Interactive exploration of uncertainty in fuzzy classifications by isosurface visualization of class clusters

Uncertainty and vagueness are important concepts when dealing with transition zones between vegetation communities or land-cover classes. In this study, classification uncertainty is quantified by applying a supervised fuzzy classification algorithm. New visualization techniques are proposed and presented in order to come to a better understanding of the relationship between uncertainty in the spatial extent of image classes and their thematic uncertainty. The thematic extent of a class is visualized as a three-dimensional (3D) class cluster shape in a feature-space plot, and the spatial extent of the class is highlighted in an image display based on a user-defined uncertainty threshold. Changing this threshold updates both visualizations, showing the effect of uncertainty on the spatial extent of a class and its shape in feature space. Spheres, ellipsoids, convex hulls, -shapes and isosurfaces are compared for visualization of 3D class clusters. Isosurfaces are implemented to facilitate real-time rendering and interaction with class clusters in feature space. The visualization tool is illustrated with a fuzzy classification of a Quickbird image of Macquarie Island, one of the unique sub-Antarctic World Heritage Areas that is characterized by vegetation transition zones. This study shows that visualization techniques are valuable for the interpretation and exploration of image classification results and associated uncertainty.