Date of Award

2025-12-01

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

Advisor(s)

Miguel Velez-Reyes

Abstract

Hyperspectral Imaging (HSI) captures hundreds of contiguous narrow wavelength bands across the optical region of the electromagnetic spectrum collecting the spectral signature of materials in the field of view of the sensor enabling detailed analysis of each pixel's spectral signature. Satellite or airborne remote sensing systems often capture imagery with low to moderate spatial resolution (LMSR). At these resolutions, the measured spectral signature is a mixture of the signatures of the materials within a single pixel. This mixing of spectral information makes analysis and material identification difficult. Hyperspectral unmixing is an analysis technique that decomposes a pixel's spectrum into constituent signatures (or endmembers) and their relative proportions (or abundances). This has been an important analysis technique in hyperspectral remote sensing for decades. Advancements in hyperspectral imaging technology have enabled application of this imaging technology across multiple platforms in laboratory, industrial and remote sensing using UAVs. In many applications, these systems perform Very High Spatial Resolution Hyperspectral Imaging (VHSR-HSI) with resolutions ranging from millimeters to centimeters. Although it is expected that some spectral mixing may occur in these applications, most pixels in these images will contain a single material on it. A challenge is still to extract the spectral signatures of the materials in the image for further analysis in the image exploitation pipeline in an unsupervised manner. This research studied the application of endmember extraction algorithms for unsupervised signature extraction in in VHSR-HSI. For this task, we propose the Modified Pixel Purity Index (ModPPI) algorithm which is a modified version of the traditional PPI algorithm. Key modifications include counting both the number of times a pixel appears as a maximum or a minimum when projected to the random skewers and the use of clustering techniques to cluster spectral endmembers into endmember classes capturing material spectral variability and changing the determination of the number of endmembers as the number of endmember classes (or cluster). The application of ModPPI to the analysis of VHSR-HSI is studied in medical, vegetation and drone imagery. Results show that ModPPI outperforms traditional endmember extraction algorithms for this application while capturing materials spectral variability and able to map their spatial distribution. We also applied ModPPI to moderate resolution imagery and results point to comparable performance to published results for the image used in the study. ModPPI can be used to mine VHSR-HSI to create spectral signature libraries. Furthermore, the method could serve as a tool to assist in automated material identification in VHSR-HSI. By improving spectral signature extraction in VHSR-HSI, this study supports diverse applications such as ground-based imaging of resident space objects, agriculture, vegetation analysis, and organ spectral identification.

Language

en

Provenance

Received from ProQuest

File Size

143 p.

File Format

application/pdf

Rights Holder

Ana C. Chavez Lopez

Download

Share

COinS