Domain localization used in data assimilation allows computationally simplerand faster algorithms, particularly on parallel computers. Domain localizationis widely used in combination with the ensemble based Kalman Filter algorithmsthat use the analysis error covariance matrix for calculation of the gain insteadof the forecast error covariance matrix. However, since the assimilations areperformed independently on each local subdomain, smoothness of the analysisfields across the subdomain boundaries becomes an issue of concern. In order to address the problem of smoothness, we modifed the algorithm by using oneach subdomain observations from an area larger than the subdomain and theSchur product with an submatrix of isotropic matrix. On a simple example wedemonstrate that such a modification can produce results comparable to thoseproduced using direct forecast error localization.