Reservoir characterization is needed for estimating reservoir properties and forecasting production rates in a reliable manner. However, it is challenging to figure out reservoir properties of interest due to limited information. Therefore, well-designed reservoir models, which reflect characteristics of a true field, should be selected and fine-tuned. We propose a novel scheme of generating initial reservoir models by using static data and production history data available. We select representative reservoir models by projecting reservoir models onto a two-dimensional (2D) plane using principal component analysis (PCA) and calculating errors of production rates against observed data. These selected models, which will have similar geological properties with the reference, are used to regenerate models by perturbing along the boundary of the different facies. These regenerated models have all the different facies distributions but share principal characteristics based on the selected models. We compare cases using 400 ensemble members, 100 models with unbiased uniform sampling, and 100 regenerated models by the proposed method. We analyze two synthetic reservoirs with different permeability distributions: one is a typical heterogeneous reservoir and the other is a channel reservoir with a bimodal permeability distribution. Compared to the cases using all the 400 models with ensemble Kalman filter (EnKF), the simulation time is dramatically reduced to 4.7%, while the prediction quality on oil and water productions is improved. Even in the more complex reservoir case, the proposed method shows great improvements with reduced uncertainties against the other cases.