Hyperspectral image classification based on mixed similarity graph convolutional network and pixel refinement

The traditional convolutional neural network cannot extract Non-Euclidean spatial information in hyperspectral image classification, while the superpixel-based graph convolutional network relies on node aggregation and superpixel segmentation accuracy. To solve these problems, a mixed similarity graph convolutional network and pixel refinement method (MSGCN-CRF) is proposed in this paper. Firstly, homogeneous superpixel-level features are obtained as nodes using a superpixel segmentation algorithm, and then a mixed similarity method is designed to aggregate nodes. This method combines spectral intensity similarity and spectral curve similarity to obtain more discriminant node features. Secondly, a two-layer convolutional layer is used to remove noise from the original hyperspectral image. Then node features are constructed and input into the two-layer graph convolutional network, which using the adjacency matrix generated by mixed similarity to guide superpixel node aggregation. Finally, pixel refinement is implemented by fully connect CRF, which is used to correct the false prediction caused by the segmentation error of superpixels, and to obtain more accurate classification results using spatial and spectral information between pixels. Experimental results with a small amount of training samples on three different datasets show that the proposed MSGCN-CRF can obtain better classification results than the seven state-of-the-art classification methods.