Graph cut segmentation of sparsely sampled images with application to InSAR-measured changes in elevation

In this paper, we outline an algorithm for the automatic segmentation of sparse data in order to detect possible terrain-deformation phenomena. Segmentation is accomplished through a graph cut technique. In the graph structure, for each edge, we derive a unique energy by combining multiple independent energies tailored toward accurately locating the boundaries of spot-like, subsiding regions. We then find the series of cuts with minimum total energy and fit splines to these cuts for smooth segment boundaries. The segmentation approach is applied to the problem of localizing sinkholes in karst regions. Test results indicate efficacy for a sufficient density of InSAR features.