Figure 5 shows results from the observing system sensitivity experiments. Note that it is not practical to extend these tests prior to 1986 because of lack of TAO data which does weaken the significance of the results. Perhaps the most striking aspect is the apparent lack of strong sensitivity. Neither the TAO-alone nor the VOS XBT-alone experiments have anomaly correlations that stand apart from that obtained ingesting all data. This point notwithstanding, the finer detail of the results do suggest that at short lead times (less than 6 months) it is the TAO data which is slightly more effective for initialisation and that in the range 6-12 months it is the VOS network which is more important. While the shortness of the hindcast record and the size of the impact make it impossible to conclude anything with confidence, these results are at least consistent with the results of the previous section. TAO resolves the equatorial regions within which reside the fast modes, and hence the short memory of the ocean-atmosphere system. VOS resolves the off-equatorial, slower moving modes which potentially provide the long memory and longer lead-time signals of the coupled system.
Figure 5 Anomaly correlation (top) and rms error (bottom) of the coupled model hindcasts for various subsuiface data sets, verified over the period 1986-94. FCST and PERS are the same as in Figure 4 while VOS and TAO refer to the VOS XBT and TAO-only data sets respectively (see text and Fig. 2).
Figure 6 shows results of the hindcast experiments using the 70% reduced data set compared with the ALL DATA experiment (FCST), persistence and the control run. While there is a consistent negative impact at all lead times due to the data base reduction in both the anomaly correlation and rms error scores, the interesting result is that it is relatively small compared to the impact of withdrawing all the subsurface data (that is, CNTL; see Fig. 4). Again, the shortness of the test period (14 years, or 56 realisations) and the limitations of the experimental design (a simple model and a rudimentary skill test) warn against over-interpretation of this result, but it is at least suggesting that perhaps a
Figure 6 Anomaly correlation (top) and rms error (bottom) of the coupled model hindcasts for the FCST, 70%" and PERS experiments verified over the period 1981-94.
