Forecasts of Tropical Pacific SST Using a Simple Coupled Ocean-Atmosphere Dynamical Model

Since the middle to late 1980s, forecasts of the Nino 3 SST anomaly have been regularly made at Lamont-Doherty Earth Observatory of Columbia University using a simple coupled ocean-atmosphere dynamical model (Cane et al. 1986, Cane and Zebiak 1987, Zebiak and Cane 1987). This represented the beginning of a strong movement toward physical approaches to the diagnosis and prediction of climate and its short-term fluctuations. Here, we present a few details of this model's current forecasts of Niño 3 and the tropical Pacific basin. Forecasts using a new version of the model with improved initialization are now additionally available; these will be highlighted following a discussion of the forecasts of the standard version of the model.

Figure 1 shows forecasts of the SST anomaly in Niño 3 for 3 to 15 months lead using data through January 1996, and the observed Niño 3 SST over the past 2 years. These forecasts are actually ensemble means of forecasts from six consecutive months ranging from August 1995 to January 1996. Forecasts are adjusted to have the same mean and standard deviation as the observed data for each calendar month and lead time. The vertical bars show the error standard deviations. These do not necessarily become larger with increasing lead time, because even with an unchanging expected fraction of SST variance explained by the forecasts, the natural interannual SST variability changes with season. (The vertical axis is in anomalous C as opposed to standardized anomaly in which case the size of the error bars would reflect only the skill of the forecasts.) In this case the forecasts describe a continuation of somewhat below normal Nino 3 conditions, with strongest negative anomalies for winter 1996-97.

A closer look at the forecast integrations is provided in Fig. 2, where six individual SST forecasts beginning from 1-month-apart initial conditions from August 1995 to January 1996 are shown along with the ensemble mean which is used in Fig. 1. The forecasts shown in Fig. 2 may not correspond exactly to those shown in Fig. 1, because in Fig. 2 the forecasts are adjusted to have the same mean and standard deviation as observed data only on an overall basis rather than for each calendar month and lead time. The spread among the individual ensemble members is small from the initialization times to boreal fall 1996, but becomes large from that point onward.

Figure 3 shows 6, 9 and 12 month lead SST anomaly forecasts for the tropical Pacific Basin, verifying in July and October of 1996 and January of 1997, respectively. These forecasts are adjusted to have the same mean and standard deviation as observed data on an overall basis (not for each calendar month and lead time), so they are exactly comparable with those shown in Fig. 2 but not necessarily Fig. 1. Like those for other target periods or for just Niño 3, the forecasts are adjusted for systematic biases. One such bias is an underestimation of amplitude of anomalies in the central (but not eastern) Pacific, which causes anomaly maxima to be placed too far east or prevents the central Pacific from fully participating. A statistical correction using singular value decomposition (SVD) is carried out to do the adjustment. The 6, 9 and 12 month lead forecast shown in Fig. 3 indicate a continuation of the present cool conditions without abatement through winter 1996-97.

Recent research at Lamont has shown that the skill of the SST forecasts can be increased significantly by improving the intialization system (Chen et al. 1995). The existing system has used wind stress anomalies (derived at Florida State University) to initialize the forecast runs, without current analyzed SST data. A newly developed system allows observed SST data to participate in the intialization process. Skill is found to increase not only in the early part of a forecast run but at intermediate and long leads as well. The "spring barrier" in skill that is present in the existing initialization scheme is substantially reduced using the improved system.

When the new initialization system is applied to the current SST forecast, the result is as shown in Fig. 4. The improved scheme produces a forecast generally similar to the customary Lamont forecast. This has not always been the case, however; e.g. several months ago the new scheme's forecast was for considerably colder conditions than the old. Now, both schemes favor a cold forecast.

Cane, M., S.E. Zebiak and S.C. Dolan, 1986: Experimental forecasts of El Nino. Nature, 321, 827-832.

Cane, M. and S.E. Zebiak, 1987: Prediction of El Nino events using a physical model. In Atmospheric and Oceanic Variability, H. Cattle, Ed., Royal Meteorological Society Press, 153-182.

Chen, D., S.E. Zebiak, A.J. Busalacchi and M.A. Cane, 1995:An improved procedure for El Nino forecast-ing: Implications for predictability. Science, 269, 1699-1702.

Zebiak, S.E. and M.A. Cane, 1987: A model El Nino-Southern Oscillation. Mon. Wea. Rev., 115, 2262- 2278.

Figures

Figure 1. Forecasts for the SST anomaly (oC) in the Ni¤o 3 region (5oN-5oS, 90o-150oW), based on the simple coupled model of Cane and Zebiak. Filled squares at the midpoints of the vertical forecast boxes represent the predictions, and the vertical boxes (lines) show the one (two) error standard deviations. The solid line represents the observed three month mean SST anomaly in Ni¤o 3 up to the most recently available data. The bars show forecasts for 1 month mean SST anomalies at leads of 3, 6, 9, 12 and 15 months. See text for additional detail.

Figure 2. Time series of forecasts of Ni¤o 3 SST from the Cane and Zebiak coupled model, for individual 1-month-apart initial conditions from August 1995 to Janaury 1996 (dashed lines) and the ensemble mean (solid line) used to form Fig. 1. However, an overall adjustment for the mean and standard deviation is used rather than lead- and season-specific adjustments as was done for Fig. 1. The thick solid line on left side shows the observed one month mean SST over the temporal range of the initial conditions.

Figure 3. Cane and Zebiak coupled model SST anomaly forecast fields for July and October 1996, and January 1997, made at 6, 9 and 12 month lead times, respectively. The forecasts are ensemble averages of 6 forecasts with 1-month-apart initial conditions ranging from August 1995 to January 1996. Adjustments for the mean and standard deviation are applied, based on lead time but independent of start time.

Figure 4. As in Fig. 3 except for the forecast made using the new initialization procedure (LDEO2) (Chen et al. 1995).