Semantic Data Integration across Different Scales: Automatic Learning of Generalization Rules

In this paper we present an approach realizing the integration of data sets of different origin and with different resolution levels. The underlying idea is to reveal semantic correspondences between object classes of different geo-ontologies only by analysis of spatial and geometrical characteristics of instances of the data sets. As a result we derive transformation rules with Data Mining methods, which subsequently allow the semantic connection between data sets. For our case study we use data sets with similar thematic focus, but different semantic and geometric resolution: on the one hand building objects from cadastral data in a scale about 1:1K (detailed data set) and on the other hand settlement areas from topographic data in a scale about 1:25K (less detailed data set). To derive links between instances from the detailed data set to the more general one by geometric overlay it is required that the data sets are available in the same geographical extent. Then we generalize the detailed data set by using the object boundaries in the less detailed data set as a constraint for the generalization and generate an ’intermediate data set’ that is in a similar spatial resolution. We enrich the given information with additional attributes representing spatial relations and implicit or intrinsic given instance properties (e.g. object size), in order to derive transformation rules. These rules can be further used for classification of settlement areas of unknown regions in the target data set.