Outline Ensemble Kalman filter (EnKF) implementation for Poseidon OGCM

Online Bias correction and Altimetry Online Bias correction and Altimetry Assimilation into a High Resolution OGCM Assimilation into a High Resolution OGCM

with an Ensemble Kalman Filter and with an Ensemble Kalman Filter and Impact on Seasonal Forecasts Impact on Seasonal Forecasts

Christian L. KeppenneChristian L. Keppenne11, Michele M. Rienecker, Michele M. Rienecker22, Nicole P. Kurkowski, Nicole P. Kurkowski33 and David D. Adamec and David D. Adamec55

1,31,3Science Applications International Corporation Science Applications International Corporation San Diego, CA 92121San Diego, CA 92121

2,42,4Global Modeling and Assimilation OfficeGlobal Modeling and Assimilation OfficeNASA Goddard Space Flight CenterNASA Goddard Space Flight Center

Greenbelt, MD 20771Greenbelt, MD 20771

JCSDA Science MeetingJCSDA Science MeetingApril 20-21, 2005April 20-21, 2005

[email protected], [email protected], [email protected] , [email protected]

OutlineOutline• Ensemble Kalman filter (EnKF) implementation for Poseidon Ensemble Kalman filter (EnKF) implementation for Poseidon

OGCMOGCM• Assimilation experiments with online bias estimationAssimilation experiments with online bias estimation• Impact of initializating OGCM with EnKF on CGCM seasonal Impact of initializating OGCM with EnKF on CGCM seasonal

hindcast skill hindcast skill

OGCM (ocean component of GMAO CGCMv1 forecasting OGCM (ocean component of GMAO CGCMv1 forecasting system)system)

• Poseidon (Schopf and Loughe, 1995)Poseidon (Schopf and Loughe, 1995)• Quasi-isopycnal vertical coordinateQuasi-isopycnal vertical coordinate• Prognostic variables are Prognostic variables are hh, , tt, , ss,, u u and and vv

• Sea surface height (SSH) is diagnostic: h = Sea surface height (SSH) is diagnostic: h = iibuoyancy(tbuoyancy(tii, s, sii) h) hii

• About 30 million prognostic variables at About 30 million prognostic variables at 1/3 x 5/8 x L27 resolution resolution

S/I Forecasting ObjectivesS/I Forecasting Objectives• Replace Temperature OI with Temperature + SSH EnKF in ODASReplace Temperature OI with Temperature + SSH EnKF in ODAS• Initialization of ocean component of CGCM ensembles with EnKFInitialization of ocean component of CGCM ensembles with EnKF

Multivariate updateMultivariate update• Multivariate compactly supported covariances

• Update T, S, u & v

• Layer thicknesses (h) adjust between analyses

• Incremental update

• Process SSH and T observations separately

Massively parallel implementationMassively parallel implementation(Keppenne and Rienecker, MWR 128, 1971-1981)(Keppenne and Rienecker, MWR 128, 1971-1981)

(Covariance scales determine size of PE regions and # of local observations)(Covariance scales determine size of PE regions and # of local observations)

• 4 hours/month on 256 PEs for 16 members on current platform: Compaq 4 hours/month on 256 PEs for 16 members on current platform: Compaq SC 45SC 45

EnKF implementationEnKF implementation

Online bias estimationOnline bias estimation• Used in SSH assimilationUsed in SSH assimilation

T+SSH EnKF also improves T

field and preserves salinity Equatorial Pacific too cold near surface

and below thermocline in free-model run

T OI (without S(T) correction) improves T field over free-

model run but S is degraded

Monthly mean TAO temperature 01/1998

T+SSH EnKF vs. T T+SSH EnKF vs. T OIOI

Online bias estimation in SSH assimilationOnline bias estimation in SSH assimilationProblemProblem

• T/P along-track data are anomaliesT/P along-track data are anomalies

• Model climatology changes as data are assimilatedModel climatology changes as data are assimilated

• Introduction of systematic errorsIntroduction of systematic errors

(particularly severe in eastern Pacific)(particularly severe in eastern Pacific)

SolutionSolution

• Assimilate unbiased innovations: Assimilate unbiased innovations: (z - (w(z - (wf f - b))- b))– Update bias estimateUpdate bias estimate– Update modelUpdate model

– Run bias estimate and ocean state side by sideRun bias estimate and ocean state side by side

observations

forecast

bias estimate

true climate

model climate

obse

rvab

le

a) Standard assimilation

Model is pulled away from model climatology towards

observed climate

true climate

model climate

obse

rvab

le

b) Assimilation with Online bias estimation (OBE)

est.

biasEstimates of variability (biased

model state) and climatological error (bias) are

evolved in parallel

).(][

)],([)(1

1

ftwtwfa

ka

kf tttt

HwzRHHPHPww

wFw

a) Standard update equation

)).((][

)),((][

),()(

)],([)(

1

1

1

1

aftwtwfa

fftbtbfa

ka

kf

ka

kf

tttt

tttt

bwHzRHHPHPww

bwHzRHHPHPbb

bb

wFw

b) Online bias estimation

•Pw and Pb are estimated from same ensemble distribution

•Covariances are localized but Pb is allowed to have larger scales than Pw

Assimilation experimentAssimilation experiment• 16-member EnKF16-member EnKF

• Assimilate T/P anomalies + TAO & XBT temperature profiles 1/1/93-Assimilate T/P anomalies + TAO & XBT temperature profiles 1/1/93-12/31/9312/31/93

• Online bias estimation in SSH assimilationOnline bias estimation in SSH assimilation

• Compare with no-assimilation control and with EnKF run without OBE Compare with no-assimilation control and with EnKF run without OBE

Marginal change to depth of 20C isotherm in response to a 0.1m SSH

innovation at (0N, 100W) on two different dates

Marginal change to the SSH bias estimate in response to the same 0.1m

SSH innovation

• RMS T OMF is 34% below control and RMS SSH OMF is 7% below control in EnKF run without bias estimation in SSH assimilation

• RMS OMF are 39% lower than control for T and 23% lower for SSH in run with bias estimation in SSH assimilation

Root mean square OMF differences for T & SSH

T SSH

Estimated SSH bias after 6 months Estimated SSH bias after 12 months

Impact of assimilation on CGCM hindcast skillImpact of assimilation on CGCM hindcast skill• Assimilate T + SSH for February-April of each year since 1993Assimilate T + SSH for February-April of each year since 1993• 16-member ensembles16-member ensembles• Start 12-month hindcasts using ocean states at the end of assimilation Start 12-month hindcasts using ocean states at the end of assimilation

runsruns• Assess impact of assimilation on SST and SSH hindcast skillAssess impact of assimilation on SST and SSH hindcast skill• Compare to history of CGCMv1 May-start hindcastsCompare to history of CGCMv1 May-start hindcasts

(to save CPU time, only 5 EnKF ensemble members are used to initialize hindcasts)(to save CPU time, only 5 EnKF ensemble members are used to initialize hindcasts)

Niño-3.4 SSTNiño-3.4 SST

EnKF T+SSH, OBEEnKF T+SSH, OBE

missed EN missed EN

NSIPP CGCMv1 tier 1NSIPP CGCMv1 tier 1

false LN alert false LN alert

missed EN

False EN alert

Niño-3 SSHNiño-3 SSH

EnKF T+SSH, OBEEnKF T+SSH, OBE

GMAO CGCMv1GMAO CGCMv1

SSH hindcastsSSH hindcasts

Although minor 1995 warm event is Although minor 1995 warm event is missed and 1998 EN is missed and 1998 EN is underestimated, the correlation with underestimated, the correlation with TOPEX is generally higher in the TOPEX is generally higher in the hindcasts initialized with EnKF T+SSHhindcasts initialized with EnKF T+SSH

Niño-3.4 SSHNiño-3.4 SSH

ConclusionsConclusions• Although only 16 members are used, use of EnKF has positive

impact on seasonal hindcast skill

• Use of online bias estimation in SSH assim. markedly improves the correlation with TOPEX of model SSH hindcasts

• EnKF-initialized hindcasts produce less false El Niño or La Niña alerts than the CGCMv1 hindcasts initialized with temperature OI

• Future: increase ensemble size on NAS Columbia system