Figure 1 shows the total number of observations in the data base for the period 1970-present (see Smith [1995] and Smith and Meyers [1996] for more details on the data base). The peak around 1987 in the Pacific Ocean was due to efforts in the TOGA program; the peak around 1990 is due to recent World Ocean Circulation Experiment/Global Temperature-Salinity (Pilot) Project initiatives to gather delayed-mode data; the steady increase from the late 1980's reflects the gradual implementation of the TAO array. Case (5) above is based roughly on the sampling rate changes shown in Figure 1 which suggest that during the 1970's sampling was around 20-30% of that during TOGA. The 70% reduction was achieved by randomly eliminating 70% of the stations and/or lines, each defined by individual identifiers (for example, a ship call sign) attached to the data. So around 70% of the TAO mooring locations and 70% of the VOS lines were withdrawn.
Figure 1 (a) The number of original single observations available for the analysis of the 400m depth-averaged temperature for the Pacific (thick) and Indian (thin) Oceans in the BMRC data base for each month from January 1970 through to December 1994. (b) As (a), but the number of effective observations. This Count is obtained by estimating the number of independent units of subsurface temperature information (see Smith and Meyers [1996]).
3.The analysis system and prediction model
The analysis system has been presented in Smith et al. [1991] and Smith [1995] and will not be discussed in any detail here. It is a univariate, level-by-level temperature (or isotherm depth or heat content) statistical interpolation system, using a statistical forecast combined with climatology as the initial forecast. For analyses of the depth of the 200C isotherm the analysis period is 10 days. For the heat content analyses (400m depth-averaged temperature) the analysis period is 1 month. The analysis method fits a crude statistical model to the unresolved modes of variability, the temporal scale being set at 5 days, an assumption recently verified by Kessler et al. [1996]. This assumed model, plus the defined analysis period and the assumed spatial and temporal scales for the signal, in effect jointly determine the level of redundancy in the data set(s).
The coupled model with data assimilation was introduced by Kleeman et al. [1995]. Both the atmospheric and oceanic models retain only one vertical model and the ocean model expresses its meridional dynamic structure in terms of a few meridional modes (in effect the transients are represented by a Kelvin wave and the first 6 equatorial Rossby waves). The SST equation includes a local feedback term plus a remote thermocline forcing term. The assimilation of the heat content analysis (which is used as a proxy for the model thermocline variable) is performed using the adjoint of the ocean
