Pattern Recognition Of Road Network Based On Graph Convolutional Network

As an important part of geospatial database,road network is the foundation of spatial phenomenon representation,spatial information service,and spatial analysis and planning.Road network pattern recognition plays an important role in urban planning,municipal management,road travel decision-making and other fields.Traditionally,the research on this problem is mainly based on geometric measures and statistical analysis methods,lacking effective cognitive inference and intelligent decision-making methods.Road network is a kind of complex spatial network,and the expression of its pattern is determined by the spatial cognition of the cognitive subject.Therefore,it's necessary to make a highly abstract summary of the characteristics of its pattern,which is difficult to obtain results through definite rule reasoning and index calculations,it increases the difficulty of the problem.In recent years,deep learning technology in artificial intelligence has achieved rapid development in simulating the cognitive level of human brain cognition with its unique learning ability.On tasks that rely on human cognitive experience such as face recognition and pattern recognition,deep learning technology has achieved outstanding results through sample training.The pattern recognition of the road network also relies on human cognition,so it is suitable for using deep learning.However,the road network data is a kind of spatial data,which cannot meet the requirements of data standardization for deep learning.In this regard,this paper introduces Graph Convolutional Network(GCN),which is specially used to process graph data,and applies it to solve the problem of road network pattern recognition.The main results of this paper include:(1)The recognition of the multi-pattern mixture of the road network was achieved,and we integrated multiple patterns such as orthogonal grid pattern,radiation pattern,and irregular pattern,and proposed a road network pattern recognition method.At present,the road network pattern recognition methods can only recognize one single pattern,and use specific methods to extract the characteristics of the road network.For example,the orthogonality is extracted when recognizing grid pattern,which lacks scalability and mobility.When there are new patterns to be recognized,it is necessary to redesign new features.This study summarized the road network visual cognitive parameters based on the Gestalt principle,extracted road network features from the perspective of visual cognition,and used Graph Convolutional Network model to realize the recognition of multiple road network patterns.This method is no longer limited to the characteristics of the pattern,it draws on the human visual cognitive process and is closer to the principle of human eye recognition,which can realize the hybrid recognition of multiple patterns.(2)Based on the idea of graph structure classification,we realized intelligent road pattern recognition based on GCN model.This method is similar to image classification.It classifies and recognizes the entire road samples,after the road network samples are modeled as graph,each graph node is equivalent to a pixel unit,and the feature of the graph nodes are equivalent to the RGB value of the image.The classification of road samples is equivalent to the classification of images.This method classifies and recognizes a variety of road network patterns in the GCN model at one time,which changes the strategy that a specific pattern needs to be identified by a specific method.Experimental results show that,compared with machine learning methods such as Gradient Boosted Decision Tree and Random Forest,this method does not need to artificially construct the overall characteristics of road network samples,but only needs to pay attention to the characteristics of its graph nodes,it is an end to end's solution.(3)Based on the decision of graph node classification,deep learning is used to realize the extraction of multi-scale road network patterns.Road network patterns in different scales have different forms,in this study,the road pattern is defined as a local regional structure composed of a series of adjacent units,namely a series of adjacent graph nodes.The segmentation of the road network is understood as the classification of graph nodes.The classification of road segments is realized by classifying graph nodes,and finally the grid pattern of the road network can be extracted..Experiments show that this method can extract grid patterns at multiple scales,and its classification performance is better than that of machine learning methods such as Gradient Boosted Decision Tree and Random Forest.Based on the visual cognition process,this paper summarized the visual cognition parameters of the road network,and used them as the input value of the graph convolutional network to realize the classification recognition and segmentation recognition of the road network.Based on experimental analysis and quality evaluation,the effects of the two models show good performance,the recognition results are at the level of human eye recognition,and they have sufficient distinguishing ability for different models.