In order to reduce the influence of the change of muscle contraction force on EMG pattern recognition, this paper proposes the feature of DCSP. Firstly, the spatial projection matrix that maximizes inter-class distance is obtained by the CSP. Then, the new signal after projection is differentiated and normalized. Finally, the data are projected into the low-dimensional space with the smallest intraclass distance and the largest inter-class distance by the uncorrelated linear discriminant analysis, the EMG gesture based on DCSP feature is verified on two datasets. The first dataset contains data from 10 complete limb subjects, and the second dataset contains data from 9 upper limb amputees. Among the four schemes of recognition rate testing, the recognition accuracy of the DCSP feature in this paper is higher than that of the CSP feature, and the highest recognition rate is achieved in all force training and all force testing schemes ( dataset1: 95. 83% , dataset2: 86. 93% ). Compared with CSP feature ( dataset1:89. 01% , dataset2: 70. 03% ), the classification accuracy rates are increased by 6% and 16% , respectively. In the two test schemes of feature spatial distribution, the DCSP feature has a smaller intra-class distance and a larger inter-class distance than the CSP feature. In the comparison results of other studies, the DCSP feature improves the recognition accuracy by about 5% (dataset1) and 8% ( dataset2) compared with the existing robust features, and the performance does not depend on the classifier.