 This paper proposes a novel approach for learning informative representations from electromyogram, EMG, signals. The approach uses symmetric positive definite, SPD, manifolds to capture spatial structural information within multiple EMG channels. The proposed method was tested on 11 gestures collected from 10 subjects and achieved an accuracy of 84.85% with an improvement of 4.04%. Additionally, the proposed method outperformed the contrast method in terms of accuracy and F1 score. Furthermore, the computational cost is less than the contrast method, making it more suitable for low-cost systems. This article was authored by Dijun Xiong, Daohui Zhang, Xinang Zhao, and others.