Abstract: The dimension reduced projection four dimensional variational data assimilation (DRP-4DVar) approach utilizes the historical forecas based initial perturbation samples to estimate the background error covariance (BEC). Then the new idea of pre analyzing the initial perturbation samples is presented to improve the flow-dependent property of BEC in the DRP-4DVar system. Before analyzing the background field, each sample is pre-analyzed by using the DRP-4DVar system so that the newly estimated BEC will evolve with the actual weather situation. Consequently, the real flow-dependency of BEC can be achieved and meanwhile the “filter divergence” can be effectively prevented in cycling runs. The experiment results show that,by ameliorating the spatical structure and flow-dependency of the BEC, the pre-analyzed samples indeed help to further improve the performance of DRP 4DVar, which provides more accurate initial field for the numerical forecast model. Not only the forecasting precision of basic model variables can be improved, but also the simulated precipitation is ameliorated. By contrast, the assimilation run with all of the initial perturbation samples pre-analyzed obtains the optimal analysis and forecasts.