A block structure Laplacian for hyperspectral image data clustering

Abstract

Over the past decade, the problem of hyperspectral data clustering has generated a growing interest from various fields including the machine learning community. This paper presents an analysis of the traditional spectral clustering approach and points to new directions that boost unsupervised pattern classification. In particular, the paper offers design insights on the generation of a well structured graph Laplacian based on an affinity function that induces context-dependence to create compact neighborhoods. A novel bilateral-kernel (affinity) function exploits the spatial information to generate a diagonal-block structured Laplacian. Experimental validations through the analysis of eigenvalues and eigenvectors demonstrate the benefits of seeking block structured affinities in hyperspectral image clustering and visualization.