Brain-computer interface (BCI) is a new promising technology for control and communication, the BCI system aims to decode the measured brain activity into a command signal. This paper proposes a hybrid approach to improve the classification performance of motor imagery BCI (MI BCI). Our proposed method aims to take the advantage of two principal feature extraction approaches. The first approach named Multi-Band common spatial patterns (MBCSP) consists of decomposing the MI trial into multiple sub-bands, for each sub-band CSP is applied to extract the features. Then, the subject-specific frequency bands are selected. Simultaneously, the selected frequency bands are used as input to the second approach named boosted tangent space mapping (BTSM), which extracts the features from each sub-band. An automatic feature selection algorithm is introduced to select the subject-specific frequency bands and to reduce the high dimensionality space of extracted features. Finally, the LogitBoost (LB) classifiers are learned on the extracted features by each approach and the linear combination of these classifiers is used to identify the class of MI trial. The proposed model is evaluated on three public MI-BCI datasets, including multiclass motor imagery datasets (BCI competition IV dataset 2a and BCI competition III dataset 3a) and binary class motor imagery dataset (BCI competition III dataset 4a), The average accuracy attained on the three datasets are respectively 73.61%, 86.66%, and 86.68%. The conducted comparative study between the proposed approach and some state-of-the-the art methods showed a statistically significant improvement (p-value < 0.05) of the classification performance of MI BCI. The experiment results showed a significant improvement when using the proposed hybrid approach against the use of a single-feature extraction method.