Detection and attribution of Atlantic salinity changes



[1] An analysis of observed and modeled oceanic salinity changes shows that significant changes of salinity, which are predicted in the World's oceans as a result of human influence, are beginning to emerge. A significant increase in salinity has been observed in recent decades in the 20N–50N latitude band of the Atlantic ocean, although changes at sub-polar latitudes of the Atlantic, and in other ocean basins, are not found to be significant compared to modeled internal variability. An optimal detection analysis of spatial patterns of salinity trends detects a human influence on the observed salinity increases in the Atlantic ocean. These results indicate the growing potential for using observations to constrain important properties of the climate system's response to anthropogenic forcing.

1. Introduction

[2] Documenting and understanding the changes taking place in the global hydrological cycle is arguably one of the most important challenges in contemporary climate science. In this context, the properties of ocean water masses are an important indicator, and changes in salinity are of special interest since they can be substantially influenced by changes in the surface fresh water flux [e.g., Stark et al., 2006].

[3] The linear trends of salinity in the upper 500 m of the World's ocean basins since 1955 were documented by Boyer et al. [2005], and the changes discussed by Bindoff et al. [2007] included a freshening of high-latitude regions and increases in salinity in near-surface waters in the more evaporative regions (notably the sub-tropical regions) in almost all ocean basins. Bindoff et al. [2007] concluded that the observed changes are consistent with a change in the Earth's hydrological cycle, notably changes in precipitation and inferred larger water transport in the atmosphere from low to high latitudes and from the Atlantic to the Pacific. Prominent changes in the salinity of the best observed ocean basin, the Atlantic, over the last 50 years have been extensively documented [e.g., Curry et al., 2003; Boyer et al., 2005; Polyakov et al., 2005]. Greatest attention has focused on freshening of the sub-polar North Atlantic, which proceeded most rapidly in the 1970s [e.g., Dickson et al., 2002; Curry and Mauritzen, 2005], but increases in the salinity of the upper North Atlantic at low latitudes [e.g., Curry et al., 2003; Boyer et al., 2005] are another notable feature of the observations.

[4] Increases in net precipitation (including runoff) favour freshening at high latitudes, and decreases in net precipitation favour salinification in the tropics and subtropics [Held and Soden, 2006; Zhang et al., 2007; Wu et al., 2005] and this apparent agreement with observed Atlantic salinity changes has led to speculation that the observed changes are indeed a signature of anthropogenic climate change [e.g., Curry et al., 2003]. However, recent studies using the HadCM3 climate model have cast doubt on the idea that observed ocean freshening at high latitudes might be attributable to anthropogenic forcing. Pardaens et al. [2008] suggest that the high latitude freshening is most likely dominated by natural unforced variability, closely associated with the Meridional Overturning Circulation. Furthermore, Pardaens et al. [2008] show that in HadCM3 anthropogenic forcing does not lead to high latitude freshening, because additional fresh water input from enhanced precipitation is balanced by increased northward transport of salinity from lower latitudes. Wu and Wood [2008] also suggest that sub-polar North Atlantic freshening can come from a freshwater re-distribution within the Arctic/sub-polar North Atlantic, without invoking large-scale changes in surface freshwater flux. For low latitudes, however, Pardaens et al. [2008] found that anthropogenically forced decreases in net precipitation do tend to give decreasing freshwater content (i.e. increasing salinity), and they suggest that the observed changes in this region are likely to have exceeded the range of natural unforced variability.

[5] Motivated by this previous work, the purpose of this study is to investigate in more detail the extent to which an anthropogenic fingerprint of ocean salinity changes can be detected in observational data sets. An important component of this analysis is the use of climate model predictions to determine how the anthropogenic fingerprint might strengthen in future.

2. Observational Data and Model Simulations

[6] In this paper we compare two observational data sets, that of Boyer et al. [2005] and the ocean analyses of Smith and Murphy [2007], with data from the HadCM3 climate model.

[7] The ocean analyses of Smith and Murphy [2007] were used to initialise the DePreSys decadal forecasting system described by Smith et al. [2007]. These analyses are available on the HadCM3 ocean grid and were created by four-dimensional (horizontally, vertically and in time), multivariate optimal interpolation of salinity and sub-surface temperature observations and analysed sea surface temperature from HadISST [Rayner et al., 2003], using covariances from HadCM3. The analyses are for the full depth of the ocean, and will be close to climatology in the deep ocean where data are particularly sparse. The other data set we analyse is the Boyer et al. [2005] data set derived from the World Ocean Database 2001 of quality controlled salinity measurements. Salinity anomalies were calculated for each 1o by 1o grid box according to the methodology described by Boyer et al. [2002].

[8] The HadCM3 simulations we analyse are a set of four simulations that include anthropogenic forcings and that are initialised in 1860 and continue to 2100 (denoted ANT), a set of four naturally forced simulations from 1860 to 2000 (denoted NAT) and a long control run of HadCM3 in which external forcings are held constant to represent the changes possible due to internal variability only. The natural forcings considered before 2000 are changes in total solar irradiance and changes in stratospheric aerosols following explosive volcanic eruptions [Stott et al., 2000]. Anthropogenic forcings considered in simulations up to 2000 are observationally based estimates of changes in well mixed greenhouse gases and changes in aerosols and consequent changes in cloud brightness. For this analysis we require simulations that continue seamlessly through the period of the observational data sets and on into the future, so that an assessment of the potential for exploiting the growing signal of climate change can be made. The only set of four simulations available to us, that satisfy this requirement and which all include common anthropogenic forcings to 2000, make different assumptions about emissions in future anthropogenic and past natural forcings. Anthropogenic forcings in future follow the SRES B2 scenario after 2000 in two of the simulations and follow the SRES A2 scenario in the other two simulations. In addition, one of the simulations also includes natural forcings to 2000. These different assumptions are found not to have a noticeable impact on the evolution of anthropogenically forced salinity anomalies analysed in this paper (see Figure 1).

Figure 1.

Observed and modeled salinity changes averaged over the top 500 m and over three latitude bands (20S–20N, top; 20N–50N, middle; 50N–66N bottom) in practical salinity units (psu). For the model simulations, solid lines show the mean of four simulations that include anthropogenic forcings (red) and the mean of four simulations that include only natural forcings (green). Broken lines show the individual simulations.

[9] Models contain long term drifts in ocean data that could contaminate any climate change signal. Since each transient simulation is initialised from a particular point in the control run and is then run forward in parallel with an equivalent segment of the control simulation, removing any long term drift that is not externally forced is enabled by subtracting the polynomial fit to the corresponding control segment at each grid box. This detrending procedure will remove some component of internal variability, but, to avoid biasing the results, that same second order polynomial fit is also subtracted from parallel segments of the control simulation.

3. Analysis

[10] Figure 1 shows the observed and modeled changes in salinity averaged over the top 500 m and three latitude bands (20S–20N, 20N–50N and 50–66N) of the Atlantic ocean, the best observed of the ocean basins. These regions were chosen to represent the main regions of the relatively well observed upper ocean, including the tropics, and the sub-tropical and sub-polar gyres in the northern Atlantic. The models predict large salinity increases in future in the 20S–20N and the 20N–50N bands with the ensemble mean showing an increasing trend in both latitude bands. The observational data sets show evidence that the predicted trends are beginning to emerge, with both data sets showing the trend starting to increase slightly sooner than modeled in the 20–50N latitude band. The 40-yr trend to 2006, when averaged over this 20N–50N region (0.070 psu over 40 years), is significantly greater than expected from natural internal variability, being greater than the 95th percentile of 40-yr trends from the HadCM3 control run (0.052 psu). This finding is robust to not removing a second order polynomial from the control trends. Significant 40-yr trends have not been observed in other ocean basins. The model simulations predict freshening in both Pacific and Indian oceans although such trends are not expected to emerge clearly until later this century (auxiliary material Figure S1). We therefore focus the remainder of the analysis on the Atlantic ocean.

[11] There is broad agreement between the Atlantic salinity trend patterns from the two observational data sets (Figure 2), both showing freshening at high latitudes and salinity increases at lower latitudes. The model ensemble means do not capture the observed pattern although there is considerable internal variability at high latitudes and some model ensemble members do show freshening in this region. At lower latitudes, model simulations show an increasingly strong signal of salinity increases in the tropics and subtropics (Figure 2), such that by the 2007–2046 period large parts of the tropics and NH sub-tropics show significant changes in individual model simulations over 5 degree by 5 degree areas (auxiliary material Figure S2).

Figure 2.

Observed and modeled 40-yr trend patterns of salinity (in psu) averaged over the top 500 m. (top) Trends for 1957–1996. (bottom) from left to right: observed and modeled trends for 1967–2006, and modeled predicted trends for 1987–2026 and 2007–2046.

[12] While the large internal variability in the high latitudes precludes the identification of significant trends in Atlantic salinity up to present, model simulations do indicate that the North Atlantic is expected to systematically freshen after about 2010, although the large amount of internal variability means that individual realisations can vary greatly from the ensemble mean freshening trend. The freshening trend reverses in the middle of the century, likely as a consequence of transport of salinity from lower latitudes [Pardaens et al., 2008].

[13] To analyse further the extent to which significant changes could be detected as being significantly different from those expected from natural internal variability and have a significant contribution from anthropogenic or natural forcings, a standard optimal detection analysis is carried out [e.g., International Ad Hoc Detection and Attribution Group, 2005]. Such techniques have been applied extensively to atmospheric temperature changes, and are being increasingly applied to other climate variables [Hegerl et al., 2007]. We calculate spatial patterns of the expected changes (the “fingerprints”) by averaging the 40-yr salinity trends over the top 500 m (i.e., the patterns shown in Figure 2) for 10 degree latitude bands from 50S to 50N. The model fingerprints are then regressed against the corresponding observed patterns in a standard “total least squares” optimal detection analysis [e.g., Stott et al., 2006]. Simulated and observed spatio-temporal patterns are projected onto the leading EOFs of the first half of the HadCM3 control simulation, and uncertainties in the regression coefficients are estimated using the second half of the control simulations, as is standard practice in such analyses [e.g., see Hegerl et al., 2007, Appendix 9.A.1]. The optimal detection analysis calculates scaling factors by which the modeled response can be scaled up or down while still remaining consistent with the observed response. Where the 5th percentile of a scaling factor is greater than zero, the postulated climate change signal is said to be detected (with a 5% chance of a type 1 error, ie that the null hypothesis of no influence of the postulated climate change signal is correct).

[14] Figure 3 (top) shows the results for both the 1957–1996 and 1967–2006 periods when the anthropogenic and natural fingerprints are regressed separately against the observed data sets. For the earlier period neither anthropogenic nor natural factors are detected (i.e. the 5th percentiles of the scaling factors are not greater than zero and we cannot reject the null hypothesis that the observed changes are consistent with internal variability). Consistent results are obtained for the different observational data sets (solid and dashed lines in Figure 3) for the anthropogenic scaling factors while for natural scaling factors very large ranges are found, indicating that the contribution of the noisy natural fingerprint in the observations is very uncertain. Ten years on, anthropogenic factors are (just) detected in the observed 1967–2006 salinity trends, and the anthropogenic scaling factor has an amplitude consistent with one.

Figure 3.

(top) Scaling factors obtained from optimal detection for one way regressions on the anthropogenic and natural signals for (left) 1957–1996, calculated using the Smith and Murphy [2007] data set, solid line, and Boyer et al. [2005] data set, dashed line, and (right) 1967–2006. (bottom) Scaling factors for simultaneous two way regression on anthropogenic factors (x axis) and natural factors (y axis) for the 1967–2006 period where the anthropogenic fingerprint is estimated from the modeled trend for 2007–2046.

[15] When both anthropogenic and natural spatial fingerprints are regressed simultaneously against the observed changes, large ranges of scaling factors are found for both anthropogenic and natural factors and neither is detected. Whilst it is not possible to detect the influence of anthropogenic factors in the presence of natural factors when a spatial fingerprint is used that is derived from model simulations of trends to present, detection of anthropogenic factors is found if a spatial fingerprint pattern is taken from model simulations for future trends when the anthropogenic response has strengthened and the noise contamination is reduced. Results are shown in Figure 3 (bottom) for an anthropogenic fingerprint estimated from the modeled trend for 2007–2046 and the observed 1967–2006 trend pattern. Note that the best estimate of the anthropogenic scaling factor is less than one as is expected to be the case for the amplitude of a future fingerprint in the observed spatial trend pattern to present. This approach requires that such anthropogenic changes grow approximately linearly with time which model predictions indicate is expected to be the case, at least until the mid 21st century (Figure 1 and auxiliary material Figure S1).

[16] For all successful detection results, a residual consistency test [Allen and Tett, 1999] is passed, indicating no evidence for inconsistency between the estimated attributed response and the observed response. A further check on the ability of the model to simulate internal variability (auxiliary material Figures S3a–S3c) shows no inconsistency between modeled simulation of internal variability and that observed over the Atlantic regions analysed.

4. Summary and Discussion

[17] An analysis of observed and modeled oceanic salinity changes shows that significant changes of salinity, which are predicted in the World's oceans as a result of human influence, are beginning to emerge. A significant increase in salinity has been observed in recent decades in the 20N–50N latitude band of the Atlantic ocean, and an optimal detection analysis of spatial patterns of salinity trends detects a human influence on the observed salinity increases in the Atlantic ocean. Earlier indications of freshening trends in the high latitudes of the North Atlantic are not borne out by more recent observations which show salinity increases in recent years, but in any case the observed changes at high latitudes are consistent with natural internal variability. Significant salinity changes in other ocean basins are not predicted to be observed until later in the 21st century. However, in the sub-tropics of the Atlantic, significant trends appear to have emerged slightly sooner than predicted, underscoring the importance of monitoring the continuing evolution of the World's oceans. Salinity changes could be used, alongside other data sets [e.g., Santer et al., 2007], to infer changes in the hydrological cycle over the ocean, where there is less direct observational data measuring changes in precipitation and evaporation. Our analysis shows that the human fingerprint of salinity changes is likely to become more distinct over the next few years. There are therefore good prospects for further exploitation of ocean data, including, for example, the global array of ARGO floats and satellite data from the SMOS (Soil Moisture and Ocean Salinity) and Aquarius missions, in order to better constrain characteristics of the climate system's response to anthropogenic forcing.


[18] P.A.S. and D.M.S. were supported by the Joint Defra and MoD Integrated Climate Programme GA01101, CBC/2B/0417 Annex C5, and R.T.S. by a Royal Society University Research Fellowship. We thank Jon Robson for help with data, and two anonymous referees, Anne Pardaens, Richard Wood and Peili Wu for helpful comments.