Research On Automatic Selection Of Road Network Based On Depth Graph Convolutional Neural Networks

The road plays an important role as a key element in the map,both in social life and in the military sphere.Due to the difference of road shape,the diversity of level and type,the complexity of spatial structure and relationship,the problem of automatic selection of road has become the focus and difficulty in cartographic generalization.Therefore,the study of automatic selection of road networks has important theoretical and practical implications.From the theoretical level,the research on the automatic selection of road network will be beneficial to enrich the theory of automatic cartographic generalization method,and at the same time,it can provide new ideas for the automatic selection of other line elements in the map.From a practical point of view,the automatic selection of the road network can not only reduce a lot of human and material resources,improve the efficiency of map production,but also improve the speed of map updating to meet the more complex and changing needs of the people.At present,the road network selection is less automated and relies more on human subjective factors.Automatic selection based on graph theory requires an in-depth understanding of the interactions between the various features of a road and the human formulation of selection rules.The traditional neural network-based approach requires artificial extraction of spatial information about the road,greatly increasing the complexity and subjectivity of the selection process,resulting in spatial information not being fully utilized.Graph convolutional networks(GCNs)effectively combine graph theory with neural networks to automatically extract spatial information about a road through graph convolution,reducing the process of artificially constructing spatial features and enabling end-to-end learning on the graph.Based on the fact that the nature of road network is also a graph data structure consideration,this paper translates the road network selection problem into a graph node classification problem and applies GCNs for the automatic selection of road networks.The main studies in this paper include:(1)Depth of the graph convolutional networks.The over-smoothness of GCNs allows the superimposed convolutional layers to be generally not too deep,which greatly limits the ability of GCNs to exploit road space information.In this paper,we study the depth methods based on Jumping Knowledge Networks(JK-Nets),residual connections and dense connections,and analyze the effectiveness and applicability of these depth methods to the road network selection problem through comparative experiments.(2)Construction of the model.With the rapid development of GCNs in recent years,numerous graph convolution methods have been derived.In this paper,we first introduce the basic principles of various graph convolution methods from the perspective of graph signal processing.The architecture of the model was then designed based on the graph convolution principle and the depth strategy,and different depth graph convolution models were implemented,trained,predicted and evaluated in the experimental phase.(3)Evaluation of selected results.This paper evaluates the prediction results of the model in two ways.One is to evaluate the generalization ability of different models,considering the unbalanced distribution of positive and negative samples,this paper uses AUC(area under ROC Curve),which is not affected by the distribution of sample categories,as an evaluation indicator.The second is to evaluate the reasonableness of the selected results,using the expert's selected results as a criterion to calculate evaluation indicators such as the line density and accuracy of the road.The results show that the graph convolution not only takes advantage of the existing semantic geometric features of the road,but also extracts the spatial information of the road effectively,thus making more accurate predictions when the road has less information.The depth approach allows the selected model to overlay deeper layers of the convolutional accumulation to extract a wider range of spatial information.In this paper,a comparative test was conducted on a small-scale road network with different depth convolutional network models,in which graph attention network(GAT)has better selection results than other models,and JK-Nets has more significant improvement effect compared with other depth methods,with the highest accuracy rate of 88.12%.