LIP6 2002/005
- Thesis
Apprentissage Supervisé pour la Généralisation Cartographique - S. Mustière
- 244 pages - 06/08/2001- document en - http://www.lip6.fr/lip6/reports/2002/lip6.2002.005.pdf - 8,517 Ko
- Contact : sebastien.mustiere (at) nullign.fr, Jean-Daniel.Zucker (at) nulllip6.fr
- Ancien Thème : APA
- Keywords : Machine Learning, Abstraction, Cartographic Generalisation, Mapping
- Publisher : Ghislaine.Mary (at) nulllip6.fr
The context of this work is the automation of cartographic generalisation, which is the process of creating maps from over-detailed geographic databases. Many generalisation algorithms exist to transform the geometry of geographic objects to be represented on the map, but none of them is generic. We use a step by step, adaptive and focalised approach, where one geographic object must be transformed by the mean of several algorithms on adapted working spaces. In this context, rules must be determined to choose which algorithms to apply on an object, according to a description of it by the mean of a set of numeric measures. A process to chain algorithms is empirically developed for road generalisation. The efficiency and the limits of this process incite to use supervised machine learning techniques to acquire the knowledge necessary to a cartographic expert system. Our learning problem can be characterised as the search for efficient and understandable rules from few, noisy and large examples. Then, a classical learning algorithm provides rules with a poor quality.
In order to improve the results quality, background knowledge is used to guide the learning process while decomposing the learning problem in several simpler sub-problems : we successively learn to abstract and to choose the transformation to apply on the geographic objects. The abstraction phase changes the object representation from a large set of numeric measures to a small set of new symbolic attributes. The transformation choice phase determines which transformation to apply according to the abstract description of the object. The introduction of this abstraction phase enables to learn more efficient and understandable rules than a direct learning. It so enables to improve the cartographic quality of the results.
In order to improve the results quality, background knowledge is used to guide the learning process while decomposing the learning problem in several simpler sub-problems : we successively learn to abstract and to choose the transformation to apply on the geographic objects. The abstraction phase changes the object representation from a large set of numeric measures to a small set of new symbolic attributes. The transformation choice phase determines which transformation to apply according to the abstract description of the object. The introduction of this abstraction phase enables to learn more efficient and understandable rules than a direct learning. It so enables to improve the cartographic quality of the results.