RESEARCH ON MAP GENERALIZATION BASED ON RULES AND NEURAL
NETWORK
Y. Cai, Q. Guo
caiyx2002cn@yahoo.com.cn
Cartographic researchers find that intelligent methods is a trend in map generalization, because map generalization is a extra-complicated process (Wang Guangxia,1996), the experiences of cartographers play important roles in the process of generalization. Generalization operations concern both the obtainment of knowledge and analysis of the data, which have been researched by cartographers for years. Neural Networks have been successfully applied in solving some pattern recognition problems for many years as an intelligent method, self-organizing map (SOM), one of the unsupervised learning networks, is most suitable for pattern recognition and coarse classification. Back-Propagation network (BP), one of the supervised learning networks, is suit for numerical computer and refined classification, but it needs some typical samples for training.
In the paper, we utilize the advantages of SOM and BP networks and present a new application of combining NNS with expert knowledge as a method of automated map generalization; it is called Neural Network Expert System (NNES). Firstly, some attributes involving topological, geometric and semantic properties are input to a SOM to categorize similar objects in the map. Secondly, some typical objects are determined from these categories as training samples, results about whether selection or not are identified by expert knowledge of specific features as expectation outputs, and a BP network is trained by samples. Thirdly, the trained BP network is used to select objects; the selection threshold is relative with the target scale. Finally, the selected objects must be generalized further under constraints and guidelines of some rules, which are derived from cartographic specification and the sum-up of cartographers’ experiences.
An example of buildings generalization based on NNES is given in this paper, selection of streets are based on NNS, some attributes such as length, connectedness, grade and closeness of streets are considered in the input of SOM, SOM categorizes streets into five classes, 15 typical objects have been selected from five classes as training samples, and a BP network is trained by these samples, then the trained BP network is used to select streets. Then, some operations such as line simplification, buildings combination, displacement, exaggeration, typification of streets and buildings are selected and used according to rules of specific features. An example shows the suitability of the NNES approach for map generalization. It has some advantages over other methods in that it combines the virtue of NNS with knowledge rules, and can be used for selection of general objects as well the abstraction of specific feature. And it can also be used as a general tool for map generalization.