Hyperspectral Band Selection with the N-dimensional Spectral Solid Angle and Its Utilization in the Discrimination of Spectrally Similar Targets

• Author / Creator

  Long, Yaqian

• The discrimination of earth surface materials using hyperspectral sensing can be facilitated by selecting a subset of spectral bands that focuses on essential features. Materials in detailed classes such as rock types and tree species often present great spectral similarity, producing challenges for band selection. A method named the N-dimensional Solid Spectral Angle (NSSA) was proposed to select the most dissimilar spectral regions amongst targets for their maximum spectral separation, however, the use and performance of this method in practical application needed to be explored.
  In chapter 2, the NSSA method was applied to two real datasets of geologic relevance to establish guidelines for the selection of parameters that will allow non-expert users to exploit this method. This study demonstrated that the NSSA method is a robust tool for feature identification, since bands selected from the two datasets not only captured absorption feature position, and shape and depth, but also showed improved class separation.
  In chapter 3, the NSSA method was applied in a hierarchical manner to address the inter- and intra-class variability among materials. Two datasets were analyzed, including airborne image endmembers for geological mapping and leaf spectra for tree species discrimination. Bands were separately selected from different hierarchies of those categorized materials using the NSSA and combined into a single band set. The agreement between bands selected by the hierarchical strategy and by experts suggested that the hierarchical band selection using the NSSA method is both practical and effective in addressing the spectral variability.
  In chapter 4, an ensemble of multiple band selection methods encompassing random forest, minimum redundancy maximum relevance, and the NSSA was used to select and characterize longwave infrared features of leaves for the discrimination of tree species which display great spectral similarity. The selected features could be related to leaf constitutional compounds such as cellulose and oleanolic acid. Meanwhile, the band selection improved the classification using a regularized logistic regression method by 3%. These results can be useful to future image mapping of tree species at large scales. The ensemble strategy was recommended for the band analysis of vegetation.
  Chapter 5 proposed a strategy that simultaneously employed band selection and endmember selection by incorporating the NSSA into the Spatial Spectral Endmember Selection (SSEE) method in order to select bands that enhance the spectral contrast of endmembers and hence improve the estimation of fractional abundances from hyperspectral images. The detailed methodology was described and an airborne image that was acquired for the mapping of mafic and ultramafic rocks was used to evaluate the proposed method. The results showed that the integration of NSSA and SSEE automates band selection in spectral mixture analysis and reduces the efforts in field investigations for feature identification.
  The results of this thesis demonstrated that the NSSA method, whether used in a hierarchical manner or integrated with other methods, was robust in the analysis of spectral libraries collected from field samples or from hyperspectral imagery collected from laboratory or airborne imaging systems. Its effectiveness also spans the visible near-infrared, shortwave infrared to the thermal infrared range of the data. The band selection results were evaluated by both classification performance and the physical meaning of spectral features, which balanced the need for high accuracies in statistical learning algorithms and application significance highlighted by remote sensing experts. The proposed method, guidelines, and experimental designs provided in this thesis contribute in identifying meaningful features from data encompassing a small number of labeled samples for the discrimination of spectrally similar material in a variety of fields including geology, ecology, urban and agriculture.

• Subjects / Keywords

  • hyperspectral band selection

• Graduation date

  Fall 2019

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-4ap0-ye51

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.