Hyperspectral Image Classification using Band-Group Non-negative Tensor Factorization

In this paper, we propose a classification framework for 3D hyperspectral data. Discriminative features are extracted through applying Non-negative Tensor Factorization (NTF) technique to the image tensor. The factorized components indicate the spectral signatures and 2D abundance maps of the constituent materials. We use a composite kernel Multinomial Logistic Regression (MLR) classifier. The obtained abundance matrices for training samples, constitute the spatial features which are fed to the classifier. Applying NTF, the spatial structure of the image is preserved in contrast to matrix factorization methods. We have also analyzed the effect of exploiting band group NTF; Instead of decomposing the image over whole spectral bands, we split the spectra into several band groups and apply the NTF algorithm to each sub-band. The abundance maps obtained for these band groups construct the spatial features. The original image cube makes the spectral features. The spatial and spectral kernels are acquired and stacked to form the training feature vector. This way, both spectral properties and spatial structure are effectively exploited to achieve higher classification accuracy. The experiments are performed on a widely studied hyperspectral dataset. Superior classification performance is attained using the proposed training features compared to NMF. We also compare the MLR with SVM classifier.