Techniques for the graph representation of spectral imagery

Many techniques from graph theory and network theory have been applied to traditional images, and some techniques are now being applied to spectral imagery. Contrary to the typical approaches of utilizing the first order statistics, mixture models, and linear subspaces, the methods described in this paper utilize the spectral data structure to generate a graph representation of the image. By ignoring any reliance on the shape of the data, graph based methods can succeed where typical methods break down, such as in high resolution scenes with very high clutter. Before graph theory techniques can be utilized on an image, it must be represented as a graph. Because images contain only measured nodes (pixels) and no edges, edges are drawn between pixels based on some similarity measure. With a specific focus on creating graphs for clustering, several graph creation techniques are compared with two novel methods described: the locally weighted k-nearest neighbor approach and the density weighted k-nearest neighbor approach. By applying two different clustering techniques to the resulting graphs, the various graph creation techniques are compared using real world data.