Can oceanic reanalyses be used to assess recent anthropogenic changes and low-frequency internal variability of upper ocean temperature?
- First Online:
- Cite this article as:
- Corre, L., Terray, L., Balmaseda, M. et al. Clim Dyn (2012) 38: 877. doi:10.1007/s00382-010-0950-8
- 234 Views
A multivariate analysis of the upper ocean thermal structure is used to examine the recent long-term changes and decadal variability in the upper ocean heat content as represented by model-based ocean reanalyses and a model-independent objective analysis. The three variables used are the mean temperature above the 14°C isotherm, its depth and a fixed depth mean temperature (250 m mean temperature). The mean temperature above the 14°C isotherm is a convenient, albeit simple, way to isolate thermodynamical changes by filtering out dynamical changes related to thermocline vertical displacements. The global upper ocean observations and reanalyses exhibit very similar warming trends (0.045°C per decade) over the period 1965–2005, superimposed with marked decadal variability in the 1970s and 1980s. The spatial patterns of the regression between indices (representative of anthropogenic changes and known modes of internal decadal variability), and the three variables associated with the ocean heat content are used as fingerprint to separate out the different contributions. The choice of variables provides information about the local heat absorption, vertical distribution and horizontal redistribution of heat, this latter being suggestive of changes in ocean circulation. The discrepancy between the objective analysis and the reanalyses, as well as the spread among the different reanalyses, are used as a simple estimate of ocean state uncertainties. Two robust findings result from this analysis: (1) the signature of anthropogenic changes is qualitatively different from those of the internal decadal variability associated to the Pacific Interdecadal Oscillation and the Atlantic Meridional Oscillation, and (2) the anthropogenic changes in ocean heat content do not only consist of local heat absorption, but are likely related with changes in the ocean circulation, with a clear shallowing of the tropical thermocline in the Pacific and Indian oceans.
KeywordsAnthropogenic climate changeInternal variabilityOceanic reanalysesObservationsUpper ocean temperature
Identifying and quantifying the impact of anthropogenic forcing on the global ocean have become of critical importance. A large fraction of the energy gained by the Earth climate system during the twentieth century has accumulated in the world subsurface oceans (Bindoff et al. 2007). In order to achieve a rigorous estimation of the Earth’s energy balance, monitoring the variations of ocean heat content at both global and regional scales is essential. The future rate of ocean heat uptake is also of primary importance as it dictates the lagged response of surface temperature rise. Recent observational studies show that some of the expected modifications due to anthropogenic forcing may already have been emerging over the last few decades. The most prominent signal is the increase in global ocean heat content as a result of anthropogenic forcing (Barnett et al. 2005; Levitus et al. 2005). A number of observational estimates show a gradual warming of the upper ocean (0–700m) heat content superimposed with decadal variations (Ishii et al. 2003; Levitus et al. 2005; Bindoff et al. 2007). Beyond the general analyses consistency, accounting for their differences is of great interest in order to shed light on uncertainties in the recent ocean evolution. Differences may come from many factors such as the choice of input data, quality control, infilling method, and/or instrumental bias correction (Palmer et al. 2010). The latter have been produced to minimize instrument-related biases due to changing technologies used to collect ocean temperature profiles. The most well-known instrument bias is the significant, time varying, warm bias in the abundant but low accuracy temperatures collected by expendable bathythermograph (XBT) (Gouretski and Koltermann 2007). This bias is responsible for some of the observed temperature fluctuations in the 1970s. Different methods for correcting XBTs have been proposed, based on correcting fall-rates (Wijffels et al. 2008; Ishii and Kimoto 2009), or reported temperatures (Levitus et al. 2009). Most corrections result in a reduction of the global ocean heat content interdecadal variability. However, effect on the long-term trend estimates vary from one correction type to another. According to Levitus et al. (2009), their corrections have little effect on the trend estimate. On the contrary, Domingues et al. (2008) find an ocean warming rate, from 1961 to 2003, about 50% larger than equivalent rate of non-corrected estimates when using Wijffels et al. (2008) corrections. By comparing the ocean heat content from different observed datasets, Lyman et al. (2010) estimated uncertainties arising from the choice of climatology, the choice of mapping methodology, the effects of irregular and sparse sampling, the method of XBT correction and XBT quality control. They find that uncertainties in XBT bias corrections are the dominant error source over the period 1993–2008.
Most of the studies cited above are objective analyses based solely on observations and use statistical methods to fill the data sparse regions. Yet, statistical infilling introduces uncertainties as different infilling methods lead to different results (AchutaRao et al. 2007). Oceanic reanalyses are an alternative to purely statistical approaches which are used to reconstruct the 4-dimensional structure of the global ocean. Constructing an oceanic reanalysis consists of simulating the evolving ocean state with an ocean model driven by the best available atmospheric surface forcing and constrained by assimilating direct observations. In that case, the infilling of the data sparse regions is also based on the physics of ocean models. Ocean reanalyses thus provide a time-continuous, dynamically-interpolated estimate of the ocean state and include the influence of the evolving atmospheric state on subsurface ocean variables. This physically-based approach has proved worthwhile in providing spatially complete fields. Carton and Santorelli (2008) examined the evolution of the temperature in the upper 700 m of the global ocean as represented by two objective analyses and seven model-based oceanic reanalyses. They discussed the relative role of natural variability, analysis method and data inadequacies in the main uncertainties associated with the recent upper ocean warming. Within the European ENACT and ENSEMBLES projects, global multi-decadal ocean reanalyses have been produced to provide initial conditions for coupled model forecasts on seasonal to decadal time scales (Davey et al. 2006). This set of reanalyses provides a probabilistic estimate of the changing ocean state over the past decades. A first set of ENACT reanalyses has been compared in terms of globally averaged temperature over the top 300 m from 1960 to 2000 (Davey et al. 2006). Results highlight the substantial progress that has been made in reconstructing recent changes in upper ocean heat content. In most regions, differences between reanalyses are substantially less than the amplitude of interannual variability. Nevertheless, understanding similarities and differences between reanalyses is needed to improve each product and increase the value of ocean state estimates. This is a difficult task as many factors can be involved such as model errors, problems associated with the assimilation systems and/or with the atmospheric forcings, biases and scarcity of the observations.
This paper further examines recent upper ocean changes using five model-based ocean reanalyses and a model-independent objective analysis. We use the mean temperature above the 14°C isotherm (Tiso14 thereafter) as a measure of the recent ocean warming. Based on the analysis of the EN2 observational dataset (Ingleby and Huddleston 2007) between 1965 and 2005, Palmer et al. (2007) show an average warming trend of 0.043°C per decade for this variable. As Tiso14 is independent of the absolute reported depth, the analysis of this variable in products derived only by the observations is insensitive to the XBT’s fall-rate errors. Moreover, compared with more classical fixed depth mean temperatures, Tiso14 has been proved valuable to isolate externally forced air–sea heat fluxes changes by filtering temperature modifications induced by dynamical processes (Palmer and Haines 2009). By reducing the influence of ocean circulation changes, Tiso14 also improves observation-model comparison and allows a robust detection of both natural and anthropogenic influences on ocean subsurface temperatures changes (Palmer et al. 2009). The detection of climatic changes is based on the ability of separating the influence of each of the different climatic forcings. Different approaches have been proposed to isolate the low frequency signal associated with anthropogenic forcing. The first commonly used is to represent it by a linear trend (Enfield et al. 2001; Sutton and Hodson 2005; Knight et al. 2006). This approach assumes that the anthropogenically forced signal is uniform and linear over time, thereby including the non-linear part of the trend in the noise. A second method consists of characterizing the anthropogenically forced signal by a simple index, using for example the global mean sea surface temperature (Trenberth and Shea 2006; Mann and Emanuel 2006). In order to account for the spatial non uniformity of the anthropogenically forced signal, Ting et al. (2009) regress the two dimensional fields they want to analyze on the time series of globally averaged SST. Here we will follow a similar approach, but construct an anthropogenically forced signal index with the Tiso14 variable from climate model scenarios [based on Ribes et al. (2009) method]. This method as well as the observed and reanalysis data used are briefly described in the next section. The time evolving upper ocean temperature is simply investigated in terms of linear trend and decadal variability in Sect. 3. Section 4 focuses on the spatial pattern of the changes, isolating both the anthropogenically forced signal and internal variability. Section 5 is a summary of our main results.
2 Data and method
Temperature fields from the EN3-OA observational product are compared with three ENSEMBLES reanalyses (CERFACS, ECMWF and INGV) and the SODA reanalysis. The ENSEMBLES reanalyses are based on the assimilation of in-situ temperature and salinity profiles from EN3_v2a dataset. Atmospheric forcings fields consist of daily fluxes of momentum, heat and fresh water derived from the ERA-40 reanalysis (Uppala et al. 2005). Additional bias correction terms are applied to precipitation (Troccoli and Källberg 2004). From September 2002 onwards, when ERA40 terminates, ECMWF operational surface fluxes are used. The sea surface temperature is strongly restored to analyzed SST maps derived from the OIv2 SST product (Reynolds et al. 2002). A globally uniform relaxation coefficient of −200 W m−2 K−1 is used, which corresponds to a relaxation time-scale of 12 days for a mixed-layer depth of 50 m. A common initialization sequence (‘spinup’) is also prescribed, starting on 1 January 1953. It uses ERA40 climatological surface fluxes up to 1 January 1958, and then the daily ERA40 surface fluxes from 1 January 1958 to 31 December 1961 (Davey et al. 2006). During spinup, relaxation terms to prescribed temperature and salinity fields are imposed, as described further for each reanalysis.
Reanalyses considered in this study
SODA v. 2.0.2-4
0.25° × 0.4°
GPCP freshwater flux
OPA v. 8.2
ORCA 2°: 2° × 2°−0.5°
ERA-40 daily flux corrected precipitations after 2001: ECMWF fluxes
ENACT/ENSEMBLE dataset v. EN3_v2a
OPAVAR 3-dimensional variational
OPA v. 8.2
ORCA 2°: 2° × 2°−0.5°
ERA-40 daily flux corrected precipitations after 2001: ECMWF fluxes
ENACT/ENSEMBLE dataset v. EN3_v2a
Reduced order optimal interpolation
1° × 1°−0.3°
ERA-40 daily flux corrected precipitations after 2001: ECMWF fluxes
ENACT/ENSEMBLE dataset v.EN3_v2a altimeter data
3D optimal interpolation
HOPE 1° × 1°−0.3°
ERA-40 daily flux corrected precipitations after 2001: ECMWF fluxes
ENACT/ENSEMBLE dataset v.EN3_v2a XBT corrections (Wijffels et al. 2008) altimeter data
3D optimal interpolation
The Simple Ocean Data Assimilation (SODA) product, version 2.0.2-4 (Carton and Giese 2008) uses a model based on the Parallel Ocean Program numerics (Smith et al. 1992) whose resolution is eddy-permitting. It is driven by ERA-40 winds and surface freshwater flux from the Global Precipitation Climatology Project monthly satellite-gauge merged product (Adler et al. 2003). A relaxation to the World Ocean Atlas 2001 (Boyer et al. 2002) climatological sea surface salinity is applied with a 3-month relaxation time scale. The mixed layer temperature is updated with near-surface temperature observations. The assimilated dataset consists of approximately 7 × 106 temperature and salinity profiles. Two-thirds of the data are from the World Ocean Database 2001 (Boyer et al. 2002; Stephens et al. 2002), completed by real-time temperature profile observations from the National Oceanographic Data Center (NODC)/NOAA temperature archive. The correction to the XBT’s fall-rate equation provided by Hanawa et al. (1995) is included.
As mentioned in the Introduction, observational datasets are affected by instrumental biases. With the aim of reducing those biases, a new version of the EN3_v2a dataset has been produced (http://hadobs.metoffice.com/en3/). It has removed the cold water tapering (Thadathil et al. 2002) that is applied to some XBT corrections. Then, the corrections to XBT depths from Table 1 of Wijffels et al. (2008) have been applied. This corrected dataset has been used in a new reanalysis hereafter referred to as ECMWF_update. Model, data assimilation technique, atmospheric forcing and restoring coefficients are similar to those described for the ECMWF reanalysis. Comparison between the two products gives an estimate of the impact of the XBT warm bias on upper ocean heat content decadal variability and trends.
2.3 Statistical method
AOGCMs used together with their originating group
Center and location
Bjerknes Centre for Climate Research (Norway)
Canadian Centre for Climate Modelling and Analysis (Canada)
Canadian Centre for Climate Modelling and Analysis (Canada)
Météo-France, Centre National de Recherches Météorologiques (France)
CSIRO Atmospheric Research (Australia)
NASA/Goddard Institute for Space Studies (USA)
Instituto Nazionale di Geofisica e Vulcanologia (Italy)
Institut Pierre Simon Laplace (France)
Hadley Centre for Climate Prediction and Research (USA)
3 Global and basin-averaged upper-ocean temperature changes
Linear trend and amplitude of the decadal variability of the Tiso14 time series as seen by the MEAN (bold), EN3-OA, ECMWF and ECMWF_update products, over the period 1965–2005, in individual basins
Tiso14 trend (°C per decade)
Tiso14 decadal variability amplitude (°C)
0.045 (+0.005, −0.010)
0.028 (+0.004, −0.002)
0.063 (+0.007, −0.016)
0.033 (+0.013, −0.001)
0.047 (+0.009, −0.012)
0.048 (+0.007, −0.002)
0.026 (+0.007, −0.006)
0.039 (+0.022, −0.011)
3.1 Decadal variability
The origin of decadal variability of the global ocean temperatures has long been debated. Here we use the comparison of Tiso14 with T250m, Diso14, and the sea surface temperature (SST) from the HadSST2 data set (Rayner et al. 2006) (Fig. 4) to discuss it. Decadal variability has been shown to be influenced among other factors by external natural forcing due to volcanoes (Church et al. 2005). During the estimation period, major volcanic eruptions occur in 1963 (Mount Agung), 1982 (El Chichon) and 1991 (Pinatubo). The first two eruptions are followed by few years coolings of both surface and subsurface integrated temperatures, especially marked in the Pacific basin. This is the region where the maximum cooling induced by volcanic aerosols occurs, as suggested by the coupled model study of Delworth et al. (2005). The impact of the Pinatubo eruption on the Pacific SST is not as clear. The tropical Pacific SSTs are dominated by El Niño conditions between 1990 and 1995. The resulting warming seems to mask the effects due to volcanic forcing. On the contrary, the two integrated temperatures (Tiso14 and T250m) do show a cooling phase after 1991. In the subsurface, the main El Niño effect is in the same direction as the volcanic forcing. Indeed, in a large part of the equatorial Pacific basin, extending from West to the center, the isotherms move upwards (not shown), resulting in a subsurface cooling. At the basin scale, this cooling dominates the warming associated with the deepening thermocline in the eastern equatorial Pacific. The examination of the Diso14 time series shows no clear response to volcanic forcing. Consistent with Palmer and Haines (2009), this result confirms that isotherm depth changes are not sensitive to changes in air–sea heat flux.
Upper ocean temperature decadal variability may also result from internal ocean and/or coupled variability modes. The presence of large decadal fluctuations in the Pacific and Indian Ocean basins suggests an influence of El Niño Southern Oscillation (ENSO) and the Interdecadal Pacific Oscillation (IPO), on both SSTs and integrated variables. In agreement with Palmer et al. (2007), the T250m decadal variability is larger than that of Tiso14 (Tables 3, 4). It suggests that the use of Tiso14 increases the signal to noise ratio due to the filtering of dynamical changes associated with internal variability modes. The North Atlantic Oscillation (NAO) and Atlantic Multidecadal Oscillation (AMO) also play significant roles in the Atlantic basin. In particular, the Diso14 deepening associated with T250m warming observed since 1980 may be related with predominantly positive phases in the NAO during this period (Palmer and Haines 2009). More recently, the AMO shift from negative to positive phase may also have reinforced those Atlantic trends. The spatial signature of the AMO and IPO modes on the three variables (Tiso14, T250m and Diso14) is investigated in Sect. 4.2.
Finally, a fraction of the decadal variability could also be artificial and induced by changes and/or biases in the observational datasets. In particular, the increased oceanic warming, occurring between 1970 and 1980, has been shown to be partly linked to a warm bias in XBT temperature profiles (Gouretski and Koltermann 2007). Comparison between the ECMWF and ECMWF_update products allows us to quantify the XBT bias impact on the ocean heat content estimates. At the basin scale, the impact of the XBT correction on the amplitude of the decadal variability is weak (Tables 3, 4, 5). This is true in all the basins, and for the three variables. However, the specific details of the time evolution in T250m and Diso14 are strongly affected by the XBT correction especially during the 1970s and 1980s. We recall that the XBT bias has been attributed to erroneous reported depths of the profiles (Gouretski and Koltermann 2007). The bias thus affects isotherm depths directly. Understandably, the assimilation of data with the XBT correction of Wijffels et al. (2008) strongly impacts the evolution of Diso14. In particular, it suppresses the enhanced Diso14 deepening between 1970 and 1980, resulting in a damped Diso14 variability from 1970 to the late 1990s. The Diso14 weak variability exhibited by ECMWF_update in the Pacific basin confirms that there is no clear influence of volcanic forcing in the isotherm depth time series. As a result of the reduced Diso14 deepening, the XBT corrections also clearly modify the T250m time series. Whereas the T250m variable as seen by the ECMWF and MEAN products warms gradually between 1970 and 1980, ECMWF_update exhibits a strong cooling centered around 1975. Consistency between the T250m and SST is then increased and the strong SST cold episode during 1974–1976 in the Pacific and Indian Oceans also manifests itself in a reduced upper ocean heat content. Correlations between global T250m and SST time-series equal 0.85 and 0.97 when computed over the 1965–2005 period using ECMWF and ECMWF_update respectively. Between 1970 and 1980, the effect of the XBT corrections for Tiso14 opposes the one for T250m. At the global scale, the ECMWF_update anomalies in Tiso14 are slightly warmer than the ECMWF anomalies. This difference arises mainly from the Pacific basin. A possible mechanism for Tiso14 warming associated with Diso14 shallowing is related with an increased stratification which would reduces the vertical mixing and consequently warms the surface layers. Despite a few Tiso14 differences between ECMWF and ECMWF_update, the XBT corrections affect less this variable than the two others. This suggests that Tiso14 is relatively immune to the XBT bias even in the assimilation products.
The assimilation of XBT bias-corrected data also clearly affects the upper ocean evolution at the end of the estimation period. Whereas the effect on Diso14 varies from one basin to one other (shallowing in the Atlantic and Indian basins, and deepening in the Pacific basin), the corrections enhance the T250m warming after 2000 in all the basins. T250m warming associated with Diso14 shallowing may be the result of an increased warming near the surface, together with an increased stratification at depth. However those results need to be taken with caution, as examination of local changes in Diso14 and T250m show that the spatial patterns are very contrasted (not shown). On the contrary, the effect of the XBT corrections between 1970 and 1980 is very consistent across different regions (not shown). Moreover, over the last decade, the effect of the XBT correction is mixed with the effect of altimeter data assimilation. Comparison between the MEAN and ECMWF curves suggests two effects of the introduction of altimeter data. First, Diso14 deepens inducing an enhanced T250m warming. This could be explained by a pushing down of the water column (i.e. a vertical displacement without any change in the stratification) resulting from the Cooper and Haines (1996) scheme used for assimilation of altimeter in the ECMWF system. This scheme translates sea level changes into vertical displacement of the water column (Vidard et al. 2009). Second, the warming in Tiso14 is reduced, which is not straightforward to understand. As the vertical displacement of the profiles by the assimilation of altimeter data is local in space and time, explaining the global effect (average in space and time) is not immediate. One possibility is that although in principle the altimeter only moves the water column up or down, the cumulated effect leads to additional vertical mixing, which might produce a stretching down of the water column (i.e. a vertical displacement leading to a reduced stratification). If the cold subsurface layers stretch more than the warm surface layers, it results in a Tiso14 cooling. This additional mixing may or may not be realistic. In order to confirm this hypothesis, we have compared the ECMWF product with an experiment based on the same model and assimilation system, but without altimeter data, as described in Balmaseda et al. (2007) (not shown). The comparison confirms that the assimilation of altimeter data indeed deepens the equatorial thermocline and cools the waters above the thermocline.
3.2 Linear trends
Linear trend and amplitude of the decadal variability of the T250m time series, as seen by the MEAN (bold), EN3-OA, ECMWF and ECMWF_update products, over the period 1965–2005, in individual basins
T250m trend (°C per decade)
T250m decadal variability amplitude (°C)
Linear trend and amplitude of the decadal variability of the Diso14 time series, as seen by the MEAN (bold), EN3-OA, ECMWF and ECMWF_update products, over the period 1965–2005, in individual basins
Diso14 trend (m per decade)
Diso14 decadal variability amplitude (m)
Linear trend and amplitude of the decadal variability of the SST time series computed from the HadSST2 dataset, over the period 1965–2005, in individual basins between 437.5°S and 37.5°N
SST trend (°C per decade)
SST decadal variability amplitude (°C)
ECMWF_update time series show that whereas the XBT corrections have little impact on the Tiso14 trend, it clearly increases the warming in T250m. Note however the strong dependence of the linear trend on the period chosen for the computations. Since the XBT correction induces a cooling period around 1965, it results in an increased linear trend when computed from 1965 to 2005. At the global scale, the T250m trend ranges from 0.070°C per decade (ECMWF) to 0.095°C per decade (ECMWF_update); i.e. a 34% increase. However, when computed over the 1960–2005 period, the ECMWF_update trend (0.067°C per decade) is only 24% larger than the ECMWF trend (0.054°C per decade). Another limitation of the trend calculation is to assume that the change is linear over time. Yet, Figs. 3 and 4 show that the upper ocean heat content time evolution varies among the different basins with an earlier warming of the Indian and Atlantic oceans. In the next section, we attempt to isolate the upper ocean temperature change due to anthropogenic forcing. Instead of using linear trend diagnostics, we use the regression methodology described in Sect. 2.3.
4 Causes and spatial patterns of upper-ocean temperature changes
4.1 Anthropogenically forced signal
In general, the spatial structure of the Tiso14 changes is smoother than the spatial structure in T250m and Diso14, shows more uniform pattern, and seems more consistent between reanalysis products. The spatial patterns of T250m and Diso14 are more complex, since they are affected by local heat storage and redistribution of heat due to changes in the ocean circulation. The dispersion among reanalyses is larger in these two variables. Although the cooling in the equatorial Pacific and southern Indian oceans appears to be robust, there is some uncertainty about its magnitude and the east–west gradients. In the Atlantic basin, the ECMWF and ECMWF_updated reanalyses exhibit stronger T250m warming and Diso14 deepening than other products. This result may suggest that the warming is penetrating deeper over time, and so causing the depth of the 14°C isotherm to increase. This may be due to the weaker relaxation to climatology used in the ECMWF reanalyses. Another possibility is that the strong changes in T250m and Diso14 shown by those two products are linked with the introduction of altimeter data which pushes down the water column. Spatial correlations between the change in Diso14 and the pattern of the differences in changes between Tiso14 and T250m (not shown) are indicated in Fig. 7. Most of them are in the range 0.6–0.7, smaller than the 0.77 correlation found by Palmer et al. (2007). The difference between Tiso14 and T250m trends, together with the resemblance between T250m and Diso14, would suggest that changes in the upper ocean heat content can not be explained without considering changes in the global ocean circulation. Finally, the strong similarity between the Tiso14 change patterns as seen by ECMWF and ECMWF_updated confirms that Tiso14 is not very sensitive to the XBT bias. Impact of the XBT corrections is more noticeable in T250m change maps. For example, the ECMWF_update change map exhibits more intense warming around 10°N in the Atlantic and in the eastern Pacific basins. Note however that in general, the two analyses show very similar patterns of change. A possible explanation is that the corrections are mostly applied as a function only of depth and time (Wijffels et al. 2008) and thus have a small impact on the local changes. This result is consistent with Carson and Harrison (2010) who find that the strength and spatial patterns of regional interdecadal variability, at the sub-basin to basin scales, are unaffected by the application of bias correction to the XBT fall rates.
4.2 Signature of internal variability modes
Once the anthropogenic forced signal has been isolated, we can interpret the residual as an estimate of the recent upper ocean temperature internal variability. Here we choose to focus on the signature of the two following low-frequency internal modes: the Interdecadal Pacific Oscillation (IPO) (Power et al. 1999), in the Pacific and Indian Ocean basins and the Atlantic Multidecadal Oscillation (AMO) (Kerr 2000) in the Atlantic Ocean. The IPO index corresponds to the standardized PC time serie associated with the first EOF of the SST internal variability in the Pacific basin between 40°S and 60°N. The AMO index is the standardized time serie of the area weighted SST internal variability in the Atlantic basin between 0° and 60°N. SST data are from the ERSST2 dataset (Reynolds et al. 2002). To compute internal variability, we first constructed a CMIP3 Multi-Model temporal pattern characterizing the response of the SST to anthropogenic forcing. This pattern is obtained by applying the procedure described in Sect. 2.3 to the global 10-year low-pass filtered monthly SST from the nine CMIP3 AOGCMs listed Table 2, and by averaging the nine resulting curves. The regression of the observed SST on this temporal pattern can be considered as an estimate of the SST spatial change associated with the anthropogenic forcing. The SST internal variability we use here corresponds to the regression residual.
5 Summary and conclusions
This paper has examined the recent evolution of the mean temperature above the 14°C isotherm in the global ocean, as viewed by five model-based reanalyses and a model-independent objective analysis. The most recent reanalysis assimilates data with new time-varying XBT bias corrections Wijffels et al. (2008). We define the MEAN reanalysis as the average of the four other reanalyses, whose global spread is always lower than 0.04°C. Over the 1965–2005 period, the MEAN reanalysis exhibits a global warming trend of 0.045°C per decade superimposed with decadal variability of 0.028°C standard deviation. Results are very similar for the model-independent objective analysis with a global trend of 0.046°C per decade and decadal variability of 0.030°C standard deviation. In all the basins, the trend of the objective analysis is contained within the spread of the reanalyses trends. Spatially, estimated change associated with anthropogenic forcing is largely dominated by a warming pattern. For most products, cooling zones are limited to small patches in the equatorial band, in the central Pacific around 30°N, and along the African west coast. In a few regions, the anthropogenically forced signal may have been reinforced by internal variability associated with the AMO and IPO. In the Pacific basin (in the western equatorial region and in an area centered around 30°N and 140°W,) both IPO and anthropogenic related changes exhibit cooling patterns. On the contrary, both contributions lead to upper ocean warming in the southern subtropical Pacific. In the North Atlantic basin, the positive AMO phase may also have reinforced the recent warming. With the exception of those regions, the signature of the AMO and IPO internal variability modes differs clearly from the anthropogenically forced signal. The question remains whether there are other important internal decadal variability modes not included in this analysis which can project onto the signature of the long-term changes.
The recent evolution of Tiso14 has been compared with the one of the 250 m mean temperature. Two major advantages of the Tiso14 analysis have been emphasized. First, it filters out temperature change due to vertical isotherm motion, associated both with internal variability and anthropogenic forcing. In the Pacific basin, internally varying isotherm depths are related to the IPO induced fluctuations in the east–west tilt of the thermocline. Anthropogenically forced isotherm motions result in long-term shallowing in the western and central equatorial Pacific and around 10°S in the Indian Ocean. On the contrary, isotherms deepen in the North Atlantic basin, suggesting a net accumulation of heat water in this region. The water balance between the Indian and Pacific low latitudes on the one hand, and the North Atlantic on the other hand may be related to changes in the AMOC (Palmer and Haines 2009). However, our 40-year period of study is too short to draw confident conclusions linked with the global ocean circulation on secular time scales and to determine whether the long-term circulation changes are wind or thermo-haline driven.
The second advantage of Tiso14 is that it is not very sensitive to the XBT bias fall-rate. Our results show that the impact of the XBT’s bias on the spatial pattern of change is quite weak. On the contrary, this bias is a major source of uncertainty for the time evolution of a mean temperature above a fixed depth. The main impact of the XBT correction is the suppression of the accentuated warming between 1970 and 1980, inducing a 34% increase in the T250m trend computed over 1965–2005. Our result is consistent with Domingues et al. (2008) who examine the ocean upper 300 m in an objective analysis based on the EN3 profiles after applying the correction by Wijffels et al. (2008). They found an ocean warming trend about 50% larger than previous estimates (Levitus et al. 2005). Nevertheless, other groups obtain different results using different XBT bias corrections (Ishii and Kimoto 2008; Levitus et al. 2009). The main difficulty in resolving the XBT bias seems to be the lack of accurate metadata, with approximately half of XBTs being of unknown type (Palmer et al. 2010). Research is still underway to establish a consensual correction.
Beyond the issue of the XBT bias, the intercomparison of different ocean analyses allows us to identify other uncertainties associated with the recent upper ocean temperature evolution. The consistency between the observations and the reanalyses results, as well as the spread among the reanalyses, has been used as a simple estimate of ocean state uncertainties. We found that at the end of the estimation period, introduction of altimeter data within the assimilation scheme is responsible for the ECMWF divergence from the other ENSEMBLES reanalyses. Although it is clear that altimeter assimilation has a noticeable effect in the ocean heat thermal structure, and therefore in the representation of ocean heat content trends, it is not easy to say if it produces a more reliable estimation. Therefore, the impact of assimilation on the ocean heat content signals deserves more attention. We would like to underline the limits of such an intercomparison exercise. When possible, we have attempted to propose reasons for the origin of the discrepancies. However, identifying the detailed causes remains difficult. Indeed, many factors can be involved such as model errors, problems associated with the assimilation systems and/or with the atmospheric forcings, biases and scarcity of the observations. Within the ENSEMBLES reanalyses, while the atmospheric forcing fields and the assimilated observations used are the same, the number of possible causes for discrepancies is reduced. Nevertheless, their clear understanding would require additional intercomparison studies currently planned within the Climate Variability and Predictability (CLIVAR) Global Synthesis and Observations Panel (GSOP) (Stammer 2006).
In general, the results presented here on the comparison of Tiso14, T250m and Diso14 support the findings of Palmer et al. (2007). The use of reanalyses, as a dynamically based means of combining the in situ observations with the best available estimates of forcing fields, makes it possible to extend the Tiso14 analysis in the data sparse region. Moreover, we argue that the regression based method presented here allows an optimal separation of the anthropogenically forced changes from internal variability. Future work will require determining whether the temporal pattern characterizing the anthropogenic climate change (μ) is significantly contained in the observations and reanalyses, by using a statistical test developed by Ribes et al. (2009). This next step will constitute a detection study of climate change in the upper ocean. In order to carry out the detection study on a larger latitude domain, careful examination of the appropriate variable to examine will be needed. For example, using the temperature integrated throughout the mixed layer depth could present the same advantage as Tiso14 as a dynamical filter, but without the need to exclude the high latitudes.
The ENSEMBLES data used in this work was funded by the EU FP6 Integrated Project ENSEMBLES (contract number 505539) whose support is gratefully acknowledged. In particular, the authors thank Philippe Rogel for his assistance and advice with the reanalyses. The SODA data were obtained from the IRI/LDEO Climate Data Library Web site (http://ingrid.ldeo.columbia.edu/SOURCES/.CARTON-GIESE/.SODA/.v2p0p2-4/). We are very grateful to Simon Good who provided the EN3_v2a version of the ENACT/ENSEMBLES quality-controlled dataset, and Matthew Palmer who provided the code we used to filter the data. We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP’s Working Group on Coupled Modelling (WGCM) for their roles in making available the WCRP CMIP3 multi-model dataset. Support of this dataset is provided by the Office of Science, U.S. Department of Energy. We thank Gilles Reverdin, Thierry Delcroix, Sophie Cravatte and Gael Alory for stimulating discussions on this work, and anonymous reviewers whose comments helped to improve this manuscript. The figures were produced with the NCL software developed at NCAR.