Experiments assimilating the RAPID dataset of deep temperature and salinity profiles at 26.5°N on the western and eastern Atlantic boundaries into a 1° global NEMO ocean model have been performed. The meridional overturning circulation (MOC) is then assessed against the transports calculated directly from observations. The best initialization found for this short period was obtained by assimilating the EN3 upper-ocean hydrography database prior to 2004, after which different methods of assimilating 5-day average RAPID profiles at the western boundary were tested. The model MOC is strengthened by ∼ 2 Sv giving closer agreement with the RAPID array transports, when the western boundary profiles are assimilated only below 900 m (the approximate depth of the Florida Straits, which are not well resolved) and when the T,S observations are spread meridionally from 10 to 35°N along the deep western boundary.
The Atlantic meridional overturning circulation (MOC hereafter) plays a key role in controlling climate because it largely determines the northward heat transport in the Atlantic Ocean, therefore influencing sea-surface temperatures (Knight et al., 2005). Changes in the MOC, and the related climate impacts, may also be predictable on decadal time-scales, as inferred by modelling studies (Griffies and Bryan, 1997; Collins and Sinha, 2003; Collins et al., 2006; Pohlmann et al., 2009). The strength of the overturning at 26.5°N has been determined from trans-ocean sections on five occasions in the past (Bryden et al., 2005), and it has been continuously monitored using a mooring array since April 2004 (Cunningham et al., 2007, henceforth RAPID data). The past sampling of the MOC, and the present-day RAPID monitoring, provide a stringent test of ocean and coupled climate models to see if they can reproduce aspects of the MOC variability, as this would potentially allow the models to be used for climate predictive purposes.
The Rapid Climate Change Programme (RAPID) observing system at 26.5°N (Cunningham et al., 2007) consists of an array of moorings spanning the zonal extent of the Atlantic Ocean between the Bahamas and Africa. The moorings are located close to the western and eastern margins, as well as on the western and eastern flanks of the Mid-Atlantic Ridge (Figure 1). The array provides temperature and salinity time series at fixed depths which are filtered to remove fluctuations with periods shorter than one day, and interpolated to 20 dbar levels (details in Cunningham etal., 2007). Additional current meter, bottom pressure recorder and Florida Straits cable data allow transport time series for the MOC and its components (geostrophic, Florida Straits and Ekman transports) to be calculated. To date the temperature, salinity and transport data are available from April 2004 to April 2009, allowing seasonal variability of the MOC to be assessed (Kanzow et al., 2010). In this article we will look at assimilating these data into an ocean circulation model, in order to better reproduce the MOC. We begin by seeking initial conditions which best reproduce the MOC during the RAPID sampling period without assimilating the RAPID data, and we then seek to show that RAPID data assimilation gives a further level of improvement. The success of this procedure might then allow coupled climate models to be initialised with better MOC transports for climate prediction purposes.
Balmaseda et al. (2007) compared the MOC in a 1° ocean simulation and an ocean reanalysis against transports calculated from past cruise sections around 26°N (Bryden et al., 2005) and showed improvements in reproducing the MOC transports in the reanalysis fields, which appear to correct a weak MOC bias of several Sverdrups (1Sv = 106 m3s−1) in the free-running simulation. They also show a remarkable reproduction of the interannual variability detected from the cruise section data, although this result has not been easy to reproduce in other ocean reanalyses (Lee et al., 2010). Baehr et al. (2009) and Sarojini et al. (2011) have made comparisons of the high-frequency variability of the MOC in coupled climate and ocean-only models and found reasonably good agreement in the strength and amplitudes of daily to seasonal variability. Pohlmann et al. (2012) make some interannual MOC comparisons in ocean reanalysis products and also find a level of agreement for interannual variability, despite large differences in the mean MOC values, with many models having a low MOC bias to a greater or lesser degree. Smith et al. (2010) found that assimilation of temperature and salinity in the upper ocean alone (down to 2000 m, the depth of Argo floats) was insufficient to correct for the low MOC bias in their Nucleus for European Modelling of the Ocean (NEMO) 1° model, because the assimilation spins up the subtropical gyre rather than the MOC. Baehr (2010) also found that the effect of the RAPID array profiles was to redistribute the upper ocean circulation without changing the MOC strength, in 1° 4D-Var Consortium for Estimating the Circulation and Climate of the Oceans (ECCO) assimilation experiments using only the first year of monthly averaged RAPID data.
In seeking to assimilate RAPID data this article tackles two distinct technical challenges: how to assimilate deep observations, and how to treat observations close to a boundary. The article first shows that specific boundary-focussed covariance or regression information is needed to assimilate the RAPID array data effectively. This is most clearly shown with experiments which assimilate RAPID data alone. Once this boundary methodology is developed, the article tests whether assimilation of RAPID array data, along with other hydrography data particularly from the EN3 database, gives improved flows, compared with assimilating EN3 data alone. Some operational ocean and seasonal forecasting systems currently filter out such observations which are too close to ocean boundaries, for various reasons (e.g. Balmaseda et al., 2008), and therefore this work potentially shows that these data could be more effectively used operationally. Also, observations below 2000 m are scarce and yet such observations are needed because many models show deep drifts due to inadequate deep water formation or mixing parametrizations. We therefore anticipate large impacts of the RAPID array data at these depths. The NEMO model being used here is quite close to the NEMO system currently in operational use in France, and at the UK Met Office and the European Centre for Medium-range Weather Forecasts (ECMWF), so we believe these results are highly relevant to the operational ocean predictions.
This article presents a first attempt at the assimilation of five years of RAPID mooring data into a 1° global NEMO model. Section 2 introduces the model and also presents a brief description of the method of RAPID transport calculation, and the importance of deeper density assimilation. Section 3 discusses a number of experiments designed to identify the best initial conditions for the assimilation of RAPID data after 2004. This is necessary because the RAPID data present only a very short time series and the circulation will be strongly influenced by the initial model conditions at the beginning of the runs. We then discuss how the RAPID data are prepared for assimilation. Section 4 presents the results of a number of different assimilation experiments demonstrating the key importance of improving the covariances near the western boundary in order to have a positive impact on the MOC stream function over a range of latitudes, and also demonstrating the limitations of the approach in an ocean-only model when the higher latitude water formation areas become impacted. Section 5 provides some discussion and conclusions about the effect of assimilating RAPID array data into the ocean model.
2. Model description
2.1. The code and forcing
The numerical model used is the NEMO coupled ice–ocean model (Madec, 2008) version 2.3, based on the OPA9 ocean model (Madec et al., 1998) and the LIM2.0 sea ice model (Louvain sea Ice Model: Fichefet and Maqueda, 1997; Goosse and Fichefet, 1999). The ocean model is a primitive equation z-level model making use of the hydrostatic and Boussinesq approximations. The model employs a free surface (Roullet and Madec, 2000) with partial cell topography (Adcroft et al., 1997). The version used here has a tri-polar ‘ORCA’ grid and 46 levels in the vertical, with thicknesses ranging from 6 m at the surface to 250 m at the ocean bottom.
The model configuration has a global 1° resolution with a tropical refinement to 1/3° (ORCA1). The configuration has been developed through the DRAKKAR Consortium (Barnier et al., 2007) and uses model parameter settings as defined in Barnier et al. (2006) and Penduff et al. (2009). The configuration employs an energy-enstrophy conserving momentum advection scheme (Barnier et al., 2006) and a Laplacian diffusion. Horizontal viscosity is parametrized with a Laplacian operator. Additionally, ORCA1 makes use of the Gent and McWilliams (1990) mixing parametrization. Vertical mixing is parametrized using a one-equation turbulent kinetic energy scheme (Blanke and Delecluse, 1993). More details may be found in Barnier et al. (2006) and Penduff etal. (2007).
Surface atmospheric forcing is obtained from ECMWF ERA-Interim 6 h reanalysis (Simmons et al., 2007; Dee and Uppala, 2009). The ERA-Interim reanalysis provides 10 m wind, 2 m air humidity and 2 m air temperature to compute 6-hourly turbulent air/sea and air/sea-ice fluxes during model integration, using the bulk formula proposed by Large and Yeager (2004). Downwelling short- and long-wave radiative fluxes and precipitation are also provided by ERA-Interim. While biases in radiation and precipitation fields are inevitable and could have been reduced prior to use with ocean models, we found that the heat and freshwater budget of the NEMO model is able to come into global balance with the ERA-Interim forcing without further modifications being made (M. Valdivieso, personal communication). The experiments described in the article are summarized in Table 1.
Table 1. A description of experiments.
Control simulation forced with ERA-Interim and initialised from 1989 from WOA05 climatology
As CTL0 but initialised in Jan 2004 from WOA05 climatology
As CTL0 but initialised in Jan 2004 from EN3 in situ data assimilation experiment started in 1989
Assimilation of full depth data using conventional covariances
As R0 but assimilation of only below 900 m of RAPID data near western boundary with conventional covariances
R900 but using western boundary regressions to spread data
As Rreg but using regressions from 18-year control run CTL0
Recycled run with Jan 2008 CTL2 used to start the simulation/assimilation period again in Jan 2004
Recycled run with Jan 2008 Rreg used to start the simulation/assimilation period again in Jan 2004
As Rreg but including the assimilation of NAtl. EN3 data
NAtl. only EN3 data assimilation
Recycled run with Jan 2008 Rreg+EN3 used to start the assimilation period again in Jan 2004.
2.2. MOC components and the model at 26.5°N in the Atlantic
The observed MOC at 26.5°N is constructed from three components: the Florida Straits transport measured directly using electromagnetic methods (Baringer and Larsen, 2001); the Ekman transport calculated using the values of zonal wind stress; and the Upper Mid-Ocean transport, obtained from the RAPID mooring array assuming geostrophy (details in Hirschi et al., 2003; Cunningham et al., 2007). Although these are measured independently, there should be very little net volume transport across 26.5°N (Kanzow et al., 2007), so any fluctuation in either Gulf Stream or deep western boundary current transports is assumed balanced by a barotropic flow through the main basin section; see Bryden et al. (2009) for the evidence of such a flow.
In a model the total MOC can be determined either directly from modelled currents, or the components can be calculated using the same data and assumptions as for the ‘observational’ transports, and in this model the total MOC from these calculations give similar results. We always use the flow above and below 1000 m to define the 26°N MOC. However, the division of upper ocean transports between the Florida Straits and the Upper Mid-Ocean components can be very different from the observations (Smith et al., 2010). Figure 2(a) shows the NEMO 1° model bathymetry at the western boundary at 26.5°N, along with typical mean velocity contours (2007 average) from a spun-up simulation with ERA-Interim forcing. Although the Florida Straits are separated with this bathymetry, only around 20 Sv of the upper-layer northward boundary current passes through the Straits, leaving a considerable 10 Sv northward flow along the deep western wall above the southward-flowing deep western boundary current. The total northward flow at around 30 Sv is however quite comparable to the observed 32 Sv of Florida Straits transport. Figure 2(b) shows the bathymetry and currents (2004 average) in a NEMO simulation (G70: Barnier et al., 2007) showing that, even at this resolution, the Gulf Stream is still not as well confined as in the observations. The northward Florida Straits transport is weakened and the net southward Upper Mid-Ocean transport is also weakened; however, the total MOC can still be quite realistic as it depends only on the total northward flow above 1000 m or so (and the return flow at depth) and not on the exact pathway of that flow near the Florida Straits.
The top 900 m or so of the RAPID array western density profile cannot be consistent with this model's natural distribution of the Gulf Stream transport at 26°N, and this would likely be the case for any low-resolution model. Indeed Smith et al. (2010) already showed that even assimilation of open-ocean hydrographic data in upper layers affects this split between the Florida Straits and Upper Mid-Ocean components, without impacting positively on the total MOC. Assimilating the upper part of the RAPID array western moorings would therefore not be expected to have a strong impact on the MOC. Smith et al. (2010) concluded that improving the model deep return flow was more important for obtaining better total MOC transports.
The importance of deeper and shallower density distributions can also be examined from the natural variability of the model. Figure 3(a) shows the control run correlations of the monthly MOC transport at 26.5°N (without the Ekman component) with the 1000–3000 m averaged density anomalies in the NEMO 1° model for the 45-year period 1960–2004. All model variables were detrended and the seasonal cycle was removed, with the trend of the MOC transport at 26.5°N being less than 2 Sv over 45 years. The positive correlations (high densities) along the western boundary at 26.5°N indicate stronger vertical shear and hence MOC, with negative or low density anomalies at the eastern boundary. Positive density correlations extend through most of the western subtropical gyre, but at higher latitudes are more confined to the west, where they extend into the central Labrador Sea. If a high-pass filter (< 12 months) is used, positive correlations are confined to within a few degrees of the western boundary at all latitudes, reflecting that fast boundary-wave propagation is likely responsible for these correlations. In the top 1000 m the correlation patterns are quite different (Figure 3(b)) with little positive correlation in the subtropical gyre except very close to 26°N, but with some weaker positive correlations at the western boundary of the subpolar gyre. Similar results are found for correlations performed with the NEMO model (not shown), although at higher resolution the deep correlations are narrowly confined to the boundary at subtropical gyre latitudes, consistent with much stronger mesoscale noise in the subtropical gyre density fields, which is not correlated with the MOC, cf. Wunsch (2008) and Kanzow etal. (2009).
The above discussion motivates the choices we make in section 4 where we initially show results assimilating the western boundary array data at all depths, before only assimilating data deeper than 900 m, i.e. below the depth of the upper-layer Gulf Stream. This, we suggest, is a preferable option for any low-resolution model without a well-resolved Florida Straits, because it avoids disrupting the model's Gulf Stream pathway and also focuses on depths where there are the highest correlations between MOC transport and boundary density variations.
3. Assimilation of RAPID array data
In this article we mostly look at assimilating the RAPID array temperature and salinity profile data alone, located at the western and eastern boundaries, and on both sides of the Mid-Atlantic Ridge. The transports calculated from the RAPID measurements are used for comparison with model transports. Due to coarse model resolution, the western and eastern temperature (T) and salinity (S) profiles were merged before assimilation. The western boundary merged profile is a combination of data from mooring WB2, and deep measurements below 4000 m from WBH1/WBH2 (before April 2005) and WB3 (after April 2005); see Figure 1. As in Cunningham et al. (2007), the profiles from the eastern moorings EB1, EBH1, EBH2, EBH3, EBH4 and EBH5 are joined to construct a single eastern merged boundary profile (see http://www.noc.soton.ac.uk/rapidmoc for detailed information on this procedure).
We have assimilated these 5-day averaged, merged temperature and salinity data, with an assimilation cycle that compares the instantaneous model fields in the centre of each 5-day period with the observations. We note that Kanzow et al. (2007) show that the mass transport balance across 26°N closes on around a 10-day time-scale. The assimilation increments are determined at the end of each 5-day period, and a variant of the Bloom et al. (1996) incremental analysis update method is used to add increments evenly over the subsequent day (if daily time-scales were of interest the model could be re-run adding the increments during the period the observations were made, but this is unnecessarily expensive for the current study). The data assimilation method of Haines et al. (2006), which had previously been implemented within this NEMO model by Smith and Haines (2009), was used to construct the T and S increments. Referred to as the S(T) assimilation scheme, this is a two-step sequential scheme for hydrographic data based on Optimal Interpolation. Temperature profiles (T) are assimilated along with a salinity-balancing increment (Troccoli and Haines, 1999) to maintain the model's water mass properties (i.e. temperature–salinity relationships) in the absence of salinity data. In the second step, salinity profiles (S) are assimilated along isotherms (i.e. S(T): Haines et al., 2006). With the S(T) increments being spread along isotherms, this means that corrections to a particular water mass do not influence adjacent water masses, which may have uncorrelated errors. The scheme has been thoroughly tested, with details available in Smith and Haines (2009) and it is also described as part of the ECMWF system-3 reanalysis (Balmaseda et al., 2008). However, we also did a short test switching the S(T) scheme off to allow T and S to be assimilated independently, and it makes little difference to the results here.
Spatial error covariances are used to spread out the increments to be made to the model T and S properties around the region where observation–model differences are detected. The choice of these covariances turns out to be critical to recovering a good MOC. Error covariances can be estimated with sophisticated assimilation schemes such as an Ensemble Kalman Filter (Evensen, 1994); however, such schemes are expensive and they require considerable tuning with lots of available data, and many assumptions are used to justify the approach (typically model biases are hard to deal with). Many operational ocean assimilation systems therefore use simple Gaussian covariances in the open ocean. Carton et al. (2000) estimated isotropic spatial correlation scales of 375 km at the latitude of the RAPID array. The results of this choice, and the development of simple alternatives to account for the boundary nature of the RAPID data, are discussed in the following section.
4. Experiments with the RAPID data
4.1. Simulations from different initial conditions in 2004
The RAPID array data cover a very short time period and therefore the reproduction of the circulation in this period will depend critically on the initial conditions at the start of the assimilation (Baehr, 2010). We investigate the best possible initial conditions as a starting point in order to show that the assimilation of the RAPID array data can still further improve the MOC response. Figure 4(a) shows the MOC time series at 26.5°N for the 5-year period 2004–2008 inclusive, for the RAPID MOC observations and for three different free ‘Control’ runs of the NEMO 1° model with no assimilation through the period:
CTL0 was initialised with a cold start in 1989 using the World Ocean Atlas 2005 (WOA05) 1° resolution gridded climatology (Boyer et al., 2006) initial conditions and spun up through to January 2004, and on through 2008, using ERA-Interim forcing.
CTL1 was initialised with the same WOA05 initial conditions but with a cold start, i.e. no circulation, in January 2004.
CTL2 was initiated as CTL0 in 1989 but then run with hydrographic data assimilation as in Smith and Haines (2009), using the ENACT/ENSEMBLES EN3v2.1 data, (Ingleby and Huddleston, 2007), until January 2004 from which time the model was run free through 2008.
The EN3 is a global ocean database, which is regularly updated against the World Ocean Data Base (WODB) and other sources, e.g. Argo, with additional quality control (Martin et al., 2007) and maintained up to date by the UK Met Office. All experiments discussed in the article are listed in Table 1.
In Figure 4(a), run CTL0 has a considerably reduced MOC compared to observations, suggesting that the 15 years of spin-up have resulted in loss of MOC strength. Both CTL1 and CTL2 show much closer MOC agreement with the RAPID observations. Figure 4(b) shows the MOC time series at 40°N from these three control runs, and Figure 4(c) shows the Atlantic MOC stream function for CTL2, as the 2008 annual average. It can be seen that at 40°N (and at 50°N, not shown) the MOC in both CTL1 and CTL2 are clearly spinning down with time over the 5-year period, which is not obvious at 26.5°N. The MOC variability at 40°N, both from CTL2 and CTL1 experiments (∼ 2.4 Sv), is comparable with observational estimates of Willis (2010). The mean CTL1 MOC transport at 40°N is 15.6 Sv, also in a good agreement with Willis (2010) (15.5 Sv); however, CTL2 gives a higher value of ∼ 17 Sv. The 2008 stream function for CTL2 looks very reasonable, with a peak MOC of about 18 Sv located at 29°N, 800 m.
Table 2. The MOC transport and meridional Heat Transport (HT).
Mean MOC (Sv)
STD MOC (Sv)
Correlation MOC with RAPID
Mean HT (PW)
Table 2 shows the MOC statistics of the various runs described in the text (STD is standard deviation). For correlations of MOC with the RAPID time series, the first value corresponds to correlation coefficient calculated from total time series; in brackets, correlation coefficient calculated after subtraction of the annual cycle.
Comparison statistics between the model and the RAPID 26.5°N MOC transport time series (all summarised in Table 2) indicate that the model describes about 50% of the total variability and more than 25% of the interannual variability (calculated by subtraction of the seasonal cycle from the total transport). Due to model bias the absolute monthly values of the MOC transport are underestimated (by between 1 and 5 Sv, Table 2) in comparison with RAPID data. The total temporal variability in the overturning calculated from RAPID data is 3.7 Sv and this decreases to 2.6 Sv after subtraction of the seasonal cycle. The best control model run, CTL2, gives more than 60% of the total (2.4 Sv) and interannual (1.8 Sv) observed variability. The MOC range is only about half of the observed (about 18.9 for RAPID, and 10.7 Sv for CTL2). For the interannual variability the discrepancy in the range is even bigger: 13 Sv for RAPID and 6 Sv for CTL2. Correlation coefficients between the MOC variability and the RAPID array data are also given in Table 2, although these tend to be dominated by the Ekman variability, especially for CTL0. Therefore CTL0 has the highest correlations with RAPID of all the runs because it has so little variability apart from the Ekman signal.
Amplitudes of the MOC seasonal cycle in these experiments, along with the RAPID array calculated values, are shown in Figure 5. According to the RAPID data, minimum values of transport through 26.5°N are observed from February to April, with minimum ∼ 14.3 Sv in March, while from July to January the transport exceeds 21 Sv (with slight peaks of the transport in July and November). The cold start from climatology experiment, CTL1, is considerably better at reproducing the MOC seasonal cycle than the spun-up model CTL0. Control CTL2 best captures the seasonal variability, giving maximal transports for the period from July to January (∼ 19.4 Sv) and minimal transports from February to May (∼ 14.2 Sv in April). Seasonal variabilities at 40°N and 50°N (not shown) are smaller than at 26.5°N (the amplitude of the variability is about 2 Sv). However, the phase of the seasonal cycle also varies with latitude. Minimal transports at 40–50°N are observed in winter, from December to February, probably due to strong southward winter Ekman transports at these higher latitudes. Maximum MOC transports at 40°N and 50°N are in September–November and in April–June, respectively. At 50°N the MOC maximum may also be linked to the peak in water formation in the Labrador Sea in March (Bingham and Hughes, 2007).
In conclusion, for the control experiments CTL2 best captures both mean and seasonal cycle of the MOC at 26.5°N, although the amplitude of its seasonal cycle is still underestimated. We will use these results as a basis for seeking further improvements to the MOC through assimilating the RAPID array boundary data, so that all of the experiments in the following subsection start from the CTL2 initial conditions.
4.2. Assimilation experiments with the RAPID array data
In order to understand the impact of assimilating the RAPID array data we first focus on experiments assimilating this dataset alone, without introducing other data from the deep basin, e.g. Argo. Different methods of assimilating RAPID data at the eastern/western boundaries and on both sides of the Mid-Atlantic Ridge have been investigated (see the Table 1 summary).
The first experiment labelled R0 assimilated the full depth western and eastern boundary 5-day mean T,S profile data, and the profile data from either side of the Mid-Atlantic Ridge, in exactly the same way as in previous open-ocean reanalysis experiments (e.g. Smith and Haines, 2009; Smith et al., 2010), using standard isotropic error covariances (section 3). The 26.5°N MOC time series R0 is shown in Figure 6 (red) and this does give some small improvement over CTL2 (black) during the first two years, with statistics summarised in Table 2; however, a more careful analysis reveals serious problems.
Figure 7(a) shows the R0 annual mean overturning stream function after 5 years of assimilation. The MOC transport is strengthened throughout the latitude band 26–40°N; however, this is coupled with a much stronger Antarctic bottom-water cell south of 26.5°N, along with a greatly weakened upper-level cell (compare with Figure 4(c)). Figure 7(b) shows the barotropic stream-function difference (R0–CTL2) with a strengthened anticyclonic barotropic gyre confined close to the western boundary (this feature is robust but can be much stronger with other initial conditions, e.g. CTL1, not shown). This response has some resemblance to Fig. 2c,d and Fig. 3c,d in Baehr (2010), although the intensity here is larger.
We hypothesize that the circulation anomalies produced are due to inconsistency between mass transports at different latitudes that arises when data at a single latitude are assimilated, with small-scale covariances. The adjustment processes occur well beyond the region influenced by the isotropic covariance scales from the assimilation at 26.5°N (especially downstream in the western boundary currents), and are needed to compensate for the changed transport at 26.5°N. We tried varying the meridional and zonal scales of the covariance functions, keeping the same Gaussian structure, but this did not lead to improvements.
Two modifications to the assimilation procedure were used to improve this response. Section 2.2 shows that a significant component of the northward-flowing Gulf Stream flows adjacent to the deep western boundary (Figure 2), so at the western boundary we first restricted the assimilation depth range to be deeper than 900 m (that is also consistent with the correlation patterns near the western boundary in Figure 3), with these experiments denoted R900. The R900 experiments use fewer data and are at least as good as R0 for reproducing the MOC, see Figure 6 (green) and Table 2; however, the overturning cell remains disrupted (not shown).
In addition (also in accordance with correlation patterns, Figure 3) we therefore sought to spread the influence of the deeper measurements in the RAPID array data along the boundaries, recognizing that a number of other modelling studies have shown correlated density variability down the deep western boundary, e.g. Dong and Sutton (2002). Figure 8(a) shows regression coefficients between temperature time series at 26.5°N and other latitudes along the western boundary. These coefficients are based on 5-day mean fields from the CTL2 run, 2004–2008, but similar results are obtained using either monthly mean temperatures, or using the full 18 years of data from the CTL0 run, shown in Figure 8(b). All time series were detrended prior to these calculations so they represent anomalies with respect to the time mean profiles along the boundary. Figure 8 indicates strong correlations spreading at least 10° of latitude away from 26°N, especially northwards along the western boundary between 1000 and 2000 m depth where over 60% of the temperature (and salinity, not shown) variability can be related to the 26.5°N variations. These covariances are not lagged and therefore represent fast wave processes operating along the boundaries (e.g. Getzlaff et al., 2005).
These regression coefficients were used to construct T,S profiles, consisting of the 2004–2008 time-mean model profiles plus the regressed anomalies, along the deep western boundary below 900 m between 10°N and 35°N, which we denote the ‘quasi-observational RAPID dataset'. These quasi-observational profiles were then assimilated along the western boundary, with zonal-only covariances that spread the results into the basin interior, again with a zonal scale of 375 km.
The model run (Rreg) uses these quasi-observational RAPID data at the western boundary, as well as assimilating on both sides of the Mid-Atlantic Ridge and the eastern boundary, using RAPID data with conventional covariances. This experiment again demonstrates improvements over CTL2 in the MOC results despite being only based on the deep data below 900 m at the western boundary (Table 2 and Figure 9, red solid line). More importantly, Figure 10(a) shows the 2008 average overturning stream function for Rreg, now with intensification of the upper overturning cell both to the south and north of 26.5°N, without the undesirable anomalous changes in the deep Antarctic bottom-water cell circulation and the barotropic stream function (compare Figure 10(b) with Figure 7(b)). Note that the upper overturning cell is now deeper in comparison with CTL2 (compare Figure 10(a) with Figure 4(c)), which is also in a better agreement with observations. Similar (but not identical) results are obtained using boundary regressions based on 18 years of data, shown in Figure 8(b) and listed under Rreg18 in Table 2.
The seasonal cycle of the MOC transport from Rreg is shown on Figure 5 (red line). There is a small increase in the seasonal cycle with assimilation, with improvements particularly in the second half of the year when the transport is ∼ 1 Sv higher compared with CTL2. The peak MOC (∼ 20 Sv) is still lower by ∼ 2 Sv than in the RAPID observations, largely explained by the high MOC amplitudes reached in 2004 coming from the Florida electromagnetic measurements, possibly influenced by the passage of a nearby hurricane. The Rreg transports at 40°N and 50°N (not shown) are not stabilised however, with both transports continuing to decline as in CTL2 (Figure 4(b)) and with the transport at 50°N declining at an even faster rate. We will return to these features below.
4.3. Maintaining the higher MOC and basin-scale impacts
To test the robustness of the MOC changes in response to the assimilated RAPID array data, and to determine whether the improvements in the MOC can be maintained, experiments CTL2 and Rreg were each recycled to overcome the short period of available observational data. Starting from CTL2 and Rreg ocean states in January 2008, the forcing and assimilation of data was begun again from January 2004, denoted CTL2-R and Rreg-R for repetition. The MOC time series from these runs are shown with dash-dotted lines in Figure 9. The high-frequency variability of the runs is similar, indicating the dominant role of the winds. However, there are significant differences from the first run-through of the data. The CTL2-R MOC values have declined relative to CTL2, and this is consistent with knowing that the MOC will run down over time as indicated by the CTL0 results. However, despite the continued application of data assimilation, the Rreg-R MOC values also run down compared to Rreg, and they are now comparable with CTL2-R in 2007–2008 in which no data assimilation has been applied. To understand why higher MOC values are not maintained we need to look at the wider assimilation impacts.
Figure 11 parts (a) and (b) show the 2007 mean upper-ocean (300 m) temperature difference and the mean-sea-level difference in the North Atlantic (Rreg–CTL2). The assimilation increases the MOC and heat transport across 26.5°N by 30–70 TW, or up to 7% (Table 2), thus warming the area of subtropical gyre to the north, particularly along the Gulf Stream path. Large changes in ocean heat content develop at higher latitudes, in particular around the water formation areas of the Labrador and Irminger Seas leading to corresponding changes in the sea level (Figure 10(b)). This pattern of warming shows some similarities to warming patterns that develop after a period of stronger North Atlantic Oscillation (J. Robson, personal communication). We investigate whether these changes can have a negative feedback on the MOC, leading to inability to maintain the strong MOC induced by RAPID array assimilation.
Figure 12 shows the vertical structure of the differences in the Labrador and Irminger Seas, in the two boxes marked by black and red colours in Figure 11(a). Figure 12 shows the 2007 annual density profiles from Rreg (solid) and CTL2 (dashed) (a), (b), and the changes in temperature and salinity relative to the CTL2 run (c), (d), in these boxes allowing the full penetration of the differences to be seen. The warmer and cooler areas in Figure 11(a) extend quite deeply through the top 800 m and dominate the density differences, particularly in the southern box, as also seen from the steric height differences in Figure 11(b). In the northern box the stratification of the water column in the top 800 m layer is increased, making it less susceptible to wintertime convection. The same pattern is also shown by Robson et al. (personal communication) to be associated with changes in subpolar gyre convection and subsequent MOC slowdown after 1995.
Figure 13(a) shows time series for the differences (Rreg–CTL2) of the MOC transport at 26.5°N (solid) and the density averaged over the southern box in the top 500 m (dashed). Presenting differences removes the highly variable Ekman component of the transports, which are the same in each case, allowing the differences in the thermohaline transports to show up more clearly. The boost to the MOC induced by assimilation clearly falls off as the density anomaly further to the north builds up. After 2007 there is no trend in the density time series and the averaged MOC transport becomes almost the same in the two runs.
To test this further, two additional experiments are compared: EN3A in which only the North Atlantic EN3 data are assimilated through the RAPID period, with no assimilation of the RAPID data itself; and Rreg+EN3, which is the same as EN3A but also assimilates the RAPID array data as in the Rreg experiment. Figure 14 shows the MOC transport time series from these two experiments. The Rreg+EN3 experiment (red) gives higher MOC transports than the run assimilating EN3A data alone, by more than 2 Sv. The declining trend observed in the MOC time series at 40 and 50°N in the previous control and assimilation experiments (Figure 4(b)) is also absent in these experiments (not shown) due to assimilating the EN3 data. The assimilation of the EN3 data removes the upper-ocean biases in the subpolar latitudes of the North Atlantic (Figure 11) allowing the positive impact on the MOC transport due to the assimilation of the RAPID observation data to be retained, as shown in the MOC (solid line) and Labrador Sea density difference time series (Figure 13(b), dashed line).
The assimilation run above was also recycled as Rreg+EN3-R to see the assimilation impact on the 26.5°N MOC over a longer time period, and this shows that higher MOC values by about 2 Sv (compared with EN3A) are maintained through a period of at least 9 years of assimilation (green line in Figure 14).
5. Summary and discussion
We have used a simple assimilation method that successfully controls the MOC in the 1 degree resolution global NEMO model, by assimilating ocean densities (via temperature and salinity) from the RAPID Monitoring Observational Array at 26.5°N, particularly in the west near Florida and the Bahamas where the strong Gulf Stream and deep western boundary current flows.
The initial conditions in 2004 are critical to reproducing good circulation in the model, and initial conditions from a previous run assimilating open-ocean hydrographic profile data, including Argo, were found to give the best MOC results. Then assimilating the western boundary array observations below 900 m depth using a quasi-observational RAPID dataset (regressing the influence from the RAPID density data in latitude along the deep western boundary) turns out to be the key step for allowing a further improvement in model MOC results. The MOC and deep western boundary density correlations are based on long model runs, with correlations rising to include the upper layers only in the subpolar gyre, although these correlations are not presently used as part of the assimilation process. Assimilation leads to an increase of the MOC transport at 26.5°N through an intensification of the positive overturning circulation cell in upper-ocean layers, which is distributed to higher and lower latitudes in a smooth fashion.
The model mean MOC mass and heat transports, and correlation coefficients with the directly derived RAPID mass fluxes at 26.5°N, are summarized in Table 2. Assimilation of RAPID data leads to an increase in meridional heat transport at 26.5°N of 30–70 TW and to a warming of the upper layers at higher latitudes (Figure 11(a)), although no data were assimilated into the upper 900 m of the water column at the western boundary. Hence the assimilation of RAPID data has an impact on basin-scale water masses and heat content across the North Atlantic.
Although the assimilation procedure has allowed a direct impact of the RAPID data on the overturning transports at 26.5°N, this does not currently do anything to enhance deep water formation at higher latitudes which would be needed to sustain the higher MOC values in the longer term. Indeed, if RAPID data alone are assimilated in this ocean-only model, there is a lagged negative feedback on the MOC due to changes to the thermocline structure in the Labrador and Irminger Seas. However, better MOC values are more sustained when applying open-ocean assimilation, e.g. from Argo data, along with the RAPID boundary data assimilation. The open-ocean assimilation reduces any negative feedbacks and maintains better open-ocean density distributions, particularly in the subpolar gyre latitudes, allowing a decoupling of the 26.5°N MOC transports from the water formation processes further north.
One might ask whether a more sophisticated assimilation scheme could improve on these results for RAPID data assimilation. Both a 4D-Var scheme and an ensemble Kalman filter scheme would likely identify covariances, or earlier initial condition anomalies, along the western boundary as being the key to fitting variability in the RAPID MOC data. However, greater details in these covariances or the exact time-scales of connection back to the water formation regions that drive the MOC, which one might hope to exploit with a better assimilation scheme, are highly model- and resolution-dependent, e.g. Getzlaff etal. (2005). The problems of using the separate components of the MOC monitoring (the Florida Straits transport or the mid-ocean transport) in lower-resolution models has also been highlighted here, and this problem would still be present with more sophisticated assimilation schemes. The approach here using predefined covariances is reasonably effective and should be applicable in the current operational schemes which might assimilate RAPID data.
An alternative approach, that might take better account of these model biases, would be to treat the MOC transport (or anomaly) itself as the observable and seek to assimilate it through adjustments to model density. Both 4D-Var and ensemble covariance filters could immediately provide suitable schemes to do this, although again more-conventional methods using model-derived covariances can still be used (Hermanson et al. is a paper in preparation that uses this method to assimilate RAPID data with the HadCM3 coupled model).
Perhaps of greater importance to using RAPID data for predictions is: what would be the coupled model response to RAPID data assimilation? The atmosphere can then respond to the enhanced poleward heat transport from an enhanced MOC, leading to additional feedbacks. The response of a coupled model on time-scales of a year or more may therefore be very different from the ocean-only models, and further study is needed to clarify this issue. The results here are valuable on two fronts: (i) we have demonstrated the ability to assimilate deep hydrographic data near to a boundary using a simple assimilation system by taking account of boundary covariances; and (ii) this method could be used to modify the MOC in a coupled atmosphere–ocean model, which could then be used to test enhanced climate predictions on the basis that higher MOC transport will transport more heat which will have an impact on the atmosphere in the high and moderate latitudes.
Further work is under way to investigate both the effect of assimilation of RAPID data in a higher-resolution ocean model to test the robustness of the results, and also the impact of RAPID data in coupled models, where atmospheric feedbacks can occur and from where initialised climate predictions can be launched. The patterns of ocean heat content, sea-surface temperature and height that are generated by changes to the model MOC may also prove to be useful in helping to understand other ways of monitoring the MOC and the changes that it can produce in the wider basin-scale fields.
The authors would like to thank Maria Valdivieso for providing model assimilation results with the EN3 data used in our MOC analysis and some of the model runs as initial conditions. Comments from two anonymous reviewers helped to significantly improve the manuscript. This work was supported by the NCEO and the RAPID-WATCH VALOR projects.