Unsupervised Target Detection and Classification for Hyperspectral Imagery

This dissertation presents unsupervised spectral target detection and classification from a statistical signal processing point of view in the sense that the image pixels can be categorized into two groups of spectral signatures; target of interest, characterized by high-order statistics, and background, characterized by 2nd order statistics, which can be suppressed to improve the performance for target detection. The knowledge used to perform unsupervised spectral target analysis is obtained directly from the data a posteriori without pre-assumed prior knowledge. In order to generate the spectral sample signatures in these two categories, least-squares based unsupervised target sample generation (UTSG) and background sample generation (UBSG) algorithms are developed to extract spectral signatures of interest for each category with stopping rules devised for both algorithms to decide the number of sample signatures to be extracted