There are two challenging problems in classifying a single-trial EEG of motor imagery. One is spectral filter optimization - The frequency bands, in which ERD/ERS patterns reflect activation and deactivation of rhythmic activity over motor and sensorimotor cortices, are highly variable across subjects and across even trials for the same subject. The other problem is spatial filter optimization - The EEG electrodes measure the superimposed signals that originated from various sources in the brain and the EEG signals are generally contaminated with artifacts and noise that can cause performance degradation in pattern classification. In this work, we propose a novel method for class-discriminative feature extraction by means of optimizing spatio-spectral filters in a Bayesian framework for EEG-based Brain-Computer Interfaces. In our method, the problem of optimizing spatio-spectral filter is formulated as estimation of a posterior probability density function (pdf). In order to estimate the unknown posterior pdf, about which, in this paper, there is no functional assumption, a particle-based approximation method is proposed by extending a factored-sampling technique with a diffusion process. An information-theoretic observation model is also devised to measure discriminative power of features between classes. The feasibility and effectiveness of the proposed method are demonstrated by analyzing the results and its success on public databases.