Bridging the Semantic Gap for Satellite Image Annotation and Automatic Mapping Applications

This paper brings a solution for bridging the gap between the results of state-of-the-art automatic classification algorithms and high semantic human-defined manually created terminology of cartographic data. Using a recent pure-spectral rule-based fully automatic classifier to define the basic 'vocabulary', we provide a hybrid method to automatically understand and describe semantic rules that link existent mapping data according to different specifications with the end-results of unsupervised computer information mining methods. Following an agreement between the learning model and the cartographic scale implied, we exploit Latent Dirichlet Allocation model (LDA) to map heterogeneous pixels with similar intermediate-level semantic meaning into land cover classes of various mapping products. By discovering the set of rules that explain semantic classes in existent vector systems, we introduce the prototype of an interactive learning loop that uses the concept of direct semantics applied on satellite imagery. We solve a big problem in generating cartographic information layers from a fully automatic classification map and demonstrate it for the typical case of Landsat images.