306
PIERRE BRASSEUR
OPA system once the assimilation experiment had reached an asymptotic regime (figure 8). The largest forecast errors are found along the
Gulf Stream path between Cape Hatteras and 40
N, where the standard
deviation exceeds 20 cm in some places. Local maxima are also detected
in the North Atlantic Current extension and along the Azores Current
at 35
N. Using the MICOM system, Brankart et al. [2003] compared
the estimated innovation variance with the sum of the observation and
forecast error variance diagnosed by the filter in a zonal section across
the Gulf Stream region at 30
N. The comparison indicates a fairly good
agreement between the 2 estimates in most regions, reflecting a successful adaptive procedure.
Figure 8. Distribution of the 10-day forecast error diagnosed with the adaptive SEEK
filter implemented in the OPA model (reproduced from Testut [2000]).
Another interesting application of the adaptivity concept is the update of the initial structure of the error sub-spaces. Using a low-rank
KF with an idealized primitive-equation model of the Gulf Stream, the
assimilation error dynamics was investigated by Ballabrera et al. [2001].
They examined how the filter performances were aected by imperfect
specifications of the initial error statistics. They observed that the truncation error does not impact the control of the solution when at least one
PIERRE BRASSEUR
OPA system once the assimilation experiment had reached an asymptotic regime (figure 8). The largest forecast errors are found along the
Gulf Stream path between Cape Hatteras and 40
N, where the standard
deviation exceeds 20 cm in some places. Local maxima are also detected
in the North Atlantic Current extension and along the Azores Current
at 35
N. Using the MICOM system, Brankart et al. [2003] compared
the estimated innovation variance with the sum of the observation and
forecast error variance diagnosed by the filter in a zonal section across
the Gulf Stream region at 30
N. The comparison indicates a fairly good
agreement between the 2 estimates in most regions, reflecting a successful adaptive procedure.
Figure 8. Distribution of the 10-day forecast error diagnosed with the adaptive SEEK
filter implemented in the OPA model (reproduced from Testut [2000]).
Another interesting application of the adaptivity concept is the update of the initial structure of the error sub-spaces. Using a low-rank
KF with an idealized primitive-equation model of the Gulf Stream, the
assimilation error dynamics was investigated by Ballabrera et al. [2001].
They examined how the filter performances were aected by imperfect
specifications of the initial error statistics. They observed that the truncation error does not impact the control of the solution when at least one