Spectral clustering eigenvector selection of hyperspectral image based on the coincidence degree of data distribution

Abstract

ABSTRACT Spectral clustering is a well-regarded subspace clustering algorithm that exhibits outstanding performance in hyperspectral image classification through eigenvalue decomposition of the Laplacian matrix. However, its classification accuracy is severely limited by the selected eigenvectors, and the commonly used eigenvectors not only fail to guarantee the inclusion of detailed discriminative information, but also have high computational complexity. To address these challenges, we proposed an intuitive eigenvector selection method based on the coincidence degree of data distribution (CDES). First, the clustering result of improved k-means, which can well reflect the spatial distribution of various types was used as the reference map. Then, the adjusted Rand index and adjusted mutual information were calculated to assess the data distribution consistency between each eigenvector and the reference map. Finally, the eigenvectors with high coincidence degrees were selected for clustering. A case study on hyperspectral mineral mapping demonstrated that the mapping accuracies of CDES are approximately 56.3%, 15.5%, and 10.5% higher than those of the commonly used top, high entropy, and high relevance eigenvectors, and CDES can save more than 99% of the eigenvector selection time. Especially, due to the unsupervised nature of k-means, CDES provides a novel solution for autonomous feature selection of hyperspectral images.