Joint Group Sparse Collaborative Representation for Hyperspectral Image Classification

Abstract

Collaborative representation (CR) has attracted great interest in hyperspectal imagery (HSI) classification because of its efficiency. However, existing CR-based classifiers ignore the group structure characteristics among the training pixels. In this paper, a group sparse CR with Tikhonov regularization (GSCRT) classifier is proposed to consider the group prior information. In order to fully utilize both spatial and spectral information, we further propose joint GSCRT (JGSCRT) based on the idea that pixels belonging to the same class in the neighboring region should have similar group sparse constraint. Considering the limitations of traditional class decision based on the reconstruction error of a single pixel, the introduction of local decision rule can improve the overall classification accuracy by reducing the misjudgment of pixels within the class. The experimental results on University of Pavia dataset show that the proposed methods outperform other CR-based classifiers.