The difference between the two controls (C_ERA and C_FSU) is that they used different surface wind forcing fields. C_ERA used the surface wind stress fields from ERA for 1990-1993 and from ECMWF operations from 1994 onwards. C_FSU used wind stresses calculated from the FSU (Florida State University) winds, using a constant drag coefficient such that the mean of the zonal stress over the equatorial strip from 1979 to1993 matched that from ERA. The FSU winds are only available as monthly means and within 30 degrees of the equator. The high frequency variability (less than one month) from ERA/Ops was added to the FSU stress. ERA/Ops stresses were used outside the area covered by the FSU stresses, with smoothing across the boundary. This was also the difference between the two assimilation experiments A_ERA and A_FSU (see table 1).
COUPLED MODEL HINDCASTS
The analyses described above were used to initialise coupled model forecasts out to six months. Initial conditions for the atmospheric part of the coupled model were taken from ECMWF analyses (ERA from 1990-93 and operations from 1994 onwards). Coupled forecasts were started on the 1st of January, April, July and October of 1991-1997. For each start date and for each analysis described above an ensemble of five forecasts was created. The ensemble members differed only in that they had perturbations made to the EQ1 (130W to 90W and 5S to 5N), EQ2 (170W to 130W and 5S to 5N) and EQ3 (150E to 170W and 5S to 5N) SST in the initial state of magnitude ±0.01℃. The aim of these perturbations was to generate weather noise due to the chaotic evolution of the atmospheric systems. The experiment names used for the analyses and described in table 1 will also be used to represent the coupled forecasts starting from the respective analyses.
The coupled model was initialised from the oceanic and atmospheric assimilation systems (with and without sub-surface ocean data), where the initial states represent our best estimate of reality. Because of deficiencies in the oceanic and atmospheric models the state of the coupled model drifted with forecasts time. No attempt was made to reduce this during the coupled model integrations with, for example, flux correction. Instead the drift was removed from the coupled model results a postiori, following Stockdale (1997). For a given forecast, the drift in the system was calculated as the mean error in all the other forecasts, starting at the same month but for the years 1991-1996, excluding the actual forecast itself and its ensemble members. Coupled model forecasts for 1997 were not used to calculate the mean coupled model drift as the large anomalies associated with the 1997/98 El Nino may have made the coupled model behave differently to usual.
The 4-6 month mean forecasts anomalies for the coupled forecasts initialised using the ERA/Ops surface forcing without and with ocean sub-surface data assimilation are shown in fig. 1a for individual ensemble members and in fig. 1b is shown the ensemble mean. Since the ensemble members differ only in the initial state by a very small amount, differences in the evolution of the forecast arise due to the chaotic nature of the atmosphere. At this lead time the spread in the five members due to the atmospheric noise exceeds 1℃ for many of the start dates. This large spread means that, in general, an ensemble of forecasts is necessary to average out the atmospheric chaotic variability. The exact number necessary depends upon the question to be answered.
For many of the cases shown in fig. 1a the forecasts initialised with ocean data assimilation are better than those from the control without data assimilation. Clear examples are forecasts starting on: 1 Jul'92, 1 Apr'93, 1 Jul '93, 1 Apr'94, 1 Oct'95, 1 Apr'96 and 1 Oct'97. The only exception where the control is clearly better is 1 Jan'93. The benefit of the ocean data assimilation is also clear from looking at the ensemble mean in fig. 1b. In many cases the ensemble mean from initial conditions with data assimilation is closer to the observed anomaly than those from the control initial conditions. Furthermore, some of the cases initiated from the control have a large error, for example, -1.50℃ for the forecast starting on 1 Apr'94, which is reduced to about 0.5℃ with data assimilation initial conditions.
