Dimensionality reduction of hyperspectral imagery with sparse and collaborative graphs

Abstract

Hyperspectral image dimensionality reduction with graph-based approaches is considered. With available labeled samples, a graph can be formed with these samples by constructing an affinity matrix through their sparse or collaborative representations. In addition, sparse or collaborative representation can be done using within-class samples, resulting in block-sparse representation, although within each block the representation can be either sparse or non-sparse (collaborative). The experimental results show that the block-sparse plus within-block-collaborative representation can yield the best performance.