The second generation NASA GMAO ocean data assimilation system (ODAS-2) uses an “augmented” ensemble Kalman filter (EnKF) approach to assimilate ocean observations into an ensemble of coupled GEOS-5 GCMs (GEOS-5 AGCM coupled to the MOM-4 OGCM). The AGCM component of the coupled system is constrained by “replaying” the GMAO atmospheric analysis into the AGCM. The replay procedure involves the inclusion of random perturbations to maintain a representative ensemble spread.
The augmented EnKF approach relies on a background error-covariance model which combines information from the following sources: (a) a dynamic ensemble of model trajectories, (b) high-pass filtered past instances of the state of each model trajectory, (c) a steady state ensemble of background-error estimates obtained from an analysis of short-term forecast-error growth and (d) an idealized analytical error covariance model. The EnKF analysis is localized by allowing each observation to only influence the state variables inside a local volume in 5-dimensional (x, y, z, t, density) space. The localization in density space confers flow-adaptive properties upon the covariance localization. An iterative data adaptive algorithm is used to separately optimize the error-covariance localization scales involved in the processing of each observation. The representation error associated with each observation is also estimated iteratively to satisfy specific global and local constraints.
Prior to each EnKF analysis , the ensemble of ocean state estimates obtained by integrating the ensemble of coupled models since the time of the previous analysis is first redistributed using a particle filter approach. The prior particle filter redistribution generally results in posterior ocean-state estimates that more resemble a validation set of control observations than ocean-state estimates obtained with a traditional EnKF approach (i.e., without a prior redistribution).
Results from the assimilation of remotely sensed sea surface height observations and in situ temperature and salinity data into ensembles of MOM-4 OGCMs (on a 720×410×40 grid) and GEOS-5 AGCMs (on a 288×144×72 grid) will be shown and the relative performance of the various components of the ensemble assimilation system will be contrasted using closeness to independent unassimilated observations as the main performance metric.