Research On Adjacent Scale Road Data Matching Based On Graph Neural Network

Roads are the vital infrastructure of a country and plays an important role in the development of national economy.Road network data as an important geospatial data type is widely used in intelligent transportation system(ITS),location-based service(LBS)and other fields.With the development of social economy,application departments in various fields have higher and higher requirements on the quality and currency of road data.It is necessary to update the road data network in order to maintain the current trend of road data.However,the road data of different scales are stored separately and hierarchically,and the data of different scales need to be updated separately in the process of data update.Such update method obviously takes a lot of time,which seriously affects the currency of geospatial data.Therefore,the association of multi-scale data will contribute to the efficient utilization of geospatial data,alleviate the contradiction between the rapid development of social and economic development and the reduction of the currency of geospatial data.At present,the road matching algorithm has low matching accuracy and low automation degree.The process of matching often depends on the influence of subjective factors such as similarity weight and matching threshold.In order to improve the matching accuracy and automation of road data,Inspired by the graph expression ability of graph neural network,a road matching model of adjacent scale road data graph neural network is proposed in this paper.This method can effectively utilize simple geometric features without constructing complex similarity measures,and can learn potential rules in geospatial data from training data,thus achieving adaptive road matching.The main research of this paper includes:(1)Construction of dual graph and feature extraction of road dataModeling vector linear road data as graph data is the basis of road matching model based on graph neural network.This paper presents a new dual graph construction method for road data.The road is divided into M sub-roads,each sub-road is abstracted as nodes in the dual graph,and the connection relation and proximity between sub-roads are abstracted as edges in the dual graph.This dual graph construction method can better describe the morphological characteristics of line elements and facilitate the use of curve feature description as node features in the graph.The nodes in the dual graph represent the sub-paths in the road.This paper describes the morphological characteristics of the sub-paths from three aspects: distance,direction and shape characteristics,and defines the characteristics of nodes in the graph.(2)Road data matching based on graph embedding modelIn this paper,a graph embedding model is designed by combining siamese network and graph neural network to judge whether vector road pairs match or not.Graph embedding model uses two sub-networks of siamese neural networks to extract features from road pairs,and then the road pairs are independently embedded into a high dimensional graph level vector,and then the two graph level vectors are combined and passed to the full connection layer for matching prediction.However,due to the lack of node and graph level interaction,the graph embedding model can only compare the similarity between pairs to be matched on the whole,which may lose some fine-grained features.(3)Road data matching based on graph matching modelBecause the graph embedding model lacks node-level feature comparison information,the graph matching model can capture fine-grained features more effectively by adding graph-level interaction and cross-graph node comparison on the basis of the graph embedding model.GMN calculates the similarity between nodes across graphs through attention mechanism in node embedding stage,Sim-GNN realizes node-level comparison across graphs by computing node similarity matrix after node embedding.The similarity between nodes represents the similarity between the sub-road of two road.Therefore,the graph matching model incorporating the comparison between nodes has higher matching accuracy than the graph embedding model.The results show that the graph neural network can extract spatial information of road effectively,and the road matching model based on graph neural network can make accurate prediction only according to some simple geometric features of road data.Compared with the traditional road matching algorithm,the model proposed in this paper can avoid all kinds of complex similarity measures,avoid the setting of weights and matching thresholds under human intervention,and realize the adaptive road matching,which greatly improves the automation and matching accuracy of road matching.In this paper,matching experiments are conducted on 1:100000 and 1:250000 scale road data of various cities and counties in Henan Province,The accuracy of the graph embedding model reached 92.75%.In the experiment of comparing the graph embedding model with the same type of graph matching algorithm,the accuracy of graph embedding model and SVM is basically the same on the training data set,but the generalization performance of SVM is better than graph embedding model,the accuracy of the graph embedding model is3.5% higher than that of the MLP.In this paper,the graph matching algorithm comparison experiment,the accuracy of the graph matching model is higher than that of the graph embedding model,and Sim-GNN is better than other models with an accuracy of 95.50%.