Local Graph Convolutional Network for Hyperspectral Target Detection

A graph convolutional network (GCD) is suggested for hyperspectral target detection in this paper. To extract the spatial neighborhood information in each local region of the hyperspectral image, to extract latent spectral relationships among neighbors and to explore the irregular class boundaries corresponding to each local patch centered at the pixel under test, a small local graph is constructed for each pixel. After several graph convolutional blocks for exploration of spectral-spatial latent dependencies, the label of each pixel is predicted in the last layer. The GCD with composing a local graph correspond to each pixel and its spatial neighbors can well enhance discrimination among target and background, which leads to significant improvement in the detection results. The experimental results show superiority of GCN compared to convolutional neural network (CNN) and several other hyperspectral target detectors.