The common spatial pattern (CSP) has been widely used in the motor imagery based brain-computer interface(BCI). However, the traditional CSP has three defects, including the noise sensitivity of the sample covariance estimation, the subject-specific time window selection, and the subject- specific band selection. For the first two problems of CSP, two improved methods were proposed. In the first method, the variance of EEG signals in each channel is extracted as the feature, and Fisher linear discriminant analysis (FLDA) and Bayesian linear discriminant analysis (BLDA) methods are used to classify, then the channel weight distribution can be obtained, channels with larger weight are used to CSP transformation. The noise sensitivity of the CSP is reduced by eliminating the channel containing noise. In the second method, based on the assumption that the time and intensity of the subject's motor imagery is different, a new method of time window selection was proposed, FLDA and BLDA methods are still used for classification. In order to verify the effectiveness of the two improved algorithms, experiments were conducted using the 2005 BCI competition dataset IVa. The highest average classification accuracy of 87.77% and 81.23% are achieved by the two improved methods respectively. The experiment results show that the proposed methods are superior to the traditional CSP method. At the same time, in the two improved methods, the classification effect of the BLDA method is better than FLDA.