| Pattern recognition in hyperspectral imagery is a challenging problem. Although various algorithms have been proposed in the literature, a simple and effective solution for target detection in hyperspectral imagery is yet to be found. In this thesis, we propose a new algorithm which uses principal component analysis for preprocessing and spectral angle mapping for detecting objects of interest in hyperspectral imagery. To improve the detection rate, mean filtering is applied to the hyperspectral data, followed by the application of principal component transform. Then, spectral angle mapping is applied to detect objects of interest. Thereafter, we propose a new method for estimating the noise in each band of a hyperspectral datacube. Each band of the datacube is divided into small blocks, which are independently decorrelated. The decorrelation leaves noiselike residuals whose variance provides an estimate of the noise. A set of these variances is used to provide the best estimate of that band's noise. This method provides consistent noise estimates from images with different land-cover types. The performance of the proposed detection algorithm has been tested with three different real life hyperspectral datasets and the results show that the proposed algorithm efficiently detects object of interest. |
