Unmixing of hyperspectral data by incorporating spectral variability and spatial information

Abstract

Spectral unmixing enables quantitative information on the abundances of cover types to be estimated within each image pixel. Although many spectral unmixing methods have been developed, many of these cannot consider endmember variability, nor do they incorporate spatial information into abundance estimates. Incorporating both endmember variability and spatial information into unmixing analyses has the potential to improve significantly abundance estimates. The major contribution of this thesis is the development and validation of spectral unmixing methods that incorporate endmember variability and spatial information. First, a novel endmember extraction method, Spectral Curve-based Endmember Extraction (SCEE), is developed for extracting and clustering multiple endmember spectra representing endmember variability within each class. SCEE uses a wavelet transform and extracts endmember candidates and clusters them into endmember bundles. Second, a novel unmixing method, Spectral unmixing within a multi-task Gaussian Process framework (SUGP), is developed that can effectively incorporate endmember variability into the unmixing process. SUGP uses multiple endmember spectra representing endmember variability to estimate abundances within a fully probabilistic framework. SUGP is different to existing methods in that it estimates, not only abundances, but also the uncertainties of abundances for each endmember class. Finally, a novel framework using a Markov Random Field (SUGP-MRF) is developed in order to incorporate spatial information and improve estimates of abundances. The performance of the proposed methods is assessed and compared to existing methods using synthetic hyperspectral imagery composed of precisely known mixtures, hyperspectral imagery acquired in the laboratory and in the field, and airborne (AVIRIS) imagery of Cuprite, Nevada. Results show that the proposed methods produce more accurate estimates of abundances than do existing methods.