Reservoir characterization is a process of making models, which reliably predict reservoir behaviors. Ensemble Kalman filter (EnKF) is one of the fine methods for reservoir characterization with many advantages. However, it is hard to get trustworthy results in discrete grid system ensuring preservation of channel properties. There have been many schemes such as discrete cosine transform (DCT) and preservation of facies ratio (PFR) for improvement of channel reservoirs characterization. These schemes are mostly applied to 2D cases, but cannot present satisfactory results in 3D channel gas reservoirs with an aquifer because of complex production behaviors and high uncertainty of them. For a complicated 3D channel reservoir, we need reliable initial ensemble members to reduce uncertainty and stably characterize reservoir models due to the assumption of EnKF, which regards the mean of ensemble as true. In this study, initial ensemble design scheme is suggested for EnKF. The reference 3D channel gas reservoir system has 200 × 200 × 5 grid system (250 × 250 × 100 ft for x, y, and z, respectively), 15% porosity, and two facies of 100 md sand and 1 md shale. As the first step, it samples initial ensemble members, which show similar water production behaviors with the reference. Then, grid points are randomly selected for high and low 5% from the mean of sampled members. As a final step, initial ensemble members are remade using the selected data, which are assumed as additional known data. This proposed method reliably characterizes 3D channel reservoirs with an aquifer.