A study of initializations for ENSO forecast models

This presentation addresses two questions: How well is sea level initialized in ENSO forecast models? How is the quality of sea level initializations related to forecast skill? The sea level fields from initializations using different ocean models and initialization methods for 1980-93 are analyzed. It is found that the sea level fields from initializations using ocean data assimilation systems have the greatest accuracy. The sea level fields from initializations using OGCMs (Oceanic General Circulation Models) forced with observed winds correlate with tide gauge observations much better than those using simple ocean models. Sea level fields from initializations using ocean models forced with either model winds or combinations of model winds and observed winds generally correlate well with tide gauge observations in the equatorial belt (5OS-5ON), but not in the subtropics.

Empirical orthogonal function (EOF) analyses show that the first two EOFs account for more than 60% of the variance of sea level anomalies. The first EOF describes the mature phase of ENSO and the second EOF has a north-south dipole structure with a maximum on the equator in the central Pacific and a minimum off the equator in the northwestern Pacific. The second EOF leads the first EOF by about eight months, suggesting that the second EOF is a possible precursor for ENSO.

It is well known that the interannual variability in the 1980s was dominated by a low frequency oscillating mode, and that this mode can be described by the so-called ENSO POP (Principle Oscillation Pattern). We tested a hypothesis that the skill of ENSO forecast models is related to good representations of the ENSO POP in initializations. Compared with the standard Lamont model, the newer Lamont model LDEO2, which is forced with combinationsof model winds and observed winds, had improved forecast skill in the 1980s. This is becausein the newer model the ENSO POP is better represented and the noise in initialization is significantly reduced. The ENSO POP is also well represented in the initializations for the hybrid coupled model at Scripps and for the anomaly coupled model at the Centers for Ocean-Atmosphere-Land Studies. We propose that the moderate prediction skill of the ENSO forecast models is due to the fact that there is a dominant ENSO POP in the 1980s and this ENSO POP is well initialized in those models.

The ENSO POP becomes less dominant during 1993-96 when most of the ENSO forecast models perform poorly. However, since 1997 this ENSO POP becomes dominant again, and most of the ENSO models predicted the 1997-98 El Nino and 1998-99 La Nina events at one year advance.