Tensor partial least squares for hyperspectral image classification

Abstract

A hyperspectral image is classically a three-way (or tensor) block of data. In order to extract information from it, it has to be classified using image classifiers. Since classifiers are traditionally two-way classifiers, the hyperspectral data is unfolded into a two-way data. Processing the hyperspectral data as a two-way data often reduces the accuracy of the classification. This research explores the novel application of Tensor Partial Least Squares (TPLS) for hyperspectral image classification. TPLS has been proven to be more robust than the two-way PLS. Unlike the two-way classifiers, the TPLS utilises the hyperspectral data as a three-way (tensor) data. Two hyperspectral images of Indian Pines region in Northwest Indiana, USA and University of Pavia, Italy are used as test beds for the experiment. The results extracted by the model are the $X$ loadings, $Y$ loadings, $X$ scores, and $Y$ scores. The computed training mean r square values for Indian Pines and University of Pavia are $0.9061\pm 0.74,$ and $0.9155\pm 0.63$ respectively. The results of the experiment show that the TPLS performed better than the unfolded PLS, but fell short of the notable traditional classifiers.

Keywords:

• Remote sensing
• hyperspectral imagery
• tensor partial least squares
• partial least squares

Acknowledgement

This study was part of a series of preliminary machine learning experiment that was conceptualized and initiated by Dr Christopher E. Ndehedehe based on his Australian Research Council (ARC) proposal for the 2023 Discovery Early Career Researcher Award (DECRA). Christopher is grateful to the ARC for funding his DECRA project (DE230101327) and to Griffith University's Australian Rivers Institute for the conducive research environment supporting his strategic research in remote sensing of the environment and where the ARC DECRA is hosted.

Disclosure statement

No potential conflict of interest was reported by the authors.