By continuing to browse this site you agree to us using cookies as described in About Cookies
Notice: Wiley Online Library will be unavailable on Saturday 7th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 08.00 EDT / 13.00 BST / 17:30 IST / 20.00 SGT and Sunday 8th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 06.00 EDT / 11.00 BST / 15:30 IST / 18.00 SGT for essential maintenance. Apologies for the inconvenience.
 Previous analyses by Levitus et al.  (“Levitus”) of ocean temperature data have shown that ocean heat content has increased over the last fifty years with substantial temporal variability superimposed. The HadCM3 coupled atmosphere–ocean general circulation model (AOGCM) simulates the Levitus trend if both natural and anthropogenic forcings are included. In the relatively well-observed northern hemisphere upper ocean, HadCM3 has similar temporal variability to Levitus but, like other AOGCMs, it has generally less variability than Levitus for the world ocean. We analyse the causes of this discrepancy, which could result from deficiencies in either the model or the observational dataset. A substantial contribution to the Levitus variability comes from a strong maximum around 500 m depth, absent in HadCM3. We demonstrate a possibly large sensitivity to the method of filling in the observational dataset outside the well-observed region, and advocate caution in using it to assess AOGCM heat content changes.
 We find the same for the HadCM3 AOGCM [Gordon et al., 2000]. Figure 1 shows a comparison of five-year running means from four different HadCM3 ensembles (each consisting of four integrations) and the Levitus timeseries, which has a trend of 0.32 ± 0.04 × 1022 J yr−1 over the period (the uncertainty is one standard error calculated from the deviations from the fit). The ensemble mean for GHG (greenhouse gases only) shows a much larger trend of 0.62 × 1022 J yr−1, while NATURAL (solar and volcanic forcing only) has a slight decreasing trend of −0.07 × 1022 J yr−1. By contrast, the trends of both ANTHRO (anthropogenic forcing by greenhouse gases and sulphate aerosols) and ALL (anthropogenic and natural forcing), being 0.23 and 0.16 × 1022 J yr−1 respectively, are more similar to the Levitus trend. This improved agreement with observations would be consistent with findings for surface air temperature that a substantial portion of greenhouse warming in recent decades has been offset by cooling due to sulphate aerosols, and that natural agents alone cannot explain the recent warming [Stott et al., 2000].
 On the other hand, while the trends in the model are clear, the Levitus trend is relatively uncertain because its temporal variability is evidently much larger. The interannual RMS deviation from a linear trend to the five-year running means for world ocean heat content above 3000 m for 1957–1994 from the Levitus results is 5.35 × 1022 J. After removing long-term drifts by a cubic fit, the interannual standard deviation of 1240 years of five-year running means from the HadCM3 AOGCM is 1.93 × 1022 J. An F-test on the variances shows that they differ significantly at the 1% level. It is important to explain the discrepancy because variability of the size exhibited by the Levitus timeseries would have a substantial influence on the heat balance of the climate system, possibly affecting the conclusions of climate change detection and attribution studies. For example, the Levitus timeseries shows an increase of 7.5 × 1022 J from 1988 to 1994, corresponding to a heat flux averaged over the world of 0.7 W m−2, about half the size of the present-day radiative forcing due to carbon dioxide with respect to the pre-industrial climate. If such variability occurs it is important that climate models should simulate it.
2. Effect of Sampling on Observational Estimates
 Sampling in the World Ocean Database [World Ocean Database 1998, 1998, hereafter referred to as WOD], on which the Levitus timeseries is based, can be very sparse. Large data gaps are filled in to create almost globally complete data using interpolation methods, assigning a finite non-zero “radius of influence” to the sampled points. One might question whether any aspects of the Levitus timeseries could result from poor sampling or the processing method.
 For each month from 1941 to 1996 we regridded the WOD observations onto the ocean grid of HadCM3, which is uniform with boxes of 1.25° × 1.25° in the horizontal and 20 vertical levels ranging from 10 m to 615 m thickness. For each gridbox, we produced a monthly climatology by averaging all observations from each month over the entire period. We calculated anomalies for individual months with respect to the climatology, and annual anomalies by averaging the monthly anomalies in each year.
 In the upper 360 m of the northern hemisphere (0–65°N) after 1970 data coverage on the HadCM3 grid reaches 70% of the area, but it drops off below this depth. In the southern hemisphere (0–65°S) coverage never exceeds 40% even in the upper 360 m. Below 1200 m coverage is less than 20% in the northern hemisphere and 10% in the southern.
 To produce estimates of global ocean heat content we have tried two extreme alternatives, considering each layer individually:
 • Representative averages: assume that the average temperature anomaly of sampled points in the layer is representative of the average over the whole volume of the layer.
 • Zero anomalies: where temperature is not sampled assume that the temperature anomaly is zero and hence does not contribute to the heat content anomaly of the layer.
 The difference between results from these two methods gives an estimate of the potential uncertainty in estimates of ocean heat content. Where data coverage is complete the two estimates should be exactly equal.
Figure 2 (upper panel) shows the ocean heat content anomaly calculated from WOD over the upper 360 m of the northern hemisphere 0–65°N. Before 1970, Levitus and our estimates disagree about the sign of the anomaly and hence the sign of the trend over the full period. After 1970 our estimates are almost identical and in good agreement with Levitus because of adequate data coverage. This means that the observational dataset can be used to test the realism of the model for the upper 360 m of the northern hemisphere since 1970. For this subset of the datasets, the variability in HadCM3 is comparable with Levitus; for instance one member of the ALL ensemble by chance matches the Levitus timeseries particularly well (Figure 3). This gives us confidence in the upper-ocean variability of HadCM3. However, for this limited portion of the ocean, the trend is small compared with internally generated variability, so detection and attribution of climate change would be difficult.
 Extending the results down to 3000 m and to include the southern hemisphere 0–65°S (Figure 2, lower panel) we find much larger differences, with substantial disagreement even after 1970, arising from poor data coverage. The difference between maximum and minimum values from Levitus for the upper 3000 m of the global ocean is around 20 × 1022 J; it is about 50 × 1022 J for the method of representative averages and 6 × 1022 J for the zero anomalies.
 Repeating the calculations using a grid with 5° boxes instead of the 1.25° boxes of HadCM3 does not change the results from the representative averages but the result for zero anomalies lies between our timeseries in Figure 2. For the upper 3000 m, the range in heat content becomes 14 × 1022 J, in better agreement with Levitus. The dependence on resolution suggests there will be some sensitivity in any method of objective analysis to the radius of influence ascribed to the available measurements. This is because the procedure is not simply smoothing the observed data, but filling in the majority of the volume having few or no observations. Owing to the lack of data, local anomalies may have an undue influence; in a more complete dataset, there might be greater cancellation of anomalies.
3. Variability as a Function of Depth
 To investigate the source of the greater Levitus variability we computed the variability as a function of depth in the Levitus timeseries and HadCM3 control, both detrended, for the world ocean and in three latitude bands (south of 30°S, 30°S–30°N, and north of 30°N, Figure 4). For the Levitus anomalies our procedure was to fit a linear trend to the running means for heat content at each level for 1947–1994 (44 overlapping five-year periods), subtract it from the timeseries, and calculate the standard deviation of the residual.
 For the HadCM3 control run we calculated 1536 consecutive overlapping five-year running means of ocean temperature, divided them up into 32 non-overlapping portions of 48 years and treated each one individually like the Levitus dataset. The subtraction of the trend means that the signal of climate change in the Levitus dataset and the model drift in the control run are both removed. The use of a large number of portions of the control timeseries allows us to evaluate the uncertainty in the estimate of the model variability.
 For the HadCM3 data, we repeated the calculation with the observed data mask for the individual years 1949–1992 imposed on each set of running means; that is, we used a HadCM3 annual mean in a particular gridbox only if there was at least one observation for the corresponding year and location in the WOD. The layer heat content was obtained by the method of representative averages, which will tend to give the largest estimate of variability (see Figure 2).
 We have also computed five-year running-mean SST averages in the same latitude bands from the HadISST1 dataset [Rayner et al., 2003] for the same time-period. HadISST1 is a dataset compiled from ship-based SST measurements whose geographical coverage is completed by the use of EOFs constructed from a combination of in-situ and satellite data. HadCM3 has the same world surface variability as HadISST, whereas Levitus is 40% larger (Table 1). HadCM3 and HadISST agree in southern latitudes, the region with poorest data coverage, where Levitus has 80% greater variability. But in low latitudes HadCM3 has more surface variability than HadISST and Levitus, while in northern latitudes HadISST lies between HadCM3 and Levitus.
Table 1. Standard Deviation of Average Temperature (K) in the Regions Indicated for 44 Five-Year Running Means Covering 1947–1994
 In all regions, HadCM3 variability levels off near the surface, while in the Levitus dataset it increases towards the surface in the northern and southern region. Possibly the upper layers in HadCM3 are too strongly mixed.
 World ocean variability in Levitus is greater than in HadCM3 at most depths, but it is evident that the discrepancy mainly derives from depths between 200 and 600 m, where Levitus has more than twice the variability of HadCM3. In Levitus this range shows a peak in variability as a function of depth, principally in low latitudes, whereas HadCM3 variability declines monotonically with increasing depth, except for a small peak at 200 m. The source of most variability is the surface, but a subsurface maximum could perhaps be caused by the accumulation of subducted surface anomalies, or by vertical movement associated with internally generated instability or with larger-scale variability in the depth range of the thermocline with a strong vertical temperature gradient. HadCM3 must lack an important process if the subsurface maximum is real. There is a need for further investigation into the processes of heat uptake by the model.
 Imposing the observed data mask on HadCM3 increases the estimated variability; it doubles the variability at some levels below 200 m. In southern latitudes it produces a peak rather like that of Levitus. This sensitivity to the data mask leads us to suggest that the sparseness of the observed data could also have made a contribution to the subsurface variability in the Levitus dataset.
 Agreement between Levitus and HadCM3 at most depths is closest in the northern latitudes, the best sampled part of the ocean. This is consistent with our earlier finding, but does not by itself tell us whether disagreement elsewhere is due to deficiencies in HadCM3, Levitus, or both.
 When both natural and anthropogenic forcings are included, HadCM3 simulations and the Levitus analysis of the World Ocean Database have similar trends in global ocean heat content anomalies, but variability in the latter is larger outside the well-observed northern hemisphere upper ocean. We note that Ishii et al.  also draw a contrast between the hemispheres in their objective analysis of the heat content variability of the upper 500 m. Levitus has a strong subsurface peak around 500 m which is not present in HadCM3 simulations and appears to be sensitive to sampling.
 Previous authors have remarked on discrepancies between AOGCM and Levitus variability. In this analysis we have made some progress in identifying the sources of it, although we have not been able to reach a complete explanation. Until it is understood, we suggest that caution is required in the use of observational estimates to assess AOGCM ocean heat content changes, both variability and trend, outside the well-observed region. This caution should apply also to evaluation of salinity changes, since measurements of salinity are even sparser than those of temperature. We recommend that continuing subsurface ocean observations, for instance through the ARGO deployment of profiling floats, should be supported as it is only with sufficient data that the uncertainties associated with observational estimates and simulated processes can be reduced.
 We are grateful for useful discussions with Simon Tett, Rowan Sutton, Gabi Hegerl, Susan Lozier, James Annan, John Church and Sheila Stark and for the comments of the referees. This work was supported by the UK Department for Environment, Food and Rural Affairs under contract PECD 7/12/37 and by the Government Meteorological Research and Development Programme.