First Advisor

John Lipor

Term of Graduation

Fall 2020

Date of Publication

12-7-2020

Document Type

Thesis

Degree Name

Master of Science (M.S.) in Electrical and Computer Engineering

Department

Electrical and Computer Engineering

Language

English

Subjects

Hyperspectral imaging, Algorithms, Machine learning, Signal processing

DOI

10.15760/etd.7514

Physical Description

1 online resource (x, 71 pages)

Abstract

Aerial target detection is often used to search for relatively small things over large areas of land. Depending on the size and signature of the target, detection can be a very easy or very difficult task. By capturing images with several hundred color channels, hyperspectral sensors provide a new way of looking at this task, both literally and figuratively. Hyperspectral sensors can be used in many aerial target detection tasks such as identifying unhealthy trees in a forest, searching for minerals at a mining site, or finding the sources of chemical leaks at a factory. The high spectral resolution of hyperspectral imagery makes it well suited for these tasks, but the inherent high dimensionality of these images poses a unique set of challenges.

The motivation of this work is to investigate the use of data clustering to improve our ability to detect targets within hyperspectral images. Target detection algorithms operate by identifying locations that are likely to contain a target when compared with the background. We propose a new clustering-based target detection method that allows multiple background models to be used. This new method pairs a clustering algorithm with an array of spectral matched filters. We then analyze the performance of various clustering algorithms when used with this method to detect targets in aerial hyperspectral images.

We evaluate the performance of our clustered target detector on several aerial hyperspectral images when using clusters generated by several popular algorithms, specifically k-means, spectral clustering, Gaussian mixture models, and two variants of subspace clustering. We show empirically that clusters generated by Gaussian mixture models provide the best performance, obtaining a pAUC score of 0.192 in the true positive detection rate on the RIT Radiance image for false positive rates of 1% or less, providing over a 12-fold increase when compared to the pAUC score of 0.0148 obtained for target detection without clustering. We then tune a Laplacian-regularized Gaussian mixture model (LapGMM) algorithm specifically for the task of aerial hyperspectral target detection. We show empirically that our tuned algorithm outperforms all others when used for this task, outpacing the traditional Gaussian mixture model with a pAUC score of 0.219 for the same case above, thereby offering over a 14-fold improvement in performance. We offer several hypotheses to explain these results. We then discuss some of the features, most notably the versatility provided by the regularizer, that make the tuned LapGMM algorithm well suited for this application.

Considering future work, we propose a number of potential applications for our tuned LapGMM algorithm, as well as several potential improvements or modifications to the clustered target detector that may be worth further investigation. The contributions of this thesis are a detailed investigation and analysis of the use of clustering algorithms when used for target detection, and an analysis of the performance of several clustering algorithms when used in an aerial hyperspectral image application. Additionally, we contribute an algorithm tuned specifically for clustering aerial hyperspectral images, which to the best of our knowledge is state of the art.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Persistent Identifier

https://archives.pdx.edu/ds/psu/34674

Available for download on Tuesday, December 07, 2021

Share

COinS