Multi-feature hyperspectral image classification with L_2,1 norm constrained joint sparse representation

Abstract

This paper proposes a novel multi-feature hyperspectral image (HSI) classification framework that utilizes joint sparse representation (JSR) to combine pixel-wise and superpixel-wise features (SMFSR). In this framework, a multi-feature sparse representation algorithm is proposed to exploit different kinds of pixel-wise features. In the meantime, Entropy rate segmentation is utilized to acquire HSI superpixels, which can get harmonious neighbourhood and distinct boundary. SMFSR combines two types of spatial information and is trained for HSI classification. A new solution for SMFSR is proposed which can convert the NP-hard problem of JSR to a convex optimization one. Experimental results on well-known hyperspectral data sets demonstrate that the proposed SMFSR outperforms other commonly used methods.