GMAO Ocean EnKF Error Covariance Modeling

1
GMAO Ocean EnKF Error Covariance Modeling

• Ocean EnKF
• Multivariate compactly supported background
  covariances updates T, S, u v
• System-noise representation with forcing
  perturbations and vertical momemtum perturbations
• Online bias estimation for SSH
• (note SSH is diagnostic in Poseidon 4)
• Focus is on maximizing the performance with
  small ensembles

• Spatio-temporal filtering of background
  covariances
• New focus Adaptive error-covariance
  localization/deflation

GSOP/GODAE Synthesis Evaluation Meeting September
24-25, 2007
2
Spatio-temporal filtering of background
covariances

• Objective Improving the performance for small
  ensembles
• Temporal filter applied to Xf integration
  (exponential moving average)
• Spatial filter applied to (Xf -ltXgt) deviations
  (Gaussian filter)

Compactly supported EnKF (bias estimation omitted)
Time filter
3
Effect of spatial filtering Marginal Kalman
gain T obs _at_(0n,156E,150m) on 12/31/01 horiz.
section through ltT,Tgt covariances
EnKF-9
EnKF-17
EnKF-33
Unfiltered, compactly supported
Filtered, compactly supported
Unfiltered, not compactly supported
4
SSH OMF and OMA statistics
SSH passive
mean OMF
RMS OMF
mean OMA
RMS OMA
5
T OMF and OMA statistics
active
passive
passive
(a)
(c)
Control
EnKF
-
9
EnKF
-
9
mean OMF
mean OMF
RMS OMF
RMS OMF
mean OMA
mean OMA
RMS OMA
RMS OMA
(b)
(d)
active
passive
active
OIS(T)
EnKF
-
17
EnKF
-
17
(e)
passive
active
EnKF
-
33
EnKF
-
33
(f)
passive
active
EnKF
-
65
EnKF
-
65
6
Analysis validation
Comparison with December 2001 ADCP zonal currents
7
CTD
Comparison to CTD Salinity July 2005
OI-T,S no ARGO S
OI-T,S
ENKF T,S, SSH
Analysis validation
8
Analysis validation
T,S diagrams Comparison with ARGO/Levitus
OI-TS
Best above 20oC
Best below 20oC
9
Uniform localization

• SSS increment from processing Jan. 1 2000, XBT T
  data
• on ODAS 360x205 tripolar grid
• (MOM4 running on 720x410 tripolar grid)
• Each square corresponds to the location of a XBT
  profiler

adaptive localization and deflation
Principle Larger/smaller Kalman gain amplitude
and covariance localization scales where the
ensemble covariances at the observation location
are considered more/less reliable
Adaptive covariance localization/deflation