Unit:
Τομέας Επικοινωνιών και Επεξεργασίας ΣήματοςLibrary of the School of Science
Author:
Θεμελής Κωνσταντίνος
Dissertation committee:
Θεοδωρίδης Σέργιος Καθηγ.(Επιβλέπων), Αθανάσιος Ροντογιάννης Ερευνητής Β΄, Κωνσταντίνος Κουτρούμπας Ερευνητής Β΄
Original Title:
Bayesian signal processing techniques for hyperspectral image unmixing
Summary:
This thesis presents a framework of novel Bayesian signal processing techniques
developed for the unmixing of hyperspectral data. Hyperspectral data consist of
hundreds or thousands of spatially coregistered images, each one produced by
sampling at a specific wavelength. Modern remote sensing systems exploit the
spectral information of hyperspectral data in order to detect targets of
interest and to identify materials. Spectral unmixing is the process of
decomposing mixed hyperspectral pixels to their constituent materials, or
endmembers, and their corresponding proportional fractions, or abundances.
Exploiting the Bayesian framework, we develop a hierarchical Bayesian model
which utilizes the Laplace distribution as a prior for the abundance
parameters. The Laplace prior is widely used in the Bayesian compressive
sensing literature to meet the l1 norm of the celebrated lasso operator, which
is known to promote sparsity. This original concept is extended in the proposed
model, which utilizes an independent Laplace prior for each coecient of the
abundances' vector. The proposed model can then be viewed as a Bayesian
analogue to the adaptively weighted lasso. To perform Bayesian inference, an
efficient method, termed Bayesian inference iterative conditional expectations
(BI-ICE) is developed. BI-ICE is a greedy approximation to the Gibbs sampler,
but it can be also viewed as a first-moments approximation to variational
Bayesian inference techniques. Experimental results on simulated and real
hyperspectral data show that the proposed method converges fast, favors
sparsity in the abundances' vector, and oers improved estimation accuracy
compared to other related methods.
Keywords:
spectral unmixing, statistical signal processing, Bayesian analysis, sparse regression, hyperspectral data
Number of index pages:
1-11
Number of references:
129
File:
File access is restricted only to the intranet of UoA.
document.pdf
8 MB
File access is restricted only to the intranet of UoA.