Separating patterns and finding the independent components of mixed signals based on non-Gaussian distribution properties

The effect of assuming and using non-Gaussian attributes of underlying source signals for separating/encoding patterns is investigated, for application to terrain categorization (TERCAT) problems. Our analysis provides transformed data, denoted as "Independent Components," which can be used and interpreted in different ways. The basis vectors of the resulting transformed data are statistically independent and tend to align themselves with source signals. In this effort, we investigate the basic formulation designed to transform signals for subsequent processing or analysis, as well as a more sophisticated model designed specifically for unsupervised classification. Mixes of single band images are used, as well as simulated color infrared and Landsat. A number of experiments are performed. We first validate the basic formulation using a straightforward application of the method to unmix signal data in image space. We next show the advantage of using this transformed data compared to the original data for visually detecting TERCAT targets of interest. Subsequently, we test two methods of performing unsupervised classification on a scene that contain a diverse range of terrain features, showing the benefit of these methods against a control method for TERCAT applications.