1

Measuring oxygen concentrations improves the detection capabilities of

an ocean circulation observation array

Catherine E. Brennan1, Richard J. Matear2, and Klaus Keller1*

1Department of Geosciences, Penn State, University Park, PA

2CSIRO Marine and Atmospheric Research, Hobart, Tasmania, Australia

*corresponding author e-mail: kkeller@geosc.psu.edu

Running Title:

BRENNAN ET AL.: OXYGEN OBSERVATIONS IMPROVE MOC DETECTION

2

Abstract

The North Atlantic meridional overturning circulation (MOC) may weaken or even

collapse in response to anthropogenic climate forcing, with potentially nontrivial

socioeconomic impacts. One currently implemented MOC observation system uses

temperature and salinity (as well as other) observations along a zonal transect in the

North Atlantic. The resulting MOC estimate has, however, a relatively low signal-to-

noise ratio due to large internal variability and observation errors. Observations of

hydrographic tracers that are mechanistically linked to MOC changes may increase the

signal-to-noise ratio. An MOC slowdown is associated in model simulations with a

shoaling of the boundary between North Atlantic Deepwater (NADW) and Antarctic

Bottom Water (AABW). This shoaling results in detectable trends in water mass tracers.

Here we deploy a virtual observation array into a numerical model starting in model year

2006 to test whether observing the apparent oxygen utilization (AOU) in addition to the

MOC estimate improves detection capabilities. Our detection method accounts for

observation errors, autocorrelated variability, and uncertainty about the initial conditions.

Neglecting the effects of observation errors and the uncertainty about the initial

conditions results in artificially early detection times. The MOC signal alone enables

reliable detection in roughly five decades. Adding AOU observations reduces this

detection time by approximately 40%.

Index Terms: ocean observing systems, abrupt/rapid climate change, thermohaline,

water masses.

Additional Keywords: meridional overturning circulation, apparent oxygen utilization,

climate change detection, ocean observation system.

3

1. Introduction

The geologic record and model simulations suggest that the North Atlantic meridional

overturning circulation (MOC) may weaken or even collapse in response to climatic

forcing [Alley, et al., 2003; Dansgaard, et al., 1993; Gregory, et al., 2005; Manabe and

Stouffer, 1994; Stommel, 1961]. The current predictions about the future fate of the MOC

are, however, deeply uncertain [Gregory, et al., 2005; Latif, et al., 2000; Marotzke, 2000;

Schmittner, et al., 2005; Wunsch, 2006]. MOC changes could be associated with

considerable ecological and economic impacts [Keller, et al., 2007b; Link and Tol, 2004;

Tol, 1998; Vellinga and Wood, 2002]. Reducing the uncertainty about potential

anthropogenic MOC changes has potentially large economic value [Keller, et al., 2007a],

but poses nontrivial scientific and operational challenges [Baehr, et al., 2007; Bryden, et

al., 2005; Santer, et al., 1995].

Detecting MOC changes is complicated by the large internal variability of the MOC

[Bentsen, et al., 2004], sparse spatiotemporal observations [Bryden, et al., 2005], sizeable

observation errors [Ganachaud, 2003; Talley, et al., 2003], and uncertainties in the

expected spatiotemporal fingerprint of MOC changes [Gregory, et al., 2005]. Studies

analyzing the task of detecting potential MOC changes typically focus on a single

quantity – the maximum meridional water flow in the North Atlantic (or across a given

latitude, such as 26 °N) [Baehr, et al., 2007; Keller, et al., 2007a; Santer, et al., 1995].

The recently installed MOC observation array at 26 °N [Marotzke, et al., 2002], for

example, provides MOC estimates using oceanic temperature (T) and salinity (S)

observations (as well as ancillary information, e.g., the flow through the Florida Strait)

[Baehr, et al., 2007; Hirschi, et al., 2003]. Additional hydrographic tracers (such as

4

CFC, oxygen, phosphate or derived quantities) contain important information about

ocean circulation changes [Broecker, et al., 1998; Ganachaud and Wunsch, 2000; Keller,

et al., 2002; Schlosser, et al., 1991]. In particular, changes in the formation rate of North

Atlantic Deepwater (NADW) – an important MOC component – are associated with

temporal trends in water mass tracers [Broecker, et al., 1998]. Oxygen observations have

been used to characterize decadal-scale variability in subtropical Atlantic water masses

[e.g., Garcia, et al., 1998]. Here we analyze whether adding deepwater observations of

apparent oxygen utilization (AOU) would improve the detection capabilities of an MOC

observation system. We choose AOU (the difference between the oxygen concentration

at saturation for a particular water mass and the observed oxygen concentration) as

previous work suggests that this tracer is (i) a sensitive indicator for circulation changes

[e.g., Bopp, et al., 2002; Matear and Hirst, 2003; Plattner, et al., 2002] and (ii) can be

observed with observation errors that are small relative to the expected trends [Joos, et

al., 2003; Keller, et al., 2002]. AOU trends have the advantage over oxygen trends in

that they are – to first approximation – not affected by changes in oxygen solubility due

to potential changes in temperature or salinity.

Changes in NADW formation rates affect the water mass composition in the

subtropical North Atlantic. The subtropical deepwaters in the North Atlantic are a

complex mix of Antarctic Bottom Waters (AABW) from the Southern Ocean and

overlying NADW [Mantyla and Reid, 1983, Bryden et al, 1996]. The boundary between

these water masses has previously been hypothesized to be a useful indicator for the

relative strength of deepwaters originating from the Northern and Southern hemispheres

[Sutherland, et al., 2001]. A decrease in the relative contribution of NADW to the

5

subtropical deepwaters would result in an upward shift (shoaling) of the AABW/NADW

boundary. Such a shift would result in distinct temporal trends in hydrographic tracers.

For example, AABW is colder, fresher, and has higher values of AOU than NADW

[Reid, 2005]. One prediction from this simple model is that a shoaling of the

AABW/NADW boundary would result in decreased temperatures and salinities and

increased AOU concentrations in a fixed volume around the current AABW/NADW

boundary.

Detection of anthropogenic MOC changes addresses the question of whether the

current MOC observations are outside the range of natural variability, thus implying

anthropogenic change. Previous MOC detection studies have broken important new

ground [Baehr, et al., 2007a,b; Hu, et al., 2004; Keller, et al., 2007a; Santer, et al., 1995;

Vellinga and Wood, 2004] but remain silent on several important questions. Hu et al.

[2004], Santer et al. [1995] and Vellinga and Wood [2004] neglect the effects of

observation errors and scenario uncertainty. Observation error is a measure of how

accurately a quantity is observed, while scenario uncertainty accounts for the fact that a

single model realization is only one possible representation for a given set of uncertain

initial conditions. Detection systems that assume perfect observations are inherently

overconfident. Baehr et al. [2007a] and Keller et al. [2007a] account for observation

error and internal variability, but consider only an MOC signal. Baehr et al [2007b]

analyze a statistical fingerprint using T and S observations, but are silent on the potential

utility of oxygen observations to improve detection.

Here we improve on previous studies in three main ways. First, we select signals that

are arguably feasible to observe in the ocean. Second, we account for observation errors

6

and scenario uncertainty. Third, we use a hydrographic tracer (deepwater oxygen in the

form of AOU) that might have a higher signal-to-noise ratio and hence enable an earlier

detection. We analyze how much the addition of AOU observations might improve the

MOC change detection capability of an MOC observation array at 26 °N [Baehr, et al.,

2007; Marotzke, et al., 2002]. In our study, we sample the maximum transport at 26 °N

directly from the analyzed model, a quantity that in the actual RAPID array is

reconstructed from vertical T and S profiles (and other observations, cf. Marotzke et al.

[2002] and Hirschi et al. [2003]). The MOC signal in our analysis hence contains

information that would be provided in the actual RAPID array from T and S

observations. The key point of our study is to address the question how adding

information derived from oxygen observations to the information derived from T and S

observations may improve the detection capabilities.

We analyze the detection capabilities of virtual observations from two experiments in

the Commonwealth Scientific and Industrial Research Organization (CSIRO) climate-

biogeochemical model [Matear and Hirst, 2003]: (i) a ‘control’ scenario with greenhouse

gas concentrations fixed at pre-anthropogenic levels and (ii) a ‘forced’ scenario with

increasing greenhouse gas concentrations according to the IS92a scenario. We assess the

detection capabilities in terms of detection time – the year when a statistically significant

detection occurs. We analyze the detection skills of hypothetical observation arrays at 26

oN that observe (i) just the MOC, (ii) just deepwater AOU in the western basin, or (iii)

both signals. We show that detection times increase as observation errors and scenario

uncertainty are considered. Combining the AOU and MOC signals in a statistically

optimal fingerprint can improve detection times by almost two decades in the model.

7

2. Data

We analyze simulation results of the CSIRO climate model [Gordon and O'Farrell,

1997; Hirst, et al., 2000]. This model couples modules representing the atmosphere,

ocean, sea-ice, and land components as well as ocean biogeochemistry [Matear and

Hirst, 1999]. The climate model includes flux adjustments between its atmospheric and

ocean components. The ocean component is based on Cox [1984] with a 5.6° x 3.2°

(longitude x latitude) resolution, 21 vertical levels, Gent and McWilliams [1990] eddy

parameterization, isoneutral diffusivity of 1000 m2 s-1, and vertical diffusivities following

Gargett [1984]. The ocean biogeochemical model uses archived monthly mean values of

temperature, salinity, currents, wind speed, and sea-ice coverage to simulate the cycling

of dissolved oxygen and phosphate in the ocean. (The use of monthly resolved forcings

is important in this analysis to approximately resolve seasonal variability and the

potential artifacts due to aliasing of the seasonal cycle.) Production and export of

particulate organic matter (POM) occurs in the euphotic zone as a function of solar

radiation, mixed layer depth, euphotic layer depth, concentration of surface phosphate,

and sea surface temperature [Matear and Hirst, 2003]. POM is remineralized below the

euphotic zone as a function of depth and Redfield stochiometry [Redfield, et al., 1963].

The control experiment is driven by a fixed atmospheric equivalent CO2

concentration of 330 μatm, while the forced experiment applies the IS92a radiative

forcing scenario, increasing the atmospheric equivalent CO2 concentration through year

2083 when a tripling of pre-industrial equivalent CO2 is achieved [Matear and Hirst,

2003]. Beyond 2083, atmospheric equivalent CO2 levels are held constant at 990 μatm.

The analyzed time series span the (model) years 1880 to 2100.

8

Meridional and zonal maps of AOU (Figure 1) show the approximate location of the

boundary between NADW and AABW along the 90 μmol kg-1 contour line in model year

2006 (solid line). This water mass boundary shoals over the next century, displacing the

90 μmol kg-1 contour line upwards by model year 2100 (dashed line). The size of the

shoaling at 26 oN in the model ranges from approximately 90 to 500 m, increasing from

the eastern to the western edge of the basin. The prediction of the simple model was that

(i) a reduction of the MOC intensity leads to a shoaling of the interface between NADW

and AABW and (ii) that this shoaling would result in a cooling, freshening, and an

increase in AOU close to the interface. The trends simulated by the CSIRO model over

the next century are consistent with this prediction (Figure 2).

The strength of the MOC is represented by the maximum of the meridional stream

function at 26 °N in the North Atlantic basin (Figure 3). The deepwater AOU signal is

defined as the zonal mean AOU value between 2000 and 4500 meters in the western

basin (45 °W – 70 °W) at 26°N. We correct for potential artifacts due to model drift by

estimating linear signal trends in the control run and subtracting these trends from the

control and the forced runs. 26 °N was selected because is the approximate location of

the recently deployed NERC/RAPID mooring array [Marotzke, et al., 2002].

3. Methods

We approximate natural variability by the variability found in the control simulation.

The forced signal is the slope of the forced simulation time series, estimated by a least

squares linear fit. We expand on the trend detection method pioneered by Santer et al.

[1995] and refined by Baehr et al. [2007]. Specifically, we estimate trends of varying

9

chunk length for the control and forced signals. The probability density function (pdf) of

the unforced trends is derived using a bootstrap method with random starting points.

This distribution then defines the 95% confidence limits for the unforced trends. A

forced signal trend of equal chunk length is estimated, with the year 2006 as the starting

point – the first year of the hypothetical observation system. A detection of a statistically

significant trend (p < 0.05) occurs when the forced signal is outside the 95% confidence

limits of the control (Figure 4).

Uncertainty in the analyzed signals affects the detection time. We estimate the effects

of observation errors and scenario uncertainty by superimposing representations of those

errors. Estimating realistic observation errors for the MOC and AOU signals is an area of

active research [e.g., Ganachaud, 2003; Keller et al, 2002; Min and Keller, 2005, Baehr et

al, 2007c]. For this proof-of-concept study, we choose illustrative values derived from

the published literature. For the MOC signal, we adopt the results from previous

modeling studies which suggest MOC errors around 1 Sv (standard deviation) [Baehr, et

al, 2004; Baehr et al, 2007c]. For the AOU signal, we derive an arguably conservative

estimate based on a simple error analysis. The error of an average AOU concentration

depends on (i) uncertainties introduced by oxygen measurement errors and eddy induced

variability, (ii) uncertainties due to the calculation of the derived AOU tracer, and (iii) the

effects of aggregating several independent observations to an average AOU

concentration. A quite conservative upper bound for the combined effects of oxygen

measurement errors and eddy-induced variability is 5 μmol kg-1 [Gouretski and Jancke,

2001]. Combining this with an upper estimate of 3 μmol kg-1 for the error introduced by

calculating the oxygen solubility [Weiss, 1970]) results in an error for single AOU

10

observations of roughly 6 μmol kg-1. (assuming uncorrelated errors). One independent

check for this upper bound estimate of 6 μmol kg-1 for a single AOU observation is the

standard deviation of the trends in AOU observations at close-by (“crossover”) locations

between the GEOSECS and WOCE observations in North Pacific intermediate waters,

which has been estimated as roughly 4 μmol kg-1 [Keller et al, 2002]. The last step in

estimating the error for an average AOU concentration is to account for the effects of the

averaging process. This is important because the standard deviation of an average of N

statistically independent observations of a quantity decreases with 1/ N . Comparing

the typical station spacing of past hydrographic transects [Cunningham and Alderson;

2007] with the decorrelation length scale in the subtropical North Atlantic [Roemmich,

1983] suggest that N very likely exceeds ten. Hence we adopt a conservative error

estimate of 2 μmol kg-1 for the average AOU signal. Note that adopting smaller estimate

of this observation error would strengthen our forthcoming conclusions. We represent the

potential MOC and AOU observation errors by superimposing random draws from a

normal distribution with zero mean and a given standard deviation.

We additionally account for scenario uncertainty by creating multiple (103) time

series with the same trend and autoregressive properties as the original model MOC and

AOU signals. By drawing an ‘observation’ from a randomly-selected series, we

essentially superimpose the effects of varying initial conditions on the model MOC and

AOU signals (which otherwise would represent only one set of initial conditions). In

order to create the multiple time series, a smoothed fit (derived by a locally weighted

regression [Cleveland and Devlin, 1988]) is removed from the original MOC and AOU

signals. We fit autoregressive (AR) models to the MOC and AOU residuals such that the

11

selected AR model coefficients result in a minimized Akaike Information Criterion (AIC)

using a maximum likelihood method [Gilgen, 2006, pp. 266-269]. Using the selected AR

models, we generate multiple series with the same AR properties (or “red noise”) as the

original signal. The smoothed fit is subsequently recombined with the series to produce a

set of time series with the same trend and autoregressive properties as the original MOC

and AOU model signals. We approximate the effects of scenario uncertainty by

randomly sampling these time series. The detection method is applied to the analyzed

period with varying levels of uncertainty for MOC observations alone, AOU observations

alone, and the combined MOC-AOU signal (described below). The detection frequency

of the resulting detection time provides a reliability level: a 95% reliable estimate

corresponds to the detection frequency of 0.95.

We analyze a statistically optimal fingerprint constructed from the MOC and AOU

signals, in addition to analyzing the separate MOC and AOU signals. The fingerprint is

optimized by choosing a time-independent weighting to maximize the signal-to-noise

ratio. Specifically, we approximate the MOC and AOU signals from the analyzed period

(2006 – 2070) as linear trends with random noise (the standard deviation of this noise is

estimated from the residuals of the linear fit). We derive an approximately optimal

fingerprint by varying the weighting (w and 1-w, w ∈[0,1]) on the approximated MOC

and AOU signals and plotting the signal-to-noise ratio of the resulting fingerprints

(Figure 5). The signal-to-noise ratio is calculated by dividing the magnitude of the

difference in the fingerprint over the analyzed period (model years 2006 to 2070) in the

forced experiment (i.e. the signal) by the standard deviation of the fingerprint over the

entire model run in the control experiment (i.e. the noise).

12

4. Results and Discussion

In the control run, the MOC has a mean of 12.7 Sv (1 Sv = 106 m3 s-1) with a standard

deviation of 0.5 Sv (and a range of ~3.3 Sv) (Figure 3). The mean and the variability of

the MOC are within the range of other models [Gregory, et al., 2005; Schmittner, et al.,

2005]. The AOU signal shows less variability on interannual timescales than the MOC

signal. We hypothesize that the relatively low variability of the AOU signal on

interannual timescales results from the averaging over a relatively large region and the

analysis of a deepwater signal (where the interannual variability can be lower compared

to more shallow signals). There is no statistically significant correlation between MOC

and AOU signals in the control run (p > 0.05, accounting for serial correlation [Ebisuzaki,

1997]).

In the forced experiment (Figure 3), the MOC strength is relatively steady until 1980,

but then weakens to approximately 8 Sv by 2070. The average AOU concentration (26

°N, 2000 – 4500 m, 45 °W – 70 °W) has a mean value of 80 μmol kg-1 with a standard

deviation of 0.6 μmol kg-1 in the control run. In the forced run, the AOU signal begins to

considerably increase in the 1980s, roughly coinciding with the MOC weakening, and

rises by approximately 10 μmol kg-1 by 2070.

The relative changes in the forced AOU and MOC signals (Figure 3) are

approximately similar between the 1980’s and the 2030’s. Beyond the 2030’s, the

relative AOU changes exceed the relative MOC changes. The AOU and MOC signals in

the forced experiment show a statistically significant negative correlation (p < 0.05,

accounting for serial correlation [Ebisuzaki, 1997]). One possible explanation for the fact

13

that the AOU and MOC signals are statistically significantly anticorrelated in the forced

run but not in the control run is that the anticorrelation observed in the forced run is

largely driven by the anthropogenic forcing, which is missing from the control run.

Detection time is a random variable as it depends on random realizations of

observation errors and scenario uncertainty. If one neglects observation errors and

scenario uncertainty, detection occurs by 2025 for the MOC signal alone (Figure 5, panel

a). With observation errors of 1 Sv on the MOC signal, detection time is a probabilistic

variable, and reliable detection occurs later: at 95% reliability, detection occurs in 2054.

Considering both observation error (1 Sv) and scenario uncertainty results in a 95%

reliable detection time in 2060. The detection frequency of the AOU signal alone (Figure

5, panel b) has the same pattern of increasing detection times with increasing uncertainty,

and is shifted to earlier detection times. If one neglects observation errors as well as

scenario uncertainty, the AOU signal is detectable by 2012. When observation errors of

2 μmol kg-1 are considered, 95% reliable detection occurs by 2045. Additionally

superimposing the effects of scenario uncertainty to the effects of 2 μmol kg-1

observation errors results in a 95% reliable detection time of 2050.

The analysis so far analyzes the MOC and the AOU signals separately. Combining

these two signals in a fingerprint can improve the signal-to-noise ratio and the detection

capabilities [Hasselmann, 1993]. The signal-to-noise ratio of this fingerprint as a

function of the weight (w) for the OUT signal is shown in Figure 6. A fingerprint with a

weighting of w = 0.5 is close to the maximum value and has a higher expected signal-to-

noise ratio than either the AOU or the MOC signal (Figure 6). This optimal fingerprint

(w = 0.5) enables an earlier detection than either signal alone (Figure 7). Whereas the

14

95% reliable detection time using MOC observations alone is approximately five

decades, the deepwater AOU signal is detectable more than a decade earlier (Figure 7).

The optimal fingerprint achieves a 95% reliable detection time in three decades – roughly

two decades earlier than an observation system based on the MOC signal alone.

5. Caveats

Our simple analysis relies on several approximations. First, we analyze a single

model and are silent on the question of how robust the optimal fingerprint might be

across the range of structural uncertainty. Second, we neglect information contained in

tracers such as chlorofluorocarbons [Schlosser, et al., 1991; Smethie and Fine, 2001] or

129I [Edmonds, et al., 2001], which might well provide useful additional observations.

Third, potential changes in export production (different from those projected by the

model) will affect the ability to link changes in AOU to changes in MOC and hamper

detection. Fourth, our analysis focuses on the question of MOC change detection and is

silent on the arguably more relevant (but also much more complex) task of projecting the

future MOC intensity [Marotzke, 2000; Keller and McInerney, 2007]. Fifth, we use a

very simple approach to derive a statistical fingerprint. In addition, the link between

trends in hydrographic tracer concentrations and MOC changes is clearly more complex

than just a shoaling of the AABW/NADW boundary [cf. Matear and Hirst, 2003; Baehr et

al, 2007]. Last, but not least, the analyzed model likely underestimates the internal

variability of the signals as it does not resolve eddies [Hirst, et al., 2000].

15

6. Conclusions

Given the aforementioned caveats, we draw two main conclusions. First, adding AOU

observations to an MOC observation system can improve the signal-to-noise ratio and

result in earlier detection of anthropogenic MOC trends. The effect of including AOU

can be sizeable: the reliable detection times in our model study improve by almost two

decades. AOU may hence be a valuable hydrographic tracer for increasing understanding

of changes in the MOC. Second, the consideration of uncertainties introduced by

observation errors and scenario uncertainty results in later detection times. Conclusions

of previous studies neglecting these arguably important uncertainties may hence need to

be revisited.

The current debate about the future fate of the MOC can be informed by an effective

monitoring system. We show that additionally observing AOU, a water mass tracer that

is mechanistically tied to circulation changes, can improve the detection capabilities of an

MOC observation system.

Acknowledgements

We thank Joshua Dorin, Dong-Ha Min, David McInerney, Ray Najjar, and Johanna

Baehr for helpful discussions. The comments of Raghu Murtugudde and two anonymous

reviewers considerably improved the presentation of the paper. We gratefully

acknowledge support from the National Science Foundation (SES #0345925) and the

Penn State Institute for the Environment. Opinions, findings and conclusions expressed

in this work are those of the authors, and do not necessarily reflect the views of funding

entities.

16

References

Alley, R. B., J. Marotzke, W. D. Nordhaus, J. T. Overpeck, D. M. Peteet, R. A. Pielke, R. T. Pierrehumbert, P. B. Rhines, T. F. Stocker, L. D. Talley, and J. M. Wallace (2003), Abrupt climate change, Science, 299, 2005-2010.

Baehr, J., J. Hirschi, J. O. Beismann, and J. Marotzke (2004), Monitoring the meridional overturning circulation in the North Atlantic: A model-based array design study, Journal of Marine Research, 62, 283-312.

Baehr, J., K. Keller, and J. Marotze (2007a), Detecting potential changes in the meridional overturning circulation at 26 οN in the Atlantic, Climatic Change, published online on Wednesday, January 10, 2007, DOI 10.1007/S10584-10006-19153-Z, http://dx.doi.org/10.1007/S10584-10006-19153-Z

Baehr, J., H. Haak, S. Alderson, S. A. Cunningham, J. H. Junclaus, and J. Marotzke (2007b), Timely detection of changes in the meridional overturning circulation at 26 oN in the Atlantic, Journal of Climate, in the press.

Baehr, J., D. McInerney, K. Keller, and J. Marotzke (2007c), Optimization of an observing system design for the North Atlantic meridional overturning circulation, Journal of Atmospheric and Oceanic Technology, in the press, available at: http://www.geosc.psu.edu/~kkeller/publications.html.

Bentsen, M., H. Drange, T. Furevik, and T. Zhou (2004), Simulated variability of the Atlantic meridional overturning circulation, Climate Dyn., 22, 701-720.

Bopp, L., C. LeQuere, M. Heimann, A. Manning, and P. Monfray (2002), Climate induced oceanic oxygen fluxes: Implications for the contemporary carbon budget, Global Biogeochemical Cycles, 16.

Broecker, W. S., S. L. Peacock, S. Walker, R. Weiss, E. Fahrbach, M. Schroeder, U. Mikolajewicz, C. Heinze, R. Key, T. H. Peng, and S. Rubin (1998), How much deep water is formed in the Southern Ocean?, Journal of Geophysical Research-Oceans, 103, 15833-15843.

Bryden, H. L., M. J. Griffiths, A. M. Lavin, R. C. Millard, G. Parrilla, and W. M. Smethie (1996), Decadal changes in water mass characteristics at 24 degrees N in the subtropical North Atlantic ocean, Journal of Climate, 9, 3162-3186.

17

Bryden, H. L., H. R. Longworth, and S. A. Cunningham (2005), Slowing of the Atlantic meridional overturning circulation at 25 oN, Nature, 438, 655-657.

Cleveland, W. S., and S. J. Devlin (1988), Locally weighted regression: An approach to regression analysis by local fitting, Journal of the American Statistical Association, 403, 596-610.

Cox, M. D. (1984), A primitive equation, three-dimensional model of the ocean, Technical report, 141 pp., Geophys. Fluid Dyn. Lab. Ocean Group, Princeton Univ., Princeton, N.J., Princeton University.

Cunningham, S. A., and S. Alderson (2007), Transatlantic temperature and salinity changes at 24.5 oN from 1957 to 2004, Geophysical Research Letters, 34, L14606, doi:14610.11029/12007GL029821.

Dansgaard, W., S. J. Johnsen, H. B. Clausen, D. Dahl-Jensen, N. S. Gundestrup, C. U. Hammer, C. S. Hvidberg, J. P. Steffensen, A. E. Sveinbjörnsdottir, J. Jouzel, and G. Bond (1993), Evidence for general instability of past climate from a 250-kyr ice-core record, Nature, 364, 218-220.

Ebisuzaki, W. (1997), A method to estimate the statistical significance of a correlation when the data are serially correlated, Journal of Climate, 10, 2147-2153.

Edmonds, H. N., Z. Q. Zhou, G. M. Raisbeck, F. Yiou, L. Kilius, and J. M. Edmond (2001), Distribution and behaviour of anthropogenic 129I in water masses ventilating the North Atlantic Ocean, Journal of Geophysical Research, 106, 6881-6894.

Ganachaud, A. (2003), Error budget of inverse box models: The North Atlantic, J. Atmos. Oc. Technology, 20, 1641-1655.

Ganachaud, A., and C. Wunsch (2000), Improved estimates of global ocean circulation, heat transport and mixing from hydrographic data, Nature, 408, 453-457.

Garcia, H., A. Cruzado, and J. Escanez (1998), Decadal-scale chemical variability in the subtropical North Atlantic deduced from nutrient and oxygen data, J. Geophys. Res., 103, 2817-2830.

Gargett, A. E. (1984), Vertical eddy diffusivity in the ocean interior, J. Mar. Res., 42, 359-393.

18

Gent, P. R., and J. C. McWilliams (1990), Isopycnal mixing in ocean circulation models, J. Phys. Oceanogr., 20, 150-155.

Gilgen, H. (2006), Univariate time series in geosciences: Theory and examples, pp. 266-269 pp., Springer, New York.

Gordon, H. B., and S. P. O'Farrell (1997), Transient climate change in the CSIRO coupled model with dynamical sea ice, Mon. Wea. Rev., 125, 875-907.

Gouretski, V. V., and K. Jancke (2001), Systematic errors as the cause for an apparent deep water property variability: Global analysis of the WOCE and historical hydrographic data, Progress in Oceanography, 48, 337-402.

Gregory, J. M., K. W. Dixon, R. J. Stouffer, A. J. Weaver, E. Driesschaert, M. Eby, T. Fichefet, H. Hasumi, A. Hu, J. H. Jungclaus, I. V. Kamenkovich, A. Levermann, M. Montoya, S. Murakami, S. Nawrath, A. Oka, A. P. Sokolov, and R. B. Thorpe (2005), A model intercomparison of changes in the Atlantic thermohaline circulation in response to increasing atmospheric CO2 concentration, Geophysical Research Letters, 32, doi:10.1029/2005GL023209.

Hasselmann, K. (1993), Optimal fingerprints for the detection of time-dependent climate change, J. Clim., 6, 1957-1971.

Hirschi, J., J. Baehr, J. Marotzke, J. Stark, S. Cunningham, and J.-O. Beismann (2003), A monitoring design for the Atlantic meridional overturning circulation, Geophys. Res. Lett., 38, doi: 10.1029/2002GL016776.

Hirst, A. C., S. P. O'Farrell, and H. B. Gordon (2000), Comparison of a coupled ocean-atmosphere model with and without oceanic eddy-induced advection. Part I: Ocean spinup and control integrations., J. Clim., 13, 139-163.

Hu, A. X., G. A. Meehl, and W. Q. Han (2004), Detecting thermohaline circulation changes from ocean properties in a coupled model, Geophysical Research Letters, 31, doi: 10.1029/2004GL020218.

Joos, F., G.-K. Plattner, T. F. Stocker, A. Körtzinger, and D. W. R. Wallace (2003), Trends in marine dissolved oxygen: Implications for ocean circulation changes and the carbon budget, EOS, 84, 197-204.

19

Keller, K., C. Deutsch, M. G. Hall, and D. F. Bradford (2007a), Early detection of changes in the North Atlantic meridional overturning circulation: Implications for the design of ocean observation systems, Journal of Climate, 20, 145-157.

Keller, K., M. Schlesinger, and G. Yohe (2007b), Managing the risks of climate thresholds: Uncertainties and information needs (An editorial essay), Climatic Change, doi:10.1007/s10584-006-9114-6, published online Tuesday, January 23, 2007, http://dx.doi.org/10.1007/s10584-006-9114-6.

Keller, K., and D. McInerney (2007), The dynamics of learning about a climate threshold, Climate Dynamics, published online 11 July 2007, doi: 10.1007/s00382-007-0290-5, http://dx.doi.org/10.1007/s00382-007-0290-5.

Keller, K., R. Slater, M. Bender, and R. M. Key (2002), Possible biological or physical explanations for decadal scale trends in North Pacific nutrient concentrations and oxygen utilization, Deep-Sea-Research II, 49, 345-362.

Latif, M., E. Roeckner, U. Mikolajewski, and R. Voss (2000), Tropical stabilization of the thermohaline circulation in a greenhouse warming simulation, Journal of Climate, 13, 1809-1813.

Link, P. M., and R. S. J. Tol (2004), Possible economic impacts of a shutdown of the thermohaline circulation: an application of FUND, Portuguese Economic Journal, 3, 99-114.

Manabe, S., and R. J. Stouffer (1994), Multiple-century response of a coupled ocean-atmosphere model to an increase of atmospheric carbon dioxide, Journal of Climate, 7, 5-23.

Mantyla, A., and J. L. Reid (1983), Abysal characteristics of the World Oceans waters, Deep-Sea Research Part a-Oceanographic Research Papers, 30, 805-833.

Marotzke, J. (2000), Abrupt climate change and thermohaline circulation: Mechanisms and predictability, Proceedings of the National Academy of Sciences of the United States of America, 97, 1347-1350.

Marotzke, J., S. A. Cunningham, and H. L. Bryden (2002), Monitoring the Atlantic meridional overturning circulation at 26.5οN, Proposal accepted by the Natural Environment Research Council (UK), available at: http://www.nerc.ac.uk/funding/thematics/rcc/Scienceplan.shtml.

20

Matear, R. J., and A. C. Hirst (1999), Climate change feedback on the future oceanic CO2 uptake, Tellus, Series B, 51, 722-733.

Matear, R. J., and A. C. Hirst (2003), Long-term changes in dissolved oxygen concentrations in the oceans caused by protracted global warming, Global Biogeochemical Cycles, 17, doi: 10.1029/2002GB001997.

Min, D. H., and K. Keller (2005), Errors in estimated temporal tracer trends due to changes in the historical observation network: A case study of oxygen trends in the Southern Ocean., Ocean and Polar Research, 27, 189-195.

Plattner, G.-K., F. Joos, and T. F. Stocker (2002), Revision of the global carbon budget due to changing air-sea oxygen fluxes, Global Biogeochem. Cycles, 16, doi:10.1029/2001GB001746.

Redfield, A. C., B. H. Ketchum, and F. A. Richards (1963), The influence of organisms on the composition of sea-water, in The Sea, edited by M. N. Hill, pp. 26-77, Wiley-Interscience, New York.

Reid, J. L. (2005), On the world-wide circulation of the deep water from the North Atlantic Ocean, J. Mar. Res., 63, 187-201.

Roemmich, D. (1983), Optimal Estimation of Hydrographic Station Data and Derived Fields, Journal of Physical Oceanography, 13(8), 1544-1549.

Santer, B. D., U. Mikolajewicz, W. Bruggemann, U. Cubasch, K. Hasselmann, H. Hock, E. Maierreimer, and T. M. L. Wigley (1995), Ocean variability and its influence on the detectability of greenhouse warming signals, Journal of Geophysical Research-Oceans, 100, 10693-10725.

Schlosser, P., G. Bonisch, M. Rhein, and R. Bayer (1991), Reduction of deepwater formation in the Greenland sea during the 1980s: Evidence from tracer data, Science, 251, 1054-1056.

Schmittner, A., M. Latif, and B. Schneider (2005), Model projections of the North Atlantic thermohaline circulation for the 21st century assessed by observations, Geophysical Research Letters, 32, doi:10.1029/2005GL024368.

Smethie, W. M., and R. A. Fine (2001), Rates of North Atlantic Deep Water formation calculated from chlorofluorocarbon inventories, Deep Sea Research, 48, 189-215.

21

Stommel, H. (1961), Thermohaline convection with two stable regimes of flow, Tellus, 13, 224-230.

Sutherland, S. C., W. S. Broecker, and T. Takahashi (2001), Stability of the boundary separating Antarctic Bottom Water from North Atlantic Deep Water in the western South Atlantic, Geophys. Res. Lett., 28, 4219-4222.

Talley, L. D., J. L. Reid, and P. E. Robbins (2003), Data-based meridional overturning streamfunctions for the global ocean, Journal of Climate, 3213-3226.

Tol, R. S. J. (1998), Potential slowdown of the thermohaline circulation and climate policy, Discussion Paper DS98/06 Institute for Environmental Studies Vrije Universiteit Amsterdam.

Vellinga, M., and R. A. Wood (2002), Global climatic impacts of a collapse of the Atlantic thermohaline circulation, Climatic Change, 54, 251-267.

Vellinga, M., and R. A. Wood (2004), Timely detection of anthropogenic change in the Atlantic meridional overturning circulation, Geophysical Research Letters, 31, Art. No. L14203.

Weiss, R. F. (1970), The solubility of nitrogen, oxygen and argon in water and seawater, Deep-Sea Res., 17, 721-735.

Wunsch, C. (2006), Abrupt climate change: An alternative view, Quaternary Research, 65, 191-203.

22

Figure Captions

Figure 1. Meridional (panel a) and zonal (panel b) sections of AOU along 30 oW and 26 oN

(respectively) from model year 2006 in the CSIRO climate model [Matear and Hirst, 2003].

AOU is defined as the oxygen solubility at the in-situ temperature and salinity minus the

observed oxygen concentration. The contour lines of 90 μmol/kg AOU illustrate the

boundary between Antarctic Bottom Water (AABW) and North Atlantic Deep Water

(NADW). The dashed contour lines show the location of the boundary after approximately a

century of continued anthropogenic forcing (i.e., the model year 2100). See text for details.

Figure 2: Zonal sections of the projected changes over the twenty-first century (i.e., model

years 2100-2000) in temperature (panel a), salinity (panel b), and AOU (panel c) along 26 oN

in the CSIRO climate model [Matear and Hirst, 2003].

Figure 3. Annual maximum overturning (Sv, 1 Sv = 106 m3 s-1) at 26 °N (solid line) and

deepwater apparent oxygen utilization (AOU) (μmol kg-1) at 26 °N (averaged over 2000 to

4500 meters and 45 °W to 70 °W) (dashed line). Shown are the unforced (control) (a) and

the forced (b) experiments of the CSIRO climate model [Matear and Hirst, 2003]. We

perform the detection analysis from 2006 through 2070.

Figure 4. Illustration of the detection method. A statistically significant detection of forced

changes occurs when the forced signal (solid line) leaves the 95% confidence limits of the

unforced system (dashed lines). This illustration uses the MOC signal shown in Figure 3,

analyzes annual observations starting in 2006, and neglects the effects of observation errors

and uncertainty in the initial conditions (cf. Figure 6).

23

Figure 5. Effect of observation errors and scenario uncertainty on the detection of forced

MOC (panel a) and AOU (panel b) changes. Observation errors are approximated by

superimposing random perturbations drawn from an identical and independently distributed

Gaussian distribution with zero mean and varying standard deviation on the model signal.

Scenario uncertainty, representing uncertain initial conditions, is simulated by randomly

drawing from an estimated time series model (see methods text for details). Shown are

results for the MOC (AOU) with zero Sv (zero μmol kg-1) standard deviation without

scenario uncertainty (dotted line), 1 Sv (2 μmol kg-1) standard deviation without scenario

uncertainty (dashed line), and 1 Sv (2 μmol kg-1) standard deviation with scenario

uncertainty (solid line).

Figure 6. Signal-to-noise ratio (S N-1) of the fingerprint resulting from varying the

weighting on the AOU (w) (plotted on the x-axis) and the MOC (1- w) (where the weights

sum to 1) for the cases of 1 Sv and 2 µmol kg-1 observation errors with scenario uncertainty.

The signal is the magnitude of the difference between the combined signal observed in 2006

and in 2070. The noise is the standard deviation of the resulting combined signal. The

maximum of the expected signal-to-noise ratio occurs with a w of approximately 0.5.

Figure 7. Frequency of the detection times (across 104 states of the world) with observation

errors (1 Sv and 2 μmol kg-1) and scenario uncertainty for the MOC signal alone (dashed

line), the deepwater AOU signal alone (solid line), and the optimal fingerprint (solid line

with open circles). A reliable detection occurs when 95% of the realizations detect a change.

24

����

����

����

����

��

��

��

������

���������

����������

����� ��� ���� ��� ��� � �� �

��

���� �� �!"���������� #��

���

��$"����

��

��

��

����

����

����

����

%� # �� #�� #�� #�� #�� # �� #

���

���

%�

��

��

��

�

��

���� �� �!"���������� ���

���

Figure 1

25

����

����

����

����

�%��

����

����

����

����

����

�%��

����

����

����

����

����

�%��

����

����&�$�� �� �!"���'�$�&(�����

��$"������ # �� # �� #�� #�� #�� #

��$"������ # �� # �� #�� #�� #�� #

��$"������ # �� # �� #�� #�� #�� #

)���&��&��&�$�� *�'�$�&(������

���������

������������

�+%�+��+��+�

���+���+���+���+%

�+�%�+���+���+��

���+����+����+����+�%

��

��

��

��

��

��

��

��

��

��

,�

'�

�� ��

�� �$�(�&�$�� -��'�$�&(�������

Figure 2

26

��������

�����%%�%�%�%�%%����

��

��

��

�!"�

���

.��� �(��&

/�&'��

���%

��������

.�

0�

�1�

��������

��������

.��� �(��&

0�$&�

����

��.�0

���

�� ,�

Figure 3

27

���� ���� ���� ���� ���� ������+�

��+�

��+�

��+�

�

�+�

�+�

�+�

�+�

��2�'�$3���$'�� ���-��������3����'�$&�

/�&'����-�"$�

���'��$��������

.��� �(��&

4-

���

���&

�$��

�1�

(&���

���

Figure 4

28

���� ���� ���� ���� ���� ���� �����

�+�

�+�

�+�

�+�

�+�

�+�

�+�

�+%

�+�

�

.��� �(��&

/&�

5��$

'(��

3���

�'��

$

.�0�-�"$�

���� ���� ���� ���� ���� ���� �����

�+�

�+�

�+�

�+�

�+�

�+�

�+�

�+%

�+�

�

.��� �(��&

/&�

5��$

'(��

3���

�'��

$

����-�"$�

�� ,�

Figure 5

29

� �+� �+� �+� �+% �

�

�

�

�

�

#��"���$����

���

��

���$6�����&�� �7���$-�+�6���+���&'�$� �

Figure 6

30

���� ���� ���� ���� ���� �����

�+�

�+�

�+�

�+�

�+�

�+�

�+�

�+%

�+�

�

.��� �(��&

.�0

/&�

5��$

'(��

3���

�'��

$

���

0��,�$��-�"$�

Figure 7