Multisource clustering of remote sensing images with Entropy-based Dempster-Shafer fusion

In this paper, we propose a strategy for fusing clustering maps obtained with different remote sensing sources. Dempster-Shafer (DS) Theory is a powerful fusion method that allows to combine classifications from different sources and handles ignorance, imprecision and conflict between them. To do so, it attributes evidences (weights) to different hypothesis representing single or unions of classes. We introduce a fully unsupervised evidence assignment strategy exploiting the entropy among cluster memberships. Ambiguous pixels get stronger evidences for union of classes to better represent the uncertainty among them. On two multisource experiments, the proposed Entropy-based Dempster-Shafer (EDS) performs best along the different fusion methods with VHR images, when the single class accuracies from each source are complementary and one of the sources shows low overall accuracy.