Union of random subspace-based group sparse representation for hyperspectral imagery classification

Abstract

ABSTRACTThis paper proposes a new spectral-spatial hyperspectral imagery classification approach, integrating union of random subspace projection (URsub) with group sparse representation (GSR). The URsub is to project the scene to a lower-dimensional subspace which is spanned by groups of vectors randomly selected within each class, while the GSR classifier (GSRC) is to represent the testing set as a sparse linear combination of group-structure labelled samples. The proposed approach, abbreviated as GSRCusub (URsub-based GSRC), is evaluated by two real hyperspectral data sets. Experimental results demonstrate that it provides better characterization of spectral features and consideration of spatial coherence, and brings steady improvement to the other related methods.