Wheeler, Bradley J.
(2025)
Adaptive ensemble learning for anomaly detection in hyperspectral imaging.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Hyperspectral images capture detailed electromagnetic radiation across ultraviolet, visible, and infrared wavelengths, providing critical data for remote sensing tasks where visible light alone is insufficient. These images have applications in domains such as agriculture, surveillance, disaster recovery, and environmental monitoring. A key challenge in leveraging hyperspectral data is anomaly detection, which aims to identify spectral signatures that differ from the background. Hyperspectral anomaly detection (HAD) has been the focus of extensive research, resulting in a variety of algorithms designed to exploit the rich information within hyperspectral images.
Most HAD algorithms stem from a limited set of modeling biases, leading to three key challenges: selection bias, performance disparities, and singular modeling bias. The restricted range of modeling biases results in correlations between datasets and algorithms, making it easier for certain algorithms to perform better on specific datasets. This creates a risk of skewed results if the datasets are not selected carefully, providing an inherent advantage to some algorithms. Performance disparities exist where some algorithms excel on individual datasets while others generalize better across multiple datasets. Lastly, using a singular modeling bias can limit model flexibility, leading to issues such as underfitting or overfitting depending on the scenario.
To address these challenges, this thesis focuses on three main research tasks. First, I develop a framework to identify significant correlations between HAD modeling biases and datasets. This framework helps predict which biases are likely to perform best on a given dataset, reducing selection bias when choosing datasets and algorithms. Next, I design an adaptive ensemble learning algorithm that integrates multiple HAD modeling biases. This ensemble approach bridges the gap between specialized and generalized performance by combining diverse biases, thus reducing disparities. Finally, I conduct a systematic study of how different error quantification methods in ensemble learning influence the contribution of each modeling bias to the final solution. This study provides valuable insights into the utility of various modeling biases across different datasets.
Together, these contributions highlight the importance of incorporating diverse modeling biases in HAD and demonstrate how ensemble learning can effectively integrate them for better performance across hyperspectral datasets.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
7 January 2025 |
| Date Type: |
Publication |
| Defense Date: |
5 November 2024 |
| Approval Date: |
7 January 2025 |
| Submission Date: |
22 November 2024 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
127 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Computing and Information > Information Science |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
anomaly detection
ensemble learning
adaptive learning
hyperspectral imaging |
| Date Deposited: |
07 Jan 2025 19:38 |
| Last Modified: |
07 Jan 2025 19:38 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/47126 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}