Comparison of Local Kernel and Covariance Matrix Descriptors for Spatial-Spectral Classification of Hyperspectral Images

Hyperspectral sensors collect information from the earth's surface in the form of images with a large number of electromagnetic bands. Accurate classification of hyperspectral images has been one of the hot topics in remote sensing. Spatial information as a complementary source for spectral information helps increase the classification accuracy of hyperspectral images (HSI). Local covariance matrix descriptor (LCMD) is the new spatial-spectral feature generation method for HSI classification. Although the LCMD is easy to use and performs well in HSI classification, it has some limitations, such as discarding the nonlinear relationships between features, which are useful in HSI classification. To address these issues, we propose a local kernel matrix descriptor (LKMD) for the classification of HSIs. In this study, the performance of LCMD is compared with LKMD with two widely used kernels, RBF and polynomial, and final classification results on two real HSIs, Indian Pines and Pavia University, proved the superiority of LKMD over LCMD.