An approach is proposed to estimate the location-dependent correlation scales for background error covariance matrix used in the 3D-Var data assimilation, and the impact on ocean assimilation is examined. By a nonlinear fitting procedure, the horizontal scales are computed at each model grid point using the outputs of a tropical Pacific OGCM. The derived correlation scales are comparable to those obtained from observations, but present more complicated structures quite different from empirical functions. They vary largely in the horizontal and show a distinct distribution from surface to subsurface. Three assimilation experiments from 1982 to 2000 are performed to assess the improvement due to the location-dependent correlation scales. The overall root-mean-square errors in the upper ocean are reduced about 0.5¿C in the eastern equatorial Pacific and about 0.4¿C in the western equatorial Pacific. The improvement shows the location-dependent correlation scales can afford more information of the background error structures.