Detection of Anthropogenic Climate Change in the World's Oceans

See allHide authors and affiliations

Science  13 Apr 2001:
Vol. 292, Issue 5515, pp. 270-274
DOI: 10.1126/science.1058304


Large-scale increases in the heat content of the world's oceans have been observed to occur over the last 45 years. The horizontal and temporal character of these changes has been closely replicated by the state-of-the-art Parallel Climate Model (PCM) forced by observed and estimated anthropogenic gases. Application of optimal detection methodology shows that the model-produced signals are indistinguishable from the observations at the 0.05 confidence level. Further, the chances of either the anthropogenic or observed signals being produced by the PCM as a result of natural, internal forcing alone are less than 5%. This suggests that the observed ocean heat-content changes are consistent with those expected from anthropogenic forcing, which broadens the basis for claims that an anthropogenic signal has been detected in the global climate system. Additionally, the requirement that modeled ocean heat uptakes match observations puts a strong, new constraint on anthropogenically forced climate models. It is unknown if the current generation of climate models, other than the PCM, meet this constraint.

Almost all rigorous studies attempting to find the impact of anthropogenic forcing in today's climate system have used air-temperature observations as the data set of choice for the detection study (1). These studies most often use near-surface air temperature [e.g., (2–6)]. Changes in sea ice (7) and the vertical temperature structure of the atmosphere obtained from radiosonde data (8–11) have also been investigated for evidence of an anthropogenic signal.

Climate models predict that there will be substantial anthropogenic changes in variables only weakly related to near-surface air temperature. For example, model-predicted signals have been detected in the magnitude of the annual cycle and wintertime diurnal temperature ranges (12), but there are few such studies at present. It seems imperative that statements about the detection and attribution of model-predicted anthropogenic climate change, such as made in the most recent Intergovernmental Panel on Climate Change (IPCC) Assessment (13), be placed on a stronger foundation by quantitatively identifying such signals in other elements of the climate system.

A major component of the global climate system is the oceans; covering roughly 72% of the planet's surface, they have the thermal inertia and heat capacity to help maintain and ameliorate climate variability. Although the surface temperature of the oceans has been used in detection and attribution studies, apparently no attempt has been made to use changes in temperature at depth. A recent observational study (14) has shown that the heat content of the upper ocean has been increasing over the last 45 years in all the world's oceans, although the warming rate varies considerably among different ocean basins. We show in this study that at least one global climate model, forced by a combination of observed and estimated anthropogenic gases, has reproduced the observed changes in ocean heat content with surprising accuracy.

The decadal changes over the last 45 years in the heat content of the upper 3000 m of the water column estimated from observations are shown in Fig. 1(14). Only the very low frequency component of the signal is of interest here, so we chose to work relative to this depth and use decadal averages to filter out noise associated with eddies and interannual or decadal natural variability. A similar set of heat-content changes, relative to a 300-year control run climate, was computed from five different realizations of the Parallel Climate Model (PCM) forced by observed and estimated concentrations of greenhouse gases and the direct effect of sulfate aerosols on the atmosphere (15). This state-of-the-art global climate model, which uses no flux-correction scheme, is a cooperative effort between a number of universities and government laboratories in the United States (16). A brief summary of the model components, forcing scenarios, and current results are given in (17–19), while a more detailed description can be found at

Figure 1

Decadal values of anomalous heat content (1022 J) in various ocean basins. The heavy dashed line is from observations (14), and the solid line is the average from five realizations of the PCM (16–19) forced by observed and estimated anthropogenic forcing. Both curves show significant warming in all basins since the 1950s. The shaded bands denote one (heavy shading) and two (light shading) standard deviations about the model mean signal estimated from the standard deviation in the scatter of the five-member ensemble. The heat content is computed over the upper 3000 m of the water column. The space/time sampling was identical for both model and observations. Basin averages for the northern oceans are defined between 60°N and the equator. The southern ocean averages are between the equator and 77°S.

Figure 1 shows an unexpectedly close correspondence between the observed heat-content change and the average of the same quantity from the five model realizations. These results were obtained by subsampling the model data at the same locations and times where observations existed (20). When the scatter between the multiple model runs is included (shaded regions on Fig. 1), it becomes apparent that there is little or no significant difference between model and observations, even though the heat-content changes vary among ocean basins. The main exception occurs in the 1970s, when the observations show a decadal anomaly that the model runs do not reproduce. We speculate that the anomaly is associated with the apparent regime-like shift in the North Pacific Oscillation and other regional climate modes that occurred at that time (21–24). It is not possible, given the manner in which it was forced, that the model could have captured this specific decadal signal. However, the model does produce, in both its anthropogenically forced runs and control run, decadal fluctuations that have the same magnitude and time scale as those associated with the observed anomaly of the 1970s. In any event, the anomaly does not alter the close correspondence between model and observations.

In summary, the PCM, forced by anthropogenic constituents, produced changes in heat content in each of the major oceans over the last 45 years that are highly similar to those observed.

The vertical development of the oceanic warming signal in the PCM for the world's oceans is shown in Fig. 2. Here we show the time evolution, from the start of the integrations (1870) through the year 2000, for the average of the five-member ensemble. The scatter among the five realizations allowed us to estimate a standard deviation that was used to filter the results so that temperature anomalies exceeding a 90% confidence limit are indicated by the gray shaded areas.

Figure 2

Decadal temperature anomalies (°C) in various ocean basins since 1870 from the PCM. Gray-shaded regions indicate signals statistically distinguishable from zero. The timing of the warming, as well as its vertical structure, varies substantially among the basins. The deep changes in the “World” oceans are seen to be derived solely from changes in the Atlantic. The contour interval is 0.05°C.

The nature of the warming in the various oceans is markedly different. The Atlantic, particularly the South Atlantic, shows strong vertical convection taking the signal to depth quite rapidly. The South Atlantic regional average includes portions of the model's deep-water formation region in the Weddell Sea, so the rapid penetration to depth is expected. The signal is larger in the North Atlantic, but does not appear to penetrate as rapidly, possibly because the regional averaging area in that ocean is large relative to the space scales of deep vertical mixing. In the other oceans, the signal is more consistent with what one would expect from a purely diffusive process. It is important to note that there is little deep water formed in the South Pacific regional average, which otherwise would be expected to resemble the South Atlantic regional average. At any rate, it is apparent that the signal in the deep ocean labeled “world” oceans comes from the Atlantic in the model simulations.

The temporal evolution of the model's vertical temperature structure is compared with observations in Fig. 3, for the North Pacific and North Atlantic Oceans. These are oceans where the data density is highest. In this case, we have concentrated on temperature changes in the upper 2000 m to emphasize the region of maximal signal strength and applied the model-derived significance factors used in Fig. 2 to filter noise from the data.

Figure 3

Modeled and observed temporal and vertical changes in the temperature in the upper 2000 m of the data-rich North Pacific and North Atlantic Oceans. Near the surface, where interannual and decadal changes and external forcings strongly affect the thermal structure, any agreement is largely due to chance. Gray-shaded regions denote areas where changes are statistically different from zero. The model broadly reproduces the main features of the vertical structure and its temporal evolution over the last 40 years. The contour interval is 0.05°C.

In the North Atlantic, the observations show a near-surface warming since about 1980, whereas the model begins warming at about 1950. This result may be partially due to the presence of a single, noisy observation set compared with the smoother, ensemble average. However, the penetration of warming with time and depth is otherwise similar between the model and observations. The more-or-less diffusive penetration in the North Pacific is captured in the model, although the very near-surface structure is again somewhat different. These discrepancies in near-surface behavior may be due to the large interannual and decadal variability that the model, running with no real-world input, cannot be expected to capture (unless it were to be forced by the observed fluxes of heat, momentum, and moisture). The model also contains no natural external forcings such as solar or volcanic mechanisms (25). The key point is that the substantial differences in the way the observed warming has penetrated to depth in the two oceans is reasonably well captured by the PCM, albeit with the caveats noted above.

We examined whether the model-produced changes in ocean heat content are different from those expected by chance, e.g., different from those expected in a long control run of the PCM. This indicates whether a significant climate change has been observed in the forced model runs. We further examined whether the model-produced changes are statistically consistent with the observations. The first question addresses the detection of climate change, whereas the second examines whether anthropogenic forcing could be responsible for the observed changes.

The detection and attribution method that we use here is termed “optimal detection” and is well documented (3,26–28). The basic idea is to compute the space/time pattern of change predicted by the model—the anthropogenic fingerprint—and its strength. Normalization of the data before estimating these signals by the natural variability or noise, determined from one of two PCM control runs (300 years long), is necessary for “optimal” detection (29,30). The space/time distribution of the observations are projected onto this “fingerprint,” and the observed pattern strength estimated. We then used the second PCM control run (270 years long) to estimate the level of expected natural variability under conditions where no anthropogenic forcing exists. This latter information allows us to estimate a confidence level on the observations and thus make significance statements about difference between the observed signal strength and the model-predicted signal strength (31)

The analysis showed that a single space/time pattern or fingerprint captured the model's anthropogenic signal well enough that both the optimal and nonoptimal detection analyses gave basically the same result. From here on, we use a nonoptimal analysis so that both control runs can be used to estimate levels of natural variability. The signal strength of this pattern from each of the five realizations and their mean value are shown in Fig. 4. Also shown is the signal strength for the observations projected onto the model fingerprint. The projection of the control run data onto this fingerprint allows a noise estimate for computing significance. These latter values were expressed as 95% confidence intervals and centered on the mean model value to facilitate discussion (stippled region).

Figure 4

Detection and attribution diagram. The strength of the anthropogenic pattern of model-predicted changes in depth-averaged ocean temperatures (°C) is shown for each of the five realizations (+) and their ensemble average (•). Also shown is the strength of the model-predicted signal in the observations (X). The lightly stippled region corresponds to the 95% confidence region, centered on the ensemble average, associated with natural variability as estimated from the PCM control runs [see (31)]. The individual signal-strength estimates for independent 45-year chunks of the control run are indicated (★). All forced runs, their mean, and the observations fall within the 95% confidence region and so are indistinguishable from each other, i.e., the model-produced signal and the observations are “consistent” with each other. The uncertainty region does not include the origin, so that natural variability, as estimated by the model, cannot explain the signal in the observations. The results are for the nonoptimal detection method, but do not differ appreciably from that for the full optimal approach.

The confidence limits do not include the origin (the “no signal” or no climate change case), and so the mean of the anthropogenic runs is highly significant and detection of a climate change signal in the model has occurred, i.e., the mean is not expected to occur by chance in the control run. Further, the differences between the various model runs and the observed state all fall within the confidence limits, so that we cannot distinguish between any of the five realizations or the observations. They are all identical from a statistical point of view. Hence, we say that the observations are consistent with the anthropogenically forced model results. This suggests, with confidence exceeding 95%, that one may accept anthropogenic forcing as one possible explanation for the observed changes in heat content of the global oceans. There may be other possible explanations that were not included in the model simulations—hence the word “consistent” used above.

The work reported here yields strong results, but some caveats are necessary. The observational data are sparse in the southern oceans before the 1980s, which should be kept in mind when comparing the observations with the model and when using model and observed data from these regions in a detection methodology. However, our subsampling method of using model data only where comparable observations occurred should largely overcome this problem for detection purposes. Finally, the estimates of natural variability used in the detection work were derived from simulations—not observations, which do not exist in sufficient quantity for this purpose.

A very low frequency variability, perhaps drift, can be seen in the model control run that varied in magnitude and sign from ocean to ocean. The time scale of this variability was on the order of hundreds of years, much longer than the 45-year time slices used in this study. Sensitivity studies suggest that it has not affected our analysis. We suggest that the control run be extended to an order of 1000 years in future studies, as has been done by other centers, to better evaluate natural variability in the model and the role of ultralow-frequency change. In addition, the forced runs we used for detection did not include indirect sulfate, ozone, so-called black aerosols, and other anthropogenic factors, nor did they include external forcing due to solar variability and volcanic activity [e.g., (25)]. Most of the excluded factors introduce a cooling that might negate the fact that the model is a little too warm.

The detection and attribution study used regional and vertical averages. A more complete analysis should consider the full three-dimensional structure of the signal and noise fields. Such an analysis is currently in progress, although we do not expect it to change the major conclusions presented here. Further confirmation of our results comes from a similar analysis of anthropogenically forced runs at the Max Planck Institute in Hamburg, Germany (32).

Perhaps the most important aspect of this work is that it establishes a strong constraint on the performance and veracity of anthropogenically forced climate models. For example, a climate model that reproduces the observed change in global air temperature over the last 50 years, but fails to quantitatively reproduce the observed changed in ocean heat content, cannot be correct. The PCM has a relatively low sensitivity (less anthropogenic impact on climate) and captures both the ocean- and air-temperature changes. It seems likely that models with higher sensitivity, those predicting the most drastic anthropogenic climate changes in the future, may have difficulty satisfying the ocean constraint. To our knowledge, the PCM is the only model currently able to do this and still accurately reflect the changes in surface air temperature over the last 50 years. Future studies should take into account the ocean constraint when deciding which future climate summaries are most reliable.

  • * To whom correspondence should be addressed. E-mail: tbarnett{at}


View Abstract

Stay Connected to Science

Navigate This Article