Hyperspectral Feature Extraction by Tensor Modeling and Intrinsic Decomposition

Abstract Recently, hyperspectral tensor modeling has been assessed by its capacity of determining more compact as well as by its useful intrinsic data representation. In this paper, to enhance the hyperspectral tensor data representation and to eliminate non-relevant spatial datum, we integrated the intrinsic decomposition (ID) as a pre-processing step. The suggested approach acts in agreement with the joint use of spectral and spatial features provided in hyperspectral scenes, and it incorporates more usefully with the spatial data in the dimensional reduction step. The suggested framework consists of three steps: firstly, the intrinsic decomposition is employed to remove useless spatial data from hyperspectral image (HSI). Secondly, after modelling ID results as a tensor structure, the tensor principal component analysis is used to reduce tensorial data redundancy. Finally, we evaluated the proposed approach during classification tasks using real hyperspectral data sets. Compared to other methods, experiment results have proved that our approach can pave the way for the best classification accuracy.