New data analysis and dimensionality reduction methods for hyperspectral imagery

Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2022
Thesis identifier
  • T16336
Person Identifier (Local)
  • 201860303
Qualification Level
Qualification Name
Department, School or Faculty
  • Hyperspectral data contains rich spectral information and so have become very useful in data classification. However, hyperspectral data contains several spectral bands (usually in hundreds) which bring about curse of dimensionality and limits its potential in classification applications. With a focus on addressing this problem, this thesis applies Linear Discriminant Analysis (LDA) for hyperspectral data dimensionality reduction and proposes novel extensions of LDA. LDA is a supervised technique which can reduce the number of dimensions in data. One problem with LDA is that the number of features it can produce is limited to c−1 where c is the number of classes in the data. Also, LDA gives sub-optimal performance when applied on small training samples, which limits its use on hyperspectral data since such data does not always contain enough samples for training. Firstly, this thesis applies LDA on spectral features extracted from hyperspectral data to reduce its dimensionality and combines the LDA outputs with spatial features from RGB images. Comparative performance analysis of LDA and PCA is also performed. Results show that LDA can perform better than PCA and that combining spectral features with spatial features from RGB images can improve performance of classification models. Secondly, Folded LDA (F-LDA), a novel extension of LDA, is proposed for hyperspectral data dimensionality reduction. F-LDA is based on a mathematical ‘trick’ (folding the pixels) which was inspired by previous work to extend PCA using a similar innovative step. Results show that F-LDA achieves higher accuracy than LDA and other state-of-the-art methods when applied on small training samples. When compared with LDA, F-LDA achieves reduction in computational complexity and memory requirement, and can extract many more features. Finally, F-LDA is applied on optimal spectral features selected by Genetic Algorithms. Results show that a novel combination of GA and F-LDA can achieve further reduction in computational complexity and memory requirement in certain applications.
Advisor / supervisor
  • Murray, Paul
  • Zabalza, Jaime
Resource Type