As one of the key techniques determining the overall system performances, efficient and reliable algorithms for improving the classification accuracy of motor imagery (MI) based electroencephalography (EEG) signals are highly desired for the development of brain-computer interface (BCI) systems. In this study, we propose, for the first time to the best of our knowledge, a novel data adaptive empirical wavelet transform (EWT) based signal decomposition method for improving the classification accuracy of MI based EEG signals. Specifically, to reduce the system complexity and execution time, the proposed method selects 18 electrodes out of 118 to analyze the non-stationary and nonlinear EEG signal behaviors. Meanwhile, the method adopts the Welch power spectral density (PSD) analysis method for single mode selection out of the total 10 for each channel, and the Hilbert transform (HT) method for both instantaneous amplitude (IA) and instantaneous frequency (IF) signal components extraction for each selected mode. With seven commonly used machine-learning classifiers adopted, extensive experiments were conducted with the benchmark dataset IVa from BCI competition III to evaluate the performance of the proposed method. Results show that with the IA and IF component features being tested using the least-square support vector machine (LS-SVM) classifier, the EWT method achieves an average classification accuracy of 95.2% and 94.6% respectively, which is higher as compared with the existing methods. While for every participant, a classification accuracy of at least 80% could be achieved by employing a single feature only. Results also show that a combination of EWT and higher order statistics features, which contain both kurtosis and skewness of the extracted instantaneous components, help achieve a higher success rate. The better performances of EWT over those of the existing methods demonstrate the effectiveness and great potential of EWT for BCI system applications.