In this paper, we present a novel motor imagery classification method in electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) using locally generated CSP features centered at each channel. By favoring the channels with the local CSP features exhibiting significant eigenvalue disparity in the classification stage, improved performance in classification accuracy can be achieved in comparison with the conventional globally optimized CSP feature, especially for small-sample setting environments. Simulation results confirm the significant performance improvement of the proposed method for BCI competition III dataset Iva using 18 channels in the motor area.