Hybrid coupled models (HCMs), used in several ENSO simulations and predictions, consist of an ocean GCM coupled to a simplified (usually steady-state) atmospheric model. This reasonably approximates the coupled system in which the ocean has the memory of the system and the atmosphere has a relatively shorter time scale. In our HCM, an empirical atmospheric model estimates the correlation between the wind and SST anomaly fields. This wind stress-SST contempor-aneous correlation is based on an assumption, for monthly or longer time scales, of the atmosphere's rapidly adjusted response to the SST pattern nonlocally throughout the basin.

The HCM is similar to the one documented in Syu et al. (1995) and used in Waliser et el. (1994) and Blanke et al. (1997). The empirical atmospheric model is estimated from observations using a singular value decomposition (SVD) technique. The model contains the first seven SVD modes of the covariance matrix calculated from the time series of pairs of observed monthly mean Reynolds SST and FSU pseudo-stress fields for the 19-year period of 1970-88. Atmospheric spin-up time, which was neglected in the previous version, is parameterized, albeit crudely, in the current version within the coupling procedures. A 60-day spin-up time scale is chosen for all ENSO simulations and predictions. Heat flux is parameterized according to Oberhuber's (1988) formulation using climatological data, with the negative feedback on SST estimated following Seager et al. (1988). The OGCM is a version of the GFDL Modular Ocean Model (Pacanowski, Dixon and Rosati, 1991, personal communication) for the Pacific basin. The vertical resolution is 27 levels, with 10 levels in the upper 100 meters. A Richardson-number-dependent vertical mixing scheme is combined with a surface mixed layer parameterization, as in Latif et al. (1994).

The HCM reasonably simulates ENSO spatially and temporally, with ENSO periods of 3 to 4 years. Model performance in "retroactive real-time forecasts" (hindcasts hereafter) is shown in Figs. 1 and 2 from initial times within 1980-1992. The ocean climatology used in the hindcast experiments is specified as the averaged model SST, forced by FSU wind stress over 1978-93 without modification by the data assimilation scheme. The climatological wind stress used in the hindcast experiment is also specified as the average of the FSU wind stress over the same period. The hindcast /forecast results are verified against the observations from Reynolds' (1988) SST data set.

Since the SVD atmospheric model is estimated from 1970-88 observed data, partially overlapping with the hindcast period, a set of the SVD models derived from non-overlapping times is used. For example, when the years 1981-83 are being hindcasted, the SVD model is obtained only from data of 1970-80 and 1984-88, excluding the hindcast period.

Two initialization schemes are used here. The first is the widely used Cane et al. (1986) scheme, in which the coupled model is initialized by the model ocean state forced by the history of observed wind stress. We refer to this as the stress-only scheme. Florida Sate University (FSU)-converted wind stress for 1976-93 with the drag coefficient of 1.2 x 10^-3, is used to integrate the ocean model up to the start of the hindcast.

The second initialization scheme makes use of the ocean model data assimilation product from the Climate Prediction Center (CPC) (Ji et al. 1995). In addition to the specified FSU wind forcing, the CPC reanalyzed anomalous ocean temperature field is "injected" into the ocean model (27 layers) every month from 1980 to the start of the hindcast. The scheme is hereafter referred to as the injection scheme. Because our ocean model (the GFDL MOM) is in a version reasonably close to that used by CPC, approximate consistency is assumed in injecting the CPC reanalyzed data. To make a distinc-tion between this procedure and raw-data injection, we refer to it as a "piggy-back" data assimilation scheme, because it uses the CPC data assimilation product.

The initial SST anomaly fields for the stress-only scheme and the injection scheme show a month-to-month (i.e., not 3-month running mean) correlation with the observations of 0.71 for the stress-only scheme and 0.92 for the injection scheme.

Hindcast correlation skills (Fig. 1) are computed for Niño 3 (5S-5N, 150W-90W) SST as a function of lead time, defined as the model time span from the beginning of the hindcast integration. The correlations and root mean square (RMS) errors for 1980-92 use an ensemble of 156 for the stress-only initialization and 155 for injection initialization. Three-month running means are used; the mean of each set of data for the same lead month is removed before calculating the correlations and RMS errors for all hindcasts. The same procedures are also applied to the observed SST anomalies. The results for the stress-only initialization (dotted line) and the injection scheme (thin solid line) are shown in Fig. 1 as a function of lead month.

With the stress-only initialization, the correlation for short lead times is similar to that of persistence (thick solid line), but starts to beat persistence after a lead time of 3 months. Starting from 0.75, it remains greater than 0.5 to a lead time of 8 months, with the RMS errors remaining lower than 1C over the entire hindcast period. With the injection initialization scheme, skill is improved. Starting at 0.93, it remains greater than 0.5 up to a lead time of 17 months. The predicted (gray solid line) and observed (solid line) Niño 3 indices are displayed in Fig. 2 as a function of year, at lead times of 3, 6 and 9 months. Each vertical bar indicates plus and minus one RMS error with respect to the prediction. In general, the predicted SSTA captures the variation of observed SSTA up to month 9. However, the magnitude of the 1982-83 El Niño is not well predicted for leads longer than 9 months.

The hindcast skill is competitive with that of other prediction results, as noted in the review paper by Latif et al. (1997). We note that no "MOS" (Model Output Statistics) correction, which is constructed statistically based on the difference (error) between model fields and observations, is used in this forecast scheme. This differs, for instance, from Barnett et al. (1993), in which such correction is used within the coupling (e.g., getting MOS-corrected SST from model SST before coupling to the atmosphere). The "piggy-back" data assimilation scheme gives a substantial improvement in hindcast skill, and thus appears to be a viable and economical component of our forecast method.

Figures 3 and 4 present the Niño 3 index for forecasts from 1992 to present, using the injection scheme. Observations through July 1997 are used. Figure 3 shows Niño 3 observed SSTAs (3-month running average; thick black curves) and forecasts (gray curves) at 3-, 6- and 9-month lead. Averages of each lead month based on forecast verification over the 1980-1992 time span are removed before plotting the curves. Vertical bars represent plus and minus one RMS error about the forecast. Forecasts at 3-, 6- and 9-month leads predict a warming for fall/winter 1997-98. The rapidity of the warming over recent months was not predicted.

Figure 4 shows the latest two forecast results (starts from June and July, 1997, respectively, for 12 months), with the mean over the forecast verification time span removed as in Fig. 3. The observations and model control run since 1992 are also displayed. Both forecasts predict an ongoing El Niño event with a peak in Sep-Oct 1997 and persistence of a decaying warming through spring 1998.

Barnett, T.P., M. Latif, N. Graham, M. Flugel, S. Pazan, and W. White, 1993: ENSO and ENSO-related predictability. Part I: Prediction of equatorial Pacific sea surface temperature with a hybrid coupled ocean-atmosphere model. J. Climate, 6, 1545-1566.

Blanke, B. J.D. Neelin and D. Gutzler, 1997: Estimating the effect of stochastic wind stress forcing on ENSO irregularity. J. Climate 10, 1473-1486.

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

Ji, M., A. Leetmaa and J. Derber, 1995: An ocean analysis system for seasonal to interannual climate studies. Mon. Wea. Rev., 123, 460-481.

Latif, M., T. Stockdale, J. Wolff, G. Burgers, E. Maier-Reimer, M.M. Junge, K. Arpe and L. Bengtsson, 1994: Climatology and variability in the ECHO coupled GCM. Tellus, 46A, 351-366.

Latif, M., D. Anderson, T. Barnett, M. Cane, R. Kleeman, A. Leetmaa, J. O'Brien, A. Rosati, and E. Schneider, 1997: TOGA review paper. "Predictability and prediction." J. Geophys. Res., in press.

Oberhuber, J.M., 1988: An atlas based on the COADS data set: The budgets of heat buoyancy and turbulent kinetic energy at the surface of the global ocean. Max-Planck-Institut für Meteorologie, Report No. 15, Bundesstrasse 55, D-2000, Hamburg 13, FRG.

Reynolds, R. W., 1988: A real-time global sea surface temperature analysis. J. Climate, 1, 75-86.

Seager, R., S.E. Zebiak and M.A. Cane, 1988: A model of the tropical Pacific sea surface temperature climatology. J. Geophys. Res., 93, 1265-1280.

Syu H.-H., J. D. Neelin, and D. Gutzler, 1995: Seeasonal and interannual variability in a hybrid coupled GCM. J. Climate, 8, 2121-2143.

Waliser, D.E., B. Blanke, J.D. Neelin and C. Gautier, 1994: Shortwave feedbacks and El Niño-Southern Oscillation: Forced ocean and coupled ocean-atmosphere experiments. J. Geophys, Res., 99, 25109-25125.

Fig. 1. (a) Correlation skills and (b) RMS errors for the hindcast experiments from 1980 to 1992 with the stress-only scheme (dotted lines) and injection scheme (thin solid lines). A three-month running mean is applied and the mean for each lead month is removed in both the observation and model results before calculation.

Figures 2 and 3: Fig. 2 (left 3 panels): Hindcasts of the Niño 3 SST anomalies, from 1980 to 1992, using the injection scheme for (a) 3, (b) 6, and (c) 9 lead months. The vertical bars show the plus and minus one RMS error. The black solid line indicates the observed SST anomalies and the gray solid line indicates the predicted SST anomalies. Fig. 3 (right 3 panels): Forecasts of Niño 3 SST anomalies from 1992 to present. The solid line indicates observations. The latest forecast starts from July 1997. The mean for each lead month over the forecast verification time span (1980-92) is removed before plotting. Vertical bars represent plus and minus one RMS error, as in Fig. 2. Shown for (a) 3-month, (b) 6-month and (c) 9-month lead.

Fig. 4. Forecasts (dotted lines) of Niño 3 SST anomalies up to 12 lead months starting from the latest two initialization times--June and July 1997. Observations (solid line) and model control run (dashed line) from 1992 are also shown. The mean for each lead month is removed as in Fig. 3. Vertical bars indicate plus and minus one RMS error, as in Fig. 2.