A recursive filter or parameterized curve fitting technique is usually used in a three-dimensional variational data assimilation (3DVAR) scheme to approximate the background error covariance, which can only represent the errors of an ocean field over a predetermined scale. Without an accurate flow-dependent error covariance that is also local and time dependent, a 3DVAR system may not provide good analyses because it is optimal only under the assumption of an accurate covariance. In this study, a sequential 3DVAR (S3DVAR) is formulated in model grid space to examine if there is useful information that can be extracted from the observation. This formulation is composed of a series of 3DVARs, each of which uses recursive filters with different length scales. It can provide an inhomogeneous and anisotropic analysis for the wavelengths that can be resolved by the observation network, just as with the conventional Barnes analysis or successive corrections. Being a variational formulation, S3DVAR can deal with data globally with an explicit specification of the observation errors; explicit physical balances or constraints; and advanced datasets, such as satellite and radar. Even though the S3DVAR analysis can be viewed as a set of isotropic functions superpositioned together, this superposition is not prespecified as in a single 3DVAR approach but is determined by the information that can be resolved by observation. The S3DVAR is adopted in a global sea surface temperature (SST) data assimilation system, into which the shipboard SSTs and the 4-km Advanced Very High Resolution Radiometer (AVHRR) Pathfinder daily SSTs are assimilated, respectively. The results demonstrate that the proposed S3DVAR works better in practice than a single 3DVAR.