In motor imagery brain-computer interface (BCI), the spatial covariance matrices of electroencephalography (EEG) signals which is lying on smooth Riemannian manifolds, are well used to enhance the performance of motor imagery classification. However, in different session, due to the changes of spatial and frequency information which caused by physiological, environmental, as well as instrumental changes, model trained in calibration session often get poor performance in use session. To improve the cross-session performance, combining unsupervised domain adaptive methods with Riemannian manifold, we propose a method called as Correlation Alignment in Filter Bank Riemannian Tangent Space (FBTSCORAL). First, raw EEG signals are constructed into multiple frequency-bands spatial covariance matrices by filter bank, in each bands the spatial covariance matrices are mapped into Riemannian tangent space as a tangent vector, the final feature vector is merged by tangent vectors in all bands. Second, Correlation Alignment (CORAL) is used to align feature space in calibration session and feature space in use session. Finally, SVM is used to fit the aligned feature vectors. The experiment shows that our work achieves the state-of-the-art performance in BCI competition IV 2a dataset and the second performance in BCI competition III 3a and 4a datasets.