Spatio-spectral Image Analysis Using Classical and Neural Algorithms

Remote imaging at high spatial resolution has a number of environmental, industrial, and military applications. Analysis of high-resolution multi-spectral images usually involves either spectral analysis of single pixels in a multi-or hyper-spectral image or spatial analysis of multi-pixels in a panchromatic or monochromatic image. Although insuucient for some pattern recognition applications individually, the combination of spatial and spectral analytical techniques may allow the identiication of more complex signatures that might not otherwise be manifested in the individual spatial or spectral domains. We report on some preliminary investigation of unsupervised classiication methodologies (using both \classical" and \neural" algorithms) to identify potentially revealing features in these images. We apply dimension-reduction preprocessing to the images, cluster, and compare the clusterings obtained by diierent algorithms. Our clas-siication results are analyzed both visually and with a suite of objective, quantitative measures.