Sparse Representation Classification on Hyperspectral Imagery BasedMultiple Measurement Vectors

Abstract:

Traditional sparse representation based classifiers ignore interconnections among pixels and have high computational complexity when applied in hyperspectral imagery (HSI) field. Therefore, a multiple measurement vectors based sparse representation classifier (MMVSRC) model is proposed to solve the above problems. The model introduces a balance parameter to control the sparsity of coefficient vectors, and estimates sparse coefficient vectors of all testing pixels by minimizing reconstruction errors using the L2norm constraint. Experiments on two HSI datasets are implemented to test the performance of MMVSRC, and the results are compared with those of five stateoftheart classifiers. The results show that MMVSRC achieves best classification accuracies among all whereas taking the second shortest computational time.

Key words:multiple measurement vectors; sparse representation; hyperspectral imagery; classification