Shape-Adaptive Tensor Factorization Model for Dimensionality Reduction of Hyperspectral Images

Zhaohui Xue,Mengxue Zhang,Sirui Yang

doi:10.1109/access.2019.2935496

Abstract

Tensor-based dimensionality reduction (DR) of hyperspectral images is a promising research topic. However, patch-based tensorization usually adopts a squared neighborhood with fixed window size, which may be inaccurate in modeling the local spatial information in a hyperspectral image scene. In this work, we propose a novel shape-adaptive tensor factorization (SATF) model for dimensionality reduction and classification of hyperspectral images. Firstly, shape-adaptive patch features are extracted to build fourth-order tensors. Secondly, multilinear singular value decomposition (MLSVD) is adopted for tensor factorization and latent features are extracted via mode-i tensor-matrix product. Finally, classification is conducted by using a sparse multinomial logistic regression (SMLR) model. Experimental results, conducted with two popular hyperspectral data sets collected over the Indian Pines and the University of Pavia, respectively, indicate that the proposed method outperforms the other traditional and tensor-based DR methods.

Highlights

Hyperspectral remote sensing sensors are capable of providing land cover images with unified spectral-spatial information, which has been motivating rapid developments for hyperspectral image (HSI) processing techniques [1], [2]
HYPERSPECTRAL DATA SETS 1) The first hyperspectral image was acquired by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) sensor over the Indian Pines region in Northwestern Indiana in 1992
2) The second hyperspectral image was acquired by the Reflective Optics Spectrographic Imaging System (ROSIS) sensor over the urban area of the University of Pavia, Italy

Summary

Introduction

Hyperspectral remote sensing sensors are capable of providing land cover images with unified spectral-spatial information, which has been motivating rapid developments for hyperspectral image (HSI) processing techniques [1], [2]. HSI classification has attracted plenty of attention in the last decades [3]–[9]. The curse of dimensionality has posed great challenges for HSI classification since there is a high correlation between adjacent bands and the dimension of spectral features may be too high for classification purpose [10]. Dimensionality reduction (DR) methods have been commonly used to address this issue, and various DR approaches have been proposed in the literature [11]–[13]. Traditional popular DR methods are matrix-based, including principal component analysis (PCA) [14], local linear embedding (LLE) [15], isometric feature mapping

Methods

Results

Conclusion