Assimilation of sea surface salinity in a tropical Oceanic General Circulation Model (OGCM): A twin experiment approach

[1] Observing ocean surface salinity in the global ocean is a challenging issue for future years' oceanographic activities. It is motivated by the active role of salinity that is now well recognized in ocean dynamics and ocean/atmosphere exchanges. This is particularly evident in the case of the El Nino Southern Oscillation (ENSO) phenomenon in the tropical Pacific Ocean. Improvements to ocean state estimation and predictions will require that salinity observations must be taken into proper account in conjunction with temperature and altimetric data. The sensitivity of a primitive equation model of the tropical Pacific Ocean to sea surface salinity (SSS) is studied through the use of a data assimilation technique in the rather academic "twin experiment'' context. The data assimilation technique used, the Singular Evolutive Extended Kalman (SEEK) filter, is derived from the conventional Kalman filter theory. The paper explains why such a sophisticated technique is necessary. Indeed, an empirical scheme such as the Newtonian relaxation method, used in the same conditions, fails to constrain either the observed (surface) variable or the other components of the state vector. Within the experimental context chosen, the assimilation of SSS data with the SEEK filter is able to constrain most of the model variables linked with the SSS signal. SSS information, in particular, appears relatively successful in restoring zonal velocity, which is an important variable in warm/ fresh pool migration, and in simulating a barrier layer in the atmospheric convergence zones. The final analysis errors are small and stable over time. This is widely true when simulating satellite SSS observations based on the GODAE criteria (0.2 psu error, 200 km, 10 days), which shows the potential of these observations. To extend these results to a real context, the problems of model-data bias and unknown error covariances must be addressed as they are actually a strong limitation in assimilation performance when assimilating any real data set.