We have reached a very important juncture in the evolution of oceanography and, in particular, in the development of seasonal-to-interannual climate prediction systems and related research. The Tropical Ocean Global Atmosphere (TOGA) Experiment has been widely acclaimed for developing observing systems and modelling and data assimilation methods to the point where, for the Pacific Ocean and the El Nino-Southern Oscillation (ENSO) in particular, the potential for useful climate predictions has been realised [McPhaden et al. 1996; Latzf et al. 1996]. However, as was found several decades ago in meteorology, such a demonstration of potential forecast skill must be followed by a period of consolidation (see, for example, Bengtsson et al. [1981]). Usually the forecast model initialisation methods developed to demonstrate potential for predictive skill have the benefit of abundant data (over-sampling), a situation that is not sustalnable for the long-term. The demonstration of feasibility must be followed by research and development to identify the key elements (observations and model components) required for monitoring and prediction and establishing priority.
In this paper we wish to examine one aspect of the systems developed for ENSO monitoring and prediction, namely the dependence on the subsurface in situ thermal observing network which, for all intents and purposes, is represented by the Tropical Atmosphere-Ocean (TAO) buoy array and the Volunteer Observing System (VOS) expendable bathythermograph (XBT) network.
The methodology we wish to use, known as "observing system experimentation" (OSEs; see, for example, Smith [1993]), became popular in meteorology during the First GARP (Global Atmospheric Research Project) Global Experiment where it was used, among other things, to evaluate the impact of satellite radiance data and Southern Hemisphere drifting surface pressure buoys (see, for example, Bourke et al. [1985]). However the methodology is not without its problems in the present context. The results are more often than not very model dependent, more so if the models have simplified or inadequate physics and/or systematic biases (more often than not unquantified), all of which are true to some extent for current ENSO prediction models [Stockdale et al. 1993]. This situation is exasperated here because we only have a very limited number of ENSO events in our hindcast/forecast sample against which to test out model, so any conclusions must be weighed against the knowledge that the sample may not be representative.
For ENSO monitoring and prediction our observing system ocean temperature data are also not an exclusive source for such information. It has been known for some time (e.g., Busalacchi and O'Brien [1981]) that knowledge of the surface wind provides a proxy for subsurface heat storage and for dynamic height (usually with the aid of a suitable dynamic model). Sea level measurements (for example, from an altimeter) also give an indirect measure of the upper ocean heat content/dynamic height [Picaut et al. 1995], while sea surface temperature (SST) measurements have long provided primary knowledge of the state of ENSO [McPhaden et al. 1996]. This paper either ignores these sources of information altogether, when examining subsurface ocean analyses, or assumes they are "given" and without error in the case of model predictions.
