MACHINE LEARNING TECHNIQUES FOR DETERMINING PARAMETERS OF CARTOGRAPHIC GENERALISATION ALGORITHMS

This paper reports on research performed in the field of automated map generalization. We address the issue of determining how to set parameters of transformation algorithms. Empirical and theoretical studies have shown that, even given fixed map scale and purpose, these parameter values vary from one object to the other according to different characteristics such as the shape, size or environment of the object. Because of the complexity of cartographic rules and generalisation algorithms, we address this problem with techniques developed in the field of Machine Learning from examples. Specifically, we automatically learn, with neural networks, how to determine an algorithm parameters according to a set of measures describing the object to be transformed. We present the main issues to be addressed to use neural networks. We show that our approach is useful and that its main limit stands in the lack of good measures to describe an object. As a case study, this paper presents results of learning how to set the strength of a smoothing algorithm on a line from the French BDCarto database to represent a road on a 1/250,000 scale road map. RÉSUMÉ Cet article présente les résultats de recherches effectuées dans le domaine de la généralisation cartographique. Plus précisément, le problème traité est celui du paramétrage automatique des algorithmes de généralisation. Même pour produire une carte donnée, les algorithmes doivent être paramétrés différemment selon les caractéristiques des objet