KNOWLEDGE-BASED ROAD INTERPRETATION IN AERIAL IMAGES

This paper presents a knowledge-based method for automatic road extraction from aerial images. The method uses a hybrid control strategy in which hypotheses of roads are generated in a bottom-up process, and a top-down procedure is applied to verify the generated hypotheses. In this paper, a road model is proposed, which includes the geometric and radiometric properties of roads and relationship between roads in lowand high-resolution images. These properties and relationship are formulated as rules in PROLOG and stored in the knowledge base. The structures and relationships of roads yielded from images are stored as facts in the knowledge base. The hypotheses of roads are generated by applying the corresponding rules to the derived facts. To remove the ambiguity of the generated hypotheses, structural information of road surface and topological information of road networks are used. The missing road segments are predicted in the process of verification using topological information. An image in Hunter Valley, New South Wales, has been tested in this study. The results show that the road network is successfully extracted using the proposed method.