The Cluster Of Buildings Based On The SOM Neural Network

With the rapid development of urbanization, the demand of electronic map is also growing up rapidly in many areas. It’s a challenge for the traditional manual cartographic generalization to accelerate its producing cycle to meet the heavy demand of electronic map. Therefore, achieving automated cartographic generalization is important to fulfill the needs of the community. The clustering of vector map is an important part of cartographic generalization, which is the key step to solve building simplification problem when the map scale is transforming. Aiming to propose a more accurate clustering method to give a solid support for automated cartographic generalization, this thesis researches on automated clustering of the main element- the building polygon in urban map.The development of computer technology promotes the automation of map features clustering. These researches implement cluster result by using the algorithm analysis, software development and other steps on map’s graphical analysis and transformation. They are the main trends in the process of clustering steps. Based on the principle of SOM neural network algorithm, this thesis improves the barycentric coordinates clustering method(Boyan Cheng’s method). In the process of improvement, this thesis proposes a plurality of features which describe buildings’ shape and position, named as building factors. After first large-scale clustering by using SOM algorithm, a secondary refined clustering with similar factors it is made on the basis of the previous clustering result. On the basis of the original division of building data grouped by barycentric coordinates, it continues to group for the second-time with similar factors. This process is also called buildings’ second refined clustering.Through the secondary clustering comparative experiment for different areas, it has made a conclusion that different factors and forms of distribution make different map clustering effects. In another word, there does not exist a single building factor which works appropriately in all circumstances, because of the differences of building’s shapes and distributions. Therefore, this thesis uses factorial analysis and PCA(principal component analysis) to reduce the dimension, to explore the disadvantages of using comprehensive factors instead of different factors. The result of the experience shows that, although the method does not have a more precise effect than that of a single i with single map, it can meet the basic needs of clustering of various types of maps very well. Finally, this thesis explores the building clustering on 3D maps. By using the similarity of elevation data and barycentric coordinates, it conducts several experiments of building clustering on 3D maps, then visuals the results.