In this paper, we present an ontology-driven approach for cartographic pattern recognition in support of map generalisation. Spatial patterns are formalised by means of ontologies which are then used to deductively trigger appropriate low level pattern recognition techniques. Modelling ontologies suited for spatial pattern recognition is discussed by example of an ontology of terraced houses. The paper subsequently focuses on approaches for inferring the instances of higher level concepts. Three different approaches are employed to detect terraced houses in Ordnance Survey MasterMap® vector data: Weighted summation; Joint Bayes classifier; and Support Vector Machines. An evaluation by comparison to a manual classification reveals that weighted summation and the Joint Bayes classifier both have satisfactory prediction accuracy, but the Joint Bayes classifier has advantages when considering the calibration effort involved. In conclusion, we claim that the ontology-driven approach better captures the complex structure of spatial patterns and provides enhanced transparency and flexibility of the pattern recognition process in comparison to conventional, purely geometric and/or statistical techniques.