Hyperspectral Image Mixture Analysis Using Notions of Sparsity, Nonlinearity and Decision Fusion
4 February 2020
Stadscampus, Prentenkabinet (Hof van Liere) - Prinsstraat 13 - 2000 Antwerpen (route: UAntwerpen, Stadscampus
Organization / co-organization:
Department of Physics
PhD defence Vera Andrejchenko - Faculty of Science, Department of Physics
Hyperspectral images (HSI) contain rich spectral information, by capturing a wide range of the electromagnetic radiation. Even though this is advantageous and provides the potential of a very detailed characterization of materials, it comes with a number of challenges. Information in HSI is highly redundant due to the high number of spectral bands involved. Another source of redundancy is the large spectral variability between spectral reflectances of the same material. Spatial redundancy is introduced by mixed pixels, containing more than one materials and high correlations between neighboring spectra.
The general research objective in this work is to address these redundancies of the HSI, by employing spectral unmixing techniques. These methods have the capability of describing highly redundant spectra as (linear or nonlinear) mixtures of a very limited amount of pure materials. In this way, they capture the structure of the low-dimensional subspace in which this high redundant data lives.
In this work, we mainly address the redundancy in HSI by developing model-based techniques which employ prior information on the parameters, derived from the data. HSI are high-dimensional but intrinsically lie on a lower-dimensional subspace which will be unraveled by exploiting the proposed priors. Moreover, these low-dimensional representations are employed in a decision fusion framework to improve the classification performance of HSI.
The thesis is organized in such a way that there are three clear subdivisions. The first contains the developed spectral unmixing method with the local low rank and inter-group sparsity prior imposed on its abundance parameters simultaneously. The second embraces the model based method which considers nonlinearities, i.e., multiple reflections and the shadowing at the same time in addition to the estimation of the abundance parameters. And the last consists of two decision fusion frameworks that incorporate the sparse low dimensional feature sets to enhance the HSI classification when limited training data is available. These are naturally preceded by an a) introduction to the hyperspectral imaging area and its active research areas and b) prerequisite key elements required for the forthcoming chapters.