Local Matrix Stack Graph Convolutional Networks for Classification

Abstract

In order to make full use of the spectral and spatial information of hyperspectral images, a spatial spectral joint composition method based on nuclear spectral Angle method is designed in this paper. Graph Convolutional Networks(GCN) does not use regular convolution kernels for convolution, so it can adaptively capture geometric changes of different object regions in hyperspectral images. However, for the construction of adjacency matrix and the determination of graph structure, the traditional graph convolution network method needs very high computational cost. To solve the above problems, this paper developed a Graph Convolutional Networks based on local pixel discrimination (LS-GCN), which can predict the whole image according to part of the sampled pixels and their neighborhoods, reducing the time complexity of the composition process by an order of magnitude and improving the image recognition rate to a certain extent.