Hyperspectral image classification with a class-dependent spatial-spectral mixed metric

In this paper, we propose a class-dependent spatial-spectral mixed metric (cdSSMM) hyperspectral image (HSI) classification method, which combines the joint sparse representation USR) and a spectral mixed metric (SMM) for the purpose of exploiting both contextual correlation and spectral relationship within superpixel. The SMM is designed by combining the spectral information divergence (SID) and the spectral angle mapper (SAM) to exploit spectral discriminability of the two spectral measures. Specifically, a superpixel map is first generated, which is used for extracting the spatial information in an unfixed local region. Then, the SMMs are calculated separately by two models [i.e., SID multiplied by the tangent of SAM (SMMtg) and SID multiplied by the sine of SAM (SMMsn)] in a superpixel region. Next, the residuals between test and training samples are calculated by the JSR model. Finally, the results of SMM and the residuals obtained by JSR are combined to discriminate the class of each test sample based on a unified class membership function. Experimental results on two typical HSI datasets demonstrate that the proposed method can obtain better classification results compared with several well-known classification methods. (C) 2019 Elsevier B.V. All rights reserved.