The RM-STC (Riemannian Manifold Space Transformer CNN) model based on Riemannian manifold space was proposed to explore the true order of information flow represented by Electroencephalogram (EEG) channels and improve emotion recognition performance. Firstly, the spatial covariance matrix features of EEG signals were calculated and mapped to Riemannian manifold space. The Riemannian distance matrix between EEG channels was then computed and subjected to a non-metric multidimensional scale transformation operation to obtain the one-dimensional ranking of the channels. The Pearson correlation coefficient feature matrix was rearranged according to the calculated relative distance order of the channels, allowing the CNN network to better convolve and learn local features. The advantage of modeling long-range dependencies in Transformer networks was utilized to learn global features to supplement the CNN network perspective, and the electrode channel order based on Riemannian manifold space computation was mapped into vector encoding embedded in the position encoding of the Transformer-CNN branch network, adding additional spatial position encoding information to the network. On the DEAP database, the average recognition rates of the valence dimension and arousal dimension by the proposed method reached 90.51% and 90.98%, respectively. The experimental results demonstrated that electrode arrangement based on Riemannian manifold space and effective spatial position encoding could effectively improve the accuracy of emotion recognition.