Thesis

Multi-modality feature fusion and unsupervised hyperspectral band selection for effective classification of remote sensing images

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2020
Thesis identifier
  • T15622
Person Identifier (Local)
  • 201782280
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • In recent years, Hyperspectral image (HSI) has been widely applied in a range of applications due to its contained rich spectral information. As a fundamental topic in HSI analysis, HSI classification has attracted increasing attention. An effective classification algorithm without too much computational cost is always desired, especially under the circumstance of insufficient training samples. As a result, this thesis aims to design and implement novel techniques to reduce the high dimensionality of the HSI data and improve the classification performance with limited training samples.;In this thesis, first, a superpixel-based feature specific sparse representation framework (SPFS-SRC) is proposed for spectral-spatial classification of HSI at superpixel level, which can improve the classification performance with less training samples and better efficacy. The proposed online learning strategy can better reflect the effect of each extracted feature. Second, a superpixel-based multiple feature fusion framework has been developed to generate an effective fused feature with a reduced dimension.;In addition, two novel methods are also proposed for unsupervised band selection for dimensionality reduction in HSI. First, an adaptive distance enabled tree-based band hierarchy framework (ADBH) has been developed to obtain desired band subset of the HSI, which can help to avoid the noisy bands. With the proposed tree hierarchy-based framework, any number of band subset can be acquired. By introducing a novel adaptive distance into the hierarchy, the similarity between bands and band groups can be computed straightforward whilst reducing the effect of noisy bands.;Furthermore, a deep learning-based framework has been designed to determine the optimal band subset by utilizing the concrete autoencoder (CAE). The band subset with the most information can be chosen as the desired result. For performance evaluation, several remote sensing HSI datasets have been utilized to evaluate the proposed algorithms, where improved performance has proved the superiority of proposed methodologies.;In summary, the outcome of this thesis make contributions in the HSI community by proposing two multi-modality feature fusion algorithms and two unsupervised band selection methods for the effective dimensionality reduction and data classification in HSI, the novelty and robustness of the proposed technologies have been fully demonstrated by extensive experiments. Relevant approaches also have great potential to be applied in other signal and image analysis tasks,especially dimensionality reduction, data fusion and data classification.
Advisor / supervisor
  • Ren, Jinchang
  • Marshall, Stephen
Resource Type
DOI
Date Created
  • 2020
Former identifier
  • 9912881691802996
Embargo Note
  • The electronic version of this thesis is currently under moratorium due to a licensing issue. If you are the author of this thesis, please contact the Library to resolve this issue.

Relazioni

Articoli