SPATIAL-SPECTRAL CLUSTERING USING RECURSIVE SPANNING-TREES

The inherent contextual property of spanning trees is exploited in a nonparametric contextual clustering algorithm for multispectral satellite data. The linkage problem associated with shortest spanning trees is avoided by making extensive use of global information, and a two-stage algorithm (segmentation then clustering) is described, each stage being based upon recursive spanning trees and minimax variance techniques. A conditional entropy or 'segmentation loss' derived from mutual information is shown to provide a useful indication of the number of segments needed before clustering. The performance of the algorithm is compared with a single-pixel clustering algorithm and shows significant reduction in classification noise, both at class boundaries and within classes, while the spatial resolution of the single-pixel classifier is retained.