EnKF, 4D-Var and ECCO EnKF, 4D-Var and ECCO in a in a “ “ toy toy ” ” ocean-atmosphere model ocean-atmosphere model Tamara Singleton 1,2 , Eugenia Kalnay 1 , Kayo Ide 1 and Shu-Chih Yang 3 1 UMD, 2 Johns Hopkins, 3 Taiwan NCU
EnKF, 4D-Var and ECCOEnKF, 4D-Var and ECCO in ain a““toytoy”” ocean-atmosphere modelocean-atmosphere model
Tamara Singleton1,2, Eugenia Kalnay1,Kayo Ide1 and Shu-Chih Yang3
1UMD, 2Johns Hopkins, 3Taiwan NCU
EnKF, 4D-Var and ECCOEnKF, 4D-Var and ECCO in ain a““toytoy”” ocean-atmosphere modelocean-atmosphere model
Tamara Singleton1,2, Eugenia Kalnay1,Kayo Ide1 and Shu-Chih Yang3
1UMD, 2Johns Hopkins, 3Taiwan NCU
Simple Coupled Ocean-Atmosphere Model (Peña and Kalnay, 2004)
Coupling strengthTropical Atmosphere
Tropical OceanExtra-tropical AtmosphereOcean is vacillatingbetween a “normal”
(lasts about 2-10 years)and “El Niño” state
(lasts about a 1 year)
We compare 4D-Var and EnKF with this simple coupled model
Questions explored:-- Which is more accurate: 4D-Var or EnKF?-- Should we use short or long windows?-- Is it better to do an ocean reanalysis separately, or as
a single coupled system?-- Should we use frequent atmospheric observations in a
coupled system?-- Would RIP/QOL be beneficial in a coupled system?
ECCO is a ocean version of 4D-Var where the initialstate and the surface fluxes are both control variables.
This allows ECCO to use very long windows (decades)and estimate the surface fluxes that give the bestanalysis.
ECCO provides a single, continuous reanalysis--Is ECCO the best approach?
Simple Coupled Ocean-Atmosphere System
Ocean
Tropical atmosphere
Extratropical atmosphere
Model Parameter Definitions
3 coupled Lorenz models: A slow “ocean”component strongly coupled with a fast“tropical atmosphere component”, in turnweakly coupled with a fast “extratropicalatmosphere” (Peña and Kalnay, 2004).
Model State:k1=10k2 = -11
Uncenteringparameters
k1,k2
σ=10,b=8/3, andr=28
Lorenzparameters
σ, b, and rτ = 0.1time scaleτ
c,cz = 1ce = 0.08
Couplingcoefficient
c,cz,ce
ValuesDescriptionVariables
!xe = ! (ye " xe ) " ce(xt + k1)!ye = rxe " ye " xeze " ce(yt + k1)!ze = xeye " bze
!xt = ! (yt " xt ) " c(X + k2 ) " ce(xe + k1)!yt = rxt " yt " xtzt + c(Y + k2 ) + ce(ye + k1)!zt = xt yt " bze + czZ
!X = !" (Y # X) # c(xt + k2 )!Y = !rX # !Y # !XZ + c(yt + k2 )!Z = !XY # !bZ + czzt
[xe, ye, ze, xt , yt , zt ,X,Y ,Z]T
Simple Coupled Ocean-Atmosphere Model (Peña and Kalnay, 2004)
Coupling strengthTropical Atmosphere
Tropical OceanExtra-tropical AtmosphereOcean is vacillatingbetween a “normal”
(lasts about 2-10 years)and “El Nino” state
(lasts about a 1 year)
We compare 4D-Var and EnKF with this simple coupled model
Time series of the x-component
Simple Coupled Ocean-Atmosphere Model (Peña and Kalnay, 2004)
fast tropicalatmosphere
slow ocean“normalyears”
fastextratropicalatmosphere
Δt=0.01
We compare 4D-Var and EnKF with this simple coupled model
slow ocean“El Niñoyears”
Data Assimilation Experiment Design
• Simple Coupled Ocean-Atmosphere Model (perfect model)– Used to create the “true” trajectory
• Observations– Generated from the nature run plus “random errors” with s.d.– Every 8 time steps of a simulation
• Perform coupled and uncoupled ocean data assimilationswith several EnKF, 4D-Var, and ECCO-4D-Var
• Compute RMS errors of the difference between the analysisand the true solution.
• Lengthen assimilation windows, from 8 to 320 steps
• Perform fully coupled data assimilation (ETKF, 4D-Var),and just ocean assimilation (LETKF, 4D-Var and ECCO)
2
EnKF-Based Methods
Atmos: Available atanalysis timeOcean: Availablethroughout anassimilation window
Available at analysistime
Available at analysistime
Available throughoutan assimilationwindow
Available at the endof a window (analysistime)
Observations
4-dimensional
Subsystemlocalization
Fast and slowvariables separately
4D-LETKF(Separate Ocean)
Subsystemlocalization
Fast and slowvariables separately
LETKF(Separate Ocean)
Uses quasi-outerloop to improve theinitial analysis mean
Fast and slowvariablessimultaneously
ETKF-QOL(Fully coupled)
4-dimensionalFast and slowvariablessimultaneously
4D-ETKF(Fully coupled)
Fast and slowvariablessimultaneously
ETKF(Fully coupled)
Special FeaturesAssimilatingMethod
Description of EnKF-based methods
Coupled ocean-atmosphere ensembles:ETKF, 4D-ETKF, ETKF-QOL
RMS error as a function of assimilation window length
The fully coupled ETKF data assimilations work well.The shortest assimilation window (8 steps) is the best.
4D-ETKF (assimilating all the obs) is better than ETKF for longerwindows. ETKF-QOL has the best performance (short windows).
Variational Data AssimilationExperiments:
Fully Coupled 4D-VarOcean only 4D-Var
ECCO-like Ocean 4D-Var
Fully coupled 4D-Var : the Cost Function• In 4D-Var, a cost function is minimized to produce an optimal analysis.
– The cost function measures the distance between the model withrespect to the observations and with respect to the background state.
• The analysis is obtained by minimizing the cost function given by
J(x t0
) =12
x t0! x t0
b"# $%TB0
!1 x t0! x t0
b"# $% +12
H(x ti)! y ti
o"# $%TR ti
!1
i=1
N
& H(x ti)! y ti
o"# $%
Jo- “observation” cost functionJb - ”background” cost function
where the control variables are the initial 9 model variables:
x0 = xe0,ye
0,ze0,xt
0,yt0,zt
0,X0,Y0,Z0( )T
Initial model state for oceanInitial model state for tropical atmos.Initial model state for extratropical atmos
4D-Var: Quasi-static Variational DataAssimilation (QVA)
• For longer windows, multipleminima are a problem for 4D-Var minimization (Pires et al.,1996).
• Also for longer assimilationwindows, non-Gaussianperturbations of theobservation error andbackground error -> in non-quadratic cost functions
• Pires et al. (1996) proposedthe Quasi-static VariationalData Assimilation (QVA)approach.
x
t
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1 3 5 7 9
ETKF-QOL
LETKF
4D-LETKF
4D-Var
Obs. Error
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1 3 5 7 9
ETKF-QOL
LETKF
4D-LETKF
4D-Var
Obs. Error
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1 3 5 7 9
ETKF-QOL
LETKF
4D-LETKF
4D-Var
Obs. Error
Fully coupled 4D-Var (+QVA) and EnKF: shorter windows
8 16 24 32 40 48 56 64 72 80
assimilation window (time-steps)
8 16 24 32 40 48 56 64 72 80
assimilation window (time-steps)
8 16 24 32 40 48 56 64 72 80
assimilation window (time-steps)
Extratropics Tropics
OceanETKF-QOLprovides thebest analysisfor very short
windows
4D-Var competeswith EnKF-based
methods forlonger windows
Obs. Error Obs. Error
Obs. Error
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
1 2 3 4 5 6
ETKF-QOLLETKF4D-LETKF4D-VarObs. Error
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1 2 3 4 5 6
ETKF-QOLLETKF4D-LETKF4D-VarObs. Error
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
1 2 3 4 5 6
ETKF-QOLLETKF4D-LETKF4D-VarObs. Error
Fully coupled 4D-Var (+QVA) vs. EnKF: longer windows
88 96 120 160 200 240
assimilation window (time-steps) 88 96 120 160 200 240
assimilation window (time-steps)
88 96 120 160 200 240
assimilation window (time-steps)
Extratropics
Tropics
OceanCoupled 4D-Var
and EnKFcompetitive forlonger windows
Fully coupled 4D-Var vs EnKF summary
• We developed fully coupled 4D-Var and EnKF systems for thesimple coupled ocean-atmosphere model
• Lengthening the assimilation windows and applying QVAimproves the 4D-Var analysis because 4D-Var “forgets” B.But longer windows are more expensive…
• Fully coupled EnKF are optimal for short windows. Shortwindows are less expensive…
• EnKF+QOL works best (short windows).
• The optimal configurations (short windows for EnKFand long windows for 4D-Var) have similar accuracy.
ECCO-like 4D-Var
• The Consortium for Estimating the Circulation and Climateof the Ocean (ECCO) is a collaboration of a group ofscientists from the MIT, JPL, and the Scripps Institute ofOceanography
• The main characteristic of ECCO is that they includesurface fluxes as control variables.– This allows them to have exceedingly long assimilation windows in
4D-Var (e.g. 10 years or even 50 years).– They used NCEP Reanalysis fluxes (Kalnay et al, 1996) as a first
guess.
• ECCO used 4D-Var to estimate the initial ocean state andsurface fluxes (Stammer et al., 2004; Kohl et al., 2007) ina 50-year reanalysis
Carton and Santorelli (2008) plot of the First Empirical Orthogonal Eigenfunction of monthly heat content anomalyin the latitude band 20N-60N Explained variance is shown on the title line. Lower panel shows the correspondingcomponent time series annually averaged along with the Pacific Decadal Oscillation Index of Mantua et al. (1997)in black.
ECCO is the only one of the analyses for whichneither the first nor second heating EOF resemblethe Pacific Decadal Oscillation Pattern
ECCO
ECCO
Motivation: Comparison of Ocean Analyses
ECCO-like 4D-Var: Cost Function includesall surface fluxes as control variables
J =
12
[x0,f ! xb,nfe ]T (B0,nfe )!1[x0,f ! xb,nfe ]+12
[Hx ti- y ti
o ]T(R ti
!1)[Hx ti- y ti
o ]i=1
N
"
x0,f = X0,Y0,Z0 f1
1, f21, f3
1 f12 , f2
2 , f32 ... f1
n, f1n, f1
n( )T
Initial model state Fluxes for first 8 time steps Fluxes for last 8 time steps
Background state for the ocean
NCEP-like flux estimates for first 8 time steps
NCEP-like flux estimates for last 8 time steps
xb,nfe = Xb,Yb ,Zb f1
nfe,1, f2nfe,1, f3
nfe,1 f1nfe,2 , f2
nfe,2 , f3nfe2 ... f1
nfe,n, f1nfe,n, f1
nfe,n( )T
where the controlvariables are:
B0,nfe =B 00 Q
!
"#
$
%&
Comparison of ECCO-like & Ocean 4D-VarObs. s.d. error = 1.41 for oceanQVA APPLIED
ECCO improves the 4D-analyses
0
1
2
3
4
5
6
7
8
9
10
1 6 11
4D-VarECCO 4D-VarObs. Error
8 16 24 32 40 48 56 64 72 80 120 160 200 240 320
assimilation window (time-steps)
RMSE : Ocean State
ECCO (ocean only) remains satisfactory
OCEAN ONLY
4D-Var (ocean only) fails
By using sfc fluxes as control variables, ECCO can use very long windows
Are the ECCO fluxes more accurate?
RMS Errors (Flux 3 Estimate)
0.51
1.52
2.53
3.54
4.55
5.56
1 6 11
assimilation window (time-steps)
NCEP-like FluxEstimatesECCO Flux Estimates
ECCO does not improve the flux estimates over the first guess
8 16 24 32 40 48 56 64 72 80 120 160 200 240 320
assimilation window (time-steps)
Answers to the Research Questions
Questions:-- Which is more accurate: 4D-Var or EnKF?Fully coupled EnKF (with short windows) and 4D-Var (with longerwindows) have about the same accuracy. Both can handlefrequent atmospheric observations.
Answers to the Research Questions
Questions:-- Which is more accurate: 4D-Var or EnKF?Fully coupled EnKF (with short windows) and 4D-Var (with longerwindows) have about the same accuracy. Both can handlefrequent atmospheric observations.-- Is it better to do the ocean reanalysis separately, or as a singlecoupled system?Both EnKF and 4D-Var are similar and most accurate whencoupled, but uncoupled (ocean only) reanalyses are fairly good.
Answers to the Research Questions
Questions:-- Which is more accurate: 4D-Var or EnKF?Fully coupled EnKF (with short windows) and 4D-Var (with longerwindows) have about the same accuracy. Both can handlefrequent atmospheric observations.-- Is it better to do the ocean reanalysis separately, or as a singlecoupled system?Both EnKF and 4D-Var are similar and most accurate whencoupled, but uncoupled (ocean only) reanalyses are fairly good.-- Is ECCO 4D-Var with both the initial state and the surfacefluxes as control variables the best approach?In our simple ocean model 4D-Var cannot remain accurate withvery long windows. Our “ECCO” reanalysis remained satisfactorywith very long windows but at the expense of less accuratefluxes.
Practical Conclusions for CoupledAssimilation
• Since EnKF is as accurate as 4D-Var, and is optimal for shortwindows, it is more efficient than 4D-Var, which requires verylong windows for maximum accuracy.
• Contrary to our expectations, the best results included frequentatmospheric observations.
• ECCO is 4D-Var including surface fluxes as control variables.This allows ECCO to have very long windows (decades).
• Since the estimated surface fluxes “adapt” in order to force theocean model to be close to the observations, they are notguaranteed to be more accurate than the background fluxes.
• In our toy coupled model, the estimated surface fluxes wereless accurate than the background fluxes.
Extra: no-cost LETKF smoother allows a comparison of EnKFinitial and final increments: the initial 4D-Var increments are
sensitive to the norm, the final increments are similar to EnKF
“Smoother” reanalysis
LETKF Analysisxna = xn
f + Xnfwn
aLETKF analysis
at time n
Smoother analysis at time n-1 !xn!1
a = xn!1f + Xn!1
f wna
This very simple smoother allows us to go backand forth in time within an assimilation window:it allows assimilation of future data in reanalysis
Initial and final analysis corrections(colors), with one Bred Vector (contours)
LETKF
4D-Var
Initial increments
Initial increments
Final increments
Final increments
LETKF
4D-Var