Hyperspectral image classification using spectral mixing metrics representation

ABSTRACT

This letter presents new classification methods that effectively fuse the spectral mixing metrics and joint sparse representation (JSR) (SMMR) for hyperspectral image (HSI), in which SMMR combines the spectral information divergence (SID) with transformed spectral angle mapper (SAM), namely that SID is multiplied by the tangent of the SAM (SMMRtg) and the sine of SAM (SMMRsn), respectively. The proposed methods SMMRtg and SMMRsn mainly include the following steps. First, the SID$\times$tan(SAM) and SID$\times$sin(SAM) of the spectral bands among test and training samples are calculated that respectively depend on two different function versions. Second, the JSR model is utilized to obtain the representation residuals of different pixels for the test set. Third, a regularization parameter $\lambda$ is employed to achieve a balance in SMMR. Finally, the label of the test pixel is determined by the eventual decision function of two proposed methods. Experimental results on the Indian Pines and Washington DC datasets demonstrate that the proposed methods achieve superior performance compared with several previous methods.

Acknowledgments

The authors would like to thank the Editor-in-chief, the anonymous Associate Editor and the reviewers for their insightful comments and suggestions that have improved this letter.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [Grant No. 51704115] and by the Science and Technology Program of Hunan Province [Grant No. 2016TP1021].