Date of Award


Document Type


Degree Name

Master of Science (MS)


College of Science and Mathematics


Computer Science

Thesis Sponsor/Dissertation Chair/Project Chair

Stefan A. Robila

Committee Member

Angel Gutierrez

Committee Member

John Jenq


Hyperspectral imaging is a branch of remote sensing which deals with creating and processing aerial or satellite pictures that capture wide range of wavelengths, most of which are invisible to the naked eye. Hyperspectral images are composed of many bands, each corresponding to certain light frequencies. Because of their complex nature, image processing tasks such as feature extraction can be resource and time consuming. There are many unsupervised extraction methods available. A recently investigated one is Nonnegative Matrix Factorization (NMF), a method that given positive linear matrix of positive sources, attempts to recover them. In this thesis we designed, implemented and tested parallel versions of two popular iterative NMF algorithms: one based on multiplicative updates, and another on alternative gradient computation.

Our algorithms are designed to leverage the multi-processor SMP architecture and power of threading to evenly distribute the workload among the available CPU’s and improve the performance as compared to their sequential counterparts. This work could be used as a basis for creating even more powerful distributed algorithms that would work on clustered architectures. The experiments show a speedup in both algorithms without reduction in accuracy.

In addition, we have also developed a java based framework offering reading and writing tools for various hyperspectral image types, as well as visualization tools, and a graphical user interface to launch and control the factorization processes.

File Format