Spectral Angle-Based Feature Extraction and Sparse Representation-Based Classification of Hyperspectral Imagery
Remote sensing involves measuring and analyzing objects of interests through data collected by a remote imaging modality without physical contact with the objects. Hyperspectral sensors have become increasingly popular for a variety of remote sensing applications. Hyperspectral data are composed of densely sampled reflectance values over a wide range of the electromagnetic spectrum. Such a wealth of spectral information can provide unique spectral signatures of different materials present in a scene, which makes it especially suitable for classification tasks. In this dissertation, we present new dimensionality reduction (feature extraction) and classification algorithms for high-dimensional hyperspectral data. Specifically, we develop the theory and validate a new dimensionality reduction approach that maximizes angular separation in the lower dimensional subspace. We also propose and develop its
local'' and nonlinear kernel'' variants for robust feature extraction of hyperspectral data. By preserving angular properties, the resulting subspaces demonstrate robustness to a variety of sources of variability that are commonly encountered in remote sensing applications. We also extend this approach to its ``spatial variant'' by incorporating spatial-contextual information along with spectral information from the hyperspectral images. We also optimize and develop a suitable sparse representation based classification framework for hyperspectral images. By extensive experiments on several real-world hyperspectral datasets, we demonstrate that the proposed algorithms significantly outperform the state-of-the-art methods. Further, we also demonstrate the applicability of the proposed methods for a practical environmental remote sensing task.