Investigation of Target Detection in Hyperspectral Images for Aided Target Recognition

Author

• Mazharul Hossain

Date

2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

Committee Chair

Lan Wang

Committee Member

Aaron Robinson

Committee Member

Chrysanthe Preza

Committee Member

Deepak Venugopal

Committee Member

Xiaolei Huang

Abstract

Aided/Automatic Target Recognition (AiTR/ATR) enhances human decision-making by improving the detection, localization, and interpretation of targets in sensor imagery. Hyperspectral (HS) imaging provides rich spectral information for AiTR that can significantly improve target detection; however, HS detectors often perform unreliably in real-world conditions due to high data dimensionality, limited labeled training data, class imbalance, sensitivity to environmental and sensor variations, and a shortage of pre-trained models. These challenges frequently lead to degraded performance when models trained in one setting are deployed in different scenes, sensors, or operating conditions. The overarching goal of this dissertation is to make hyperspectral target detection for AiTR systems reliable under real-world conditions. This work focuses on improving robustness to sensor variability, environmental changes, and limited labeled data while maintaining computational efficiency for deployment on resource-constrained airborne platforms such as unmanned aerial vehicles (UAVs). To address these challenges, the dissertation improves the quality of HS data, combines complementary detectors via ensemble learning, incorporates physically interpretable spectral features, and develops domain adaptation techniques to maintain performance across different sensors and scenes. Finally, the work extends beyond anomaly detection for target localization to downstream scene understanding through hyperspectral semantic segmentation to support AiTR interrogation systems. This dissertation first examines the importance of HS image processing and enhancement techniques for enhancing data quality, improving feature extraction, and interpretability for both humans and machine learning models. Furthermore, this dissertation investigates robust anomaly and target detection in airborne HS imagery, with an emphasis on unsupervised and weakly supervised learning paradigms suitable for label-scarce operational settings. Individual detectors often have limitations, as the backgrounds in which they operate may not meet their underlying assumptions and can be cluttered. To overcome these challenges, a Greedy Ensemble Anomaly Detection (GE-AD) framework is proposed that systematically selects and weights heterogeneous statistical and machine-learning-based detectors via supervised and unsupervised stacking. To enhance robustness and interpretability, physically grounded spectral features derived from spectral unmixing are integrated into GE-AD. The resulting ensemble methods are designed to be lightweight and CPU-efficient, enabling deployment on resource-constrained unmanned aerial systems while maintaining competitive detection performance. To improve generalization under domain shift, the dissertation further explores transfer learning and unsupervised domain adaptation (DA) within a unified MoE stacking ensemble architecture. These approaches mitigate performance degradation caused by sensor variability, environmental changes, and acquisition inconsistencies without requiring retraining on target-domain data. Finally, the work extends beyond anomaly detection or target localization to broader scene understanding through HS semantic segmentation. A channel-attention-based adaptation of existing deep segmentation models enables effective processing of high-dimensional HS data with minimal architectural modifications, as demonstrated by agricultural crop-mapping case studies. Taken together, this dissertation demonstrates that robust hyperspectral target detection under real operational variability can be achieved by combining improved hyperspectral data processing, ensemble anomaly detection, physically interpretable spectral features, and adaptive generalization strategies while preserving computational efficiency for practical AiTR deployment.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest/Clarivate.

Notes

Embargoed until 2026-10-03

Recommended Citation

Hossain, Mazharul, "Investigation of Target Detection in Hyperspectral Images for Aided Target Recognition" (2026). Electronic Theses and Dissertations Archive. 4001.
https://digitalcommons.memphis.edu/etd/4001