The back error covariance controls the magnitude and structure of adjustment to the observations. Therefore, it plays a very important role in the quality of the assimilation. For the ensemblebased data assimilation methods, the ensemble determines the distribution of the background error covariance. This study investigates the impacts of choice and update of static ensemble for Ensemble Optimal Interpolation (EnOI) on the structure of the backerror covariance based on the longtime HYCOM model integration. The result shows that the static ensemble consisting of original model states amplify the ensemble correlation. The ensemble estimated from the model anomalies obtained by subduction of seasonal signals may reflect a reasonable distribution of the back errors. In the region dominated by the monsoon, the ensemble with season changes describes the flowdependence of backerror covariance better than the static ensemble. A series of experiments are carried out to validate the effects of different ensembles on the assimilation.