Hyperspectral image classification using non-negative tensor factorization and multinomial logistic regression

摘要

We aim to classify every pixel of a hyperspectral image. Toward this goal, we first decompose the image data. The three-dimensional (3-D) tensor of an image cube is decomposed to the spectral signatures and abundance matrix using non-negative tensor factorization (NTF) methods. In contrast to the matrix factorization where the pixels' spectra are stacked in columns of the data matrix, the NTF techniques preserve the spatial structure of the image data. Therefore, the obtained abundance maps provide discriminant spatial features for classification. Morphological attribute profiles are also computed for abundance maps and their effect on the classification performance is studied. Using the original spectral image and the obtained spatial features of the abundance matrices, we construct a composite kernel framework. We apply a multinomial logistic regression classifier to the kernels. Experiments show that, in terms of the classification accuracy, the proposed feature sets acquired through NTF techniques can lead to a better performance compared to the principal component analysis and non-negative matrix factorization feature sets. (C) 2019 Society of Photo-Optical Instrumentation Engineers (SPIE)