11
4-Dimensional Variational Assimilation of Satellite Temperature and Sea Level Data in the
Coastal Ocean and Adjacent Deep Sea
John WilkinJavier Zavala-Garay
Julia Levinand Weifeng Gordon Zhang
Institute of Marine and Coastal Sciences Rutgers, The State University of New Jersey
New Brunswick, NJ, USA
ROMS User MeetingGrenoble, 6-8 October 2008
http://marine.rutgers.edu
22
ROMS* models of two Western Boundary Current regimes: East Australian Current (EAC) and Middle Atlantic Bight (MAB)
EAC:EAC:• deep ocean adjacent to deep ocean adjacent to
narrow continental shelf narrow continental shelf influenced by proximity of influenced by proximity of boundary currentboundary current
• large (250 km) mesoscale large (250 km) mesoscale eddies generated locally eddies generated locally by boundary current by boundary current separation process separation process
• 2000 x 1600 km model 2000 x 1600 km model domain; 25 km resolution domain; 25 km resolution
MAB:MAB:• wide shallow shelf wide shallow shelf
separated from Gulf separated from Gulf Stream by the Slope SeaStream by the Slope Sea
• Shelf/Slope Front (~0.3 Shelf/Slope Front (~0.3 m/s) at shelf edgem/s) at shelf edge
• Gulf Stream rings Gulf Stream rings frequently enter Slope frequently enter Slope Sea and impact shelfSea and impact shelf
• 800 x 300 km model 800 x 300 km model domain; 5 km resolutiondomain; 5 km resolution
* http://myroms.org
33
44
NERACOOS
Altimetry: Jason-n, ERS
55
NERACOOS
Altimetry: Jason-n, ERS
MARCOOSCODAR, gliders …
66
IS4DVAR*
• Given a first guess (the forward trajectory)…
• and given the available data…and given the available data…
( )oR x
ox
*Incremental Strong Constraint 4-Dimensional Variational data assimilation
77
IS4DVAR
• Given a first guess (the forward trajectory)…
• and given the available data…
• what change (or increment) to the initial what change (or increment) to the initial conditions (conditions (ICIC) produces a new forward trajectory ) produces a new forward trajectory that better fits the observations?that better fits the observations?
ox
( )oR x
( )o oR x x
88
The “best fit” becomes the analysis
assimilation window
ttii = = analysis analysis initial time initial time
ttff = analysis = analysis final time final time
The strong constraint requires the trajectory satisfies the physics in ROMS. The Adjoint enforces the consistency among state variables.
99
The final analysis state becomes the IC for the forecast window
assimilation window forecast
ttff = analysis = analysis final time final time
ttff + + = forecast = forecast horizon horizon
1010
Forecast verification is with respect to data Forecast verification is with respect to data not yet assimilatednot yet assimilated
assimilation window forecast
verification
ttff + + = forecast = forecast horizon horizon
1111
Basic IS4DVAR procedure:
Lagrange function
Lagrange multiplier
1
( ) ( )N
T ii i i
i
dL J
dt
xx λ N x F
( )i i t F F
( )i i t x x
( ) ( )i it i t λ λ λ
The “best fit” simulation will minimize L: model-data misfit is small, and model physics are satisfied
1 1
1
1 12 2( )
obsNT T
b b i i i i i ii
bo
JJ
J x
x x B x x H x y O H x y
J = model-data misfit
1212
Basic IS4DVAR procedure:
Lagrange function
Lagrange multiplier
1
1
0 ( ) 0
0
0 (0) (0) &(0)
0 ( ) 0 . .( )
ii i
i
TTi
i im m mi
b
dLNLROMS
dt
dLADROMS
dt
Lcoupling of NL AD
Li c of ADROMS
xN x F
λ
λ Nλ H O Hx y
x x
B x x λx
λx
1
( ) ( )N
T ii i i
i
dL J
dt
xx λ N x F
( )i i t F F
( )i i t x x
( ) ( )i it i t λ λ λ
At extrema of L
we require:
The “best” simulation minimizes L:
112
112
1
( )
obs
b
o
T
b b
NT
i i i i i ii
J
J
J x
x x B x x
H x y O H x y
J = model-data misfit
1313
Basic IS4DVAR procedure:
(1) Choose an
(2) Integrate NLROMS and save
(a) Choose a
(b) Integrate TLROMS and compute J
(c) Integrate ADROMS to yield
(d) Compute
(e) Use a descent algorithm to determine a “down gradient” correction to that will yield a smaller value of J
(f) Back to (b) until converged
(3) Compute new and back to (2) until converged
(0) (0)bx x
[0, ]t
(0)x
[0, ]t
[ ,0]t (0)(0)oJ
λx
1 (0) (0)(0)
J
B x λ
x
(0)x
(0) (0) (0) x x x
( )tx
Out
er-l
oop
(1
0)
Inne
r-lo
op
(3)
NLROMS = Non-linear forward model; TLROMS = Tangent linear; ADROMS = Adjoint
1
1
1
12
12
( )
b
o
T
b b
J
NT
i i i i i ii
J
J x
x x B x x
H x y O H x y
J = model-data misfit
14
Adjoint model integration is forced by
the model-data error
xb = model state (background) at end of previous cycle, and 1st guess for the next forecast
In 4DVAR assimilation the adjoint gives the sensitivity of the initial conditions to mis-match between model and data
A descent algorithm uses this sensitivity to iteratively update the initial conditions, xa, (analysis) to minimize Jb+ Jo
Observations minus previous forecast
x
0 1 2 3 4 time
previous forecast
xb
1515
Three ways: Three ways:
(1) (1) The Adjoint ModelThe Adjoint Model
(2) (2) Empirical statistical correlations to generate Empirical statistical correlations to generate “synthetic XBT/CTD”“synthetic XBT/CTD”
In EAC assimilation get T(z),S(z) from In EAC assimilation get T(z),S(z) from vertical EOFs of historical CTD vertical EOFs of historical CTD observations regressed on SSH and SSTobservations regressed on SSH and SST
(3) (3) Modeling of the background covariance matrixModeling of the background covariance matrix e.g. via the hydrostatic/geostrophic relation e.g. via the hydrostatic/geostrophic relation
How is observed information (SLA, SST) transferred tounobserved state variables (velocity) andprojected from surface to subsurface?
1616
Mid-Atlantic Bight ROMS Model for IS4DVAR
5 km resolution for IS4DVAR5 km resolution for IS4DVAR1 km downscale forecast1 km downscale forecast
• 3-hour forecast meteorology 3-hour forecast meteorology NCEP/NAMNCEP/NAM
• daily river flow (USGS)daily river flow (USGS)• boundary tides (TPX0.7)boundary tides (TPX0.7)
• nested in ROMS MAB-GoM nested in ROMS MAB-GoM (which is nested in Global-(which is nested in Global-HyCOM*)HyCOM*)– nudging in a 30 km boundary nudging in a 30 km boundary
zonezone– radiation barotropic mode radiation barotropic mode
(*which assimilates altimetry)(*which assimilates altimetry)
1717
Mid-Atlantic Bight ROMS Model for IS4DVAR
~20 km outer model:ROMS MAB-GoM, or…NCOM or global HyCOM+NCODA
5 km resolution IS4DVAR model embedded in …
1818
Mid-Atlantic Bight ROMS Model for IS4DVAR
5 km resolution for IS4DVAR 1 km downscale for forecast
EAC eddies are resolved by AVISO EAC eddies are resolved by AVISO multi-satellite multi-satellite SLASLA
MAB MAB SLASLA is more anisotropic with is more anisotropic with shorter length scales due to flow-shorter length scales due to flow-topography interactiontopography interaction
Use along-track altimetry:Use along-track altimetry:• 4DVar uses the data at time of 4DVar uses the data at time of
satellite passsatellite pass• model “grids” along-track data by model “grids” along-track data by
simultaneously matching simultaneously matching observations and dynamical and observations and dynamical and kinematic constraints kinematic constraints
• need coastal altimetry correctionsneed coastal altimetry corrections
1919
Mid-Atlantic Bight ROMS Model for IS4DVAR
Model variance (without Model variance (without assimilation) is comparable assimilation) is comparable to along-track in Slope Sea, to along-track in Slope Sea, but not shelf-breakbut not shelf-break
AVISO gridded SLA differs AVISO gridded SLA differs from along-track from along-track SLASLA in in Slope Sea (4 cm) and Gulf Slope Sea (4 cm) and Gulf Stream (10 cm)Stream (10 cm)
2020
The altimeter anomalies are with respect to the long-term mean The altimeter anomalies are with respect to the long-term mean and therefore contain seasonal variability, but this is small and therefore contain seasonal variability, but this is small compared to the mesoscale in the MAB. compared to the mesoscale in the MAB.
2121
ROMS assimilates total SSH. Therefore we need to add a mean dynamic topography (MDT) to the anomaly data prior to assimilation. This MDT is computed by 4DVAR analysis using a regional 3-D T-S climatology computed from historical hydrographic data
The altimeter anomalies are with respect to the long-term mean The altimeter anomalies are with respect to the long-term mean and therefore contain seasonal variability, but this is small and therefore contain seasonal variability, but this is small compared to the mesoscale in the MAB. compared to the mesoscale in the MAB.
2222
Mean Dynamic Topography (MDT) is computed by 4DVAR analysis of a regional 3-D T-S climatology computed from historical hydrographic data.
4DVAR analysis is forced with annual mean meteorology and open boundary conditions.
2323
2424
2525
High frequency variability: model and data issues
ROMS includes high frequency variability typically ROMS includes high frequency variability typically removed in altimeter processing (tides, storm surge)removed in altimeter processing (tides, storm surge)
The IS4DVAR cost function, The IS4DVAR cost function, JJ, samples this high , samples this high frequency variability, so it must be either (a) removed frequency variability, so it must be either (a) removed from the model or (b) included in the datafrom the model or (b) included in the data
Our approach:Our approach:• Run 1-year ROMS (no assimilation) forced by boundary Run 1-year ROMS (no assimilation) forced by boundary TPX0.7 tides; compute ROMS tidal harmonics TPX0.7 tides; compute ROMS tidal harmonics • de-tide along-track altimetry (developmental in MAB) de-tide along-track altimetry (developmental in MAB) • add ROMS tides to de-tided altimeter dataadd ROMS tides to de-tided altimeter data• thus the thus the observationsobservations are are adjustedadjusted to include model tide to include model tide
• assimilate – high frequency mismatch of model and assimilate – high frequency mismatch of model and altimeter is minimized and cost function is, presumably, altimeter is minimized and cost function is, presumably, dominated by sub-inertial frequency dynamics dominated by sub-inertial frequency dynamics
2626
High frequency variability: model and data issues
The IS4DVAR increment is to the initial conditions of The IS4DVAR increment is to the initial conditions of the analysis window, and this itself generates HF the analysis window, and this itself generates HF variability (inertial oscillations)variability (inertial oscillations)
2727
High frequency variability: model and data issues
The IS4DVAR increment is to the initial conditions of The IS4DVAR increment is to the initial conditions of the analysis window, and this itself generates HF the analysis window, and this itself generates HF variability (inertial oscillations)variability (inertial oscillations)
Our approach:Our approach:
• Apply a short time-domain filter to IS4DVAR initial Apply a short time-domain filter to IS4DVAR initial conditions conditions • Reduces inertial oscillations in the Slope Sea Reduces inertial oscillations in the Slope Sea butbut removes tides removes tides • Tides recover quickly Tides recover quickly
– – approach needs refinement approach needs refinement – possibly using 3-D velocity harmonic analysis of– possibly using 3-D velocity harmonic analysis of free running model free running model
2828
High frequency variability: model and data issues
Without a subsurface Without a subsurface synthetic-CTD synthetic-CTD relationship, the adjoint model can relationship, the adjoint model can erroneously accommodate too much of the erroneously accommodate too much of the SLA model-data misfit in the barotropic modeSLA model-data misfit in the barotropic mode
This sends gravity wave at along the This sends gravity wave at along the model perimeter model perimeter
Our approach:Our approach:
• Repeat (duplicate) the altimeter SLA observations at Repeat (duplicate) the altimeter SLA observations at t = -6 hour, t=0 and t = +6 hourt = -6 hour, t=0 and t = +6 hour but with appropriate time lags in the added tide signal but with appropriate time lags in the added tide signal • These data cannot easily be matched by a wave These data cannot easily be matched by a wave • We are effectively acknowledging the temporal correlation We are effectively acknowledging the temporal correlation of the sub-tidal altimeter SLA data of the sub-tidal altimeter SLA data
gh
gh
2929
High frequency variability: model and data issues
Our approach:Our approach:
• Repeat (duplicate) the altimeter SLA observations at Repeat (duplicate) the altimeter SLA observations at t = -6 hour, t=0 and t = +6 hourt = -6 hour, t=0 and t = +6 hour but with appropriate time lags in the added tide signal but with appropriate time lags in the added tide signal • These data cannot easily be matched by a wave These data cannot easily be matched by a wave • We are effectively acknowledging the temporal correlation We are effectively acknowledging the temporal correlation of the sub-tidal altimeter SLA data of the sub-tidal altimeter SLA data
gh
gh
3030
Other possible HF issues:Other possible HF issues:
• Should we include sea level air pressure in ROMS and Should we include sea level air pressure in ROMS and attempt to model the inverse barometer response? attempt to model the inverse barometer response?
• What about remote HF variability from coastal trapped What about remote HF variability from coastal trapped waves? waves? – – regional MOG2D?regional MOG2D?
High frequency variability: model and data issues
3131
Sequential assimilation of SLA and SST
Before attempting assimilation of COOS observatory in situ data for a full reanalysis, we are assimilating satellite data (SSH and SST) to tune for the assimilation parameters (e.g., horizontal and vertical de-correlation scales, assimilation window, etc.).
We use the unassimilated hydrographic data to evaluate how well the adjoint model is propagating information between variables, and in space and time. This experiment also serves as a prototype for a real-time forecast system.
3232
Sequential assimilation of SLA and SST
• Reference time is days after 01-01-2006
• 3-day assimilation
window (AW)
• Daily MW+IR blended SST (available real time)
• SSH = Dynamic topography + ROMS tides + Jason-1 SLA (repeated three times)
• For the first AW we just assimilate SST to allow the tides to ramp up.
3333
Sequential assimilation of SLA and SST
First AW (0<=t<=3 days)
Observed SST ROMS SST and currents at 200 m
3434
Sequential assimilation of SLA and SST
Second AW (3<=t<=6 days)
Observed SST ROMS SST and currents at 200 m
3535
Sequential assimilation of SLA and SST
Second AW (3<=t<=6 days)
Observed SST ROMS SST and currents at 200 m
Jason-1 data
3636
Sequential assimilation of SLA and SST
Second AW (3<=t<=6 days)
Observed SST ROMS SST and currents at 200 m
Jason-1 data
XBT transect(NOT assimilated)
3737
Sequential assimilation of SLA and SST
ROMS-IS4DVAR fits reasonably well the assimilated observations (SST and SSH), but how well does it represent unassimilated data?
ROMS solutions along the transect positions [lon,lat,time]
3838
Sequential assimilation of SLA and SST
ROMS-IS4DVAR fits reasonably well the assimilated observations (SST and SSH), but how well does it represent unassimilated data?
ROMS solutions along the transect positions [lon,lat,time]
3939
Future work
Evaluation of the forward model shows significant forecast skill error due to biases in the boundary and surface forcing.
We will therefore use climatology as an additional data source in the assimilation procedure.
After bias correction, we will assimilate all the available COOS data (CODAR surface currents, glider T-S, CTD, XBT, moorings) for 2006 to produce a MAB reanalysis.
In a true operational forecast system the state of the atmosphere and boundary forcing is not known in advance, so we are developing techniques to produce adjoint-based ensemble techniques that will allow us to place error bars to the forecast.
4040
SummaryAssimilation of gridded AVISO SLA is not appropriate in MAB
because of length/time scales of variability and anisotropy
Assimilation of along-track SSH successful but requires consideration of …– tidal signal in data-model (lest it dominate cost function)– time filtering IS4DVAR increment to reduce inertial oscillations – time correlation of SLA obs to suppress waves in adjoint solution
Subsurface projection (in addition to Adjoint) is in development using multi-variate background covariance– Less straightforward than in EAC …
– Shelf/Slope Front and shelf mean circulation reach seafloor so geostrophic balance must acknowledge the bathymetry
ROMS User MeetingGrenoble, 6-8 October 2008
[email protected] http://marine.rutgers.edu