Tesis Doctoral Nuevas Técnicas de Clasificación Probabiĺıstica de Imágenes Hiperespectrales New Probabilistic Classification Techniques for Hyperspectral Images

Hyperspectral sensors provide hundreds of images, corresponding to different wavelength channels, for the same area on the surface of the Earth. Since different materials show different spectral properties, hyperspectral imagery is an effective technology for accurately discriminating and classifying materials. However, issues such as the high dimensionality of the data and the presence of noise and mixed pixels in the data, present several challenges for image classification. Dealing with these issues, this thesis proposes several new techniques for hyperspectral image classification. Developing subspace-based techniques for probabilistic classification is the main focus of the thesis. Specifically, we propose subspace-based multinomial logistic regression methods for learning the posterior probabilities. Furthermore, in order to better characterize mixed pixels in the scene, we propose an innovative method for the integration of the global posterior probability distributions and local probabilities which result from the whole image and a set of previously derived class combination maps, respectively. Another contribution of the thesis is the integration of spatial-contextual information using a robust relaxation method, which includes the information from the discontinuity maps estimated from the original image cube. Finally, the thesis introduces a new multiple features learning method which does not require any regularization or weight parameters. We apply the proposed method for fusion and classification of hyperspectral and LiDAR (light detection and ranging) data. The effectiveness of the proposed techniques is illustrated by using several simulated and real hyperspectral data sets and comparing with state-of-the-art methods.