Summary:
In this thesis, the problem of semisupervised hyperspectral unmixing is
considered, where the spectral signatures of some materials are known and the
aim during the analysis of a specific pixel is to determine both the materials
that contribute to the composition of the pixel and the vector containing the
abundance fractions of them in the composition. Adopting the linear mixture
model for the examined hyperspectral image, a hierarchical Bayesian approach
suitable for semisupervised hyperspectral unmixing is proposed, where suitable
priors are selected for the model parameters, such that they favor sparse
solutions for the abundance vector and they ensure the non-negativity of the
abundances. Then, a new Bayesian inference iterative scheme, named
BI-ICE-single, is developed, which produces sparse estimations for the
abundance vector. Finally, a new simple technique is described, which takes
into account the possible spatial correlation between adjacent pixels of the
hyperspectral image.
Experimental results illustrate that the BI-ICE-single algorithm does not
present the estimation accuracy of BI-ICE method, in the case where the
spectral signatures of the endmembers are highly correlated. On the contrary,
the two algorithms exhibit similar performance, when the correlation is low. In
addition, the proposed technique offers significant computational savings,
without leading to inferior quality results compared to the case where the
possible spatial correlation is ignored.
Keywords:
Hyperspectral imagery, Hierarchical Bayesian model, Sparse semisupervised spectral unmixing, Bayesian inference, Spectral signature
