Abstarct: In recent years, one of the most popular research fields is the Brain Computer Interface (BCI). In Electroencephalogram (EEG)-baxsed BCI, brain signals are interpreted to control external devices as human intend. In a Motor Imagery baxsed BCI system, subject must imagine movement of one part of her/his body. This imagination produces events in her/his brain. The BCI system extracts these events from EEG signals and then recognizes the Motor Imagery type. In this thesis we are concerned with classifying EEG signals due to four class of Motor imagery, left hand, right hand, both feet and tongue. We propose four methods for classifying EEG signals. We evaluate the proposed methods using the dataset 2a of BCI Competition IV (2008). One of the most successful algorithms in Motor Imagery baxsed BCI is Common Spatial Patterns (CSP) algorithm. CSP method uses covariance matrix of two class data to extract spatial filters for discriminating them. This method fails to capture temporally local structure of samples. Moreover the noise on one sample affects the total variance. Local Temporal Common Spatial Pattern (LTCSP) method is an extension of CSP which alleviates these defects using an adjacency matrix. LTCSP has been proposed for two classes problem. In this paper we modify it using one-versus- the rest (OVR) algorithm for four-classes Motor Imagery baxsed BCI problem. In this thesis we propose a method named SEG-CSP-Var which captures time information from EEG signals. SEG-CSP-Var outperforms CSP method in two classes problem. The result of applying this method showed that the importance of EEG channels varies for different time segments and use of features of all segments is better than using features of whole signals as one segment. We propose another method so called SEG-CSP-BP which uses the power of frequency bands as features for classification. In this approach multi-interval discretisation of continuous-valued attributes is used as classifier and benchmarking attribute selection technique used as feature reduction method. For four classes problem, OVR-SEG-CSP-BP outperforms OVR-CSP, OVR-LTCSP, OVR-SEG-CSP-Var and even the best competitor of BCI competition 2008. This shows that using time and frequency domain features jointly improves the Motor Imagery classification problem.