< Back to previous page

Publication

Hyperspectral image classification using Non-negative Tensor Factorization and 3D Convolutional Neural Networks

Journal Contribution - Journal Article

© 2019 Elsevier B.V. In this paper, we address the task of hyperspectral image classification using a 3-D Convolutional Neural Network (CNN). Instead of commonly used raw spectral features, discriminative features are obtained by factorizing the hyperspectral data with Non-negative Tensor Factorization (NTF). The factors consist of the endmembers’ spectral signature matrix as well as the abundance matrix. The acquired abundance maps are exploited to extract the feature vectors representing the spatio-spectral properties of the image. Unlike Non-negative Matrix Factorization (NMF) where the pixels’ spectra are stacked in columns of the data matrix, the spatial information is preserved with NTF. Morphological attribute filters are also applied to the extracted abundance maps and construct the discriminative training features that are fed to a 3D CNN. We develop and optimize a 3D CNN framework for image classification. Using joint spatio-spectral features and parameter sharing, a 3D CNN yields promising classification performance. It is shown in the experiments that the proposed feature sets can lead to better results in terms of classification accuracy compared to both raw spatio-spectral data and to NMF. We also demonstrate the effectiveness of the proposed CNN by comparing the results with a SVM classifier.
Journal: Signal Processing. Image Communication
ISSN: 0923-5965
Volume: 76
Pages: 178 - 185
Publication year:2019
BOF-keylabel:yes
IOF-keylabel:yes
BOF-publication weight:1
CSS-citation score:2
Authors:International
Authors from:Higher Education
Accessibility:Open