LAUSR

Hyperspectral remote sensing images (HRSIs) contain rich spectral information, and an HRSIs feature extraction method, i.e., higher-order spectra linear discriminant analysis (HOS-LDA) subspace method, is presented, which demonstrates that the calculation of HOS is a kernel mapping with the inner product of HOS as the kernel function and uses linear discriminant analysis (LDA) to extract the HOS feature (HOSF) from the HOS of the spectral pixels of HRSIs according to a presented strategy of kernel space methods. To verify the efficiency of the deep feature of HOSF, a two-branch convolutional neural network (CNN), i.e., Two-CNNHOSF-spa, is designed with each branch fed with HOSF and spatial feature, respectively. A 3D-CNN, i.e., spatial-higher-order spectra-spectral (spa-hos-spe) 3D-CNN, is designed, which is fed with the spa-hos-spe feature to evaluate the deep joint spa-hos-spe feature. The experimental results of three measured HRSIs show that: the presented HOS-LDA method has improved the classification rates compared with the LDA, spectral regression discriminant analysis (SRDA), kernel LDA (KLDA) methods by using support vector machine (SVM), Bayesian, and ${K}$ -nearest neighbor ( ${K}$ -NN) classifiers, which shows the efficiency of the presented HOSF; the presented Two-CNNHOSF-spa outperforms the two-branch CNN fed with spectral and spatial features and the two-branch CNNs fed with LDA, SRDA, and KLDA extracted feature and spatial feature, which verifies the effectiveness of the deep HOSF; the presented spa-hos-spe 3D-CNN outperforms the 3D-CNN fed with spatial-spectral (spa-spe) feature, which shows that the deep spa-hos-spe feature is better than the deep spa-spe feature.