| Modern imaging sensors produce vast amounts of data, overwhelming human analysts. One such sensor is the Airborne Visible and Infrared Imaging Spectrometer (AVIRIS) hyperspectral sensor. The AVIRIS sensor simultaneously collects data in 224 spectral bands that range from 0.4{dollar}mu{dollar}m to 2.5{dollar}mu{dollar}m in approximately 10nm increments, producing 224 images, each representing a single spectral band. Autonomous systems are required that can fuse "important" spectral bands and then classify regions of interest if all of this data is to be exploited. This dissertation presents a comprehensive solution that consists of a new physiologically motivated fusion algorithm and a novel Bayes optimal self-architecting classifier that processes the outputs of the fusion algorithm. The fusion algorithm which uses a contrast sensitivity weighted wavelet-based multiresolution analysis is shown to outperform other fusion algorithms in both visual aesthetics and signal-to-noise ratios. The self-architecting classifier is a Radial Basis Function (RBF) Iterative Construction Algorithm (RICA) that is designed to autonomously determine the size of its network architecture for optimal classification performance. RICA is shown to outperform several neural network algorithms, including a fixed architecture multi-layer Perceptron (MLP), a fixed architecture RBF, and an adaptive architecture MLP. A proof is also presented demonstrating that RICA produces a network which is a minimum mean squared-error approximate to Bayes optimal discriminant functions. Finally, it is shown that this combination of image fusion and self-architecting classifier provide an excellent means to detect targets in hyperspectral sensor data. |
