Are machine learning based methods more robust to poor image quality when performing hyperspectral imagery anomaly detection compared with conventional signal-processing methods? : A comparative discussion of the impact of image quality on model performance

Thesis

Download PDF

Creator

Rights statement

Awarding institution

Date of award

Thesis identifier

Person Identifier (Local)

Qualification Level

Qualification Name

Department, School or Faculty

Abstract

Hyperspectral imaging (HSI) is being increasingly used for the detection of anomalies in various applications, for example, remote sensing for environmental monitoring. However, the quality of HSI sensors can vary significantly due to noise and resolution, which effect the accuracy of anomaly detection methods applied to the data. To address this, this thesis evaluates how different levels of image degradation and enhancement affect the detection capability of a traditional signal processing algorithm (RX detection), and a Machine Learning (ML) based approach (AETNet). Following analysis of the utility of both ML and conventional models, a series of novel experiments were developed to understand the effect of varying image qualities. The results of these experiments demonstrate that image quality affects detection performance at varying levels, depending upon which technique is being used. ML techniques demonstrated slightly higher performance when compared to conventional methods suggesting they may be less degraded by image noise. These findings demonstrate the importance of considering image quality when developing and applying HSI anomaly detection techniques and the approaches that can be taken to ensure optimal pre-processing to achieve the best results.

Advisor / supervisor

Resource Type

DOI

Funder

Relations

Thumbnail	Title	Date Uploaded	Visibility	Actions
	PDF of thesis T17983	2026-04-30	Public	Download