Dissertation Announcement for Vineetha Menon
10/11/16 at 12:00 PM
October 3, 2016
Dear Faculty, Graduate and Undergraduate students,
You are cordially invited to my Ph.D. dissertation oral defense.
Dissertation Title: Dimensionality Reduction of Hyperspectral Imagery Using Random Projections
When: Tuesday, October 11, 2016, 12:00 PM
Where: Simrall 228
Candidate: Vineetha Menon
Degree: Ph.D., Electrical and Computer Engineering
Committee:
Dr. James E. Fowler
Professor, Graduate Program Director and Associate Director (DASI)
Department of Electrical and Computer Engineering
(Major Professor/Graduate Coordinator and Director of Dissertation )
Dr. Jenny Q. Du
Professor and Bobby Shackhouls Endowed Professorship
Department of Electrical and Computer Engineering
(Co-Major Professor and Director of Dissertation)
Dr. Nicolas H. Younan
Professor, Department Head and James Worth Bagley Chair
Department of Electrical and Computer Engineering
(Committee Member)
Dr. Robert J. Moorhead Professor and GRI/NGI Director
Department of Electrical and Computer Engineering
(Committee Member)
Abstract:
Hyperspectral imagery is often associated with high storage and transmission costs. Dimensionality reduction serves as an indispensable tool to reduce the time and space complexity of hyperspectral imagery. Dimensionality-reduction methods based on transforms are widely used and give a data-dependent representation that is unfortunately computationally expensive. For this reason, of particular prominence are random projections which are attractive alternative for dimensionality-reduction, due to their computational efficiency and simplicity of implementation.
This dissertation explores computationally fast and efficient dimensionality reduction for hyperspectral imagery driven by random projections. For this we present three significant contributions for dimensionality reduction and classification of hyperspectral data. Firstly, we introduce Hadamard matrix (HM) as an alternative to traditional Gaussian matrix (GM) for random projections for the dimensionality reduction of hyperspectral imagery. We find that HM-based random projections were not only computationally faster but also gave better classification performance than traditional GM-based projections even with far fewer dimensions. Secondly, we investigate dimensionality reduction via random projections in conjunction with feature selection using a fast variant of singular value decomposition (SVD) known as FSVD. In particular, HM-based projections are coupled with FSVD in order to implement a two-stage dimensionality reduction that marries the computational benefits of the data-independent random projection to the structure-capturing capability of the data-dependent singular value transform. It is shown that random projections with FSVD gave better classification performance along with reduction in both time and space complexity in comparison to SVD performed on the original dataset. Finally, we propose a new framework that addresses the commonly encountered spectral-unmixing phenomenon in hyperspectral imagery by introducing HM- and GM-based nonnegative least squares (NNLS) that combines dimensionality reduction with estimation of endmember abundances. In this effort, we consider both our proposed HM- and HM-FSVD-based projections. It is demonstrated again that the HM-based dimensionality reduction methods gave more accurate abundance estimation and superior classification performance with low errors at fewer dimensions as compared to using the original data without dimension reduction.
Best Regards,
Vineetha