provides error
characterization can sample non-Gaussian PDF
(e.g., CO2-CO-CH4 inversion). Disadvantages the
size of the ensemble can be large (12x1441).
8

• Regional flux definitions based on TransCom 3
  regions

Control calculation 911 land regions, 411
ocean regions and 1 snow region (cf T3 11 land
and 11 ocean regions)

• Uncertainties based on TransCom 3
• We assume NO correlation in prior estimates
• Assume model error of 2.5 (1.5) ppm over land
  (ocean)

9
Mean Error Reduction from 2-Month Control
Inversion of 8-Day Surface Fluxes
Jan - Feb
Example South American Tropical A priori
err 3.2 Gt C/y A posteriori err 1.9-0.5 Gt
C/y Error reduction0.45-0.85.
10

• Error reductions are obviously sensitive to
  number of clean (aerosol and cloud free)
  observations

?
11

• Because of large assumed model error results are
  insensitive to observation error of single OCO
  retrieval

?
12

• Glint observations over ocean are more effective
  at constraining continental fluxes than nadir
  measurements

Results for 8-day mean flux estimates during May
to June
?
13

• Sensitivity to the spatial resolution of control
  variables from TransCom3 to Model Grid

South American Tropical Region
1
Avg Error Reduction
0.3
9x1/9 Transcom3
4x5 degree model grid
4x1/4 Transcom3
Transcom3
Correlations between neighbouring regions get
progressively larger using regions smaller than
1000x1000 km2.
14
Sensitivity to the spatial resolution of control
variables from TransCom3 to Model Grid

• Inversions at high spatial resolutions are
  under-determined, and usually show strong
  negative spatial correlation in the resulting
  error covariances

15
Concluding Remarks

• We have an EnKF assimilation tool for
  interpreting XCO2 data
• Realistic XCO2 distributions and associated
  errors will significantly reduce the uncertainty
  of continental CO2 fluxes on 8-day timescales BUT
  some consideration must be given to the lag
  window (not shown)
• Perturbing random and systematic components of
  measurement error lead to results consistent with
  4DVAR studies (not shown)
• Results are sensitive to assumed model error
• The number of clean observations impacts the
  quality of the flux estimates
• Glint observations offer the most leverage to
  reduce uncertainty in estimated continental CO2
  fluxes implications for 16-16 duty cycle?
• The spatial resolution of independently estimated
  CO2 fluxes from realistic XCO2 distributions is
  close to 1000x1000 km2