Dimensionality reduction and coloured noise removal from hyperspectral images

Abstract

Hyperspectral image (HSI) classification requires spectral dimensionality reduction and noise reduction. While common dimensionality reduction (DR) and denoising methods are based on linear algebra, we propose a multilinear algebra method to jointly achieve denoising reduction and DR. Multilinear tools consider multidimensional data as whole entity by processing jointly spatial and spectral ways. However, it cannot cope with the HSIs distorted by non-white noise which is the most realistic case and cannot preserve rare signals. First, we propose a new method for whitening the noise (W) in HSI. Then we propose a method based on multidimensional wavelet packet transform (MWPT) and multiway Wiener filter (MWF) which performs both non-white noise and spectral DR, referred to as W-MWPT-MWF_dr-(). The Classification algorithm support vector (SVM) machines are applied to the output of the following DR and noise reduction methods to compare their efficiency: The proposed W-MWPT-MWF_dr-(); prewhitening method associated with MWF (PMWF), principal component analysis (PCA_dr), minimum noise fraction (MNF_dr), PCA_dr associated with Wiener filtering (PCA_dr-Wiener), and MNF_dr associated with Wiener filtering (MNF_dr-Wiener).

Acknowledgement

The authors would like to thank the anonymous referees for their careful reading and helpful comments which improved the quality of this letter.