Assimilation of Argo float positions in the north western Mediterranean Sea and impact on ocean circulation simulations



[1] Assimilation of Argo float trajectories (350m parking depth) in a high resolution model of the Mediterranean Sea has been performed to test the importance of this type of data for operational systems. Starting from a state corresponding to the 28th year of the simulation, two experiments were performed, one with assimilation and one without, during three months in winter 2005. Four Argo floats released in the context of the MFSTEP project are used. The results are very encouraging, showing significant and consistent changes in the ocean circulation. Comparison with independent data of transport through the Channel of Corsica indicates that the results are more realistic with assimilation, showing a non negligible impact of the assimilation process on the large scale circulation of the basin.

1. Introduction

[2] Argo floats are an integral part of the ocean observing system ( They are autonomous profiling buoys that freely drift at a prescribed parking depth for time intervals of days, then rising at the surface to communicate via satellite information on temperature and salinity (TS) profiles and on position. Their main advantages are that they are autonomous, relatively inexpensive, and provide information on the water column which cannot be obtained by satellite products. For this reason a large number of such instruments have been released in the world oceans and in marginal seas as part of the international Argo program.

[3] Argo float data have an enormous range of applications ( They have been used to provide information on the general circulation and to validate models. Since they provide near real time information on water mass properties and currents, they can also be used directly in an operational mode via assimilation. Temperature and salinity profiles are presently assimilated in a number of operational systems ( Assimilation of their positions r, however, has not yet been performed operationally.

[4] The main difficulty in assimilating positions is that the integrated variable r is not an Eulerian state variable and is not related linearly to the Eulerian velocity state variable u. When the interval Δt between successive positions is small with respect to the Lagrangian time scale TL (order of 1–3 days at the ocean surface and 5–10 days in the subsurface), the problem can be circumvented by approximating u as the finite difference Δr/Δt. For Argo floats, though, Δt is of the order of TL, so the simplified approach does not hold and its application is expected to lead to significant errors and biases [Molcard et al., 2003]. In the last few years, many papers have addressed the question of developing an appropriate methodology to assimilate r [e.g., Kamachi and O'Brien, 1995; Kutsnetsov et al., 2003; Molcard et al., 2003, 2005]. They have shown the potential of assimilating positions, but tests were mostly focused on simplified configurations. More recently, the method by Molcard et al. [2005] has been further improved and implemented in a realistic high resolution circulation model [Taillandier et al., 2006; V. Taillandier and A. Griffa, Implementation of a position assimilation method for Argo floats in a Mediterranean Sea OPA model and twin experiment testing, submitted to Ocean Science, 2006, hereinafter referred to as Taillandier and Griffa, submitted manuscript, 2006], and the results have been tested using a twin experiment approach in the North Western Mediterranean Sea.

[5] In this paper we present a first example of assimilation of in-situ position data from Argo floats in the Mediterranean Sea. The floats have been released as part of the European MFSTEP program starting in 2004 [Pinardi et al., 2003], with a parking depth zp near 350m and a cycling interval Δt of 5 days [Poulain, 2005]. Our goal is to investigate whether the assimilation provides a significant and consistent correction to the modelled circulation. In order to specifically address this question, in the following experiments only position assimilation is performed, while no TS data from the floats nor satellite information are assimilated.

2. Position Assimilation Method

[6] The modelled circulation can be corrected by assimilating Argo float positions in a sequential manner. Estimates of the model velocity field u and mass field (T, S) are then provided every Δt as detailed below.

[7] Two successive float positions are used to estimate the velocity field at the parking depth zp. The variational method developed by Taillandier et al. [2006] provides a bi-dimensional velocity correction Δu(zp) by minimising the distance between observed float positions at the end of the sequence and positions of “prior” trajectories computed inside the model velocity field. This approach is of potential interest given the high model resolution (1/16 degree) and the low data sampling (5 days), since it expands the velocity correction all along the prior trajectories. Drifts occurring on ascent and descent float motions between zp and the sea surface are taken into account in the computation of the prior trajectories in order to adapt the purely Lagrangian approach to Argo data sets, which are quasi-Lagrangian in nature. In this way, surface positions can be transposed at depth zp at the accuracy of vertical float motion effects.

[8] The horizontal velocity correction is then used to estimate the three-dimensional correction Δu. As discussed by Molcard et al. [2005], Δu(zp) is projected on the water column

equation image

where R represents the vertical projection operator, computed from seasonal and spatial averages (in quasi-homogeneous regions) of the linear regression coefficients linking each velocity component at depth z with velocity at zp.

[9] Finally, the mass field can be estimated from data positions on the following inverse problem. The velocity variation Δu (equation (1)) can be geostrophically maintained with respect to mass variations (δT, δS) relative to their prior values [Oschlies and Willebrand, 1996]. More precisely, the vertical velocity shear is expressed as δσ = M.(δT, δS), where M is governed by the thermal wind equation for isopycnal slope adjustment and by the equation of state for density specification. The shear variation ∂zΔu is fitted to the geostrophic velocity shear variation δσ by finding the best combination (δT, δS) that minimises the cost function

equation image

In numerical practice, J is minimised along the steepest descent drawn by the gradient

equation image

where MT is governed by the adjoint equations of state and of the thermal wind; B is classically defined by background error covariances, simplified in this case to standard deviation profiles for T and S.

[10] As described by Taillandier and Griffa (submitted manuscript, 2006), this assimilation method has been implemented to satisfy operational assimilation processing and low computational cost requirements. It has been tested with a number of twin experiments using a realistic model and synthetic “Argo type” float positions. Sensitivity to data distribution and data accuracy associated to float drifts during vertical motions has been investigated. Such observational errors appear to influence the results at high data coverage, while for low but realistic coverage the results do not significantly deteriorate. Even with sparse coverage, the assimilation efficiently acts on the location, intensity and shape of velocity structures.

3. Argo Float Data and Circulation Model

[11] The experiments presented here occur in winter 2005 and they focus on the Levantine Intermediate Water (LIW) circulation as indicated by the float parking depth (zp = 350m). Patterns and evolution of the circulation in the North Western Mediterranean Sea constitute our experimental framework.

[12] The data of four Argo floats in the Western Mediterranean are considered during the period December 31 2004 – March 26 2005. Their successive positions at the sea surface form trajectories sampled at Δt = 5d, as represented in Figure 1 (see web page for details, The most western float deployed at (2°E, 40.5°N) drifts slowly while trapped inside an anticyclonic eddy, then quickly exits from the Balearic Sea. Its mean displacement during 5-day sequences is equal to 25 km. In the central area (6°E, 40°N) between the Balearic Islands and Sardinia, a second float describes a cyclonic circle loop with mean sequential displacement of 18 km. In the eastern area, two floats are partially observed during their northward drift with mean amplitude of 15–16 km. These four trajectories are in agreement with documented cyclonic paths of the LIW [Millot, 1999] in the North Western Mediterranean Sea. Despite some gaps in the data, they show the presence of meanders and recirculations.

Figure 1.

Map of the monthly mean streamfunction for January 2005 in the reference run, with four Argo float trajectories superimposed (initial positions indicated by crosses) as observed during December 31 2004 – March 26 2005. The Balearic and Corsican regions are drawn, as well as the CCS section (dashed).

[13] The numerical model is an extended version of the primitive equation model OPA configured in the Mediterranean Sea (K. Béranger et al., Interannual variability of water formation derived from a high resolution model of the Mediterranean Sea, submitted to Ocean Dynamics, 2005). The horizontal resolution at 1/16 degree is well adapted to study circulation features of lengths larger than 30 km. The vertical discretisation is set up on 43 levels stretching from 6m at the sea surface to 200m at the bottom. The model is initialised at rest by climatological hydrological conditions from MEDATLAS. It is forced by daily winds and heat fluxes from the ECMWF atmospheric forcing fields during the period 1987–2005.

[14] Two experiments have been performed from the same initial state obtained after 28 years of the model integration. They start 1st January 2005 and the integration is performed during 3 months. The “reference” run is done without assimilation, while the “analysis” run is done with the assimilation of the Argo float positions. The general circulation in the reference run is in good qualitative agreement with the float trajectories, as shown in Figure 1 where the trajectories are superimposed to the mean streamfunction for January 2005. The deep anticyclonic eddy near the Balearic Islands (2°E) can be seen in the Balearic float clockwise trajectory. Meanders drawn by the Corsica float follow the streamlines of the cyclonic general northward path of the LIW tongue along the Sardinia and Corsica coasts, while the cyclonic gyre centred at (6°E, 40°N) matches the generally cyclonic Sardinian float path.

4. Assimilation Results

[15] The assimilation of Argo float positions allows a model state update at the final time of each 5-day sequence. The analysis model run provides sequential averages of the velocity and mass fields, as well as sequential predictions of float trajectories from their observed positions. These model outputs are to be compared to the ones obtained from the reference run, i.e., without assimilation, in the same experimental framework.

[16] A first qualitative comparison is performed considering the circulation at the float parking depth. Examples of snapshots of current and salinity fields at 350m are shown in two different regions indicated in Figure 1: the Corsican region (Figure 2) and the Balearic region (Figure 3). The observed float positions and “tails” during two sequences are superimposed on the fields. The float motion appears in better agreement with the velocity field in the analysis run than in the reference, indicating that the velocity correction is consistent. Also, after two months of assimilation, significant differences appear in the velocity and salinity fields of the reference and analysis. In the Corsica region (Figure 2), the impact of assimilation is to modify the meander patterns and to move the current closer to the coast in northern Corsica, while the Northern Current appears reinforced along the French and Spanish coasts. In the Balearic region (Figure 3), the retroflection of the Northern Current appears more to the east in the analysis (3.5°E), and a stronger eastward jet can be seen north of the Baleares, in agreement with the float drift. For both regions, in the analysis run the boundary currents get more confined to the coasts and appear geostrophically maintained with respect to a redistribution of the mass field.

Figure 2.

Maps of salinity (shades) and currents (arrows) at 350m in the Corsican region in the middle of March 2005, for (a) the reference run and (b) the analysis run. Argo float trajectories corresponding to the 10 days before the middle of March are superimposed (initial positions indicated by crosses).

Figure 3.

Same as in Figure 2 for the Balearic region.

[17] The quantitative consistency of the analysis velocity field with the assimilated positions can be assessed using float positions as observed and predicted during each sequence. If we define ro as the observed position of the float at the end of a sequence, and rr and ra as the predicted positions computed from the reference and analysis velocity field respectively, a simple error metric can be defined as Er = equation imagerorrequation image, Ea = equation imageroraequation image. Note that particle trajectories are sensitive to the details of the Eulerian flows [e.g., Griffa et al., 2004] and are highly chaotic [e.g., Aref, 1984], so that prediction errors can easily reach the order of the observed float displacements. The time evolutions of Er and Ea are shown in Figure 4 for the representative case of the float in the Balearic Sea. As it can be seen, aside from the first sequence, Ea is always significantly lower than Er indicating that particle prediction is improved in the analysis run. The improvement appears especially evident when the float tends to move with a regular drift direction, such as, for instance, in the mid-March event in Figure 3. When instead the float is entrained in smaller scale structures resulting in looping trajectories, the correction appears less effective. This is due to the analysis time scale (5 days as dictated by the trajectory sampling), which does not allow for a full resolution of the structures even though they are compatible with the model space resolution [Testor et al., 2005]. Results for the other floats are qualitatively similar to Figure 4, with a mean ratio between the Ea and Er ranging between 60% to 80%.

Figure 4.

Evolution of the Lagrangian prediction errors for the float in the Balearic Sea: dashed (solid) line indicated the reference (analysis) error Er (Ea).

[18] A final assessment of analysis consistency is carried out considering model corrections in the water column. We consider time series of transport across the Corsica Channel section (CCS, indicated in Figure 1), which is not directly crossed by any float, and where independent transport observations are available from long term current meter measurements [Astraldi et al., 1999]. Time series of net transport computed from the analysis and reference runs and from the in-situ measurements (courtesy of G.P. Gasparini) are shown in Figure 5 for the period of interest. As it can be seen, the modelled transport generally underestimates the observed values, but the analysis run shows an increase with respect to the reference run (starting from mid-February) leading to instantaneous transport values closer to the observed ones. The increase in terms of the three month mean is approximately 10% (Figure 5). These results, even though limited by the length of the series, are suggestive that the assimilation is able to consistently correct not only the local amplitude of the transport, but also its larger scale behaviour.

Figure 5.

Net transports in Sv through section CCS (see Figure 1): dashed (solid) line indicates the reference (analysis) transport, while dotted line indicates the transport from current meter data (courtesy of G.-P. Gasparini).

5. Conclusions

[19] Argo float positions have been assimilated in a high resolution model of the Mediterranean Sea during 3 months in winter 2005. Despite the sparseness of the data (only 4 floats), the circulation at 350m is significantly affected by the assimilation process. Position information appear especially effective when the floats move with regular drift direction, i.e., when they are caught in mesoscale circulation structures. Position information from smaller scale recirculating structures resulting in looping trajectories are less effective since the time sampling does not allow a full resolution of the velocity field. The general circulation of the North Western Mediterranean Sea is also modified by the assimilation as illustrated by transports. In particular through the Corsica Channel, the northward transport is increased by assimilation and brought closer to observations.

[20] Overall, the results are very encouraging for future use of position assimilation from Argo floats in operational systems. Further improvements are expected to come from simultaneous assimilation of positions and temperature and salinity profiles, especially in the case of small scale structures characterised by well defined water mass signatures. Argo float data, with their subsurface information, provide an extremely valuable complement to more global but surface restricted satellite information such as Sea Surface Height (SSH) and Sea Surface Temperature (SST).


[21] We wish to thank G.-P. Gasparini for providing the current meter data in the Corsica Channel and all the scientists and crew members who helped with the float deployments, data management and processing. The model was provided by MERCATOR and calculations were performed on the IDRIS computing centre. This study was supported by the European Commission as part of the MFSTEP project (contract EVK3-CT-2002-00075) and by the Office of Naval Research (grant N00014-05-1-0094).