Ensemble Kalman filter (EnKF) is one of the widely used optimization methods in petroleum engineering. It uses multiple reservoir models, known as ensemble, for quantifying uncertainty ranges, and model parameters are updated using observation data repetitively. However, it requires a large number of ensemble members to get stable results, causing huge simulation time. In this study, we propose a sampling method using principal component analysis (PCA) and K-means clustering. It excludes poor ensemble with different geological trends to the reference so we can improve both speed and reliability of future predictions. A representative model, which is selected from candidate models of each cluster, has a role to choose proper ensemble for EnKF. For applying EnKF to channelized reservoirs, we compare cases with using 400, randomly picked 100, sampled 100 using Hausdorff distance, and sampled 100 by the proposed method. The proposed method shows improvements over the other cases compared. It gives stable uncertainty ranges and well-updated reservoir parameters after the assimilations. Randomly selected 100 ensemble members predict wrong reservoir performances, and 400 ensemble members exhibit too large uncertainty ranges with long simulation times. Even though more ensemble members are utilized, they provide worse results due to disturbance by improperly designed models. We confirm our sampling strategy in a real field case, PUNQ-S3, and it reduces simulation time as well as improves the future predictions for efficient and reliable history matching.