GRAPH-BASED URBAN LAND USE MAPPING FROM HIGH RESOLUTION SATELLITE IMAGES

Abstract. Due to the dynamic character of urban land use (e.g. urban sprawl) there is a demand for frequent updates for monitoring, modeling, and controlling purposes. Urban land use is an added value that can be indirectly derived with the help of various properties of land cover classes that describe a certain area and create a distinguishable structure. The goal of this project is to extract land use (LU) classes out of a structure of land cover (LC) classes from high resolution Quickbird data and additional LiDAR building height models. The study area is Rostock, a German city with more than 200.000 inhabitants. To model the properties of urban land use a graph based approach is adapted from other disciplines (industrial image processing, medicine, informatics). A graph consists of nodes and edges while nodes describe the land cover and edges define the relationship of neighboring objects. To calculate the adjacency that describes which nodes are combined with an edge several distance ranges and building height properties are tested. Furthermore the information value of planar versus non-planar graph types is analyzed. After creating the graphs specific indices are computed that evaluate how compact or connected the graphs are. In this work several graph indices are explained and applied to training areas. Results show that the distance of buildings and building height are reliable indicators for LU-categories. The separability of LU-classes improves when properties of land cover classes and graph indices are combined to a LU-signature.