Unsupervised Feature Extraction for Hyperspectral Imagery Using Collaboration-Competition Graph

Abstract

Signal processing on graph offers the ability to define relationships of high-dimensional data on graph. In this paper, an unsupervised feature extraction method using graph for hyperspectral imagery is proposed, which incorporates collaborative representation using $\ell _2$ -norm regularization with locality constrained property into graph construction, named collaboration-competition preserving graph embedding. First, an undirected and weighted graph is constructed to exploit the data structure. Then, a weight matrix of edge in graph is built by formulating the combined collaborative-competitive representation into a convex optimization problem. The constructed graph is expected to reveal local intrinsic manifold and global geometry information of hyperspectral data. The superiority of the proposed graph-based unsupervised feature extraction method, compared with other traditional and state-of-the-art methods, is demonstrated by verifying the classification accuracy on four typical hyperspectral datasets.