| Remote sensing technology has advanced tremendously in recent decades. An important driver for this development has been the offering of wide spatial and temporal coverage by space- and airborne platforms, as well as the ever-improving capability of their sensors to record images with high spatial and spectral resolution. A modality that produces a bulk of data for remote sensing is hyperspectral imaging. This modality records the reflected solar radiation in contiguous and often numerous spectral bands, thereby extending the standard photography by enabling to treat each pixel individually as a spectrum discernible for each class of materials. One limitation of such imaging, where the spatial and spectral resolutions are inherently traded against each other, is the occurrence of mixed pixels and spectral mixing. The unraveling of spectral mixtures has been widely studied as spectral unmixing, where two main aspects are of interest: the estimation of the constituent spectra, and of their fractions or abundances, in the mixture.;The work described in the thesis regards spectral unmixing from two objectives: advancement of methodology and introduction of unmixing in new applications. The first part, specifically, is concerned with the development of data-driven methods for spectral unmixing that can mitigate the dependency on physical parameters and models, and reduce high computational complexity due to the typical use of optimization techniques. A concrete realization consists of several algorithms that reformulate the known geometrical framework of spectral unmixing by introducing linear and nonlinear distance-based and analytical formulations. The second part introduces or elaborates spectral unmixing for detection of the atmospheric adjacency-effect and the estimation of quality of inland and coastal waters. The presented unmixing-based approaches in this context have been validated through theoretical and empirical comparison using available datasets and reference methods. |
