Experiments were done using a hybrid ensemble Kalman filter / 3D-var data assimilation scheme on a doubly periodic barotropic vorticity model. The hybrid technique uses all the other members of an ensemble to produce covariances for a particular member, and then perturbed observations are used with these covariances (as well as climatological covariances) to make a new analysis for each member. This is a computationally costly approach as ensemble sizes get large. Methods which produce perturbations about a control member, thus making only one new analysis (to the control member) for each data assimilation cycle, are explored.

Several ensemble generation techniques are investigated in the framework of the simple barotropic model: Perturbed Observations, Singular Vectors, Bred Vectors, Monte Carlo methods, as well as using eigenvectors of the climatological covariance matricies. The different perturbation methods are compared to determine which approach yields the best results. This is done with a 99 day simulation, building increments each day, and seeing which scheme produces the lowest daily global analysis and forecast error.