1. Introduction
The study of the global ocean interannual and lower-frequency variability requires the accurate knowledge of the surface and subsurface ocean state. This is usually obtained by optimally combining in situ and satellite observations with models. The latter are useful especially in those areas where the network of conventional observations is poor (e.g., in the extratropical Southern Hemisphere).
In the last three decades, many altimetric satellite missions have been launched with the aim of monitoring variations of sea level height, at both regional and global scales. Sea level heights measured by space-borne altimeters refer to a reference ellipsoid and not to the spatially varying elevation of the earth geoid. Therefore, the measurements are provided in terms of anomalies from the time-averaged sea level height referenced to the earth geoid, which is usually called the mean dynamic topography (MDT). The information brought by the sea level anomaly data (SLA), in terms of temporal and spatial variabilities of sea level height, has been successfully exploited within data assimilation systems to adjust surface and subsurface temperature and salinity fields. For instance, Zheng et al. (2007) demonstrated that assimilating SLA observations improves the predictions of the sea surface temperature (SST) anomaly in the equatorial Pacific. In other works (Fischer et al. 1997; Ji et al. 2000), the prediction of ENSO in seasonal forecasts is improved by using the sea level information only to correct the first corresponding subsurface temperature levels. However, evidence of local importance of the halosteric contribution to the sea surface height trend (Ivchenko et al. 2008) suggests that sea level anomaly data can be fruitfully assimilated to also correct the subsurface salinity fields. This was also confirmed by the positive impacts of the bivariate assimilation of altimetric data found by Vossepoel and Behringer (2000), although they used synthetic observations.
A global mean dynamic topography, added to the sea level anomaly observations, is required in order to compare the observations with the model equivalents and to assimilate the data. This is a nontrivial issue if one considers that even though the very precise recent gravimetric data available from the Challenging Minisatellite Payload (CHAMP), the Gravity Recovery And Climate Experiment (GRACE) or even from the higher-resolution recently launched Gravity Field and Steady-State Ocean Circulation Explorer (GOCE) can provide a precise estimate of the geoid, ocean models usually represents the geoid as a sphere, neglecting its spatial variability. Moreover, even with a precise representation of the geoid, they would need very accurate sea ice, river runoff, and steric variability parameterizations to correctly estimate the absolute value of the sea surface height and correctly follow its seasonal and interannual variations.
As explained in details by Vidard et al. (2009), an observed MDT, such as the one from gravimetric missions, might be improper as it bears a geodetic bias that varies locally and cannot simplify to a global offset. This is even more dramatic since the impacts of the choice of the MDT can be nonnegligible. Segschneider et al. (2000), for instance, showed the sensitivity of the ocean forecasts to the MDT, which can induce variations in the Niño-3-averaged 100-m temperature up to 4°C. As a matter of course, the importance of a correctly estimated MDT and its consistency with the model sea surface height are key issues for the successful assimilation of altimetric data.
At the Euro-Mediterranean Centre for Climate Change (CMCC), a three-dimensional variational data assimilation (3DVAR) system called OceanVar has been developed by Dobricic and Pinardi (2008) within the framework of the Mediterranean Forecasting System (MFS; Pinardi et al. 2003), producing operational daily analyses and forecasts (Dobricic et al. 2008). CMCC also has a global ocean analysis system used for reanalysis and climate research applications; see, for example, Masina et al. (2010) and Lee et al. (2010). This assimilation system (Bellucci et al. 2007) includes the global implementation of a reduced-order multivariate optimal interpolation (OI) scheme (De Mey and Benkiran 2002), and all the available in situ observations were assimilated. Although the resolution of the analysis system is rather coarse (i.e., 2° with increases up to 0.5° at the equator), such a configuration is still very attractive and necessary for many applications like long-term climate studies (Masina et al. 2010) and seasonal ensemble prediction systems (Alessandri et al. 2010), where the use of higher resolutions is still too demanding in terms of computer resources. To produce global analyses with the possibility of assimilating space-borne altimetric data and, in the future, sea surface temperature and salinity measurements from satellites, we have adapted the limited-area OceanVar approach to the global ocean and implemented a scheme for locally adjusting the water-column-integrated density through the sea level anomaly observations.
As the CPU time and memory demands of an OI scheme primarily rely on the number of observations, they have the limitations of hardly being suitable for the assimilation of many space-borne instruments, which enormously increase the length of the observation vector. On the contrary, 3DVAR computing resources depend primarily on the dimensions of the computational grid. However, an increase in model resolution will lead in turn to an increase in the observation amount if the observation selection strategy (e.g., subsampling, thinning, etc.) depends on the grid resolution. Moreover, 3DVAR provides global solutions of the analysis problem without leading to large- or small-scale inconsistencies as in OI (Lorenc 1981), and allows for the implementation of complex observation operators in a more efficient way (see, e.g., Kalnay 2002, section 5.5.3).
This paper is concerned with the description and the validation of the SLA data assimilation. The structure is as follows. Section 2 briefly describes the 3DVAR/first guess at the appropriate time (FGAT) formulation and the configuration of the analysis system; in section 3 the implementation of a scheme for assimilating SLA observations is presented, together with the procedure and the results of the computation of the global MDT; section 4 presents the results obtained with the assimilation of altimetric data, while section 5 summarizes the main achievements of this paper and proposes a discussion on our results.
2. The assimilation system
a. 3DVAR/FGAT formulation
The formulation of the background term of the cost function follows Dobricic and Pinardi (2008): the background-error covariance matrix is decomposed in a sequence of linear operators that account separately for horizontal correlations and vertical covariances, which are assumed to be independent. Horizontal correlations are modeled by means of an application of four iterations of a first-order recursive filter (Hayden and Purser 1995) with horizontally homogeneous but vertically varying correlation length scale.
The vertical component of the background-error covariance matrix is eigendecomposed to obtain vertically uncorrelated eigenvectors. We use empirical orthogonal functions (EOFs), which are bivariate in temperature and salinity, at full model resolution in both the horizontal and the vertical. To reduce computational cost and avoid noisy vertical correlations, the vertical modes are truncated to 10 from the original 62 modes. The EOFs were calculated from the model seasonal climatology (Bellucci et al. 2007; Di Pietro and Masina 2009).
To impose cyclic conditions on the global ocean domain, the minimization is performed over an extended domain, with duplicated points before the western and after the eastern boundaries. Observations in the extension zones are duplicated from the corresponding area of the original domain. The extension zone is 20 grid points long on both boundaries (about 5000 km), which makes sure that analyzed fields at the inner boundaries of the extension zones are exactly the same at the western and eastern boundaries after the minimization.
The data assimilation step is performed every 10 days, using all the observations included in the temporal range of ±5 days before and after the assimilation time. The ocean model is then used to project the analyzed fields forward to the next assimilation step.
b. In situ observations
The set of in situ observations consists of vertical profiles of temperature from the expandable bathythermographs (XBTs), buoys, sea stations [temperature, salinity, and current (TESAC)], Argo floats (from late 1990s onward only), and salinity profiles from buoys, sea stations, and Argo floats. Data have been provided by the Met Office’s Hadley Centre within the framework of the European Union (EU) funded project ENSEMBLES (EN3). EN3 is a collection of in situ data that have been quality checked according to the method of Ingleby and Huddleston (2007). For XBT data, a time-dependent fall-rate correction was used (Wijffels et al. 2008).
An additional background quality check, which rejects observations with a square departure from the model equivalents larger than 3 times the sum of the background and observational error variances, is also performed to remove observations that are grossly inconsistent with the model. Finally, a thinning procedure retains only the closest observation to the analysis time in case multiple reports from the same platform or buoy are found in close proximity.
c. The ocean model
The oceanic general circulation model (OGCM) used within the analysis system is version 8.2 of the free-surface Ocean Parallèle (OPA) model (Madec et al. 1998). The spatial resolution of OPA is 2° × 2° cos(latitude) × 31 vertical levels, though in the tropical band the meridional resolution increases up to 0.5°.
Surface boundary conditions are provided by the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis dataset (ERA-40; Uppala et al. 2005) for wind stress, surface heat, and freshwater fluxes from 1990 until 2001. Starting from the beginning of 2002, the forcing fields are taken from the ECMWF operational analyses; this leads to noncontinuous forcing, but we kept this setup for comparison with the previous OI analysis system (Masina et al. 2010). Relaxations to the Reynolds SST (Reynolds and Smith 1994), and to the ECMWF operational SST from 2001 onward for consistency with the wind stress and the heat fluxes, were used with a restoring time scale of about 12 days. Furthermore, our model configuration does not allow any change in the global ocean volume, which is conserved throughout the simulations by superimposing on a daily basis the zeroed variation in the globally averaged water flux. This constraint has been used to avoid unrealistically large positive trends in the sea level that were observed without the superimposed closure of the freshwater budget. To better appreciate the impacts of sea level anomaly observations, we do not use any relaxation to climatological fields, as they may too strongly constrain the model temperature and salinity forecasts. The first 3DVAR run was initialized by using the OI analysis fields from Di Pietro and Masina (2009).
3. Assimilation of SLA observations
a. Sea level anomaly dataset
SLA observations used in this work covered the period between October 1992 and January 2006, during which time many altimetric missions have been providing sea level anomaly data. Table 1 summarizes the temporal coverage of the altimetric satellites, which have all been assimilated into the 3DVAR/FGAT assimilation system.
Along-track data were provided by Aviso, after the usual geophysical removals (tropospheric, ionospheric, electromagnetic, tidal, and inverse barometer effects) and after a multisatellite cross calibration for eliminating residual orbit errors (Le Traon et al. 1995) and large-scale biases (Le Traon and Ogor 1998; Le Traon et al. 1998). The along-track dataset is provided in delayed time, yielding a better orbital calculation, a centered time window for the multisatellite cross calibration, and was proved to have a significantly better level of accuracy (Pascual et al. 2009).
To filter out small spatial scales that are not resolved by the model, we apply a one-dimensional along-track low-pass Lanczos filter (Duchon 1979) with a latitudinally varying cutoff wavelength consistent with the model resolution, which is about 75 km from the equator and decreases poleward. The globally averaged reduction in the root-mean-square error of SLA observations minus first guesses due to the implementation of the Lanczos filter (not shown) is constantly of about 2 mm for the 1993–2005 period. Like the in situ observations, sea level anomaly observations are quality checked against model-equivalent data. Finally, a satellite-dependent thinning procedure retains only one active observation over a box 150% of the size of the model resolution. This ensures the consistency between the resolution of the assimilated SLA dataset and that of the ocean model, and decreases both the horizontal correlation of the SLA observational errors and the along-track correlation introduced by the Lanczos filter. Briefly, data thinning prevents the neglected observational error correlation from causing inaccuracies in the analyses. In the future, we plan to remove the assumption of uncorrelated SLA observations, which in turn will allow us to assimilate as many observations as possible.
b. The observation operator
In practice, our scheme splits the observation departure into its thermo- and halosteric contributions over the water column by using the adjoint version of the density operator linearized around the background fields of temperature and salinity. How the sea level increment is vertically spread on the temperature and salinity increments depends upon the bivariate definition of the vertical background-error covariances, which are spatially nonuniform. Note that this approach differs from that previously adopted within OceanVar (Dobricic and Pinardi 2008), where a barotropic model operator was used to simultaneously constrain the currents, temperature, salinity, and sea level height fields. The structure of the analysis increments for the same sea level anomaly observation minus the first-guess departure may vary significantly with the observation location. For instance, the vertical background-error signal of the temperature in the tropical Pacific reaches a maximum at around 150-m depth (not shown). The sea level anomaly observation information is spread accordingly and its temperature analysis increments have a maximum at such a depth. In extratropical regions, on the contrary, vertical background-error variances and sea level analysis increments are contained in the first 100–150-m depth, with a maximum typically located in the first 50 m.
Further, the seasonality of the background-error structures is clearly felt by the SLA observation assimilation: North Atlantic summertime assimilation of SLA causes the largest analysis increments limited in the first 50 m because of the ocean stratification, while during the winter the analysis increments are much more smoothed along the vertical.
c. Global mean dynamic topography
It is important to stress that our method leads to an MDT (hereafter MDTOI) that is consistent with both the OGCM dynamics and the in situ and altimetric observations. Therefore, it differs from gravimetry-based MDTs, which usually do not explicitly account for the ocean model dynamics and use observations mostly to increase the coarse resolution of remotely sensed gravimetry data. Furthermore, OGCM dynamics may contain some systematic errors, due, for instance, to the inaccuracies in model resolution and numerical schemes, which increase the inconsistency between the model and the true ocean dynamics and, in turn, between the model mean SSH and an observed MDT. MDTOI differs also from model only based MDTs, which usually do not account for altimetry data.
MDTOI (Fig. 1a) has features that are very similar to other mean dynamic topographies in terms of large-scale circulation and strongest gradients in correspondence with main the large-scale current systems, western boundary currents, and the Antarctic Circumpolar Current. We also show the differences between the RIO04 MDT and that computed with our two-step procedure (Fig. 1b), and the differences between a model only based MDT (hereafter MODMDT) and the MDTOI (Fig. 1c). For the latter, the MODMDT was derived from a long-term mean surface height of a model run without any data assimilation and with the surface boundary forcing as described in section 2c.
In the comparison with the RIO04 MDT, we found that the main differences are located in the North Atlantic, in correspondence with the Antarctic Circumpolar Current and in areas of large eddy activity (e.g., Kuroshio and Agulhas Current systems). The differences cannot be simplified to a global offset, and reach values of up to 40 cm; they are particularly large and irregular with many complicated features in the mid- and high-latitude Southern Hemisphere. Note that in these areas, in addition to the presence of many complicated current systems, the lack of a dense network of in situ observations decreases the accuracy of the RIO04 MDT. The main differences between the model only based MDT and the MDTOI are in general located in the same areas, although the features are more regular and smoothed. It is important to note that in many dynamically important regions such as the Gulf Stream, the subtropical South Atlantic, the South Pacific, and parts of the Kuroshio, the differences between the MODMDT and the MDTOI have opposite signs with respect to the differences between the RIO04 MDT and the MDTOI. This means that assimilating in situ observations and adjusting the MDT with the altimetric observation biases brings our MDT closer to a gravimetry-based one, and leads in turn to an MDT that accounts for the OGCM dynamical features and mitigates its inaccuracies by means of observational data ingestion.
Surface geostrophic currents (not shown) deduced from the RIO04 MDT differ from those deduced from the MDTOI, especially in the Gulf Stream region, the Caribbean, Falklands, Agulhas and Equatorial Counter Currents and, partly, in the Kuroshio area. This can be explained by the misplacement of many current systems due to the low resolution of the ocean model.
d. Characterization of sea level anomaly observational error variance
The instrumental error has been set equal to 2 cm for the Ocean Topography Experiment (TOPEX)/Poseidon and 3 cm for the other satellites [see Le Traon and Ogor (1998) and Cheney et al. (1994) for a discussion on the accuracy of the sea level anomaly observations and their preprocessing techniques].
The representativeness error arises since the signal contained in the observations cannot by resolved by the analysis mesh for some small scales. It cannot be easily eliminated by the data assimilation system (Mitchell and Daley 1997), and its quantification is nontrivial. The representativeness clearly depends on the model resolution, as it vanishes by definition when all the scales are resolved by the model. Such an error may be nonuniform on the domain, being larger in areas of strong eddy activity (Schiller et al. 2008). To derive the representativeness error, we have followed an approach similar to that of Oke and Sakov (2008). The cell-averaged sea level anomaly along-track products may be thought of as the truth in the sense that their scales are completely resolved by the data assimilation system. The standard deviation of their differences from the observations actually assimilated after a background quality check procedure has been run and spatial thinning represents, therefore, the error due to the model-data scales inconsistency. Doing this, we also include in the representativeness error the contribution of the thinning, as well as the low-pass Lanczos filter. In Fig. 2a, we show the contour map of the representativeness error thus calculated. Clearly, it is much larger in those areas characterized by a strong level of eddy activity, especially in the Gulf Stream region, in the Kuroshio region, and in the Agulhas Current region, as it is acknowledged that the sea level height small-scale variability closely follows the eddy kinetic energy signal (Kaplan et al. 2004). Its globally averaged value is 32 mm. The geographical distribution of the error signal is very similar to that computed by Oke and Sakov (2008) for a global 1°-resolution grid.
The observation operator error and its dependence on the latitude have been calculated from the differences of the SSH predicted by the model, assumed to be the truth, and the SSH computed by the diagnostic formulation of Eq. (7). The statistics have the property of being practically unbiased (not shown), and their accuracy is less than 2 cm far from the equator and less than 6 cm in correspondence of the equator.
As our mean dynamic topography was first derived from the long-term mean of the model SSH, its error intrinsically relies on the SSH variability. Therefore, the MDT error has been computed as the model SSH anomaly standard deviation. In Fig. 2b the MDT error thus derived is contoured. The variability of the SSH is larger in areas dynamically more active. The structure is therefore not very different from that of the representativeness error, though the two sources of error are conceptually different: the latter arises from the unresolved scales in the model, while the MDT error is rooted in the uncertainty in the representation of a reference ocean topography.
4. Results
In this section we present the impacts of the SLA assimilation, focusing also on the impacts of the choice of the mean dynamic topography used within the SLA observations assimilation. The experiment that uses the MDT derived from the model and adjusted via optimal interpolation of assimilation statistics is referred to as MDTOI; the experiment based on the RIO04 mean dynamic topography is referred to as RIOMDT. The control experiment without the assimilation of SLA observations is called NOSLA. Finally, MODMDT refers to the experiment where an MDT based only on the mean SSH of a model run without assimilation is used. The error variance associated with the mean dynamic topography in Eq. (12) when the RIO04 MDT is used was provided by the RIO04 MDT objective analysis. In Table 2 the nomenclature and main characteristics of the experiments are summarized.
a. Impacts of SLA assimilation on the analyses
The global impacts of the assimilation of the SLA observations on the analyses are here shown as the bias and the standard deviation of the differences in the analysis increments, in model space, with and without assimilating altimetric data for the analyses covering the period 1993–2005. In particular, following the names introduced in Table 2, we compare the experiments MDTOI and NOSLA. In Fig. 3, the effects of assimilating SLA are plotted as a contour map of the bias within the first 100-m depth, for both temperature and salinity. The analyses result in a warmer North Atlantic, especially in correspondence to the Gulf Stream current, and in the Kuroshio region. A cooling effect is also visible in the eastern equatorial Pacific, while more complicated patterns are present in the Southern Hemisphere midlatitude band. Regarding the salinity (Fig. 3b), most of the impact of the SLA observations is found in the equatorial band, especially in the Indian Ocean, and in the Gulf Stream. Further, an increase in the analyzed freshwater is found in correspondence to the Amazon River outflow, where the freshwater budget formulation in the model may be insufficient to correct the salinity fields.
The zonally averaged standard deviations of the analysis increment differences are shown in Fig. 4 as a function of the depth and latitude for the first 500-m depth. Such statistics represent the variability of the analysis due to the SLA observations with respect to the SLA-denial experiment. We found that most of the impact on the temperature fields is located above the thermocline at midlatitudes (60°S–20°S and 20°–60°N) and below the thermocline in the tropical regions, due to the deep-peaking structure of the temperature vertical covariances in those regions. Note also that, as previously mentioned, SLA observations are not assimilated when very close to the equator (2°S–2°N) and are given a larger error (i.e., a smaller weight) in the tropical region as a consequence of the weaker validity of the level-of-no-motion assumption for regions dominated by ageostrophic circulation. This means that the assimilation of SLA in the neighboring areas induces a significant modification in the density profiles and in correspondence of the equator itself. The impacts on the salinity fields are stronger in the Northern Hemisphere up to 150-m depth. For both of the variables, the variability of the analysis due to the SLA assimilation vanishes gradually below about 300 m and becomes meaningless below 600-m depth.
To study the relative impacts of the sea level anomaly observations with respect to the other in situ observation types, we computed the degrees of freedom for signal (DFSs; Cardinali et al. 2004) for each observation type and parameter. DFSs summarize the information content brought by different observation types into the analysis, and have already been applied by Montmerle et al. (2007) and Storto and Tveter (2009) to studies examining the hierarchy of observation networks within limited-area meteorological variational data assimilation systems. The procedure to compute the DFS is based on a randomization technique first explored by Chapnik et al. (2006), to whom the reader is referred for further details. The computation was performed for the period 1993–2005, perturbing one data assimilation cycle per season. The absolute DFSs (Fig. 5a) represent the global impacts of the different observation types on the analysis, while relative DFSs (i.e., DFSs divided by the number of observations; see Fig. 5b) quantify the ideal informative content brought by a single observation, regardless of the actual amount and geographical distribution of such observations. DFSs show that, besides the significant importance of the Argo network, sea level anomaly data play a role comparable to that of the other in situ observations, in both absolute and relative terms. TOPEX/Poseidon (T/P) exhibits the larger value for the absolute DFS, due to its longer data availability, while in relative terms T/P and the European Remote Sensing Satellites-1 and -2 (ERS-1 and -2) are the most informative altimetric satellites. The computation performed for the pre-Argo era only (1993–98, not shown) highlights that the sensitivity of the analysis to the SLA observing network (absolute DFS) is second only to the XBT network.
b. Impacts of different MDTs on verification skill scores
1) Sea surface height
The impacts of the mean dynamic topography have been first evaluated by comparisons of the assimilation statistics among different experiments. In terms of innovations bias and RMSE relative comparison, Fig. 6 depicts the global background bias and RMSE time series for the MDTOI, RIOMDT, and MODMDT experiments. Clearly, the use of the MDTOI reduces the positive bias that is found in the RIOMDT experiment, and the root-mean-square error decreases constantly by about 5 mm until the beginning of 2004. The RMSE decrease due to the use of MDTOI with respect to RIOMDT (not shown) is evident especially in correspondence to the Antarctic Circumpolar Current, where the difference reaches values of up to −2 cm, which is consistent with the main differences found between the two MDTs.
From about 2003 to 2004, an increase in the RMSE is visible in Fig. 6, which seems to be due to a change in the ECMWF forcing fields. In particular, an increase in the net downward water flux, as seen by the ECMWF operational analyses with respect to the ERA-40 reanalyses, has led in turn to a constant increase in the globally averaged sea level height from 2002 onward. Also the lack in the correct detection of freshwater budget changes contributes to the increased RMSE, whereas the OGCM used in this study can simulate neither the effects of variations in the steric component of sea level nor the changes in the ocean mass due to ice melting, with the latter becoming more and more important in relative terms in the last decade (Cazenave et al. 2009; Leuliette and Miller 2009). As an additional reason for the RMSE increase, note also that since June 2003 the launch of Jason-1 led to a noticeable increase in the number of SLA observations, which may cause a change in the skill score statistics.
Furthermore, the error increase due to the use of MODMDT (visible in Fig. 6), which does not account for observations, is nonnegligible from 1997 onward and highlights the benefits of the MDTOI. The error increase is relatively large (up to 1 cm; not shown) in the Gulf Stream region and in the Southern Hemisphere in general.
The impacts of the assimilation of SLA on the SSH itself are nontrivial, as our analysis system does not correct the SSH. The use of a model-derived MDT implies also that the SSH variability rather than the absolute SSH is assimilated. Model sea level height has been verified against independent measurements from many island (i.e., noncoastal) tide gauge stations, whose data availability covered the period 1993–2005. Tidal gauge observations were taken from the Research Quality Dataset of the Joint Archive for Sea Level (JASL), provided by the University of Hawaii Sea Level Center. In particular, four stations in the Atlantic, two in the Indian, and 20 in the Pacific were chosen for this study according to their data availability for the period 1993–2005 and their distance from the coast. Table 3 summarizes the impacts of the sea level anomaly observations. We found that the impacts of the SLA assimilation on the Atlantic stations verification are not very significant. More significant are the impacts on the two stations located in the Indian Ocean where the RMSE decreases by about 1 cm. Different impacts have been found for the 20 Pacific Ocean stations, where 5 stations exhibited an almost neutral impact (RMSE difference less than ±0.5 cm), 3 stations a slightly positive impact (RMSE difference between 0.5 and 1.5 cm), and 12 stations a positive impact (RMSE decrease greater than 1.5 cm). The impacts of the different MDTs are smaller compared to the impacts of the SLA observations themselves.
2) In situ observations
Verification procedures against high quality in situ observations from the Tropical Atmosphere–Ocean (TAO) project in the equatorial Pacific, the Research Moored Array for African–Asian–Australian Monsoon Analysis and Prediction (RAMA) in the equatorial Indian, and the Prediction and Research Moored Array in the Atlantic (PIRATA) have been conducted for the entire experimental period. These datasets represent a unique source of information about temperature and salinity variability from moorings in the equatorial region. While temperature data are measured by means of very precise thermistors, salinities are calculated from the conductivity and temperature values using the method of Fofonoff and Millard (1985). For verification purposes, we used averaged daily values.
The root-mean-square errors against the three observing arrays are shown in Fig. 7 for the first 150 m as a function of forecast length. In the equatorial Pacific (TAO) the impacts of SLA observations are positive on the temperature regardless of the MDT used in the experiment, while for the salinity the use of the observed MDT degrades the skill scores with respect to the experiment without SLA assimilation. A similar result is found in the Indian Ocean (RAMA), although the impacts of SLA on the temperature skill score are much slighter. Finally, the skill scores in the equatorial Atlantic are worse when the RIO04 MDT is also used for temperature, while the SLA impacts are neutral for temperature and salinity when the MDTOI is used. In these verification statistics, we have also verified the MODMDT experiment, which exhibits in general a negative impact, except for the salinity scores in the Pacific and, partly, in the Indian Oceans. To summarize, the MDTOI experiment is the one that shows the most positive impact for both temperature and salinity.
Verification of the ocean heat content has been conducted against the heat content dataset of Levitus et al. (2009). They provide an estimate of the heat content in the first 700 m for the period 1955–2008 by using bias-corrected and quality-checked in situ measurements of subsurface temperature from bathythermographs (XBT and MBT), ocean station data (OSD) casts, and conductivity–temperature–depth (CTD) sensors. The comparison was performed in terms of seasonal mean anomaly correlation. Heat content data in general fit well with model heat content in the equatorial and North Pacific, and partly in the North Atlantic and Indian Oceans, while the comparison worsens in the Southern Hemisphere extratropics (where, however, there is a much smaller number of observations). In Fig. 8a the anomaly correlation map of the heat content is contoured for the experiment MDTOI. In Fig. 8b we show the differences in the anomaly correlation between the MDTOI experiment and the NOSLA experiment (i.e., positive values indicate a model-data correlation increase when SLA observations are assimilated). We find significant improvement in the model heat content caused by SLA assimilation in the North and South Pacific, in the Indian Ocean, and partly in the Atlantic, while the impacts of the altimeter data are noisy and often negative in the high-latitude Southern Hemisphere. The use of the RIO04 MDT (not shown) shows in general features similar to the MDTOI impact, although one can note a slightly greater impact in the Pacific and a diminished impact in the equatorial Atlantic and the band between −70° and −40° of latitude, while using the MODMDT worsens the correlation scores in the North Atlantic only.
3) Surface currents
The impacts of sea level anomaly observations on the surface currents were studied by comparing the monthly means of the zonal component of surface currents against the Ocean Surface Current Analysis-Real Time (OSCAR; Bonjean and Lagerloef (2002)) dataset, which derives its surface currents from satellite altimeters and scatterometers (surface vector wind data). Although this dataset is not completely independent from our experiments, as it is obtained also from sea level anomaly data and adjusted by means of surface vector wind, it has proved to have good quality especially within tropical regions. Further, our assimilation scheme does not correct the velocity fields and, therefore, the comparison is a proxy for how the assimilation of sea level anomaly observations is able to induce a correct modification in the tropical and subtropical surface circulation patterns. The agreement of the model surface circulation is particularly evident in the tropics and in the North Pacific, while weaker elsewhere (see Fig. 9a for the surface zonal current correlation between the OSCAR data and the experiment MDTOI). The impacts of the SLA data are shown as the difference between the surface zonal current correlation of the MDTOI experiment and that of the NOSLA experiment (Fig. 9b). The largest positive impacts are found within the band 30°S–30°N in all of the oceans. An increase of up to 0.4 in the correlation is particularly clear in the tropical Pacific. A very similar positive impact was also found when the RIO04 MDT or the MODMDT were used (not shown), although a deterioration in the verification scores was found in close proximity of the equator (about 3°S–3°N) in both the Pacific and the Atlantic Oceans.
5. Summary and discussion
In this paper we have described a strategy for assimilating along-track satellite altimetric observations into a new global ocean three-dimensional variational data assimilation system and investigated its benefits.
Further, to all the quality-checked in situ observations, we assimilate along-track sea level anomaly observations from TOPEX/Poseidon, ERS-1 and -2, the Geosat Follow-On (GFO), Jason-1, and Envisat provided by Aviso and treated by means of a one-dimensional low-pass Lanczos filter to remove the signal at scales smaller than the model resolution. Along-track SLA data were assimilated via a local hydrostatic adjustment scheme, which splits the sea level increment, proportional to the water-column-integrated density increment, into its thermo- and halosteric contributions, provided that the increment is spread into vertical profiles of temperature and salinity according to the local structure of the bivariate background-error vertical covariances.
We have built an ad hoc global mean dynamic topography from model mean SSH, which was subsequently adjusted through assimilation statistics, whereas the MDT bias may be reasonably considered to be the main contributor to the observation minus first-guess bias. This approach aims at assimilating the sea level temporal and spatial variabilities rather than the absolute sea level height, which would require full consistency between an observations-based MDT and the model mean SSH. The large-scale features of the MDT thus computed are in general very similar to those derived from gravimetry and in situ data, like the MDT from Rio and Hernandez (2004), except in some areas such as the Gulf Stream region and especially the Antarctic Circumpolar Current, the former due to the different MDT resolution used and the latter to the poor amount of in situ observations.
We have devoted significant effort to rigorously define the observational error associated with the SLA observations, which was formally decomposed into (i) instrumental error; (ii) representativeness error, deduced from the differences in gridpoint-averaged SLA superobservations and raw altimetric data; (iii) observation operator error, that is, the accuracy of our local hydrostatic adjustment scheme with respect to the more accurate prognostic free-surface formulation contained in a global ocean model; and (iv) the MDT error computed from the SSH anomaly variability.
The experiment including the SLA observations assimilation was extensively compared with an SLA-denial experiment and with an experiment that used the RIO04 MDT. As a self-consistent proof of the LHA scheme, we have found good improvement in the SSH fields, which are however not corrected by the 3DVAR system when SLA data are assimilated, against both sea level anomaly data and noncoastal tide gauge stations, especially in the Pacific Ocean. For the SSH field the positive impacts of assimilating SLA seem to be quite robust for all of the MDTs that we used. Comparisons with high quality datasets in the tropical regions (TAO, PIRATA, and RAMA) prove that the SLA assimilation is clearly of benefit for both the temperature and salinity skill scores. However, the results are quite sensitive to the MDTs that we used.
We found that the definition of the mean dynamic topography is crucial for the impacts of the altimetric data. The results have shown that neglecting either the OGCM dynamics and SSH or the observational dataset can lead to inconsistencies between the spatial variations of SLA model equivalents and the observations, which in turn affects the quality of the analyses. In particular, the use of a gravimeter-based MDT is detrimental to the tropical Pacific and Indian salinity skill scores, and to the tropical Atlantic temperature scores with respect to the experiment with the assimilation of in situ observations only, while the use of the MDT that we built ad hoc improves the scores for both temperature and salinity in the Pacific and Indian Oceans. A model only based MDT, in most cases, further deteriorates the verification skill scores for both temperature and salinity in the tropics.
Another positive impact of the SLA observations assimilation strategies was found in the surface current verification against the OSCAR dataset: the assimilation of SLA significantly improves the tropical circulation, especially in the Pacific Ocean.
It is important to stress that with the present configuration of the OGCM, we have been able to simulate only the sea level variations due to advective and evaporation minus precipitation terms, and not the steric variations (due to the Boussinesq approximation in the ocean model formulation) nor the contribution of ice melting (due to the lack of estimates of continental ice melting and the use of a climatological river runoff database). The model SSH variations have much smaller variabilities as compared to those observed by altimetric satellites. Therefore, as the ocean mass variations are not properly simulated, the success of an MDT based on the model and adjusted via observations resides in the importance of simulating spatial variations of sea level height rather than absolute values of sea level height.
As a final remark on future developments, we shortly plan to increase the analysis system resolution to an eddy-permitting one (0.25° in both the zonal and meridional directions). The relative amount of sea level anomaly observational data with respect to other in situ observations will increase noticeably, and their impact is expected to grow accordingly. We also plan to remove the assumption of spatially uncorrelated observational error for the SLA observations due to the increased resolution, which will be comparable with the altimetry correlation length scale (Le Traon and Hernandez 1992; Le Traon and Dibarboure 1999).
Acknowledgments
This work has been fully supported by the European Commission-funded project MyOcean in the framework of global ocean reanalysis production (WP4). We also acknowledge the INGV study program “Programma Internazionale di Studi Avanzati sull’Ambiente e sul Clima,” funded by the “Fondazione Cassa di Risparmio di Bologna,” which initially funded the Post-doc fellowship of Andrea Storto, and the Comprehensive Modelling of the Earth System for Better Climate Prediction and Projection (COMBINE) European project, which funded the work of Srdjan Dobricic. The altimeter products were produced by Ssalto/Duacs and distributed by Aviso, with support from Cnes (http://www.aviso.oceanobs.com/duacs/). The TAO Project Office of National Oceanic and Atmospheric Administration (NOAA)/Pacific Marine Environmental Laboratory (PMEL) and the OSCAR Project Office are acknowledged for the production and distribution of the TAO/RAMA/PIRATA and the OSCAR datasets, respectively. The authors want also to thank the Aviso team for the support in the use of SLA data and Dr. Simon Good (Met Office) for the support in the use of the EN3 dataset. The authors are also very grateful to the two anonymous reviewers and the editor for their precious suggestions and valuable comments, and for the improvement of the quality of this paper.
REFERENCES
Alessandri, A. , A. Borrelli , S. Masina , A. Cherchi , S. Gualdi , A. Navarra , P. Di Pietro , and A. Carril , 2010: The INGV–CMCC Seasonal Prediction System: Improved ocean initial conditions. Mon. Wea. Rev., 138 , 2930–2952.
Alves, J. O. S. , K. Haines , and D. L. T. Anderson , 2001: Sea level assimilation experiments in the tropical Pacific. J. Phys. Oceanogr., 31 , 305–323.
Bellucci, A. , S. Masina , P. Di Pietro , and A. Navarra , 2007: Using temperature salinity relations in a global ocean implementation of a multivariate data assimilation scheme. Mon. Wea. Rev., 135 , 3785–3807.
Bonjean, F. , and G. S. E. Lagerloef , 2002: Diagnostic model and analysis of the surface currents in the tropical Pacific Ocean. J. Phys. Oceanogr., 32 , 2938–2954.
Byrd, R. H. , P. Lu , J. Nocedal , and C. Zhu , 1995: A limited memory algorithm for bound constrained optimization. SIAM J. Sci. Comput., 16 , 1190–1208.
Cardinali, C. , S. Pezzulli , and E. Andersson , 2004: Influence-matrix diagnostic of a data assimilation system. Quart. J. Roy. Meteor. Soc., 130 , 2767–2786.
Chapnik, B. , G. Desroziers , F. Rabier , and O. Talagrand , 2006: Diagnosis and tuning of observational error in a quasi-operational data assimilation setting. Quart. J. Roy. Meteor. Soc., 132 , 543–565.
Cheney, R. , L. Miller , R. Agreen , N. Doyle , and J. Lillibridge , 1994: TOPEX/Poseidon: The 2-cm solution. J. Geophys. Res., 99 , 24555–24564.
Cooper, M. , and K. Haines , 1996: Altimetric assimilation with water property conservation. J. Geophys. Res., 101 , 1059–1077.
Courtier, P. , 1997: Variational methods. J. Meteor. Soc. Japan, 75 , 211–218.
De Mey, P. , and M. Benkiran , 2002: A multi-variate 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.
Di Pietro, P. , and S. Masina , 2009: The CMCC-INGV Global Ocean Data Assimilation System (CIGODAS). CMCC Research Paper 0072, 39 pp. [Available online at http://www.cmcc.it].
Dobricic, S. , 2005: New mean dynamic topography of the Mediterranean calculated from assimilation system diagnostics. Geophys. Res. Lett., 32 , L11606. doi:10.1029/2005GL022518.
Dobricic, S. , and N. Pinardi , 2008: An oceanographic three-dimensional assimilation scheme. Ocean Modell., 22 , 89–105.
Dobricic, S. , N. Pinardi , M. Adani , C. Fratianni , A. Bonazzi , and V. Fernandez , 2008: Daily oceanographic analyses by the Mediterranean basin scale assimilation system. Ocean Sci., 3 , 149–157.
Duchon, C. E. , 1979: Lanczos filtering in one and two dimensions. J. Appl. Meteor., 18 , 1016–1022.
Fischer, M. , M. Flügel , and M. Ji , 1997: The impact of data assimilation on ENSO simulations and predictions. Mon. Wea. Rev., 125 , 819–829.
Fofonoff, N. P. , 1985: Physical properties of seawater: A new salinity scale and equation of state of seawater. J. Geophys. Res., 90 , 3332–3342.
Fofonoff, N. P. , and R. C. Millard Jr. , 1985: Algorithms for computation of fundamental properties of seawater. UNESCO Marine Science Tech. Paper 44, 53 pp.
Hayden, C. M. , and R. J. Purser , 1995: Recursive filter objective analysis of meteorological fields: Applications to NESDIS operational processing. J. Appl. Meteor., 34 , 3–15.
Ingleby, B. , and M. Huddleston , 2007: Quality control of ocean temperature and salinity profiles—Historical and real-time data. J. Mar. Syst., 65 , 158–175.
Ivchenko, V. O. , S. Danilov , D. Sidorenko , J. Schröter , M. Wenzel , and D. L. Aleynik , 2008: Steric height variability in the Northern Atlantic on seasonal and interannual scales. J. Geophys. Res., 113 , C11007. doi:10.1029/2008JC004836.
Ji, M. , R. W. Reynolds , and D. W. Behringer , 2000: Use of TOPEX/Poseidon sea level data for ocean analysis and ENSO prediction: Some early results. J. Climate, 13 , 216–231.
Kalnay, E. , 2002: Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press, 341 pp.
Kaplan, A. , M. A. Cane , D. Chen , D. L. Witter , and R. E. Cheney , 2004: Small-scale variability and model error in tropical Pacific sea level. J. Geophys. Res., 109 , C02001. doi:10.1029/2002JC001743.
Lee, T. , and Coauthors , 2010: Consistency and fidelity of Indonesian-throughflow total volume transport estimated by 14 ocean data assimilation products. Dyn. Atmos. Oceans, 50 , 201–223.
Le Traon, P. Y. , and F. Hernandez , 1992: Mapping the oceanic mesoscale circulation: Validation of satellite altimetry using surface drifters. J. Atmos. Oceanic Technol., 9 , 687–698.
Le Traon, P. Y. , and F. Ogor , 1998: ERS ½ orbit improvement using TOPEX/Poseidon: The 2 cm challange. J. Geophys. Res., 103 , 8045–8057.
Le Traon, P. Y. , and G. Dibarboure , 1999: Mesoscale mapping capabilities of multiple-satellite altimeter missions. J. Atmos. Oceanic Technol., 16 , 1208–1223.
Le Traon, P. Y. , P. Gaspar , F. Bouyssel , and H. Makhmara , 1995: Using TOPEX/Poseidon data to enhance ERS-1 orbit. J. Atmos. Oceanic Technol., 12 , 161–170.
Le Traon, P. Y. , F. Nadal , and N. Ducet , 1998: An improved mapping method of multi-satellite altimeter data. J. Atmos. Oceanic Technol., 15 , 522–534.
Leuliette, E. W. , and L. Miller , 2009: Closing the sea level rise budget with altimetry, Argo and GRACE. Geophys. Res. Lett., 36 , L04608. doi:10.1029/2008GL036010.
Levitus, S. , J. I. Antonov , T. P. Boyer , R. A. Locarnini , H. E. Garcia , and A. V. Mishonov , 2009: Global ocean heat content 1955–2008 in light of recently revealed instrumentation problems. Geophys. Res. Lett., 36 , L07608. doi:10.1029/2008GL037155.
Lorenc, A. , 1981: A global three-dimensional multivariate statistical interpolation scheme. Mon. Wea. Rev., 109 , 701–721.
Madec, G. , P. Delecluse , M. Imbard , and C. Lévy , 1998: OPA 8.1 ocean general circulation model reference manual. Note du Pôle de Modélisation 11, Institut Pierre-Simon Laplace, 91 pp.
Masina, S. , P. Di Pietro , A. Storto , S. Dobricic , A. Alessandri , and A. Cherchi , 2010: Re-analysis in the global ocean at CMCC-INGV: Examples and applications. Mercator Ocean Quarterly Newsletter, No. 36, GIP Mercator Ocean, Ramonville St. Agne, France, 28–38. [Available online at http://www.mercator-ocean.fr/documents/lettre/lettre_36_en.pdf].
Mitchell, H. L. , and R. Daley , 1997: Discretization error and signal/error correlation in atmospheric data assimilation. II: The effect of unresolved scales. Tellus, 49A , 701–708.
Montmerle, T. , F. Rabier , and C. Fischer , 2007: Relative impact of polar-orbiting and geostationary satellite radiances in the ALADIN/France numerical weather prediction system. Quart. J. Roy. Meteor. Soc., 133 , 655–671.
Oke, P. R. , and P. Sakov , 2008: Representation error of oceanic observations for data assimilation. J. Atmos. Oceanic Technol., 25 , 1004–1017.
Pascual, A. , C. Boone , G. Larnicol , and P.-Y. Le Traon , 2009: On the quality of real-time altimeter gridded fields: Comparison with in situ data. J. Atmos. Oceanic Technol., 26 , 556–569.
Pinardi, N. , and Coauthors , 2003: The Mediterranean Ocean Forecasting System: First phase of implementation (1998–2001). Ann. Geophys., 21 , 3–20.
Reynolds, R. W. , and T. M. Smith , 1994: Improved global sea surface temperature analyses using optimum interpolation. J. Climate, 7 , 929–948.
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.
Roullet, G. , and G. Madec , 2000: Salt conservation, free surface, and varying levels: A new formulation for ocean general circulation models. J. Geophys. Res., 105 , 23927–23942.
Schiller, A. , P. R. Oke , G. B. Brassington , M. Entel , R. Fiedler , D. A. Griffin , and J. V. Mansbridge , 2008: Eddy-resolving ocean circulation in the Asian–Australian region inferred from an ocean reanalysis effort. Prog. Oceanogr., 76 , 334–365.
Segschneider, J. , D. L. T. Anderson , and T. N. Stockdale , 2000: Towards the use of altimetry for operational seasonal forecasting. J. Climate, 13 , 3115–3138.
Storto, A. , and F. T. Tveter , 2009: Assimilating humidity pseudo-observations derived from the cloud profiling radar aboard CloudSat in ALADIN 3D-Var. Meteor. Appl., 16 , 461–479.
Tapley, B. D. , D. P. Chambers , S. Bettadpur , and J. C. Ries , 2003: Large scale ocean circulation from the GRACE GGM01 geoid. Geophys. Res. Lett., 30 , 2163. doi:10.1029/2003GL018622.
Uppala, S. M. , and Coauthors , 2005: The ERA-40 Re-Analysis. Quart. J. Roy. Meteor. Soc., 131 , 2961–3012.
Vidard, A. , M. Balmaseda , and D. Anderson , 2009: Assimilation of altimeter data in the ECMWF Ocean Analysis System 3. Mon. Wea. Rev., 137 , 1393–1408.
Vossepoel, F. C. , and D. W. Behringer , 2000: Impact of sea level assimilation on salinity variability in the western equatorial Pacific. J. Phys. Oceanogr., 30 , 1706–1721.
Wijffels, S. E. , J. Willis , C. M. Domingues , P. Barker , N. J. White , A. Gronell , K. Ridgway , and J. A. Church , 2008: Changing expendable bathythermograph fall rates and their impact on estimates of thermosteric sea level rise. J. Climate, 21 , 5657–5672.
Zheng, F. , J. Zhu , and R.-H. Zhang , 2007: Impact of altimetry data on ENSO ensemble initializations and predictions. Geophys. Res. Lett., 34 , L13611. doi:10.1029/2007GL030451.
(top) MDT computed as explained in the text and called MDTOI; contour interval is 20 cm. (middle) Differences between RIO04 MDT and MDTOI and (bottom) between an MDT computed only from the model long-term SSH mean and MDTOI. Units are cm.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Contributors to the SLA observational error variance: (top) representativeness error (cm) of SLA observations deduced from the differences in gridpoint-averaged SLA data and raw SLA observations and (bottom) MDT error (cm) deduced from the SSH anomaly variability.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Contour map of the mean of the differences in (top) temperature and (bottom) salinity analysis increments with and without assimilating SLA observations in the first 100-m depth.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Zonally averaged standard deviation of the differences in (top) temperature and (bottom) salinity analysis increments with and without assimilating SLA observations as a function of latitude and depth.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
(left) Absolute and (right) relative DFSs, grouped by instrument and parameter (for in situ) or satellite (for SLA) observations.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Global ocean time series of sea level observation (top) biases and (bottom) RMSEs against model-equivalent background.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
RMSEs of model (left) temperature and (right) salinity against TAO, RAMA, and PIRATA arrays as a function of forecast length, in the first 150-m depth.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Comparison of model heat content between 0- and 700-m depth with the Levitus et al. (2009) dataset for the period 1993–2005. (top) Anomaly correlation of the MDTOI experiment with the dataset. (bottom) Differences in the anomaly correlation between MDTOI and NOSLA; positive (negative) values indicate a better (worse) fit of the MDTOI model fields.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Comparison of model zonal surface current with the OSCAR dataset for the period 1993–2005. (top) Anomaly correlation of MDTOI with the dataset and (bottom) the differences of the anomaly correlation between MDTOI and NOSLA. Positive (negative) values indicate a better (worse) fit of the MDTOI model fields.
Citation: Monthly Weather Review 139, 3; 10.1175/2010MWR3350.1
Temporal coverage of altimetric satellite observations as used in the assimilation system.
List of experiments.
RMSEs of model SSH against open-ocean tide gauge measurements. Values are in cm.
