Research On Ann-based Map Intelligent Generalization For Buildings

Map generalization is one of core issues in cartography. As GIS applications expand constantly, map generalization is being applied more and more widely. Map generalization is a process of deriving a map at a reduced scale from a given map. It aims to maintain and improve the legibility of the map with less space by representing the desired map contents as distinctly as possible. As there exists a great deal of visual and inspirational thinking in map generalization, it is one of the most inventive research fields in cartography. Apparently, of those mainstream approaches to map generalization based on linear algorithms, it is difficult to achieve satisfactory qualities of automated generalization of maps with complicated areas.With the focusing on applying artificial intelligence(AI) to map generalization, the intelligente generalization technology is developed for automated building generalization, one of the most important and complicated map generalization issues. In this thesis, a novel approach for the intelligent building generalization is proposed based on neural network by simulating map expert’s thinking. Characteristics of both manual and automated generalization are analyzed, including typical automated generalization operators such as aggregation and simplification. The principles of automated building generalization are summarized as well.Building grouping is a premise of building aggregation in automated map generalization. Principles of proximity and similarity in Gestalt theory are analyzed. Then, three parameters, centroid coordinates, minimum distance between buildings, and location relationship among buildings and roads are chosen for building description. Accordingly, a method of building grouping based on self-organizing map(SOM) neural network is proposed for the building grouping.A novel back propagation neural network(BPNN) approach to building aggregation based on local perception of complicated map contexts is studied. A perceptron perceives the characteristics of building structure, orientation, distribution, and location relationships among buildings from raster maps. Also by combining rules of building aggregation and map expert knowledge, a set of mapping rules between input patterns collected by the perceptron and outputs is formulated. Once trained, the BPNN model could outline the buildings that are to be aggregated. The results are cartographically satisfactory.Similar to building aggregation, a novel approach to building simplification with raster-based local perception using the BPNN model for learning cartographer’s knowledge is studied. Cartographer’s expertise, coupled with a perceptron that perceives the characteristics of building structures in raster maps is analyzed. The relationships among local context perceived by the perceptron and outputs are formulated. After trained, the BPNN model could be applied to the simplification of buildings. Compared with the simplification tool of Arc GIS, the BPNN model could achieve better results.Aggregation and simplification are traitionally two individual operators in map generalization. However, if they are implemented simultaneously in the process of map generalization, they are coupled into one operator. The quality of generalization should exceed the traditional way. Consequently, this thesis explores a coupled mechanism of aggregation and simplification, and develops a novel aggregation and simplification combined method based on the nerual network. The results show that this coupled mechanism outperforms the traditional approach.