The NASA Seasonal-to-Interannual Prediction Project (NSIPP) at the Goddard Space Flight Center uses a coupled ocean-atmosphere-land prediction model to produce ensemble forecasts of El Nino and its global teleconnections. One of the crucial components of the coupled prediction system is the ocean data assimilation algorithm used to initialize the ocean model. At present, the assimilation method used in production mode is a univariate form of optimal interpolation (OI). It is applied to process in situ temperature measurements into the Poseidon ocean general circulation model (OGCM). To bypass inherent limitations of the OI, a parallel ensemble Kalman filter (EnKF) has been implemented and tested.

An important feature of the EnKF and of other algorithms derived from the Kalman filter is their ability to produce error estimates. Also of importance are that its cost scales linearly with the ensemble size and the associated intrinsic parallelism.

Substantial progress has been made recently towards using the EnKF in production mode as part of the NSIPP forecast system. In practice, between 30 and 40 copies of the OGCM are run in parallel on 256 CRAY T3E processors. The assimilation algorithm is entirely parallel, relying on a localization of the analysis to achieve ad hoc compromise between speedup and accuracy.

Results from recent experiments assessing the benefits of replacing the OI with the EnKF are summarized. This presentation focuses on the assimilation of satellite altimetry into the OGCM. Independent in situ temperature and current measurements are used to measure the impact of the assimilation on the modeled ocean state. The effect of the assimilation on the hindcast skill of the OGCM in standalone mode and the extent to which the EnKF corrects the forecast-model bias are also discussed.