In high-resolution ocean models, the most rapidly growing forecast errors are often associated with the small-scale (mesoscale) shears found near fronts and eddies. Information about the large-scale flow contained in point observations coexists with mesoscale "noise." The separation of information at small and large scales poses a considerable challenge for both sequential and variational data assimilation. In this paper, we compare the effectiveness of three approaches to assimilation in flows in which a non-trivial, large-scale circulation co-exists with an energetic small-scale field: the extended Kalman filter, Bayesian hierarchical modeling, and an adjoint-based variatonal approach.