In recent years, Superpixel-level Graph Convolutional Networks (SGCNs) have shown outstanding performance in hyperspectral image (HSI) classification due to their message-passing mechanism on graph structures. However, when using SGCNs for HSI classification there are several issues that they may not extract effective details within superpixel and may suffer from poor topological quality. To address these issues, this paper proposed a novel spatial-spectral graph mamba (Grama) method. Specifically, we firstly perform a multi-feature fusion operation to obtain effective spectral-spatial features, and then utilize an Mamba model to further extract spectral-spatial features from superpixels, providing GCN with more detailed local features. Secondly, during the GCN training process, the edge weights are updated based on structural similarity between superpixels. Labeled nodes are weighted according to the spectral mean of the superpixel nodes, altering the influence of nodes at different topological positions. As a result, the topological graph structure can be optimized from multiple perspectives, thereby improving the message-passing mechanism. Finally, we validated the performance of the Grama method on three publicly available HSI datasets. Comprehensive experimental results show that it outperforms recently popular methods in terms of Overall Accuracy (OA), Average Accuracy (AA), and Kappa Coefficient (KC).