Técnicas eficientes para extracción de información en imágenes obtenidas de forma remota

Abstract

The main contributions of the present thesis work are given first, by the proposal of new parallel algorithms for information retrieval from remotely sensed images of the surface of the Earth. These algorithms are based on the unmixing problem (characteristic of remotely sensed images with high spectral resolution), which allows to express the pixels of an image as a linear or nonlinear combination of spectrally pure elements, weighted by their corresponding abundance fractions. Once the theoretical foundations of the proposed study are described, the thesis work presents a comparison of different parallel architectures to address the spectral unmixing or unmixing problem with the following steps: 1) automatic identification of the number of endmembers; 2) automatic extraction; and 3) estimation of the fractional abundance of endmembers on a sub-pixel basis. In order to solve the spectral unmixing problem in satellite images with low spectral resolution (usually called multispectral), a methodology to expand the dimensionality of such images has been developed by including spectral and spatial information (using mathematical morphology concepts). Subsequently, the result is used as input to improve the interpretation of the image with respect to the case where only the original spectral information is used. Finally, the present thesis work integrates the techniques previously developed on an information system for classification of images obtained by satellite or airborne sensors for remote sensing of the Earth. The initial design as a desktop application has been improved through the use of a remote server using web technologies.