Imaging spectrometry and image registration: A singular value decomposition and wavelet based learning

This work is about supervised learning, in general, and analysis of remote sensing data for ecological studies, in particular. The original aim was to provide new insights into pattern recognition of hyper-spectral images in terms of explicit physical units based on both spectral and spatial features. However, the present work arose because of concerns about the generalization of these results. The fundamental ideas underlying the two main contributions of this dissertation, i.e., a new method for automatic image registration and a robust spatial and spectral technique for learning pattern recognition from examples (supervised learning), are motivated by the conceptual part of the dissertation (Chapter 1) which is formally presented in Chapter 2. Spectral and spatial interactions directly associated to ground units are untangled using wavelet tools and a Singular Value Decomposition (SVD) based technique, the so-called hierarchical foreground background analysis (HFBA). The idea for applying geometric and spatial properties as a tool for generalization inspired and fertilized the work in registration and finally became the heart of the method for automatic image registration: a three stage algorithm with feedback to automatically register images differing by a global rigid transformation. The significant features of the algorithm are the use of nonlinear wavelet compression to extract control points from each image and the use of SVD's geometric properties to register them. Validation and performance tests drove us to the following conclusions: (1) Nonlinear wavelet compression greatly filters extraneous or redundant information providing spatial coherence useful for generalization purposes. (2) One of the strong points of the proposed method is that we can group together samples with similar properties manifested spectrally that improves the representation of its spatial distribution (at different scales). (3) The SVD tools are significant, because they allow us to register ensembles of control points without requiring knowledge of which points specifically match up pairwise.