In ensemble data assimilation, the estimate of background error covariance is particuarly important. In general, the use of a finite ensemble size for estimating the background error covariance matrix easily introduces spurious correlations, which leads to the underestimation of covariance and filter divergence. Covariance inflation is an empirical method of correcting the underestimation of background error covariance, but it does not help to solve the problem of long-range spurious correlations. Therefore, based on the EnKF scheme, we explored two localization methods to eliminate the spurious correlations, which were the covariance localization method and the local analysis method. We analyzed their impacts on the background error covariance matrix, gain matrix, ensemble transform matrices and data assimilation results. The experimental results have been obtained. That is, the localization method not only can remove the spurious covariance in the background error covariance matrix, but also can increase the rank of the matrix. In a “weak” assimilation, the gain matrix and ensemble transform matrices of two methods are very close, but the differences of the gain matrix and ensemble transform matrices become more evident with the increase of assimilation strength. Under the different strength of assimilation, two localization methods have their own characteristics, and relatively the covariance localization method has stronger robustness on the update of ensemble mean and ensemble anomalies. This study is very helpful for the fine analysis and estimate of the background error covariance.