1. Introduction
Estimating the state of the ocean is a primary target both in the context of climate variability assessments and for operational purposes such as the initialization of the ocean for the production of seasonal and longer time-scale forecasts of the coupled ocean–atmosphere system. Several efforts aimed at extending the available observational dataset of the global ocean have been made in the past two decades, organized as international programs, such as the World Ocean Circulation Experiment (WOCE), the Tropical Ocean Global Atmosphere program (TOGA), and the Joint Global Ocean Flux Study (JGOFS). However, observations provide a patchy description of the ocean state, and a reliable state estimate can only be achieved by combining in an optimal way the existing observations with our theoretical knowledge of the ocean dynamics embodied in general circulation models. Considerable advancements have been made in the development of ocean data assimilation techniques, partly fed by the long-used expertise developed within the numerical weather prediction (NWP) community.
A widely used assimilation technique, currently employed in the production of ocean initial conditions for operational ocean and seasonal forecasting, is optimal interpolation (Klemm et al. 2000; Alves et al. 2004). OI was also the standard tool used for NWP (Daley 1991) although the major operational centers have gradually upgraded their systems by resorting to more elaborate variational assimilation methods (see Parrish and Derber 1992; Rabier et al. 2000; Lorenc et al. 2000, for a detailed description of the operational systems developed at the U.S. National Meteorology Center, the European Centre for Medium-Range Weather Forecasts, and the Met Office, respectively). Variational methods are also being developed within the ocean data assimilation community. Special efforts have been devoted to the development of sophisticated four-dimensional variational data assimilation (4DVAR) techniques, which use the ocean model dynamical equations as a constraint on the analysis increment (Weaver et al. 2003; Vialard et al. 2003). Despite their appeal, the practical implementation of the 4DVAR techniques for operational oceanographic applications does still appear to be at a pioneering stage.
The beneficial impact on the skill of seasonal and interannual forecasts stemming from the use of ocean analyses (i.e., ocean states obtained through the assimilation of oceanic data) as initial conditions for coupled ocean–atmosphere models has been assessed in a number of papers, mostly focusing on the El Niño–Southern Oscillation (ENSO) phenomenology (Ji and Leetma 1997; Rosati et al. 1997; Segschneider et al. 2001; Alves et al. 2004). The constant quest for innovative strategies aimed to improve the skill of seasonal climate forecasts paved the way to an additional development branch in ocean data assimilation (ODA) studies, focusing on the optimal combination of the available observations (mainly hydrography and altimetry data) in the assimilation process (Segschneider et al. 2001; Masina et al. 2001; Carton et al. 2000).
An aspect that has been somewhat overlooked in the ocean reanalyses production is the role of salinity. The scarcity of direct salinity observations and the belief that this tracer has a second-order impact on the tropical ocean density structure compared to temperature, were reasons why salinity has often been left unchanged during the temperature updating process. The negative effects associated with this practice have been pointed out by Troccoli and Haines (1999, hereafter TH99) and Troccoli et al. (2002) who show that assimilating temperature without simultaneously adjusting salinity may concur to alter the vertical stability of the water column whenever the background thermal stratification is weakened by the temperature assimilation. This may ultimately lead to spurious convection and mixing, thus altering both temperature (T) and salinity (S) stratification. The scheme implemented by TH99, correcting the salinity with T–S relationships built from the most recently observed (or model predicted) T and S profiles, partly compensates for this effect, also improving the corrections that can be obtained via climatological T–S relationships. Resorting to T–S relationships in the parameterization of the background error is an appealing option, although it presents the disadvantage of not being applicable to areas dominated by nonisentropic flow regimes (i.e., in the mixed layer or nearby river outflows), where temperature and salinity changes are typically uncorrelated. The shortcomings of the T–S method in areas characterized by strong ENSO-related interannual variability, such as the western tropical Pacific Ocean have been illustrated by several authors (among others, Donguy 1994). An alternative approach is to use vertical empirical orthogonal functions (EOF) of bivariate temperature and salinity profiles, either observed or derived from an OGCM output. Maes et al. (2000) successfully reconstruct the upper-ocean salinity variability in the western tropical Pacific over annual and ENSO time scales, by using combined vertical EOFs of temperature and salinity as basis functions for a least squares minimization method. Error modeling through vertical EOFs in the context of an OI scheme has been first explored by De Mey and Robinson (1987). One of the main advantages related to the use of EOFs is the possibility of operating an order reduction of the assimilation problem. Taking advantage of the redness of the eigenvalue spectrum, mode decomposition can be truncated to the leading dominant modes, explaining the largest fraction of the variance. On the other hand, a major practical difficulty in the application of this methodology is obtaining reliable EOFs for the area of interest. Sufficiently long records of simultaneous observations of temperature and salinity are only available in a few locations of the World Ocean. Therefore the use of observations for diagnosing EOFs is unfeasible in the perspective of producing global ocean reanalyses. This makes it necessary to resort to model synthetic data to build multivariate structure functions to model the background error.
In this paper we present results from a global ocean data assimilation system, designed for the assimilation of both temperature and salinity observations, based on a multivariate optimal interpolator scheme [i.e., the System for Ocean Forecasting and Analysis scheme (SOFA; De Mey and Benkiran 2002)]. The background error covariance is modeled through vertical EOFs of bivariate temperature and salinity, diagnosed from the output of an OGCM. The system [partly developed within the framework of the European Commission Enhanced Ocean Data Assimilation and Climate Prediction (ENACT) project] has been tested by producing global ocean reanalyses over the 1962–2001 period, using the ENACT in situ dataset, and the atmospheric fluxes from the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) project (Uppala et al. 2005). Results from a set of multidecadal ocean reanalyes as well as from a control experiment where no data are assimilated will be presented. A major focus of this work will be the system’s skill in reproducing the mean and temporal variability of the observed salinity. In particular, the impact of time and space resolution of the background error covariance will be addressed.
The paper is structured as follows. Section 2 describes the ODA system: details are given on the OI algorithm, with particular focus on the background error covariance matrix formulation, and on the main features of the ocean dynamical model. In section 3, the experimental strategy adopted to perform the ocean reanalyses is thoroughly described. The effects of in situ data assimilation on the subsurface thermal field and on the surface and subsurface currents are evaluated in sections 4 and 5. In section 6 additional analyses where both temperature and salinity profiles are assimilated are described, and the model salinity mean state and variability are evaluated. In particular, the sensitivity of the analyzed salinity to the spatial and temporal structure of the background error covariance is assessed. The impact of the subsurface temperature and salinity assimilation on sea level variability is analyzed in section 7. Concluding remarks are given in section 8.
2. The ocean data assimilation system
a. The OI scheme
The data assimilation system used to produce the present ocean reanalyses is based on SOFA (De Mey and Benkiran 2002) and has been designed for the assimilation of in situ (vertical profiles of temperature and salinity) data. Details on the SOFA algorithm within the framework of the OI are given below.
An important feature of SOFA lies in the multivariate structure of the background error covariance matrix, which spreads the corrections over different types of variables. In the present SOFA implementation the state vector is defined as the bivariate temperature and salinity vector. This implies that when, for example, vertical temperature profiles are assimilated, corrections are applied to salinity as well by the vertical EOFs.
The multivariate EOFs used for assimilating in situ data have been diagnosed from the synthetic dataset of vertical temperature and salinity profiles provided by an ocean model simulation (hereafter, the control run) where no data assimilation has been applied. The use of model-derived EOFs was dictated by the scarce availability of simultaneously observed temperature and salinity profiles, typically provided by conductivity–temperature–depth (CTD) data. The data scarcity (particularly in the Southern Hemisphere) poses a strong constraint on the dimensions of the regions to be used in the EOF calculation, leading to the artificial stretching of the regional boundaries so as to include a sufficiently large number of observed profiles, hence conflicting with the need to identify the dynamically homogeneous regions.
A domain splitting approach, already used in Sparnocchia et al. (2003) for the Mediterranean Sea, has been adopted for the computation of the EOFs, consisting in the partition of the global ocean domain into subregions encompassing areas characterized by homogeneous dynamical regimes. For each region an area-averaged bivariate vertical profile of temperature and salinity is calculated. Time series of the anomalous temperature and salinity are obtained by removing a regional climatology, the latter calculated as the time mean over the whole length of the control experiment. Normalization is then applied by dividing the temperature and salinity anomalies by the depth-averaged standard deviation of the corresponding variable. The use of the depth-averaged standard deviation as a normalizing factor for each level is meant to avoid large weights applied to deep layers, which are typically characterized by smaller variances. The bivariate temperature and salinity covariance matrix is then constructed, and the corresponding eigenvectors are computed. Only a subset of the full EOF space, containing the 10 dominant bivariate temperature and salinity EOFs (the so-called reduced space) has been retained so as to include the largest part of the explained variance.
An alternative procedure, where EOFs are computed at each model grid point and for each season (detailed in section 6), has been adopted for one of the performed analyses. The vertical structure of the EOFs is fairly sensitive to the procedure used to compute them. In Fig. 1 we compare the first EOF of salinity in the northeastern extratropical Pacific, obtained with the domain splitting approach, with the leading pointwise seasonal EOFs for the same region. It is evident that the procedure of averaging over large areas of the global ocean adopted for the computation of regional EOFs determines vertical covarying patterns with a considerably shorter e-folding scale, compared to gridpoint seasonal EOFs. The implications deriving from using one or the other set of EOFs for the performance of the assimilation system will be discussed in detail in sections 6–8.
Corrections on temperature and salinity are applied to the full depth of the water column. Troccoli et al. (2002) and Ricci et al. (2005) point out that in areas dominated by nonisentropic processes, such as nearby river outflows or in the mixed layer, salinity corrections derived from T–S relationships may lead to unrealistic vertical density profiles, because of the lack of correlation between temperature and salinity. To avoid this, they suggest not to update salinity in those specific regions. In the present scheme, T–S correlations are embedded in the bivariate EOFs. Thus, under weak T–S correlation conditions, salinity increments will be consistently low. This allows us to extend the present methodology to areas governed by nonisentropic processes.
The background error and the observational error covariances are assumed to be static (i.e., not dynamically evolving with time). The observational error is also assumed to be uncorrelated in space and time, with a standard deviation of 0.5°C and 0.5 psu, for temperature and salinity data, respectively. The background error correlation is an analytical Gaussian model, with an isotropic and homogenous correlation radius of 300 km for all modes. A 400-km data-selection radius (also referred to as bubble radius) has been used. Observed temperature and salinity profiles spanning the 1962–2001 period have been used in the assimilation process.
The in situ data have been collected and quality checked within the ENACT project (Ingleby and Huddleston 2007). Unfortunately a fall-rate correction has been mistakenly applied twice to expendable bathythermograph (XBT) data in the 1995–2001 period. This error mostly impacts the temperature trends for the late part of the assimilation period (Davey et al. 2006) and it should not affect the main conclusions of this paper.
b. The ocean model
The ocean general circulation model used to produce the ocean analyses is a global implementation of a free-surface version of the Océan Parallélisé (OPA) model (Madec et al. 1998). The horizontal resolution is 2° × 2° cosφ everywhere, except for the 20°N–20°S latitude belt, where the meridional grid spacing is progressively decreased to 0.5° to better resolve the equatorial dynamics. The grid is irregular, featuring two poles in the Northern Hemisphere, both located over the continents so as to avoid the North Pole singularity, while the Southern Hemisphere grid pole is naturally located over Antarctica. There are 31 unevenly spaced levels along the vertical with the top 20 levels concentrated in the upper 500 m, and the last temperature level under the ocean floor. The vertical mixing parameterization is based on a turbulent kinetic energy prognostic equation (Blanke and Delecluse 1993). The horizontal viscosity coefficient is space dependent: a typical value of 40 000 m2 s−1 is applied poleward of the 20°N–20°S belt, while it gradually decreases to 2000 m2 s−1 in the equatorial region, except for the western boundary regions. Tracers are diffused along isopycnal surfaces with an eddy-mixing coefficient of 2000 m2 s−1. Lateral diffusion is supplemented with the Gent and McWilliams (1990) eddy-induced velocity parameterization (Lazar et al. 1999). A typical strength of the eddy-induced velocity is 2000 m2 s−1, but it is decreased in the equatorial region. The time step is 1 h 36 min for both dynamics and tracers. There is no sea ice model coupled to the ocean model. A climatological ice distribution is used.
3. Experimental setup
The ODA system described in the previous section has been used to produce an extensive set of global ocean analyses, covering the data stream from 1962 to 2001. The two reference analyses, covering the long 1962–2001 stream, are identified as OIT and OITS (see Table 1). In the OIT experiment, temperature profiles are assimilated, while salinity is corrected based on the multivariate T–S statistics provided by the model EOFs. In the OITS experiment, both temperature and salinity vertical profiles are assimilated. It is worth noticing that in this experiment, the multivariate EOFs are still used to correct salinity whenever a temperature profile, which is not twinned with a collocated salinity profile (as typically occurs for CTD data), is assimilated. Similarly, salinity profiles will induce temperature corrections when only salinity observations are available. Since most instruments that provide measurements of salinity (e.g., CTDs, Argo floats) also provide measurements of temperature at the same location (while the reverse is not always true) the latter condition is rarely met. Details regarding an additional analysis (labeled as OITSHR) will be given in section 6.
To evaluate the impact of the assimilation on the ocean state, each experiment consists of 1) a control run, where the model is simply forced with daily atmospheric fluxes but without assimilating any data; and 2) an assimilation run, which (besides the atmospheric forcing) includes the assimilation of in situ data. To generate appropriate initial conditions for starting the analysis run, a suitable spinup strategy has been adopted, described in the following section.
a. Spinup strategy
The spinup has the effect of providing an initial ocean state for the start of the assimilation run, which is not too far from realistic conditions. This goal is achieved by using a strong relaxation on sea surface temperature, combined with a weaker subsurface restoring to climatological profiles of temperature and salinity so as to reduce model drift without preventing the onset of low-frequency variability. The model was spun up with ERA-40-derived climatological fluxes of momentum, heat, and freshwater, from 1953 to 1957 (included), starting from a motionless ocean, and Levitus hydrographic initial conditions (Levitus et al. 1998). Daily ERA-40 fluxes were used from January 1958 to December 1961. Sea surface temperatures were restored to an ERA-40 climatological year (1971–2000) with a Newtonian damping heat flux of 200 W (°C m2)−1, corresponding to a restoring time scale of about 12 days (assuming a mixed layer thickness of 50 m). A weaker relaxation to Levitus temperature and salinity climatology along the whole water column was also applied, with a 3-yr damping time scale. The ocean state at the end of 1961 provided the initial conditions for both the control and the assimilation run, starting on 1 January 1962. The integrations have been performed by using the same forcings and restoring parameters adopted to generate the 1958–61 nonclimatological spinup, except for the sea surface temperatures that were relaxed to Reynolds temperatures (Reynolds 1988) from 1982 onward, linearly interpolated to daily values. During the entire length of the runs, a daily adjustment was applied to the model sea surface height aimed to remove a drift associated with the nonzero residual of the globally averaged freshwater fluxes. An improved version of ERA-40 freshwater fluxes, correcting a bias in the precipitation (Troccoli and Källberg 2004) has been used. To prevent the onset of a numerical instability in the Southern Ocean, off the Antarctic coast, a full-depth relaxation to Levitus temperature and salinity climatology was applied poleward of 60°N/60°S, with a gradual reduction of the restoring time scale from 3 yr to 50 days occurring in the 60°–70°N (S) latitude belt. The restoring time scale was 50 days at the top level, gradually increasing to 1 yr at the bottom.
b. Assimilation cycle
The assimilation cycle is based on the following procedure. The OGCM is run for 14 days. During this time interval, innovations (observation–background) are computed using the background state at the appropriate measurement time. A Gaussian weighting function, with a 7-day time scale, is applied to the innovation vector in order to take into account the temporal distance of observations from the time when the model state is actually updated (the analysis time): innovations that are farther (closer) in time from the analysis time will have a lower (larger) weight. At the end of the OGCM run, the state variables of the model solution at the analysis time (taken as the central time step of the assimilation cycle), are updated by SOFA with the correction fields computed at each model grid point from the innovation vector. The updated model solution, at the beginning of the 8th day of the assimilation cycle, provides the initial conditions for the following cycle, with the OGCM restarting from the corrected model state.
4. Ocean analysis evaluation: The impact of temperature assimilation on the thermal mean state and variability
a. Consistency check
The internal consistency of the system is assessed by looking at the innovation (computed at the observation points) statistics and monitoring the fit of the analysis to the observations at different depths. In Fig. 2, the time evolution of the globally averaged root-mean-square (rms) of the temperature innovation (time mean over each assimilation cycle) at 25- and 160-m depth, is shown for the OIT experiment. A decreasing trend is clearly visible at both depths, showing that the analysis is progressively converging toward the observed state, although the near-surface innovation decreases at a faster rate compared to 160-m depth. The larger data coverage of the surface ocean, compared to deeper layers, is a likely cause of the depth dependency displayed by the innovation statistics.
b. Mean temperature
The impact of temperature assimilation on the ocean mean state will be assessed by analyzing the climatological temperature fields from experiment OIT. The Levitus dataset is here adopted as a reference global ocean observed state. We will refer to CTRL, OIT, and LEV as the climatologies derived from the control and assimilation runs (computed over the entire stream length) and the annually averaged Levitus dataset, respectively. An index of the system performance is provided by the CTRL–LEV versus OIT–LEV residuals intercomparison. Since CTRL and OIT experiments only differ for the assimilation of observed thermal profiles applied in the latter experiment, discrepancies in the residual fields can only be ascribed to the impact of data assimilation. OIT–LEV temperature residuals at 100-m depth appear considerably reduced compared to CTRL–LEV, the former residual distribution displaying anomalies smaller than 1°C over most of the Northern Hemisphere (Fig. 3). The western boundary current regions and their subpolar extension also show a beneficial impact from the assimilation: OIT–LEV anomalies having a 1°C magnitude show a much reduced extension compared to the CTRL–LEV residuals. The Southern Hemisphere exhibits a relatively modest improvement when data assimilation is applied. This is particularly evident in the Antarctic Circumpolar belt. The data scarcity, which typically characterizes the Southern Hemisphere, is the most likely cause underlying such asymmetry, although it must be mentioned that the same circumstance affects the reliablity of Levitus data in that area of the World Ocean.
We now focus on the impact of temperature assimilation on the vertical structure of the thermocline. In Fig. 4 we show the temperature root-mean-square differences as a function of depth between analysis and observations and between control and observations using data from the Tropical Atmosphere–Ocean (TAO) array (McPhaden et al. 1998), in the Niño-4 and Niño-3 areas, defined in the latitude–longitude range (5°S–5°N, 160°–210°E) and (5°S–5°N, 210°–270°E), respectively. The largest deviations from observations are detected in the thermocline region at about 100- and 200-m depth, for the Niño-3 and Niño-4 areas, respectively. The different rms maximum depth mirrors the thermocline slope existing in the equatorial Pacific region. The assimilation of vertical temperature profiles reduces the rms deviation from observations below 1°C, with a sensible improvement with respect to the control experiment. The intercomparison of time–depth temperature sections from the control and OIT experiments with the available data from the TAO array at 2°N, 110°W (not shown) reveals that the assimilation of vertical temperature profiles considerably improves the thermocline representation. In particular, the OIT analysis displays tightly packed isotherms at about 100 m resulting in a sharper thermocline compared to the rather smooth stratification exhibited by the control run. Hence, the assimilation corrects the tendency of pure simulation experiments (i.e., without data assimilation) to generate weakly stratified ocean states caused by spuriously high vertical eddy diffusion, a well-known OGCM deficiency. This feature has been found in a number of assimilation studies (Derber and Rosati 1989; Masina et al. 2001; Vialard et al. 2003; Bell et al. 2004). It is interesting to note that in Niño-3, over the 200–500-m depth range, and Niño-4, around 300-m depth, the rms in OIT is actually larger than in the control. A closer inspection of the vertical circulation structure near the equator revealed the existence of a circulation cell in the eastern Pacific featured by OIT, but not present in the control (not shown). This cell is responsible for enhanced upwelling off the Peruvian coast below the thermocline, causing the cooling of subsurface waters therein. While this process is fairly localized in space, it is responsible for a cold bias in OIT which ultimately results in the increased rms detected in the Niño-3 area, below the thermocline, overshadowing the impact of thermal data assimilation. A similar vertical velocity anomaly is also visible in the central Pacific and could be responsible for the weaker subsurface bias in Niño-4. Similar vertical circulation anomalies in the eastern equatorial Pacific have been extensively documented in a number of papers (Vialard et al. 2003; Bell et al. 2004; Burgers et al. 2002) and appear to be associated with a disruption of the main dynamical balance along the equator caused when hydrographic data are assimilated without consistently adjusting the wind forcing.
c. Additional validations
In this section the impact of thermal profile assimilation is further assessed by comparing the ocean analysis with temperature observations in the tropical and northern subtropical Atlantic areas. The observations used for this assessment are part of the assimilation dataset. Upper-ocean temperature changes in the tropical Atlantic are well documented in the Pilot Research Moored Array in the Atlantic (PIRATA) array data. Time–depth sections in the western tropical Atlantic (8°N, 35°W) of monthly mean temperatures from the control and OIT experiments are compared with data from the PIRATA array (Fig. 5). The most evident effect related to the assimilation of thermal profiles is the recovery of the periodic shoaling/deepening of the seasonal thermocline. The absence of this signal in the control experiment causes the large anomalies found at 100-m depth. The OIT analysis thermal field is here further validated against data characterized by a limited temporal scale, but wide spatial extension. The analysis skill in reproducing the upper-ocean thermal structure in the subtropical North Atlantic is evaluated by looking at the observed 1993 mean temperatures from the WOCE Upper Ocean Thermal (UOT) database, at 30°N (Fig. 6). The most evident effect of in situ data assimilation can be traced in the system’s ability to reconstruct the observed features of the subtropical mode water, also known as 18° water (Talley and Raymer 1982). This water mass can be identified as the 18°–19°C isothermal layer. The maximum layer thickness at 70°W (totally missing in the control experiment) indicates the location where the water mass generation process (intermediate depth convection) occurs. The absence of the subtropical mode water in the control run shows that the model is unable to generate this water mass, which may be possibly due to the lack of sufficiently strong cooling events in the forcing fields, or to unresolved processes of the mixed layer dynamics. A previous study (Masina et al. 2004, see their Fig. 3) show that even in an eddy-permitting model, the 18° water was reproduced only when subsurface temperature observations were assimilated.
5. Surface and subsurface currents
In the previous sections, the impact of data assimilation on the subsurface temperature (i.e., the assimilated field) has been analyzed. In the present section, the influence of temperature assimilation on a variable that is not directly assimilated will be evaluated. In Fig. 7 the Reverdin et al. (1994) climatology for the surface zonal currents in the tropical Pacific is compared to the OIT analysis and the control experiment. The Reverdin climatology was derived from drifting buoys and TAO current meters and refers to the January 1987–April 1992 period. Zonal current climatologies over the same time interval have been computed for the analysis and the control simulation. Some features of the equatorial surface circulation, such as the eastward flowing North Equatorial Counter Current (NECC) around 7°N and the northern branch of the westward South Equatorial Current (SEC) near 2°N, are captured by the control run. However, the strength of the NECC is considerably weaker, while the amplitude of the northern SEC is overestimated, when compared to the observed climatology. Moreover, the control experiment fails in correctly reproducing the southern branch of the SEC, which appears to be too weak. The assimilation of temperature partly compensates for the control experiment inaccuracies, by strengthening the NECC and enhancing the southern branch of the SEC, around 3°S. The northern SEC displayed by the analysis is still too intense and extending far too west, compared to the Reverdin climatology.
The effects of data assimilation on subsurface zonal currents is now inspected by comparing the analysis and control experiment results with climatological profiles from TAO acoustic Doppler profiler (ADCP) observations for the 1992–2001 period, at two locations in the western and eastern equatorial Pacific. In the eastern Pacific (0°, 110°W; Fig. 8, right panel) both the control and the analysis exhibit a westward bias. This indicates an overall underestimation of the Equatorial Under Current (EUC). In the central Pacific (0°, 140°W; Fig. 8, left panel) the assimilation of temperature produces an overall improvement of the zonal current profile, compared to the control, although the core of the EUC at 100-m depth is almost identical in both experiments, and similarly biased to the west with respect to the observations. The fact that the assimilation is unable to correct the model bias in the subsurface current field suggests the following. The EUC strength is governed by the dynamical equilibrium between the zonal pressure gradient and the vertical penetration of momentum associated with the equatorial easterly winds. The assimilation of temperature is potentially able to correct the biases in the zonal pressure gradients but not the inaccuracies in the wind stress forcing field, which may ultimately cause the excess of downward penetration of westward zonal momentum. Dissipative processes associated with an excessively high viscosity, and that are not directly influenced by assimilation, may also determine the persistence of the EUC bias in the analysis.
It is interesting to notice that Vialard et al. (2003) found qualitatively similar results using a 4DVAR algorithm combined with a tropical Pacific configuration of the OPA model. The authors identify spurious vertical mixing associated with the selected background error vertical correlation scales as contributing to the disagreement with the observations.
6. Salinity assimilation: Sensitivity to horizontal and temporal resolution of the background error model
In the present section the system’s ability in reproducing the observed salinity is evaluated. The scarcity of direct salinity observations, compared to in situ temperature, severely impacts the quality of this particular ocean property. Results from a hierarchy of experiments, implementing different assimilation strategies aimed to improve the salinity representation will be presented. As a first step, the direct assimilation of available salinity profiles has been introduced. An analysis based on the assimilation of both temperature and salinity observations has been produced for the 1962–2001 period. With the only exception for the inclusion of salinity profiles in the assimilation dataset, this experiment, labeled as OITS, is identical to OIT. Experiment OITS allows us to evaluate the impact of corrections as derived from the direct assimilation of observed salinity against those obtained from the multivariate EOF statistics (as it is done for experiment OIT). Finally, the sensitivity of the salinity field with respect to the horizontal and temporal resolution of the background error is inspected.
The bivariate EOFs for experiment OIT (and OITS) have been computed by splitting the model domain into a number of macro regions, which may not capture in the most appropriate way the dominant dynamical regimes and uniform T (S) relationships in a given area of the World Ocean. In particular, a too crude partition of the model domain may determine the mixing of different regimes, which may ultimately cause the spurious spreading of some of the covarying T–S patterns beyond the boundaries of their actual extension. An additional source of inaccuracy may be associated with the temporal resolution of the EOFs. The computation of the EOF set adopted for experiment OIT has been based on the full length of the control experiment, and no time evolution has been introduced. This is equivalent to assuming that the T–S variability modes (hence, the background error statistics) are stationary. To clarify the impact of time and space resolution of the background error, an additional experiment, indicated as OITSHR, is designed making use of EOFs computed at each grid point, using coupled T–S profiles from the control experiment. For this experiment, prior to the EOF computation, the profiles are averaged over a bubble centered on the corresponding grid point. This additional processing is equivalent to apply a space filtering to the vertical EOFs, enabling a smooth lateral transition in the background error correlation patterns embodied by the EOFs. A further essential difference with respect to the approach adopted for OIT is in the way the temporal resolution is dealt with. In OITSHR the computation is performed by grouping the T–S synthetic data by season, using the full length of the 40-yr control experiment. Hence, seasonal EOFs are obtained. During the assimilation, the EOFs are changed according to the current season, thus introducing time variability in the background error covariance.
a. Mean salinity
In Figs. 9 and 10, the climatological surface salinity patterns diagnosed from OIT, OITS, OITSHR, and the control experiment are compared to Levitus climatology, using the same approach already used in section 4b for the mean temperature fields. The intercomparison shows that when only temperature is assimilated (OIT), the largest bias corrections occur in the North Pacific area. The fresh anomaly exhibited by the control experiment in the subpolar North Atlantic and in the Gulf Stream extension regions is largely corrected in the OIT analysis, but the latter displays a 1-psu fresh anomaly, which was absent in the control, in the subtropical and tropical North Atlantic, suggesting the onset of new biases as a result of the assimilation of only temperature profiles. On the other hand, the OITS–Levitus residuals for the climatological surface salinity reveal that the most evident effect deriving from the assimilation of salinity profiles is a considerable reduction of the biases in the Gulf Stream region and in the subtropical North Atlantic areas, when a comparison with OIT is done. Differences with Levitus climatology appear to be reduced in the Indonesian area as well. When a higher-resolution background error is employed, a number of improvements appear with respect to OITS. The subpolar North Pacific fresh anomaly spotted in the control, OIT and OITS experiments, has now disappeared. Also the bias along the extratropical North Atlantic reveals a reduced extension.
Figures 11 and 12 show meridional sections of salinity climatology differences between CTRL, OIT, OITS, OITSHR, and Levitus for the top 270 m, across the Atlantic (30°W) and the Pacific (178°E) basins. The mid-Atlantic section displays the largest differences at 100-m depth, around the equator, with values as large as 0.6 psu, likely due to a misplaced position of the pycnocline close the western Atlantic boundary. All the analyses show a reduced vertical extension of the equatorial bias. OIT and—to a lesser extent—OITS also reveal salinity differences with Levitus up to 0.4 psu in the Northern Hemisphere, which are not found in CTRL. It is interesting to notice that OITS displays a fairly low difference with Levitus in the north subtropical gyre, around 30°N. OITSHR, on the other hand, shows smaller than 0.1-psu anomalies (in absolute value) below 100 m over most of the section. Except for the upper subtropical gyre (where differences with Levitus can be as high as 0.3 psu), there is an overall improvement with respect to the control and the other analyses in the Northern Hemisphere, particularly evident at extratropical latitudes. In the Pacific, the impact of assimilation is particularly evident in the northern extratropical belt between 40° and 50°N, and around 20°N where upper-layer fresh biases of 0.4–0.5 psu featured by the control are largely reduced in the analyses. The assimilation of salinity, particularly when combined with the use of pointwise EOFs, proves to be successful in considerably reducing the subsurface positive anomaly displayed by the control in the 40°–60°N latitude range. Overall, OITSHR shows a better skill in decreasing both the amplitude and the vertical extension of model biases across most of the northern and the southern part of the section.
b. Salinity variability
The fit of salinity analyses to the observations used in the assimilation is here estimated by comparing the salinity from experiments CTRL, OIT, OITS, and OITSHR to an objective analysis (OA), which has been provided by the Met Office during the ENACT project and has been obtained using the ENACT in situ data (the ENACT final report is available online at www.ecmwf.int/research/EU_projects/ENACT/). The OA provides a three-dimensional model-independent representation of the observations assimilated in our analyses. Space and temporal gaps in the observations are filled with Levitus climatology, using a 9-month relaxation time scale. The use of the OA as an observational reference presents the advantage, compared to Levitus, of representing a time-varying estimate of the ocean state for the period covered by the analyses. By averaging over the top 300 m and removing the monthly climatological cycle from each analysis for the 40-yr time series, we will focus on the interannual and decadal variability of salinity (hereafter S300) in the upper ocean, where most of the assimilated data are concentrated. In Fig. 13, the S300 time series for the global ocean is shown. Overall, all the analyses exhibit a larger variability over both interannual and longer time scales, compared to the control. OIT shows a negative trend that progressively reduces the salt content in the upper layers of the water column. The OA displays a salinity decline, occurring during the 1985–92 period, which is fairly well captured by all of the experiments but the control. However, there is no evidence of the large S300 negative anomalies featured by OITS and OIT during the 1990–2001 decade. Among the analyses, the OITSHR experiment appears to be the most effective in reproducing the variability revealed by the OA. In particular, OITSHR shows a more stable response during the 1975–2001 period, and a better fit, compared to OIT and OITS, to the observed salinity during the 1990s. A major discrepancy between OITSHR and OA appears during the 1992–95 period, even though the amplitude of the fresh anomaly (present in all the other analyses) is decreased in this case. In Fig. 14, the S300 time series for the northeastern Atlantic (30°–60°N, 40°W–0°) are shown. The agreement between OITSHR and OA variability is here even more striking than in the global case. The scatter between the analyses increases after year 1983. The 1990s are characterized by the largest analysis–observation discrepancies, although OITSHR performs overall better than OIT and OITS.
The ability of the ODA system in reproducing the observed salinity will now be tested against available long-term in situ observations. There are only few long-term observations of salinity in the World Ocean. Two reasonably long time series are selected: the long-term salinity record of Bermuda S station (hereafter BS) in the western subtropical Atlantic gyre (32°N, 64°W), and the sea surface salinity (SSS) series in the western tropical Pacific (165°E) from TAO moorings observations, for the 1997–2001 period. In Fig. 15, we show the mean salinity profile from observations and the model, evaluated at the BS location, for the 1962–2001 time period. The OIT experiment reveals a fresh bias of about 0.5 psu, over most of the upper-water column, which is largely reduced in the control and the other salinity analyses. The onset of a freshening signal starting in year 1995, but absent in both observations and the control run, highly deteriorates the performance of experiments OITS and OITSHR. When the 1995–2001 time segment is removed from the mentioned analyses, the corresponding mean profile appears to be closer to the observations, particularly below 50 m. To explain the late-1990s freshening detected in OITS and OITSHR it must be mentioned that most of the salinity data from Bermuda station did not pass the quality check from approximately 1995 onward, and therefore were not included in the assimilation dataset. This in turn implies that salinity corrections in the last segment of the OITS time series are entirely based on the T–S EOFs.
We now analyze the SSS variability in the western tropical Pacific. The 8°S–8°N meridional section centered on 165°E, from 1997 to 2001, is selected for the relatively good temporal and spatial coverage. We focus on SSS temporal variability by showing only anomalies computed with respect to the 1997–2001 mean (Fig. 16). All the time series show ENSO-related variability at interannual scales. Both the simulation and the analyses display a large negative anomaly associated with the 1997/98 El Niño event. On the other hand, none of the performed experiments captures the amplitude of the positive SSS deviation displayed by the observations at the equator during the 2000–01 La Niña episode. All of the analyses show a larger-amplitude seasonal cycle in the southern part of the section (around 5°S), compared to the much weaker seasonal variability of the control experiment. In general, the analyses show a larger meridional correlation scale in the SSS anomalies, compared to the control. Time correlation with TAO data, at several latitudes, is summarized in Table 2. When compared to OIT and OITS, the control displays a better correlation almost everywhere. On the other hand, OITSHR reveals a slighlty higher skill than the control at 8°S and in the 2°–5°N latitude range. In general, correlations are relatively low, with values never reaching 0.9. All the displayed correlation values but one are statistically significant at the 95% level.
7. Comparison with sea level
The impact of subsurface temperature and salinity corrections on the sea level anomaly (SLA) in the tropical Pacific is evaluated by comparing the SLA field from the analyses and the control experiments against the Ocean Topography Experiment (TOPEX)/Poseidon (T/P) altimetry data, for the 1993–99 period. Changes on the vertical density structure (particularly below the mixed layer) have a sizeable impact on sea level variability. Here we want to assess how the different strategies used to produce salinity increments project onto SLA temporal changes. Monthly maps of T/P SLA data (obtained with the mapping method detailed in Ducet et al. 2000) were produced and made available by the Collecte Localisation Satellites (CLS) during the ENACT project.
T/P altimetry observations represent an independent set of data, since in the present set of analyses sea level is not assimilated. To maintain consistency with the observed SLAs [which have been computed with respect to the 1993–99 Rio (Rio and Hernandez 2004) mean dynamic topography], model SLAs have been computed by removing the 1993–99 sea surface height (SSH) climatology from instantaneous snapshots of SSH. To illustrate the sensitivity of the model sea level resulting from subsurface salinity anomalies, we select two meridional sections in the Niño-4 and Niño-3 regions, centered at 165°E and 110°W, respectively, and analyze the correlation and rms differences with respect to T/P observations (Tables 3 –6). The Niño-4 section has already been analyzed in terms of sea surface salinity in the previous section. Correlations with satellite observations are higher in the analyses than in the control experiment in both Niño-3 and Niño-4, although it is worth noticing that the control experiment already produces fairly high correlations, larger than 0.9 almost everywhere. Coherency is mostly affected by the assimilation of temperature, while it is fairly unsensitive to salinity assimilation. Rms deviations from altimetry data are reduced below 3.5 cm in Niño-3 and 4.2 cm in Niño-4 in the analyses, yielding a 1–2-cm improvement with respect to the control experiment. Although most of the improvements on sea level derive from the assimilation of temperature, a hint of a positive impact on salinity is also evident. In particular, the enhanced representation of the background error implemented in OITSHR improves the fit to the observed sea level variability at the northern (and marginally at the southern) boundary of the Niño-4 region, while there is no detectable influence in Niño-3. This result suggests that the introduction of seasonality in the background error, as done in OITSHR, does impact on the quality of the analysis in regions where the seasonal cycle is stronger (i.e., around 8°S and 8°N). The fact that this effect is detected in Niño-4, but not in Niño-3, may depend on the well-documented larger influence exerted by salinity on the vertical stratification in the western tropical Pacific (Maes 1998).
8. Summary and discussion
Results from a set of global ocean reanalyses produced with a reduced-order OI scheme combined with a primitive equation OGCM for the 1962–2001 period, have been described. Ocean temperature and salinity are constrained using observations for the analyzed period, while ERA-40 atmospheric fluxes and a strong relaxation to observed SSTs are used to force the OGCM. Temperature assimilation produces a fairly large impact on the subsurface thermal structure, with the major corrections occurring in the thermocline. A typical deficiency of the state-of-the-art OGCMs is associated with a misrepresentation of mixing physics along the vertical, which in turn leads to a weakly stratified ocean, characterized by a spuriously spread thermocline. The model adopted in the current set of experiments displays a similar systematic error, which appears to be largely reduced after assimilation of temperature is switched on, particularly in the Northern Hemisphere, where most of the observations are found. Water mass formation processes also benefit from the assimilation of temperature data. A typical example is provided by the subtropical mode water, filling the intermediate depths of the subtropical North Atlantic gyre. This water mass is absent from the control simulation, due to the poor resolution of high-frequency heat loss events in the forcing fields, possibly combined with the lack of an efficient convection scheme in the ocean model. This significant feature of the midlatitude North Atlantic hydrography is fully recovered in the reanalyses.
Ocean surface circulation, although not directly constrained, is also affected by an improved representation of the subsurface thermal structure at extraequatorial latitudes, where enhanced density gradients determine (through geostrophy) stronger horizontal currents, as for the tropical Pacific NECC. At the equator, where the wind stress is a leading term of the dynamic balance, correcting the baroclinic pressure term is not sufficient to obtain a realistic current field. Biases in the wind forcing, possibly combined with errors in the model parameterization of vertical momentum fluxes concur to generate a subsurface velocity pattern that is not substantially ameliorated with respect to the control.
One of the most challenging goals to achieve in the setup of an ODA system for the global ocean is to design a system that is able to supply a reasonable analysis for salinity. In the present work a hierarchy of analyses characterized by an increasing level of sophistication in the salinity correction method has been described, and compared to a simple simulation, given by the control experiment. At the very basic level (OIT) salinity is simply updated through the T–S statistical relationships embedded in multivariate empirical orthogonal functions diagnosed from a model output over large areas of the World Ocean, following a domain splitting approach. At an intermediate level of sophistication, the direct assimilation of observed salinity profiles is introduced, while still retaining the use of EOFs (OITS). Finally, an experiment identical to the latter, but with salinity corrections based on a set of EOFs spatially (i.e., gridpointwise) and seasonally varying, has been performed (OITSHR). The large surface discrepancies between the control and observed climatology (exceeding 0.5 psu in absolute value) are noticeably reduced only after the direct assimilation of salinity is applied. The influence of increased space–time resolution of the background model error, implemented in the OITSHR experiment, becomes more evident once the vertical structure of innovations in the upper 300 m is inspected. Meridional sections across the upper Atlantic and Pacific basins reveal a progressive reduction of the vertical extension featured by climatological model-observed salinity anomalies, when the complexity of the adopted salinity correction strategy increases. This is particularly evident in the Pacific, and in the extratropical North Atlantic. The reduced penetration of the model salinity bias displayed by OITSHR could be explained by analyzing the vertical structure of the EOFs. The procedure of averaging over large areas of the global ocean adopted for the computation of regional EOFs (used for OIT and OITS) determines vertical covarying T–S patterns with a considerably shorter depth scale, compared to gridpoint EOFs (Fig. 1). Therefore, the latter provide a deeper vertical projection of corrections, hence determining a more effective reduction of the misfit across the water column. The evaluation of the upper-ocean salt content variability over interannual and decadal time scales against an objective analysis constructed with the same dataset used in the assimilation experiments reveals two important points. First, only forcing an OGCM with surface freshwater fluxes determines a low-frequency variability signal that is much weaker and mostly inconsistent with respect to the one displayed by the OA. Second, the coherency between analyses and observations undergoes a substantial improvement when high-resolution EOFs are used to model the background error. OITSHR displays a much more stable response to the assimilation process, and a greater skill in sampling the observed interannual changes. This behavior is particularly evident during the 1990s. During this decade OIT and OITS display a global decline of the salt content, followed by a slow recovery. Overall, OIT and OITS show a longer persistence of salinity anomalies, compared to both observations and OITSHR results, suggesting that the introduction of time dependency in the background error covariances induces a faster adjustment of the ocean to the time-varying observed state.
To summarize, our results show that a simple simulation of salinity performed with a state-of-the-art ocean model, is still highly unreliable, mostly due to the large uncertainties contained in the freshwater fluxes typically used to constrain the OGCM surface salinity. Moreover, the downward penetration of surface boundary fluxes to the layers underneath is still subject to deficiencies in the model vertical eddy-mixing scheme, which further compromises the three-dimensional distribution of this ocean property. We demonstrate that a combination of direct assimilation of salinity observations and a time–space-varying (at sufficiently high resolution) background error parameterization based on vertical bivariate T–S EOFs is able to reproduce reasonably well the observed salinity. The domain-splitting approach used to compute vertical EOFs, when applied to coarse partitions of the global ocean may result in spurious increments over large areas, which ultimately deteriorate the quality of the analysis.
One of the strongest assumptions underlying the current formulation of the background error covariance lies in the use of the temporal variability from a multidecadal model simulation without data assimilation to approximate the actual background error. A major limitation of this particular formulation is that it ignores the progressive error reduction induced by the assimilation. The implementation of alternative approximations of the background error covariance matrix will be addressed in future work. For example, the recalculation of the EOFs as assimilation goes on would represent a way to take into account the possible impact of assimilated observations on error reduction.
It is interesting to discuss ways the Argo program may be exploited for improving the present ODA system. The impressive amount of simultaneous temperature and salinity observations provided by the Argo profiling floats as of the year 2000, does not only represent a precious resource for better constraining an ocean model, but it also paves the way to the possibility of using direct observations for computing vertical EOFs to be adopted in the background error parameterization. However, some important caveats need to be carefully evaluated, prior to any practical implementation. If on the one hand Lagrangian data offer a good spatial coverage for vast regions of the World Ocean, on the other hand this type of observation is severely affected by temporal sparseness. Unlike Eulerian data (such as thermal and salinity profiles from moored buoys), drifters do not provide long temporal series for a given location. To build a reasonable T–S statistical relationship, profiles will need to be averaged over large areas of the ocean domain. Moreover, regions that are less densely sampled may still require some level of extrapolation. Perhaps, a promising way forward may be a hybrid approach, combining data with model-derived EOFs.
Finally, even though future satellite missions may provide a synoptic coverage for sea surface salinity, and Argo floats will supply additional information on subsurface salinity, it is likely that in the coming years we will still need to optimize the use of thermal data and the available salinity observations in order to construct large-scale estimates of this important constituent of the ocean.
Acknowledgments
We are grateful to Bruce Ingleby for providing the objective analysis, the TAO Project Office for the TAO data, and James Carton and Gennady Chepurin for supplying salinity data from Bermuda station. We wish to thank Nadia Pinardi and Srdjan Dobricic for their insightful comments. This work has been funded by the ENACT Project (Contract EVK2-CT2001-00117) for A. Bellucci and P. Di Pietro, and partially by the ENSEMBLES Project (Contract GOCE-CT-2003-505539) for A. Bellucci.
REFERENCES
Alves, O., M. Alonso-Balmaseda, D. L. T. Anderson, and T. Stockdale, 2004: Sensitivity of dynamical seasonal forecasts to ocean initial conditions. Quart. J. Roy. Meteor. Soc., 130 , 647–668.
Bell, M. J., M. J. Martin, and N. K. Nichols, 2004: Assimilation of data into an ocean model with systematic errors near the equator. Quart. J. Roy. Meteor. Soc., 130 , 873–893.
Blanke, B., and P. Delecluse, 1993: Variability of the tropical Atlantic Ocean simulated by a general circulation model with two different mixed layer physics. J. Phys. Oceanogr., 23 , 1363–1388.
Burgers, G., M. A. Balmaseda, F. C. Vossepoel, G. J. van Oldenborgh, and P. J. Van Leeuwen, 2002: Balanced ocean data assimilation near the equator. J. Phys. Oceanogr., 32 , 2509–2519.
Carton, J. A., G. Chepurin, X. Cao, and B. Giese, 2000: A simple ocean data assimilation analysis of the upper ocean, 1950–95. Part I: Methodology. J. Phys. Oceanogr., 30 , 294–309.
Daley, R., 1991: Atmospheric Data Analysis. Cambridge University Press, 457 pp.
Davey, M., and ENACT Partnership, 2006: Multimodel multi-method multi-decadal ocean analyses from the ENACT project. Exchanges, 38 , 22–25.
De Mey, P., and A. Robinson, 1987: Assimilation of altimeter eddy fields in a limited-area quasi-geostrophic model. J. Phys. Oceanogr., 17 , 2280–2293.
De Mey, P., and M. Benkiran, 2002: A multivariate reduced-order optimal interpolation method and its application to the Mediterranean basin-scale circulation. Ocean Forecasting: Conceptual Basis and Applications, N. Pinardi and J. D. Woods, Eds., Springer Verlag, 281–306.
Derber, J., and A. Rosati, 1989: A global oceanic data assimilation system. J. Phys. Oceanogr., 19 , 1333–1347.
Donguy, J. R., 1994: Surface and subsurface salinity in the tropical Pacific Ocean. Prog. Oceanogr., 34 , 45–78.
Ducet, N., P. Y. Le Traon, and G. Reverdin, 2000: Global high resolution mapping of ocean circulation from the combination of TOPEX/POSEIDON and ERS-1/2. J. Geophys. Res., 105 , C8. 19477–19498.
Gent, P., and J. McWilliams, 1990: Isopycnal mixing in ocean circulation models. J. Phys. Oceanogr., 20 , 150–155.
Ide, K., P. Courtier, M. Ghil, and A. C. Lorenc, 1997: Unified notation for data assimilation: Operational, sequential and variational. J. Meteor. Soc. Japan, 75 , 181–189.
Ingleby, B., and M. Huddleston, 2007: Quality control of ocean profiles: Historical and real-time data. J. Mar. Syst., 65 , 158–175.
Ji, M., and A. Leetma, 1997: Impact of data assimilation on ocean initialization and El Niño prediction. Mon. Wea. Rev., 125 , 742–753.
Klemm, D. D., M. J. Bell, R. M. Forbes, and A. Hines, 2000: Assessment of the FOAM global data assimilation system for real-time operational ocean forecasting. J. Mar. Syst., 25 , 1–22.
Lazar, A., G. Madec, and P. Delecluse, 1999: The deep interior downwelling, the Veronis effect, and mesoscale tracer transport parameterizations in an OGCM. J. Phys. Oceanogr., 29 , 2945–2961.
Levitus, S., and Coauthors, 1998: World Ocean Database 1998. NOAA Atlas NESDIS 1, 346 pp.
Lorenc, A. C., and Coauthors, 2000: The Met. Office global three-dimensional variational data assimilation scheme. Quart. J. Roy. Meteor. Soc., 126 , 2991–3012.
Madec, G., P. Delecleuse, M. Imbard, and C. Levy, 1998: OPA, release 8.1: Ocean general circulation model reference manual. LODYC/IPSL Tech. Note 11, Paris, France, 91 pp.
Maes, C., 1998: Estimating the influence of salinity on sea level anomaly in the ocean. Geophys. Res. Lett., 25 , 3551–3554.
Maes, C., D. Behringer, R. W. Reynolds, and M. Ji, 2000: Retrospective analysis of the salinity variability in the western tropical Pacific Ocean using an indirect minimization approach. J. Atmos. Oceanic Technol., 17 , 512–524.
Masina, S., N. Pinardi, and A. Navarra, 2001: A global ocean temperature and altimeter data assimilation system for studies of climate variability. Climate Dyn., 17 , 687–700.
Masina, S., P. Di Pietro, and A. Navarra, 2004: Interannual-to-decadal variability of the North Atlantic from an ocean data assimilation system. Climate Dyn., 23 , 531–546.
McPhaden, M. J., and Coauthors, 1998: The Tropical Ocean-Global Atmosphere observing system: A decade of progress. J. Geophys. Res., 103 , 14169–14240.
Parrish, D. F., and J. C. Derber, 1992: The National Meteorological Center’s spectral statistical analysis system. Mon. Wea. Rev., 120 , 1747–1763.
Rabier, F., H. Jarvinen, E. Klinker, J. F. Mahfouf, and A. Simmons, 2000: The ECMWF operational implementation of four-dimensional variational assimilation. Part I: Experimental results with simplified physics. Quart. J. Roy. Meteor. Soc., 126 , 1143–1170.
Reverdin, G., E. Frankignoul, E. Kestenare, and M. J. McPhaden, 1994: Seasonal variability in the surface currents of the equatorial Pacific. J. Geophys. Res., 99 , 20323–20344.
Reynolds, R. W., 1988: A real-time global surface temperature analysis. J. Climate, 1 , 75–86.
Ricci, S., A. Weaver, J. Vialard, and P. Rogel, 2005: Incorporating state-dependent temperature–salinity constraints in the background error covariance of variational ocean data assimilation. Mon. Wea. Rev., 133 , 317–338.
Rio, M. H., and F. Hernandez, 2004: A mean dynamic topography computed over the world ocean from altimetry, in situ measurements, and a geoid model. J. Geophys. Res., 109 .C12032, doi:10.1029/2003JC002226.
Rosati, A., K. Miyakoda, and R. Gudgel, 1997: The impact of ocean initial conditions on ENSO forecasting with a coupled model. Mon. Wea. Rev., 125 , 754–772.
Segschneider, J., D. L. T. Anderson, J. Vialard, M. Balmaseda, T. N. Stockdale, A. Troccoli, and K. Haines, 2001: Initialization of seasonal forecasts assimilating sea level and temperature observations. J. Climate, 14 , 4292–4307.
Sparnocchia, S., N. Pinardi, and E. Demirov, 2003: Multivariate empirical orthogonal function analysis of the upper thermocline structure of the Mediterranean Sea from observations and model simulations. Ann. Geophys., 21 , 167–187.
Talley, L. D., and M. E. Raymer, 1982: Eighteen degree water variability. J. Mar. Res., 40 , 757–775.
Troccoli, A., and K. Haines, 1999: Use of temperature–salinity relation in a data assimilation context. J. Atmos. Oceanic Technol., 16 , 2011–2025.
Troccoli, A., and P. Källberg, 2004: Precipitation correction in the ERA-40 reanalysis. ERA-40 Project Rep. Series 13, 6 pp.
Troccoli, A., and Coauthors, 2002: Salinity adjustments in the presence of temperature data assimilation. Mon. Wea. Rev., 130 , 89–102.
Uppala, S., and Coauthors, 2005: The ERA-40 reanalysis. Quart. J. Roy. Meteor. Soc., 131 , 2961–3012.
Vialard, J., A. T. Weaver, D. L. T. Anderson, and P. Delecleuse, 2003: Three- and four-dimensional variational assimilation with a general circulation model of the tropical Pacific Ocean. Part II: Physical validation. Mon. Wea. Rev., 131 , 1379–1395.
Weaver, A. T., J. Vialard, and D. L. T. Anderson, 2003: Three- and four-dimensional variational assimilation with a general circulation model of the tropical Pacific Ocean. Part I: Formulation, internal diagnostics, and consistency checks. Mon. Wea. Rev., 131 , 1360–1378.
First vertical EOFs of salinity in the northeastern extratropical Pacific. In gray, EOFs at each grid point used for experiment OITSHR are shown, for (left) winter and (right) summer. In thick black, the first EOF used in experiments OIT and OITS for the same area are shown (identical for all seasons).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
First vertical EOFs of salinity in the northeastern extratropical Pacific. In gray, EOFs at each grid point used for experiment OITSHR are shown, for (left) winter and (right) summer. In thick black, the first EOF used in experiments OIT and OITS for the same area are shown (identical for all seasons).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
First vertical EOFs of salinity in the northeastern extratropical Pacific. In gray, EOFs at each grid point used for experiment OITSHR are shown, for (left) winter and (right) summer. In thick black, the first EOF used in experiments OIT and OITS for the same area are shown (identical for all seasons).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of the globally averaged rms of temperature innovations (observation − background) near the surface (thick) and at 160 m (thin) for experiment OIT. Each value is time mean over an assimilation cycle.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of the globally averaged rms of temperature innovations (observation − background) near the surface (thick) and at 160 m (thin) for experiment OIT. Each value is time mean over an assimilation cycle.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of the globally averaged rms of temperature innovations (observation − background) near the surface (thick) and at 160 m (thin) for experiment OIT. Each value is time mean over an assimilation cycle.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Differences between (top) OIT and Levitus, and (bottom) CTRL and Levitus climatological temperature at 100 m. Only values greater (smaller) than 1°C (−1°C) are plotted. Positive values are shaded gray. Contour interval is 1°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Differences between (top) OIT and Levitus, and (bottom) CTRL and Levitus climatological temperature at 100 m. Only values greater (smaller) than 1°C (−1°C) are plotted. Positive values are shaded gray. Contour interval is 1°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Differences between (top) OIT and Levitus, and (bottom) CTRL and Levitus climatological temperature at 100 m. Only values greater (smaller) than 1°C (−1°C) are plotted. Positive values are shaded gray. Contour interval is 1°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of OIT (solid) and control (dashed) minus TAO observations rms for temperature in the Niño-3 (gray) and Niño-4 (black) areas. The rms profiles have been obtained by collecting TAO data available in 5°S–5°N, 160°–210°E (Niño-4) and 5°S–5°N, 210°–270°E (Niño-3).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of OIT (solid) and control (dashed) minus TAO observations rms for temperature in the Niño-3 (gray) and Niño-4 (black) areas. The rms profiles have been obtained by collecting TAO data available in 5°S–5°N, 160°–210°E (Niño-4) and 5°S–5°N, 210°–270°E (Niño-3).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of OIT (solid) and control (dashed) minus TAO observations rms for temperature in the Niño-3 (gray) and Niño-4 (black) areas. The rms profiles have been obtained by collecting TAO data available in 5°S–5°N, 160°–210°E (Niño-4) and 5°S–5°N, 210°–270°E (Niño-3).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time–depth sections of temperature differences (°C) at 8°N, 38°W (tropical Atlantic) between (top) control and observations, and (bottom) OIT and observations. Positive (negative) values are indicated with solid (dashed) contours. The zero line (solid thick) is also shown. The contour interval is 0.5°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time–depth sections of temperature differences (°C) at 8°N, 38°W (tropical Atlantic) between (top) control and observations, and (bottom) OIT and observations. Positive (negative) values are indicated with solid (dashed) contours. The zero line (solid thick) is also shown. The contour interval is 0.5°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time–depth sections of temperature differences (°C) at 8°N, 38°W (tropical Atlantic) between (top) control and observations, and (bottom) OIT and observations. Positive (negative) values are indicated with solid (dashed) contours. The zero line (solid thick) is also shown. The contour interval is 0.5°C.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Zonal–depth section for 1993 mean temperature (°C) at 30°N in the North Atlantic. (top) UOT observations, (middle) OIT, and (bottom) control. Dark shading indicates the 18°–19°C layer.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Zonal–depth section for 1993 mean temperature (°C) at 30°N in the North Atlantic. (top) UOT observations, (middle) OIT, and (bottom) control. Dark shading indicates the 18°–19°C layer.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Zonal–depth section for 1993 mean temperature (°C) at 30°N in the North Atlantic. (top) UOT observations, (middle) OIT, and (bottom) control. Dark shading indicates the 18°–19°C layer.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface zonal current climatologies (m s−1) for the 1987–1992 period. (top) Control, (middle) OIT, and (bottom) Reverdin et al. (1994) climatology. Gray shading indicates positive (eastward) values.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface zonal current climatologies (m s−1) for the 1987–1992 period. (top) Control, (middle) OIT, and (bottom) Reverdin et al. (1994) climatology. Gray shading indicates positive (eastward) values.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface zonal current climatologies (m s−1) for the 1987–1992 period. (top) Control, (middle) OIT, and (bottom) Reverdin et al. (1994) climatology. Gray shading indicates positive (eastward) values.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of zonal currents (m s−1) in the equatorial Pacific time averaged over the 1992–2001 period at (left) 140° and (right) 110°W from TAO (thick solid curve), OIT (thin solid curve), and the control (dashed curve).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of zonal currents (m s−1) in the equatorial Pacific time averaged over the 1992–2001 period at (left) 140° and (right) 110°W from TAO (thick solid curve), OIT (thin solid curve), and the control (dashed curve).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Vertical profiles of zonal currents (m s−1) in the equatorial Pacific time averaged over the 1992–2001 period at (left) 140° and (right) 110°W from TAO (thick solid curve), OIT (thin solid curve), and the control (dashed curve).
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface salinity climatology patterns (psu) for (top) CTRL-Levitus and (bottom) OIT-Levitus. Only values greater (smaller) than 0.25 (−0.25) psu are plotted. Negative values are shaded gray. Contour interval is 0.25 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface salinity climatology patterns (psu) for (top) CTRL-Levitus and (bottom) OIT-Levitus. Only values greater (smaller) than 0.25 (−0.25) psu are plotted. Negative values are shaded gray. Contour interval is 0.25 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Surface salinity climatology patterns (psu) for (top) CTRL-Levitus and (bottom) OIT-Levitus. Only values greater (smaller) than 0.25 (−0.25) psu are plotted. Negative values are shaded gray. Contour interval is 0.25 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 9 but for (top) OITS-Levitus and (bottom) OITSHR-Levitus.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 9 but for (top) OITS-Levitus and (bottom) OITSHR-Levitus.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 9 but for (top) OITS-Levitus and (bottom) OITSHR-Levitus.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Meridional section at 30°W of climatological salinity differences (psu) between (top to bottom) CTRL, OIT, OITS, and OITSHR and Levitus. Values in the [−0.1, 0.1] interval are shaded. Contour interval is 0.1 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Meridional section at 30°W of climatological salinity differences (psu) between (top to bottom) CTRL, OIT, OITS, and OITSHR and Levitus. Values in the [−0.1, 0.1] interval are shaded. Contour interval is 0.1 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Meridional section at 30°W of climatological salinity differences (psu) between (top to bottom) CTRL, OIT, OITS, and OITSHR and Levitus. Values in the [−0.1, 0.1] interval are shaded. Contour interval is 0.1 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 11, but for meridional section at 178°E.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 11, but for meridional section at 178°E.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
As in Fig. 11, but for meridional section at 178°E.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly globally averaged over the top 300 m.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly globally averaged over the top 300 m.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly globally averaged over the top 300 m.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly averaged over the top 300 m, in the northeastern extratropical Atlantic.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly averaged over the top 300 m, in the northeastern extratropical Atlantic.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series for salinity anomaly averaged over the top 300 m, in the northeastern extratropical Atlantic.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
(left) Mean salinity profile (psu) at Bermuda (32°N, 64°W) from OITSHR, OITS, OIT, CTRL, and BS observations for the 1962–2001 period. (right) For OITS and OITSHR, the mean profile computed over the 1962–1995 period is also shown.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
(left) Mean salinity profile (psu) at Bermuda (32°N, 64°W) from OITSHR, OITS, OIT, CTRL, and BS observations for the 1962–2001 period. (right) For OITS and OITSHR, the mean profile computed over the 1962–1995 period is also shown.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
(left) Mean salinity profile (psu) at Bermuda (32°N, 64°W) from OITSHR, OITS, OIT, CTRL, and BS observations for the 1962–2001 period. (right) For OITS and OITSHR, the mean profile computed over the 1962–1995 period is also shown.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of SSS anomaly along 165°E. (top to bottom) TAO observations, control, OIT, OITS, and OITSHR. Dashed contours are used for negative values. Contour interval is 0.2 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of SSS anomaly along 165°E. (top to bottom) TAO observations, control, OIT, OITS, and OITSHR. Dashed contours are used for negative values. Contour interval is 0.2 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
Time series of SSS anomaly along 165°E. (top to bottom) TAO observations, control, OIT, OITS, and OITSHR. Dashed contours are used for negative values. Contour interval is 0.2 psu.
Citation: Monthly Weather Review 135, 11; 10.1175/2007MWR1821.1
List of experiments.
Correlation coefficients between model and observed SSS at 165°E. Values that are not statistically significant (n.s.) at the 95% level are indicated.
SLA correlation coefficients between model and TOPEX/Poseidon data at 165°E, for the 1993–99 period.
SLA rms differences (cm) between model and TOPEX/Poseidon data at 165°E, for the 1993–99 period.
SLA correlation coefficients between model and TOPEX/Poseidon data at 110°W, for the 1993–99 period.
SLA rms differences (cm) between model and TOPEX/Poseidon data at 110°W, for the 1993–99 period.
