DATA ASSIMILATION
329
(Stammer et al., 2002.) ECCO estimates are characterized by their physical
con
es of large-scale global ocean circulation
(73º
he recursive nature of the Kalman filter and RTS smoother is
ion.
However, the
com utational requirements of evaluating the state error covariance matrix P
mak
ori, 2002).
Th
a timeinvar
the computational
cost
el integration of the state error
cov
sistency (Section 3.2) owing to smoothing algorithms (RTS smoother
and adjoint method). The estimates are based on a state-of-the-art primitive
equation model (MITgcm; Marshall et al. 1997) and employ a diverse suite
of in situ and satellite remote sensing observations including temperature
and salinity profiles and sea level.
The ECCO estimates are available from its data server at
http://www.ecco-group.org/las. In particular, ECCO has established a near
real-time analysis producing estimat
S~73ºN) on a regular basis (http://ecco.jpl.nasa.gov/external).
The
model employed is of moderate resolution (1º telescoping to 1/3º within 10º
of the equator, 10m layers within 150m of the surface with a total of 46
vertical levels) with its parameters adjusted by a Green’s function estimation
(Menemenlis, et al., 2004.) The near real-time analysis is conducted by an
approximate Kalman filter and RTS smoother. Aspects of this near real-time
assimilation are reviewed below.
5.2 ECCO near real-time analysis system
T
particularly suitable for near real-time computat
p
e direct application of these methods impractical for most state-of-theart ocean circulation models. Therefore, various methods have been put
forth that approximate the derivation of P so as to make Kalman filtering
and RTS smoothing feasible.
In ECCO, three approximations are
concurrently employed:
I. Time-asymptotic approximation (Fukumori et al., 1993),
II. State reduction (Fukumori and Rizzoli, 1995),
III. Partitioning (Fukum
e time-asymptotic approximation evaluates and employs
iant representative limit of P, thereby eliminating
associated with the continued mod
ariance matrix. Evaluation of this asymptotic limit is simplified by
partitioning and state reduction where independent elements of P are
evaluated separately from one another (partition) and within each partition
only the dominant modes of the error are estimated (state reduction). A
reduced-state model is derived for each partition to evaluate the errors while
the original fully nonlinear unapproximated model is used to integrate the
state. The smaller dimensionality of each partitioned-reduced-state model
reduces the computational cost of evaluating P. Unlike global single-stage
state reductions, the partitioning permits retaining many of the estimation
329
(Stammer et al., 2002.) ECCO estimates are characterized by their physical
con
es of large-scale global ocean circulation
(73º
he recursive nature of the Kalman filter and RTS smoother is
ion.
However, the
com utational requirements of evaluating the state error covariance matrix P
mak
ori, 2002).
Th
a timeinvar
the computational
cost
el integration of the state error
cov
sistency (Section 3.2) owing to smoothing algorithms (RTS smoother
and adjoint method). The estimates are based on a state-of-the-art primitive
equation model (MITgcm; Marshall et al. 1997) and employ a diverse suite
of in situ and satellite remote sensing observations including temperature
and salinity profiles and sea level.
The ECCO estimates are available from its data server at
http://www.ecco-group.org/las. In particular, ECCO has established a near
real-time analysis producing estimat
S~73ºN) on a regular basis (http://ecco.jpl.nasa.gov/external).
The
model employed is of moderate resolution (1º telescoping to 1/3º within 10º
of the equator, 10m layers within 150m of the surface with a total of 46
vertical levels) with its parameters adjusted by a Green’s function estimation
(Menemenlis, et al., 2004.) The near real-time analysis is conducted by an
approximate Kalman filter and RTS smoother. Aspects of this near real-time
assimilation are reviewed below.
5.2 ECCO near real-time analysis system
T
particularly suitable for near real-time computat
p
e direct application of these methods impractical for most state-of-theart ocean circulation models. Therefore, various methods have been put
forth that approximate the derivation of P so as to make Kalman filtering
and RTS smoothing feasible.
In ECCO, three approximations are
concurrently employed:
I. Time-asymptotic approximation (Fukumori et al., 1993),
II. State reduction (Fukumori and Rizzoli, 1995),
III. Partitioning (Fukum
e time-asymptotic approximation evaluates and employs
iant representative limit of P, thereby eliminating
associated with the continued mod
ariance matrix. Evaluation of this asymptotic limit is simplified by
partitioning and state reduction where independent elements of P are
evaluated separately from one another (partition) and within each partition
only the dominant modes of the error are estimated (state reduction). A
reduced-state model is derived for each partition to evaluate the errors while
the original fully nonlinear unapproximated model is used to integrate the
state. The smaller dimensionality of each partitioned-reduced-state model
reduces the computational cost of evaluating P. Unlike global single-stage
state reductions, the partitioning permits retaining many of the estimation