At lead times of two to four months there is a strong correlation between differences in NlNO 3 forecasts and analyses differences in an area between 100N and 150N in the west Pacific. This correlation exceeds 0.6 over a large area at four months lead time. This area is outside the TOGA-TAO observational array and is therefore only sampled by XBTs. The poor data distribution in this region will allow analyses using different winds to significantly differ. This is indeed the case; for some dates, FSU forced analyses and ERA/Ops forced analyses differ in the 20℃ isotherm depth by over 20 m (not shown). Figure 4, of course, only illustrates a correlation and does not necessarily mean that differences in this region have an impact on the NlNO 3 coupled forecasts. This needs to be investigated further.
At five and six months lead time there is very little correlation between differences in the initial states and the NlNO 3 forecasts. This suggests that uncertainty in the analyses has little systematic impact on the forecasts beyond four months lead time.
SUMMARY
ENSO forecasting using a fully coupled model depends critically on the initial state of the system for each forecast. This paper examined the importance of ocean initial conditions on the forecasts. Sub-surface data assimilation had a clear positive impact on the NlNO 3 SST forecasts over the 1990s. This was especially true when ERA/Ops Winds were used to force the ocean model: with data assimilation both rms errors and anomaly correlations of the NlNO 3 forecasts were significantly better than persistence over all lead times considered.
Without data assimilation, there was a strong dependence of the forecasts on the wind forcing being used to produce the initial states. With the ERA/Ops wind stresses the NlNO 3 forecasts were much worse than those using FSU wind forcing in the assimilation phase and the rms errors were bigger than persistence. Using FSU wind forcing improved the forecast statistics of the control, making the rms errors comparable with persistence and the anomaly correlations comparable with the forecasts from the assimilation initial conditions. The mean statistics of the forecasts from the two sets of assimilations with different wind stresses did not show sensitivity which wind forcing was being used during the assimilation.
The onset of the 1997/98 El Nino was well captured by the forecasts. None of the forecasts, neither from assimilation nor ocean-only initial conditions, captured the rapid warming of the east Pacific that occurred around April/May 1997, however. Once the EI Nino had developed, forecasts from assimilation initial conditions predicted its subsequent evolution better than those which did not assimilate sub-surface oceanic data.
The sensitivity of the coupled model to differences in the initial states was investigated by correlating differences in the two sets of initial states produced using sub-surface ocean data assimilation but different wind forcing, with the differences in the NIlNO 3 SST forecasts that these initial states produced. There was strong local correlation in the differences in the NlNO 3 SST forecasts and differences in the initial states in the NlNO 3 region at short lead times, particularly at the depths of the thermocline. This correlation exceeded 0.6 at short lead times (1-2 months).
At longer lead times (3-4 months) there were non-local correlations exceeding 0.6 in the west Pacific between 10N and 15N, outside the TOGA-TAO region. Results so far only show a strong correlation between initial state differences in this region and the subsequent NINO 3 forecasts. A mechanism connecting differences in this area and the NlNO 3 region has not yet been established.
