Adaptive covariance hybridization for coupled climate reanalysis

<p>Because of their very heavy computational burden, climate prediction systems that use ensemble data assimilation methods can afford only a few tens of members. Sampling error in the covariance matrix can introduce biases in the unobserved regions (e.g. in the deep ocean). Here, we assess the potential of hybrid covariance approach (EnKF-OI, Hamill and Snyder 2000) to counteract sampling error. The EnKF-OI combines the dynamical covariance to that of a large static/historical covariance (EnOI). We use the Norwegian Climate Prediction Model (NorCPM), which combines the Norwegian Earth System Model (NorESM) and the Ensemble Kalman Filter (EnKF) in an idealized twin experiment. We test the performance of reanalyses that assimilate synthetic SST observations monthly for the period 1980-2010. We use a dynamical ensemble of 30 members and a static ensemble size of 315 members sampled from a long stable pre-industrial run. We compare the performance of the EnKF to 1) an EnKF-OI with a global hybrid coefficient tuned empirically and 2) an EnKF-OI adaptive, with hybrid coefficient explicitly estimated in space and time (Ménétrier and Auligné, 2015). In the adaptive EnKF-OI, the hybrid coefficient remains stable through the course of the reanalysis with only a weak seasonal variability.  Both EnKF-OI versions show comparable performance and cure the emergence of a bias in the deep ocean in the EnKF. The assimilation updates with the EnKF-OI adaptative are smaller, suggesting that it sustains a lower error level in between the assimilation cycle. </p>