GRAPH-BASED MULTI-VIEW LEARNING FOR HYPERSPECTRAL REMOTE SENSING IMAGE CLASSIFICATION

Hyperspectral imaging has been widely used in the remote sensing due to its ability of capturing increasingly rich spectral information. However, it is still a hard work to get satisfied interpretation accuracies when dealing with such increased spectral/spatial resolution images. To take advantages of the spectral signatures and spatial information contained in hyperspectral image (HSI), this paper proposes an efficient graph based multi-view clustering for unsupervised HSI classification. The key idea of our proposed algorithm is to exploit the mutual agreement of both spectral and spatial information to obtain better classification performance than using any single view data. Firstly, spatial features and spectral features are extracted separately and the spatial features and spectral features are considered as two different views about the HSI. Secondly, with the spatial and spectral view representations, a graph-based multiview clustering algorithm is designed to generate the cluster labels since the two views admit the same underlying cluster structure of the HSI. To evaluate the performance of the proposed method, we conduct experiments on two benchmark datasets and compare with six state-of-the-art approaches. The experimental results confirm the effectiveness of our proposed method.