Abstract

Hyperspectral target detection applies algorithms to high-dimensional images to identify rare targets, or objects of interest, within cluttered scenes of diverse backgrounds. Computational demands of processing hyperspectral datasets, along with their limited availability in both public and private sectors, imposes challenges in assessing system-level limitations of detection. This research presents a methodology to quantify end-to-end sensitivities and limitations through statistical modeling across thousands of target detection scenarios. The objective is to identify specific parametric thresholds where detection begins to degrade, based on “knees” in detection curves derived from model outputs of various scenarios. The model considers subpixel targets, where an object’s signal occupies an area smaller than a pixel. Since subpixel targets are spatially unresolved, validating subpixel models is challenging with conventional datasets. To address this, a lattice-based target was developed and deployed in a UAV data collection, yielding a novel dataset with approximately 300 empirical subpixel samples for each constant fill fraction (0.2, 0.4, 0.6, 0.8). To investigate limitations, eight parameters across the imaging chain (scene, atmosphere, sensor) were considered in combination with four qualitative scenes (urban, rural, forest, desert). Broad scenario sampling of target-background combinations was guided by the Mahalanobis distance, used as a spectral contrast parameter. In total, approximately 100,000 unique scenarios were generated, with scalar detection outputs stored in a multidimensional array. These detection outputs consist of a proposed metric, the log-weighted area under the ROC curve (wAUC), which evaluates overall performance while emphasizing low false alarm rates. Results include correlation coefficients and impact scores to quantify sensitivities between system parameters and detection, revealing which parameters most influence detection under various conditions. Knees in the wAUC curves were identified using a curvature-based method, and a novel knee significance score (KSS) was introduced to rank the importance of each limitation. It concludes that aerosol visibility (2-6 km) is the leading limiter in detection, followed by the subpixel target percentage (15-35%), and background clutter, modeled using a t-distribution with 3-5 degrees of freedom. This research establishes a framework to quantify end-to-end limitations in hyperspectral target detection, extendable to additional system parameters in future studies.

Library of Congress Subject Headings

Hyperspectral imaging--Data processing; Remote sensing--Data processing; Image processing--Mathematics

Publication Date

5-7-2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science

Advisor

John P. Kerekes

Advisor/Committee Member

George Thurston

Advisor/Committee Member

Emmett Ientilucci

Campus

RIT – Main Campus

Plan Codes

IMGS-PHD

Download

Share

COinS