model, with a cost function that penalises departure of the model thermocline anomaly away from its observed proxy, weighted by the inverse of the estimated rms error of the proxy analysis. The Florida State University wind analysis [Legler and O'Brien 1984] is used for surface wind forcing.
The coupled model provides, among other things, forecasts of the NINO3 index out to lead times of 2 years from the initialisation time. Kleeman et al. [1995] showed that subsurface data had a significant positive impact on model skill for the hindcast period 1982-9 1. This model, as well as the subsurface analysis system, form part of the operational climate monitoring and prediction system at the Bureau of Meteorology. For the wind-only experiments, here extended to the period 1971-1995, the full history of the winds (from 1961) is used to initialise the model whereas in the subsurface data assimilation cases only the wind data concurrent with the assimilation period are used. Note also that the weighting in the cost function, which is derived from the rms error estimate of the heat content analysis, is critical for this model and in effect represents for the model estimate of the amount of subsurface information available. High weightings (large cost function coefficients) correspond to dense sampling and high information content, low weights to sparse sampling and low information content.
4.Results from analysis-alone experiments
Smith and Meyers [1996] conducted a series of analysis experiments for the depth of the 200C isotherm using different data inputs, as described in (1) - (3) of Section 3. They found that, for the period 1990-94, almost all the variability at the equator for periods longer than 10 days could be captured by the TAO data alone. The VOS XBT data added some detail in between TAO buoy array locations but in general the overwhelming majority of the information in the combined data set could be gleaned without recourse to the XBT network. On the other hand, in the absence of TAO data, the VOS XBT data did capture much of the significant low frequency variability and, occasionally, some of the higher-frequency variations. Figure 2 shows the evolution of the 400m depth-averaged temperature at the equator for the time periods and data sets used in the OSEs discussed in the next section. For the period 1990-95 similar results to that described in Smith and Meyers [1996] emerge. Note, however, the occasional disparity, for example west of 1500E in 1994, which might be due to bad VOS data or perhaps to vertical interpolation errors in the TAO data (the meters are 5Om apart at the level of the 200C isotherm). For the period 1985-89 the lack of TAO array sites limits their ability to follow changes in the equatorial wave-guide though, with just two buoys, the interannual east-west oscillation can still be seen.
In the period 1970-1982 even the "ALL DATA" analyses struggle to resolve the amplitude and phase of subsurface variability. However, some of the major interannual events can still be discerned, such as the 1972 ENSO. The intermittent nature of the VOS (and research cruise) XBT tracks means the true amplitude and temporal extent of low frequency variability can never be adequately determined over this period. It is interesting to compare this sampling outcome with that of the sparse TAO data period in the 1980's. TAO is sparse in space but frequent in time so, at least near the location of the moorings,
