Adaptive Coherence Estimator for Spatial/Spectral Pattern Analysis of Hyperspectral Imagery
Automated classification methods for scene analysis of hyperspectral imagery are typically full-pixel techniques that render hardened class labels at each site in an image. This often makes the identification of subtle materials or complex objects in a scene difficult for a number of reasons. The labeling might be made too soon in the overall decision process leading to results that are unreliable, or the use of categorical labels simply cannot portray soft boundaries between objects or render information about mixtures of materials. In this study, the Adaptive Coherence Estimator is applied to hyperspectral imagery resulting in "soft" class map layers for subsequent object recognition and terrain analysis. The method requires an estimate of background second order statistics in a scene that is based on global statistics of the entire scene. Unfortunately, global statistics have the disadvantage of including materials of interest in the background estimate and this implies the method assumes these materials occupy an insignificant portion of the scene. This assumption is often invalid for applications such as terrain, urban, and shallow-water mapping. To address to problem we investigate the use of supervised classification methods to restrict the regions that are considered as background to get a better estimate of the background statistics. The context of the effort is terrain, urban, and shallow-water mapping using hyperspectral imagery, where the materials of interest inherently occupy a significant portion of a scene or where certain background classes have problematic second-order statistics. Results of experiments within this context are shown.