Comparing the high-quality oxygen climatology from the World Ocean Circulation Experiment to earlier data we reveal near-global decreases in oxygen levels in the upper ocean between the 1970s and the 1990s. This globally averaged oxygen decrease is −0.93 ± 0.23μmol l−1, which is equivalent to annual oxygen losses of −0.55 ± 0.13 × 1014 mol yr−1(100–1000 m). The strongest decreases in oxygen occur in the mid-latitudes of both hemispheres, near regions where there is strong water renewal and exchange between the ocean interior and surface waters. Approximately 15% of global oxygen decrease can be explained by a warmer mixed-layer reducing the capacity of water to store oxygen, while the remainder is consistent with an overall decrease in the exchange between surface waters and the ocean interior. Here we suggest that this reduction in water mass renewal rates on a global scale is a consequence of increased stratification caused by warmer surface waters. These observations support climate model simulations of oxygen change under global warming scenarios.
 The thermohaline circulation regulates the Earth's climate by using the ocean to transport heat from low to high latitudes. While it is not practical to measure the overturning circulation directly, changes in ocean oxygen concentrations can be used to infer changes in the age of the water. Oxygen saturated surface waters are advected into the ocean interior, primarily along surfaces of constant density [Bindoff and McDougall, 1994; McDougall, 1987]. As the water spreads into the ocean interior oxygen is consumed by ocean biology. This reduces the oxygen concentration, and increases the carbon concentration. A change in oxygen concentration between two time periods can therefore be caused by a change in biological consumption rates, a change in the surface saturation levels, or a change in the age of the water since ventilation [Deutsch et al., 2005; Joos et al., 2003; Keeling et al., 2010; Plattner et al., 2002].
 The World Ocean Circulation Experiment (WOCE: ∼1989–2000) measured temperature, salinity and oxygen concentration throughout the water column. To detect historical changes in ocean oxygen, 256,078 spatially distributed profiles taken between 1940 and 1988 were objectively mapped onto the location of 38,002 profiles taken during this WOCE period. The ocean profiles from both periods came from a combination of the HydroBase2 [Curry, 2002] and the Southern Ocean databases [Orsi and Whitworth, 2002]. Stations in water depth shallower than 1000 m, and observations in the upper 100 m were removed to reduce coastal effects and seasonal variability respectively.
 To reflect the primary pathways that surface water is advected into the ocean interior, all profiles were interpolated onto neutral density surfaces using methods developed by [Jackett and McDougall, 1997]. These surfaces are simply referred to as ‘density surfaces’ throughout this manuscript. All uncertainty estimates are provided as a 90% confidence interval.
 Oxygen, temperature, and salinity measurements from 1940 to 1988 were then mapped to the locations of the WOCE data using a previously established adaptive optimal interpolation method [Aoki et al., 2005; Helm et al., 2010]. Essentially all historical data are optimally interpolated by location (xi, yi) and time (ti) to produce an historical estimate at each WOCE location (xj, yj). The normalised correlation function C_xyt used in the optimal interpolation is Guassian and of form:
 Note that the correlation patterns in the optimally interpolated data are not guassian but reflect the actual patterns of the data. All historical data (1940 to 1988) were interpolated to a mean year of 1970 (tj) using a decay timescale (yearscale = 15 years). The mean year of 1970 was chosen because the historical oxygen data have a peak in the distribution near this year. This choice allowed a comparison with the WOCE observations (mean year of 1992). The length scale varies spatially, depending on the observed spatial density of oxygen profiles. An adaptive iterative procedure was employed to determine the appropriate length scales, meaning that where data is sparse the length scales are adjusted to be longer [Helm et al., 2010]. The residuals between the observations and optimally interpolated estimates were carefully checked with the aprioriestimate of the ocean noise. The variable spatial scales means that there are always sufficient observations in each sub-region region to construct a smooth map field in the region. This approach differs from the methods used in sayLevitus et al.  and reduces biases that can occur with methods that use fixed length scales or spheres of influence in regions of sparse data. In this paper, oxygen changes are reported as a difference over the 1970 to 1992 period (negative is a decrease) and inferred net gas exchanges are expressed as a rate per year.
 All of the optimally interpolated data were first averaged in 5° × 10° grid boxes, before zonal averages with confidence intervals were calculated.
 Changes in surface density in the upper 100 m were estimated using temperature observations from the Hadley sea surface temperature dataset and surface salinity from the Levitus climatology [Levitus et al., 1994]. These surface salinity and temperature data of surface density were combined with observed ocean property changes between 100–1000 m to estimate vertical density gradient (δρ/900 m) for both time periods (∼1970 and ∼1992).
 Note that an upper-ocean warming (or freshening) displaces density layers deeper in the water column, while a surface cooling (or salinity increase) has the opposite effect. Without equivalent temperature (or salinity) changes at deeper depths, the increased stratification (δρ/δz) reduces the exchange of water between the mixed layer and the ocean [Gnanadesikan et al., 2007]. This reduction in the renewal rates of water masses allows more time for biological oxygen consumption to occur.
3. Observed Global Changes in Oxygen Concentration
 Comparing oxygen profiles centred on 1970 with oxygen profiles from the WOCE period shows widespread decreases in oxygen concentration in the upper ocean (Figure 1) that are statistically significant at the 95% confidence level. The largest oxygen decreases are in the Southern Ocean and are circumpolar in extent, while strong oxygen decreases are found in the North Pacific and North-West Atlantic Oceans. The average oxygen decrease in the equatorial regions (10°S-10°N) of all three oceans is −5.3 ± 3.6μmol l−1, consistent with the expanding oxygen minimum zones in tropical latitudes [Stramma et al., 2008]. The equatorial decreases are strongest in the East Pacific Ocean (120°W) and weaker to slightly positive in the west (170°E).
 While the dominant pattern is for a decrease in oxygen concentration, the subtropical gyres (15°–30°) in the North Pacific, South Pacific and South Indian Oceans show small regional increases in oxygen (Figures 1 and 2a). In the sub-polar North Atlantic there are oxygen concentration increases in the eastern side that are in contrast to the strong decreases to the west (Figure 1).
 Regional variability is expected and is driven by meso-scale ocean eddies and internal oscillations of the earth system. Mesoscale eddies are taken into account in thea priori estimates of the noise. The internal variability with decadal time scales are only partially included in the apriori noise estimates and include the North Atlantic Oscillation, the Pacific Decadal Oscillation and the Southern Annular Mode, as well as natural forcing such as volcanic activity (i.e., Pinatubo). The latter natural forcing could potentially mask some of the longer term oxygen decreases [Frölicher et al., 2009].
 Despite some regional variability, the consistent sign of large decreases in oxygen between ocean basins (that is with scales larger than these oscillations) suggests that a zonally averaged analysis is an appropriate means of revealing global-scale changes. Zonal averages of oxygen changes show near-global oxygen decreases on pressure surfaces in the upper 1000 m, with the strongest changes occurring poleward of 40° in both hemispheres (Figure 2a). These high-latitude oxygen decreases extend throughout the water column. The vertical extent of the oxygen increases in the narrow latitude band of the subtropical gyres (15°–30°) extend to approximately 1000 m in the Northern Hemisphere, and to approximately 500 m in the Southern Hemisphere (Figure 2a).
 The global average change in the oxygen concentration between 100 and 1000 m from 1970 to 1992 is −0.93 ± 0.23 μmol l−1. This global decrease is equivalent to an annual loss rate of −0.55 ± 0.13 × 1014 mol yr−1 from the ocean, and is slightly higher than the −0.45 ± 0.13 × 1014 mol yr−1 estimated between 1993 and 2003 [Manning and Keeling, 2006]. The Southern Ocean represents 25% of this decrease in the global average of the oxygen concentration over the upper 1000 m. There is a region of oxygen increase below 1500 m between 30°N and 30°S (Figure 2a), which largely is found in the Atlantic sector of these zonal averages and mainly in North Atlantic Deep Water (NADW).
 The oxygen changes that are evident on pressure surfaces (Figure 2a) can be caused either by an oxygen concentration change along a density surface (Figure 2b: presented on pressure surfaces), or by a vertical displacement of density surfaces in the water column [Bindoff and McDougall, 2000; McDougall, 1987] (Figure 2c). Separating these two contributions allows us to distinguish between the changes along surface driven advective pathways, and changes caused by dynamical effects (of say the winds).
 This clearly reveals that most of the zonally averaged pattern of oxygen change is caused by oxygen changes on density surfaces (Figure 2b) and is more coherent than changes resulting from the vertical movement of density surfaces (Figure 2c). The implication is that the signal of oxygen change is largely driven by changes in air-sea interaction rather than by internal readjustment of ocean properties. This conclusion is consistent with the largest decreases occurring along density surfaces in mid and high-latitude regions (Figure 2b). In these regions the Northern Hemisphere water masses, and Antarctic Intermediate Water and Circumpolar Deep Water in the Southern Hemisphere interact with surface waters either through ventilation or mixing.
 By contrast, the pattern of oxygen change caused by the vertical movement of density surfaces (Figure 2c) has a less coherent spatial pattern and is generally smaller in magnitude with no significant change in global oxygen content. Temperature, salinity and pressure observations show a downward (and poleward) displacement of density surfaces at 45°–50° in both hemispheres. This is potentially driven by strengthening westerly winds [Trenberth, 2007] and has the effect of displacing oxygen-rich water deeper in the water column (Figure 2c). The downward displacement of isopycnals causes an increase in oxygen on pressure surfaces above 1000 m (Figure 2a). The effect of temperature changes on oxygen saturation levels is between −3 and 3 μmol l−1 (Figure 2d), and is negligible when compared with the overall changes on pressure surfaces (−10 to 10 μmol l−1: Figure 2a).
 The average decrease in oxygen caused by the observed surface warming over the 100–1000 m layer is approximately −0.15 ± 0.03 μmol l−1. The decrease in saturation from warming equates to an average loss of −0.09 ± 0.01 × 1014 mol yr−1, comparable to ocean circulation model estimates of reduced saturation [Keeling and Garcia, 2002]. This accounts for just 16 ± 3% of the total observed oxygen decrease over this layer, slightly lower than the ∼25% estimated in model simulations [Bopp et al., 2002].
4. Explanation of Changes in Oxygen Concentration
 Both climate models and observations [Trenberth, 2007] suggest an intensification of westerly winds in the mid to high latitude Southern Hemisphere. These intensified winds have been combined with atmospheric CO2 observations to infer a decreased carbon uptake efficiency by the Southern Ocean [Le Quere et al., 2007]. These increased winds should increase ventilation of the subducting Antarctic Intermediate Waters and sub-Antarctic mode waters and imply higher oxygen concentrations in these waters. The increased winds also increase the upwelling of the Circumpolar Deep Waters on the poleward side of the Antarctic Circumpolar Current implying lower oxygen concentrations. However, the observed oxygen concentrations for all three of these water masses have decreased (Figure 3), suggesting that the strengthening winds cannot explain the observed decreases in the two subducting water masses. Thus it seems likely that the increased winds are not the main driving mechanism of oxygen change for these three water masses. However, the observed oxygen decreases could be due to surface buoyancy change as evidenced by warmer surface waters (Figure 3a) and freshening [Boyer et al., 2005; Helm et al., 2010], affecting both ocean stratification and hence the rate of upwelling of the CDW and ventilation of AAIW and SAMW water masses [Gnanadesikan et al., 2007].
 Stratification increases are observed in the upper 1000 m (Figure 3b) between 40°N and 50°N, and poleward of 45°S. The total layer change of oxygen is estimated by integrating the oxygen changes poleward from the equator along layers bounded by density surfaces to their intersection with the mixed layer (Figure 3c). These layer integrals reflect the interior changes and can be interpreted as a function of changes at the higher-latitude source region. By contrast the column inventory average change of oxygen shows a more complex pattern including dynamical and ventilations changes (Figure 3d).
 The statistically significant Southern Ocean stratification increase is consistent with the observed oxygen decrease in the upwelling limb of this region and with a slowing of the ventilation rate of Antarctic Intermediate Waters and sub-Antarctic mode waters. Interpreting the North Atlantic and North Pacific Oceans together between 40°N and 50°N shows a similar relationship with the decreased oxygen concentrations consistent with the observed increased stratification (Figure 3b and 3c).
 More than 90% of the global decrease in the oxygen is associated with density surfaces that are ventilated poleward of 40°. The Southern Ocean contributes 27% of this decrease and includes the changes in the upwelling Circumpolar Deep Waters and the subducting AAIW and SAMW waters. The Northern Hemisphere Oceans contribute 65% of this decrease and are associated density layers that form part of North Atlantic Deep Water and North Pacific density layers.
 In their latest report, the International Panel on Climate Change (IPCC) summarizes the literature by observing that an increase in biological consumption is an unlikely explanation for oxygen decreases over these large spatial scales [Bindoff, 2007]. Our work suggests that ocean warming plays a small role in this decrease, and that a reduction in the exchange between the mixed layer and the ocean interior remains the most likely cause of most of the observed oxygen decreases. The increase in upper-ocean stratification provides one plausible physical mechanism for large scale zonal decreases in oxygen at high latitudes, and the smaller increases that have also occurred on the poleward side of the sub-tropical gyres (Figure 3c). The increasing stratification appears to be coupled with strong oxygen decreases in the high-latitude regions of both hemispheres where a large proportion of the ocean is ventilated. In the subtropical gyres of both hemispheres the slight stratification decreases are correlated with oxygen concentration increases in the upper 500 m (Figure 2a).
 There is compelling evidence of oxygen decreases throughout much of the global ocean between the 1970s and the early 1990s. It is common to identify the equatorial region as an area of vulnerability to decreased oxygen concentrations [Keeling et al., 2010; Stramma et al., 2008], but this analysis shows that the oxygen concentrations have decreased significantly in the polar regions, particularly in the Southern and North Pacific Oceans (and to lesser extent in the North Atlantic) during this period. The majority of the oxygen decreases are likely to be driven by increases in upper ocean stratification caused by surface warming and high-latitude freshening. The reduced oxygen capacity of warming waters explains approximately 15% of global changes. These oxygen decreases suggest reduced ocean ventilation and a weakened circulation, and are an indicator that important changes are occurring in the global carbon cycle. These oxygen observations thus provide the supporting evidence for modelling studies of these gases.
 This paper is a contribution to the CSIRO Climate Change Research Program and was supported by the Australian Government's Cooperative Research Centres Program through the Antarctic Climate and Ecosystems Cooperative Research Centre. The Centre for Australian Weather and Climate Research is a partnership between CSIRO and the Australian Bureau of Meteorology. Comments from two reviewers are gratefully acknowledged.
 The Editor thanks Corinne Le Quere and Fortunat Joos for their assistance in evaluating this paper.