Alpha-shapes for visualizing irregular-shaped class clusters in 3D feature space for classification of remotely sensed imagery

In this study, we present a geovisualization tool using Alpha-shapes to visualize class clusters in a remotely sensed image classification. An Alpha-shape is an accurate representation of the shape of a cluster of points in a 2D or 3D feature space. Traditionally, spheres and ellipsoids are used to represent class clusters in a classification. These shapes, however, are rough approximations of irregular shaped class clusters. In remote sensing classification we often have to deal with these irregular clusters (e.g. concavities, pockets and voids) and Alpha-shapes will improve visualization of these classes. We argue that Alpha-shapes will also improve insight into a classification process, and related uncertainty. Uncertainty can arise from ambiguity in the attribution of class labels to pixels. This ambiguity is often caused by overlapping classes. Visualization is helpful in communicating this ambiguity as Alpha-shapes clearly show where classes overlap. In this study, we also propose and implement a novel classification algorithm based on Alpha-shapes. Most classification algorithms cannot cope with irregular and concave cluster shapes in feature space. We apply our algorithm on a Landsat 7 image scene of a study area in Southern France. We show that good classification results can be obtained with Alpha-shapes.