ORCID Identifier(s)


Graduation Semester and Year




Document Type


Degree Name

Doctor of Philosophy in Electrical Engineering


Electrical Engineering

First Advisor

Ioannis Schizas


This work discusses the problem of unsupervised classification in images. Conventional methods approached this problem with the naive assumption that the relationship among the pixels' information can be expressed sufficiently in a linear manner. However, higher accuracy was established by implementing kernel-based expressions of data to express the non-linear relationship of that data in a linear manner, when mapped in a higher dimensional space. This process allows much more effective clustering performances by increasing the informativeness of the data. Hyperspectral images, being limited in spatial resolution as a tradeoff for the significantly higher number of channels compared to traditional images, often face the challenge of having pixels that, instead of showing one material, show a mixture of multiple materials. It then becomes a challenging task of \textit{unmixing} those materials, whose challenge is greatly exacerbated with the presence of strong noise, and/or the data being corrupted due to some damage to the sensor, causing \textit{dead pixels} in the form of data entries containing zero values. Unlike a large body of work which focuses on a simpler approach, where it is assumed that the mixtures are obtained through a linear combination of the contributing materials, nonlinear mixtures are a more accurate representation of the mixtures obtained in real-life scenarios, and are thus tackled in our work. The unmixing problems were addressed via formulation of a constrained optimization problem which utilizes nonlinear mixing models, efficiently addressing the limitation of having a limited window of kernel parameters, tackling more complex mixture models, and reducing computational complexity by automatically reducing dimensions containing irrelevant data, with the added challenge of performing in a fully unsupervised setting. The design and implementation of a nonlinear autoencoder neural network further improves work in this respect, by fully customized designs of layers, which not only utilize spatial information via weighted averaging of the pixels based on their perceived similarities in the kernel space, but also have the added versatility in accommodation of higher degree nonlinear interactions, a technique unavailable in major current works. Going beyond hyperspectral remote sensing images, a unique approach was tackled in unsupervised heart disease diagnosis through observation of the mitral valve, and potential diseases affecting its ability to function effectively. A wide variety of datasets were used in measuring its efficacy, including data that was noisy and of lower resolutions, further increasing the difficulty of implementing a fully unsupervised heart disease diagnosis algorithm that detected the location of the mitral valve within the videos, and observed its movement to detect whether it was diseased or healthy.


Canonical correlations, Kernel learning, Unsupervised unmixing, Remote sensing, Nonlinear unmixing, Hyperspectral imaging, Autoencoders, Neural networks, Radial basis functions, Ultrasound, Echocardiography, Mitral valve, Apical 4 chamber


Electrical and Computer Engineering | Engineering


Degree granted by The University of Texas at Arlington