ENSO Predictions with Coupled Ocean Atmosphere Models
329
measurements may also improve ENSO predictions. SST data are not sufficient to
initialize an ENSO forecast system (Fischer and Latif 1995), since they contain no
information about the subsurface structure of the ocean. But the upper ocean heat
content anomalies impose a signature on the sea surface height anomaly, and these
data are nowadays available from satellite measurements almost continuously in
space and time. A method to extract the subsurface information from sea level
measurements has been presented in section 16.4.1. Below, three forecast ensembles are presented (Fig. 16.14): predictions initialized only with observed winds
(a), with observed winds and subsurface temperature assimilation (b), and initialized with observed winds and sea level assimilation (c) (Fischer et al. 1997). Each
forecast was initialized for at least one year, to ensure that the ocean model was
properly preconditioned. Although the SST variability in the control ensemble
(Fig. 16.14a) is underestimated by a factor of about two, the correlation between
predictions and observations is quite high. Even at lead times of one year it remains
above 0.7 (not shown here). However, the 1986/87 El Nifio was not predicted at all.
The coupled model simulated anomalous warm conditions only after the warm
event is fully developed in early 1987. If subsurface temperatures are assimilated
during the initialization, the situation improves considerably (Fig.16.14b) and the
amplitudes of the predicted SST anomalies are quite realistic representations. The
positive impact of subsurface temperature assimilation on ENSO prediction skill,
found also by other groups (e.g. Ji et al. 1996, Ji and Leetmaa 1997, Kleeman et al.
1995) is confirmed by these results. If sea level measurements are assimilated to
initialize the model comparable results as in the temperature case are obtained (Fig.
16.14c). Especially the transition from anomalous cold conditions in the mid eighties to the warm conditions which finally led to the strong El Nifio 1986/87 are well
represented. This shows that the subsurface information contained in sea level
measurements can be successfully inserted into an ocean primitive equation model,
and thus may improve the performance of ENSO prediction models.
At the GFDL experiments with three different assimilation schemes were performed (Rosati et al. 1996) with a coupled general circulation model. In two nudging experiments either SST (SDA2) or SST and surface winds (SDAl) were
nudged directly into the coupled model. In a third experiment a three dimensional
variational assimilation method was applied to insert temperature data into the
ocean model integrated in uncoupled mode (FDA). Forecasts in each experiment
started either in January or July ofthe years 1982 to 1989, which yields ensembles
of 7 members (Fig. 16.15). The correlation skill from these experiments was quite
different. In the case of mere SST nudging the skill is rather poor and only marginally better than persistence. Useful skill is only obtained up to 5 months. From this
experiment one may infer that SSTs alone are not sufficient to precondition an
ENSO forecast system. The subsurface structure is not properly initialized in this
case. In the SDAI experiment (nudging of SSTs and surf ace winds) the results are
slightly better but not satisfying. This disappointing result, however, probably must
be attributed to relatively poor wind fields. In the FDA experiment the ocean model
was forced with the same wind fields that were nudged in the SDAI case but addi-
329
measurements may also improve ENSO predictions. SST data are not sufficient to
initialize an ENSO forecast system (Fischer and Latif 1995), since they contain no
information about the subsurface structure of the ocean. But the upper ocean heat
content anomalies impose a signature on the sea surface height anomaly, and these
data are nowadays available from satellite measurements almost continuously in
space and time. A method to extract the subsurface information from sea level
measurements has been presented in section 16.4.1. Below, three forecast ensembles are presented (Fig. 16.14): predictions initialized only with observed winds
(a), with observed winds and subsurface temperature assimilation (b), and initialized with observed winds and sea level assimilation (c) (Fischer et al. 1997). Each
forecast was initialized for at least one year, to ensure that the ocean model was
properly preconditioned. Although the SST variability in the control ensemble
(Fig. 16.14a) is underestimated by a factor of about two, the correlation between
predictions and observations is quite high. Even at lead times of one year it remains
above 0.7 (not shown here). However, the 1986/87 El Nifio was not predicted at all.
The coupled model simulated anomalous warm conditions only after the warm
event is fully developed in early 1987. If subsurface temperatures are assimilated
during the initialization, the situation improves considerably (Fig.16.14b) and the
amplitudes of the predicted SST anomalies are quite realistic representations. The
positive impact of subsurface temperature assimilation on ENSO prediction skill,
found also by other groups (e.g. Ji et al. 1996, Ji and Leetmaa 1997, Kleeman et al.
1995) is confirmed by these results. If sea level measurements are assimilated to
initialize the model comparable results as in the temperature case are obtained (Fig.
16.14c). Especially the transition from anomalous cold conditions in the mid eighties to the warm conditions which finally led to the strong El Nifio 1986/87 are well
represented. This shows that the subsurface information contained in sea level
measurements can be successfully inserted into an ocean primitive equation model,
and thus may improve the performance of ENSO prediction models.
At the GFDL experiments with three different assimilation schemes were performed (Rosati et al. 1996) with a coupled general circulation model. In two nudging experiments either SST (SDA2) or SST and surface winds (SDAl) were
nudged directly into the coupled model. In a third experiment a three dimensional
variational assimilation method was applied to insert temperature data into the
ocean model integrated in uncoupled mode (FDA). Forecasts in each experiment
started either in January or July ofthe years 1982 to 1989, which yields ensembles
of 7 members (Fig. 16.15). The correlation skill from these experiments was quite
different. In the case of mere SST nudging the skill is rather poor and only marginally better than persistence. Useful skill is only obtained up to 5 months. From this
experiment one may infer that SSTs alone are not sufficient to precondition an
ENSO forecast system. The subsurface structure is not properly initialized in this
case. In the SDAI experiment (nudging of SSTs and surf ace winds) the results are
slightly better but not satisfying. This disappointing result, however, probably must
be attributed to relatively poor wind fields. In the FDA experiment the ocean model
was forced with the same wind fields that were nudged in the SDAI case but addi-