Gram Angle Field-Based Siamese Graph Convolutional Neural Network for Hyperspectral Images Classification

Previous studies have shown that hyperspectral image classification methods based on deep learning can significantly improve the classification performance due to the effective fusion of spectral spatial information. However, it remains a challenging problem in the field of remote sensing to develop new technology to extract and fuse the depth features in images more effectively to obtain satisfactory classification results. In this work, a gram angle field-based Siamese graph convolutional neural network (GAF-SGCN) is proposed for hyperspectral image classification. The proposed method first adopts gram angle field (GAF) method to extract the deep feature of each pixel and enhance the distinguishability between pixels. Then superpixel technology is used to fuse spectral-spatial information and reduce the number of node processing. Finally, Siamese graph convolutional neural network is adopted to extract the depth spectral-spatial features of the fused images. The experimental results show that, compared to several competing algorithms, the classification accuracy on Indian pines (IP) and Salinas datasets can be improved by up to 21.93% and 11.34%, respectively.