Geophysical Research Letters

The fingerprint of human-induced changes in the ocean's salinity and temperature fields

Authors

  • David W. Pierce,

    Corresponding author
    1. Division of Climate, Atmospheric Sciences, and Physical Oceanography, Scripps Institution of Oceanography, La Jolla, California, USA
    • Corresponding author: D. W. Pierce, Division of Climate, Atmospheric Sciences, and Physical Oceanography, Scripps Institution of Oceanography, Mail Stop 0224, La Jolla, CA 92093-0224, USA. (dpierce@ucsd.edu)

    Search for more papers by this author
  • Peter J. Gleckler,

    1. Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California, USA
    Search for more papers by this author
  • Tim P. Barnett,

    1. Division of Climate, Atmospheric Sciences, and Physical Oceanography, Scripps Institution of Oceanography, La Jolla, California, USA
    Search for more papers by this author
  • Benjamin D. Santer,

    1. Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California, USA
    Search for more papers by this author
  • Paul J. Durack

    1. Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California, USA
    Search for more papers by this author

Abstract

[1] The ocean's salinity field is driven primarily by evaporation, precipitation, and river discharge, all key elements of the Earth's hydrological cycle. Observations show the salinity field has been changing in recent decades. We perform a formal fingerprint-based detection and attribution analysis of these changes between 1955–2004, 60°S and 60°N, and in the top 700 m of the water column. We find that observed changes are inconsistent with the effects of natural climate variability, either internal to the climate system (such as El Niño and the Pacific Decadal Oscillation) or external (solar fluctuations and volcanic eruptions). However, the observed changes are consistent with the changes expected due to human forcing of the climate system. Joint changes in salinity and temperature yield a stronger signal of human effects on climate than either salinity or temperature alone. When examining individual depth levels, observed salinity changes are unlikely (p < 0.05) to have arisen from natural causes over the top 125 m of the water column, while temperature changes (and joint salinity/temperature changes) are distinct from natural variability over the top 250 m.

1. Introduction

[2] Changes in the planet's hydrological cycle will alter the ocean's salinity field, which in the tropics and mid-latitudes is forced primarily by evaporation (E), precipitation (P), and river discharge (as well as by the wind via advection [Yu, 2011]). The ocean's salinity field is a logical place to look for the effects of changes in the hydrological cycle, since 71% of the Earth's surface is covered by the oceans and the quality and coverage of E and P observations limit our ability to directly examine the surface freshwater flux across much of the globe. Additionally, the salinity field integrates high-frequency E and P variability, thus reducing noise [e.g.,Terray et al., 2012].

[3] In this work we examine whether human-induced climate change has affected the ocean's salinity field by performing a formal, fingerprint-based detection and attribution (D&A) analysis of observed world ocean salinity changes using 20 state of the art global climate models from the CMIP5 archive [Taylor et al., 2012]. Detection determines if observed changes are likely to have arisen from natural causes, while attribution determines if observed changes are consistent human forcing of climate [e.g., Intergovernmental Panel on Climate Change, 2007]. We expand upon previous regional D&A salinity studies [Stott et al., 2008; Terray et al., 2012] by examining salinity changes across the globe and by depth over the top 700 m, and formalize the attribution of observed salinity changes [cf. Durack et al., 2012].

[4] Previous work has shown there are trends in ocean salinity consistent with surface warming, an amplification of the hydrological cycle, and model-estimated E and P fields [Boyer et al., 2005; Hosoda et al., 2009; Durack and Wijffels, 2010; Helm et al., 2010; Durack et al., 2012]. The formal D&A methodology is useful because it evaluates the likelihood of observed changes arising due to natural variability, and whether the changes are specifically consistent with human forcing of the climate.

[5] The ocean's salinity field (S) is dynamically intertwined with the temperature field (T), which together determine the density. Simultaneously considering multiple variables that are physically linked can increase the strength of the model-predicted signal in the observations [cf.Santer et al., 1995; Barnett et al., 2008], so we also examine changes in ocean temperature, both by itself and jointly with salinity.

[6] This work extends previous D&A studies of ocean temperature [Barnett et al., 2001, 2005; Pierce et al., 2006; Gleckler et al., 2012] by examining the temperature changes on a finer spatial scale, comparing observations to the most recent generation of global climate models, and using ocean temperature datasets that have been corrected for warm instrument biases in the mid 1970s–1980s [e.g., Gouretski and Koltermann, 2007; Domingues et al., 2008; Wijffels et al., 2008; Lyman et al., 2010] (for a summary see Pierce et al. [2011]). Gleckler et al. [2012] concluded that temperature D&A is relatively insensitive to which corrected data set is used, so we use only one here [Levitus et al., 2009, 2012]. We show that eliminating the spurious 1970–80s warming in the observations slightly increases the agreement between model-predicted and observed ocean warming.

2. Data and Methods

[7] The D&A methodology follows that used in Barnett et al. [2005] and described in more detail in Pierce et al. [2006]. Details are given in the auxiliary material. Briefly, we compare three-dimensional temperature and salinity fields from 20 global climate models from the CMIP5 archive [Taylor et al., 2012] (Table S1) to 50 years of observations (1955–2004) from the National Oceanographic Data Center [Levitus et al., 2012]. S and T fields are regridded to a uniform 10° × 10° latitude/longitude grid between 60°S and 60°N over the top 700 m of the water column, then decadally averaged (1955–1964, … , 1995–2004) to reduce the effects of natural interannual variability not of interest here.

[8] Model fields are masked so gridboxes lacking observations are excluded. This avoids using infilled temperature estimates [AchutaRao et al., 2007; Gleckler et al., 2012]. Values are further weighted proportionally to the number of observations in the gridbox, so well-observed locations count more towards the result. Model fields are de-drifted using second order polynomials fit to the pre-industrial control runs. As shown byGleckler et al. [2012], the effect of fitting polynomials of different order had relatively little impact on an ocean temperature D&A analysis.

[9] A combined space-time fingerprint of anthropogenic effects on S and T is used, computed from the historical runs that include anthropogenic forcing as the single space-time pattern that maximizes the joint variance across all models. Each model's S and T field was normalized by the variance in the first half of its own pre-industrial control run. The second half of the model's pre-industrial control run was used to estimate the amplitude of natural variability. The dimensionless “signal strength” is the projection of a 50-year period of any one model run (or the observations) onto this joint space-time fingerprint (seePierce et al. [2006] for details).

3. Results

3.1. Amplitude of Natural Variability

[10] The credibility of results from a D&A analysis depends on the models' ability to simulate the observed pattern and amplitude of natural internal climate variability. This is difficult to evaluate given the relatively short record lengths and convolution of signal and noise in the observations. Here we compare the temporal standard deviations in the models' pre-industrial control run to the detrended observed time series (which assumes any anthropogenic signal is well-represented by a linear trend). The model time series were detrended in the same way for consistency. By this measure, the models have realistic salinity variability in the surface layers, but less than observed at 300 m and deeper, especially in the poorly sampled region around Antarctica (Figure S1). The models have a realistic estimate of temperature variability in the surface layers, but tend to overestimate the amplitude of temperature variability at 150 m and deeper (Figure S2). Since our analysis weights by the number of observations (which drops with depth) and the temperature and salinity biases are offsetting, there is no evidence these model deficiencies bias our signal-to-noise estimates, in accord with previous results [e.g.,Santer et al., 2011; Gleckler et al., 2012].

3.2. Model Versus Observed Trends

[11] The ensemble-averaged model vs. observed salinity and temperature trends in the Pacific, Atlantic, and Indian basins are shown inFigure 1 (zonal transects are shown in Figure S3). On broad scales the observed and multi-model ensemble averaged patterns are similar, although regional details differ. Observations show increasing salinity to ∼300 m depth in the Atlantic and freshening to ∼125 m depth in the Pacific. The ensemble averaged model trend shows similar patterns, but with weaker peak values, perhaps not surprising when comparing the highly averaged multi-model ensemble average with a single realization of the observations. Observations also show a salinity trend maximum in the Northern Indian Ocean, a feature the multi-model ensemble average does not reproduce.

Figure 1.

Basin zonal-mean trends of salinity (upper, PSS-78/50 yrs) and temperature (lower, °C/50 yrs) averaged in the (left) Pacific, (middle) Atlantic, and (right) Indian ocean basins. For each variable, (top) the Levitus observations are shown, and (bottom) the ensemble averaged model result are shown. Note the non-linear depth scale.

[12] Observations show cooling around 125–200 m depth in the Pacific and Indian oceans, along with substantial warming above. The multi-model ensemble average again reproduces these large-scale patterns, although peak values are weaker. The equatorial temperature transect (Figure S3) shows the multi-model ensemble average cooling at depth does not extend into the Indian Ocean, unlike observations.

3.3. Detection and Attribution Results

[13] Figure 2 shows results from the detection and attribution analysis of salinity (Figure 2, top), temperature (Figure 2, middle), and joint temperature/salinity (Figure 2, bottom): confidence intervals are shown as vertical bars, empirically estimated from the distributions of model values. Multiple 50-year segments of the pre-industrial control runs (Table S1) are used to provide estimates of the signal strength expected to arise from natural internal variability alone (open symbols in Figure 2; see also auxiliary material).

Figure 2.

Detection and attribution diagram for changes in the world ocean's fields of (top) salinity, (middle) temperature, and (bottom) joint temperature/salinity. Plotted values are the signal strength (nondimensional) of the multi-model fingerprint of anthropogenic climate changes in 50-yr segments of various model runs and the observations, as indicated in the legend along the bottom. Open symbols indicate values obtained from the indicated model's pre-industrial control run, where the 50-year segments partially overlap, each one starting 10 yrs after the previous segment. Closed symbols are from model runs covering the observed period 1955–2004, either including only natural forcing (solar and volcanic variability), or including both natural and human effects (primarily human-generated greenhouse gases and aerosols), as indicated. Heavy black dots and bars show the multi-model ensemble mean value and 95% confidence interval (90% c.i. for the natural-only case, due to fewer data points). Red symbols show the signal strength and 95% c.i. for the observations, using either the olderLevitus et al. [2005] or newer Levitus et al. [2012] data set, as indicated. The newer data set has been corrected for recently identified temperature biases.

[14] The signal strengths in the observations are shown as the red symbols of Figure 2. One value is computed using the Levitus et al. [2005] data set, which predates the discovery of the 1970s–80s temperature biases, and one with Levitus et al. [2012], which includes corrections. Since our period of analysis ends in 2004 but Levitus et al. [2005] ends in 2003, we appended year 2004 from the newer data set to the Levitus et al. [2005] data for this comparison. The data corrections are primarily in the 1970s–80s [Domingues et al., 2008], so this has negligible effect.

[15] The observed signal strengths are separated from zero (p < 0.05) in both salinity and temperature, as well as in the joint temperature/salinity analysis. Therefore, we conclude there is a detectable signal of climate change in the ocean's salinity and temperature fields that cannot be explained by natural internal climate variability alone.

[16] When the models are run with estimates of historical changes in solar and volcanic forcing only, the distribution of model signal strengths straddles zero (Figure 2). This is inconsistent with the observations, so we conclude that natural solar and volcanic forcing do not explain observed changes in the ocean's salinity and temperature fields over the last 50 years.

[17] The signal strengths obtained when human effects on climate are included are much stronger and well separated from zero (Figure 2). There is considerable scatter in the model results, more than seen in the pre-industrial control run, which reflects a range of climate sensitivity and differences in forcing (particularly indirect aerosol effects) across the models. The multi-model ensemble average signal strength (heavy black dots) is consistent with observations for all variables (salinity, temperature, and joint temperature/salinity), given the uncertainties. We conclude that observed changes in the ocean's salinity and temperature fields are consistent with the changes expected due to human forcing of the climate.

3.4. Joint Temperature and Salinity

[18] Comparing the individual salinity (S) and temperature (T) results to the joint T/S analysis (Figure 2), it can be seen that simultaneously using different variables that are physically linked increases the signal strength. Since S and T noise is not perfectly correlated, it becomes even less likely internal variability alone could explain the joint, multi-decadal T and S changes. The multi-model median signal strength increases from 5.5 (salinity) and 9.0 (temperature) to 9.8 (joint T/S). The observed signal strength rises as well, from 4.8 and 8.6 (for S and T) to 9.2 (for joint T/S).

3.5. Effect of Temperature Corrections

[19] Figure 2 shows that using the corrected observations results in slightly better agreement between the models and observations than using the older, uncorrected data (red symbols marked “2012” vs. red symbols marked “2005”). Biased observational values produced a mid 1970s–80s warming that fell outside the main envelope of climate model predictions [e.g., Barnett et al., 2001, Figure 1]. When a full space-time fingerprint is used, the disagreement in the 1970s–80s degrades the agreement between the models and uncorrected observations. In particular, the erroneous 1970s–80s warming in the older data sets does not bias the D&A results towards an overconfident detection of human effects on the world's oceans. Rather, the temperature detection and attribution results are relatively insensitive to uncertainties associated with measurement biases [cf.Lyman et al., 2010], consistent with Gleckler et al. [2012].

3.6. Detection and attribution as a function of depth

[20] Figure 3 shows the signal strengths when the D&A analysis is performed as a function of depth, where the fingerprint and signal strengths are calculated independently at each depth level. The salinity signal is distinct from natural variability (p < 0.05), either internal or external (solar and volcanic fluctuations) to the climate system, over the top 125 m (solid red dots). The temperature signal is distinct from natural variability over the top 300 m, and has a stronger signal than salinity in the upper 250 m of the water column. For both temperature and salinity, observed values nearly always fall within the central 25th–75th percentile of the model distribution. As previously, the joint T/S signal is stronger than found in either S or T separately.

Figure 3.

Detection and attribution diagram as a function of depth (m). Green shaded area: 95% confidence interval of signal strengths from natural internal climate variability, estimated from the control model runs with no anthropogenic forcing. Red dots: observations, plotted as solid red if the value is inconsistent with natural internal climate variability (p < 0.05), and open otherwise. (top) Blue shaded box and whiskers show the distribution of signal strengths when the models include both human (anthropogenic greenhouse gases and aerosols) and natural external (solar and volcanic) forcing. (bottom) Tan box and whiskers show the signal strength when only natural external forcing is included.

4. Summary

[21] The ocean's salinity field is a logical place to look for changes in the hydrological cycle, since it is driven by evaporation, precipitation, and runoff, integrates high-frequency weather variability, and covers 71% of the Earth's surface. We have performed a formal detection and attribution analysis of observed upper level (0–700 m) changes in the ocean's salinity (S) and temperature (T) fields, taken both individually and jointly, from 60°S to 60°N, over the period 1955–2004. The observations were compared to over 11,000 years of model simulations from 20 of the newest generation of global climate models. The results show there has been a detectible change in the ocean's global salinity field, and the observed changes are inconsistent with natural causes, either internal to the climate system (such as the El Niño/Southern Oscillation and the Pacific Decadal Oscillation) or external (the sun, volcanoes). Changes are consistent with those expected from human effects on the climate, which arise primarily from anthropogenically-induced changes in greenhouse gases and aerosols. Similar results are found for the temperature field. When salinity and temperature changes are analyzed together, an even stronger signature of human forcing on the ocean emerges. These results add to the evidence that human forcing of the climate is already taking place, and already changing the climate in ways that will have a profound impact on people throughout the world in coming decades.

Acknowledgments

[22] As a member of the International Detection and Attribution Working Group (IDAG), DWP acknowledges partial support from the US Department of Energy's Office of Science, Office of Biological and Environmental Research, grant DE-SC0004956 and the National Oceanic and Atmospheric Administration's Climate Program Office. Work undertaken at Lawrence Livermore National Laboratory is supported by the U.S. Department of Energy under contract DE-AC52-07NA27344. We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modeling groups (listed inTable S1 of this paper) for producing and making available their model output. For CMIP the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals.

[23] The Editor thanks two anonymous reviewers for their assistance in evaluating this paper.

Ancillary