In order to solve the cross-sessions variability of the motor imagery electroencephalogram (EEG) and avoid the lengthy recalibration step before using, a motor imagery classification method based on discriminative transfer feature learning (DTFL) is proposed. DTFL reduces the domain difference by jointly matching the marginal distribution and class conditional distribution of both domains. DTFL simultaneously maximizes the inter-class dispersion and minimizes the intra-class scatter as much as possible, which can preserve class discrimination information and further improve classification performance. The DTFL-based motor imagery classification method does not need the class information of EEG samples in the target domain, which can effectively avoid long-term calibration. Experimental results on brain-computer interface competition data sets verified that compared with some transfer learning methods, the proposed DTFL mitigated the cross-session variability and improved the classification accuracy of motor imagery EEG.