Research On Method Of Pattern Recognition In Urban Street Networks

Street network is one of the most foundational elements in map and the basic component of geospatial database.Meanwhile,street network plays an important role in urban spatial positioning,which provides essential spatial data support for location based service,social network service,navigation,automatic driving,et al.The data of street have increased significantly because of the rapid expansion of city and the quick development of spatial data acquisition technology.Pattern recognition in street networks can help to understand the urban structure intuitively and quickly from lots of data,and can discover the implicit rules and knowledge in the roads data,which is significant for the map generalization and multi-scale expression of spatial data.This paper mainly focuses on recognizing important pattern in street networks by the method of machine learning because of the advantage of street network pattern recognition above.The main work of the dissertation are as follows:(1)Analyze the theory of street network pattern including the formation,classification and characteristics of street network pattern and point out the advantage of using road mesh as object to recognize street network pattern;study the relevant theories of machine learning algorithms used in this dissertation.(2)Study the geometric appearance and spatial arrangement characteristics of streets in grid pattern,and use specific parameters to describe the characteristics of road mesh in grid pattern.A hybrid model combining lightGBM and logistic regression is designed.Use LightGBM and hybrid models individually to recognize the street grid pattern in cities with different street network features.Compare results of different models in different cities and analyze the factors affecting the recognition results.The comparative experiment shows that lightGBM and hybrid model are superior to traditional machine learning model in grid pattern recognition.(3)Study the characteristics of multi-lane and overpass.Analyze the geometrical and spatial distribution characteristics of road meshes and design the parameters that express these features.An adaptive clustering method for street functional structure pattern recognition using Self-organizing map(SOM)is proposed because the street functional structure patterns may differ in different cities.The experiment shows that the method can recognize multi-lane and overpass accurately and comprehensively.