Nuevas técnicas de clasificación de imágenes hiperespectrales

Resumen

The main contribution of this thesis is de development and implementation of new techniques for hyperspectral analysis which are able to incorporate the spatial component of the data when performing spectral unmixing and remote compressive sensing of hyperspectral images. Spectral mixing is one of the main problems that arise when characterizing the spectral constituents residing at a sub-pixel level in a hyperspectral scene. It consists of the fact that many pixels in the scene are �mixed� in nature, i.e. they are formed by different spectral constituents at sub-pixel levels. In this regard, one of the main contributions of the present thesis is the integration of spatial and spectral information as a previous step to the traditional endmember identification conducted by many algorithms. We accomplish this through a set of innovative spatial preprocessing modules, intended to guide the endmember identification process by including spatial information but without the need to modify the already available, spectral-based endmember identification algorithms. Furthermore in this thesis work, we develop a new compressive sensing framework for hyperspectral imaging, which exploits the spatial correlation of hyperspectral images and the spectral mixing phenomenon in order to compress the hyperspectral data in the acquisition process.