Motor-imagery-based brain computer interface has advantages in high information transition rate and has potential to achieve high portability without using a display for visual stimulation. However, electroencephalographic data is a mixture of neural activity from various brain locations and also external artifacts. This paper presents a novel method utilizing source-level features computed based on maximum contrast beamformer. The estimated source activity has the maximum contrast between the power during the reference period and the period that event-related synchronization or desynchronization emerges. Moreover, this method utilizes the classification structure based on divide-and-conquer concept to identify four classes of motor imageries, including left hand, right hand, tongue, and foot. The experimental results demonstrate the improvement in classification accuracy by using the proposed method.