Due to its impressive representation power, the graph convolutional network (GCN) has attracted increasing attention in the hyperspectral image (HSI) classification. However, the most of available GCN-based methods for HSI classification utilize superpixels as graph nodes, which ignore the pixel-wise spectral-spatial features. To overcome the issues, we propose a novel multi-feature fusion network (MFGCN), where two different convolutional networks, i.e., multi-scale GCN and multi-scale convolutional neural network (CNN), are utilized in two branches, separately. The multi-scale superpixel-based GCN can reduce the computing power requirements, deal with the problem of labeled deficiency, and refine the multi-scale spatial features from HSI. The multi-scale CNN can extract the multi-scale pixel-wise local features for HSI classification. Furthermore, we introduced a 1D CNN to extract the spectral features for superpixels (nodes), which is different from most existing methods. Finally, a concatenate operation is employed to fuse the complementary multi-scale features. In comparison with the state-of-the-art models on three datasets, the proposed method achieves superior experimental results and outperforms competitive methods.