Date of Award
5-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Engineering
Committee Chair/Advisor
Dr. Goutam Koley
Committee Member
Dr. Jon Calhoun
Committee Member
Dr. Adam Hoover
Abstract
Material classification over long distances remains a challenge, as traditional spectroscopic methods are limited by range and practicality. Modern computer vision approaches have attempted to bridge this gap, but most rely on a single imaging modality, sacrificing accuracy for usability. In this work, we present a visible and infrared fusion (VIF) pipeline deployed on a UAV platform to classify object materials from 8–10m under real-world conditions. We release a dataset of 2264 visible and infrared image pairs from two test flights, including a YOLO-ready subset of 1600 processed, fused, and annotated images. Our results show that fusion increases training accuracy by 20.2–33.6% and inference accuracy by 50.6–68.7% over single-modality approaches. Despite significant pixel misalignment inherent to real-world capture, our method achieves 85.69% classification accuracy and up to 49.61% recall on unseen images, representing a 2.44–5.04x improvement over comparable methods. These findings suggest that perfect image alignment is not a prerequisite for effective VIF, and that multimodal fusion can enable physics-aware material classification at practical aerial distances.
Recommended Citation
Benson, Joshua, "Aerial Image Based Material Classification Through Physics-Aware Visible-Infrared Image Fusion" (2026). All Theses. 4757.
https://open.clemson.edu/all_theses/4757