324 Martin Fischer
coupled system can be reduced to an initial value problem ofthe upper ocean only.
The usual way to initialize an ENSO forecast model for a prediction is as follows
(Fig. 16.11a). The oceanic model is integrated in an uncoupled mode and forced
over a period oftypically one to two years with observed wind stresses, heat fluxes,
and fresh water fluxes. Latif and Graham (Latif and Graham 1992) showed that the
ocean acts during this phase as a complicated time-space filter integrating basically
the information contained in the wind stress fields. That means that the initial oceanic state does not only depend on the actual boundary conditions, but also depends
on the history of the wind field prior to the start of a forecast. Therefore it follows
that the quality of the wind stress fields, especially the low frequency component,
is crucial for a successful initialization of a forecast. Errors in the initial state are
mainly due to model errors or to errors in the wind forcing, since it is not an easy
task to obtain high quality wind observations over the open ocean regularly in
space in time.
Ocean data assimilation can be an additional source of information during the
initialization process in order to compensate for model errors and uncertainties or
errors in the wind forcing fields (Fig. 16.11b). Since the upper ocean heat content,
down to a depth of about 300 meters is crucial, useful observations must contain
informations about the vertical structure of the upper ocean. Such data are for
example subsurface temperature measurements or velocity measurements. But the
vertical density structure can also be reconstructed from sea surface height data
(Mellor and Ezer 1991, Cooper and Haines 1996, Fischer et al. 1997). In section
16.4.2 forecast experiments from different groups with and without ocean data
assimilation are presented and compared to each other. In the following subsection
special emphasis is put on the assimilation of sea level measurements, since it is
not as straight forward as the assimilation of subsurface temperatures for instance.
Common to both methods is that the forecast system is initialized in uncoupled
mode. One problem of this approach is the so called coupling shock that occurs
immediately after coupling an degrades the correlation skill over the first 6 to 8
weeks significant1y. At the beginning ofthe coupling the two models are still not in
balance relative to each other, and the models need some time to adjust to each
other. A way to avoid this coupling shock is to initialize the coupled system
directly. In this case atmospheric and/or oceanic observations are assimilated during the initialization period of one to two years into the coupled model, in order to
constrain it (Fig. 16.11 c). This method is relatively new but first results with the
intermediate Cane/Zebiak model (Chen et al 1997) are very encouraging.
Assimilation of sea level data into an ocean primitive equation model
To assimilate observations that have a prognostic equivalent in the model formulation, like for instance temperatures, is straight forward. From the actual model
solution and the observations a correction field is computed, and this correction is
added to the model solution. Then the integration of the model is continued from
the corrected model state. With sea surf ace height (SSH) measurements this is not
possible. Most of the ocean primitive equation models used today make the so
coupled system can be reduced to an initial value problem ofthe upper ocean only.
The usual way to initialize an ENSO forecast model for a prediction is as follows
(Fig. 16.11a). The oceanic model is integrated in an uncoupled mode and forced
over a period oftypically one to two years with observed wind stresses, heat fluxes,
and fresh water fluxes. Latif and Graham (Latif and Graham 1992) showed that the
ocean acts during this phase as a complicated time-space filter integrating basically
the information contained in the wind stress fields. That means that the initial oceanic state does not only depend on the actual boundary conditions, but also depends
on the history of the wind field prior to the start of a forecast. Therefore it follows
that the quality of the wind stress fields, especially the low frequency component,
is crucial for a successful initialization of a forecast. Errors in the initial state are
mainly due to model errors or to errors in the wind forcing, since it is not an easy
task to obtain high quality wind observations over the open ocean regularly in
space in time.
Ocean data assimilation can be an additional source of information during the
initialization process in order to compensate for model errors and uncertainties or
errors in the wind forcing fields (Fig. 16.11b). Since the upper ocean heat content,
down to a depth of about 300 meters is crucial, useful observations must contain
informations about the vertical structure of the upper ocean. Such data are for
example subsurface temperature measurements or velocity measurements. But the
vertical density structure can also be reconstructed from sea surface height data
(Mellor and Ezer 1991, Cooper and Haines 1996, Fischer et al. 1997). In section
16.4.2 forecast experiments from different groups with and without ocean data
assimilation are presented and compared to each other. In the following subsection
special emphasis is put on the assimilation of sea level measurements, since it is
not as straight forward as the assimilation of subsurface temperatures for instance.
Common to both methods is that the forecast system is initialized in uncoupled
mode. One problem of this approach is the so called coupling shock that occurs
immediately after coupling an degrades the correlation skill over the first 6 to 8
weeks significant1y. At the beginning ofthe coupling the two models are still not in
balance relative to each other, and the models need some time to adjust to each
other. A way to avoid this coupling shock is to initialize the coupled system
directly. In this case atmospheric and/or oceanic observations are assimilated during the initialization period of one to two years into the coupled model, in order to
constrain it (Fig. 16.11 c). This method is relatively new but first results with the
intermediate Cane/Zebiak model (Chen et al 1997) are very encouraging.
Assimilation of sea level data into an ocean primitive equation model
To assimilate observations that have a prognostic equivalent in the model formulation, like for instance temperatures, is straight forward. From the actual model
solution and the observations a correction field is computed, and this correction is
added to the model solution. Then the integration of the model is continued from
the corrected model state. With sea surf ace height (SSH) measurements this is not
possible. Most of the ocean primitive equation models used today make the so