To address the cross-sessions variability of motor imagery electroencephalogram (EEG) and eliminate the need for lengthy recalibration step, this study proposes a motor imagery classification method based on discriminative transfer feature learning (DTFL). DTFL aims to reduce domain differences by jointly matching the marginal distribution and class conditional distribution of both domains. Simultaneously, DTFL maximizes inter-class dispersion and minimizes intra-class scatter, preserving class discrimination information and improving classification performance. This method does not require class information for EEG samples in the target domain, effectively avoiding the need for long-term calibration. Experimental results on brain-computer interface competition datasets demonstrate that, compared with some transfer learning methods, the proposed DTFL mitigates cross-session variability and improves the classification accuracy of motor imagery EEG.