A road generalization method using graph convolutional network based on mesh-line structure unit

Road network simplification is a complex decision-making process. Such a multi-factor decision and scaling operation traditionally applied rule-based methods. The establishment and adjustment of these rules involve many human-set parameters and conditions, which makes generalized results closely related to the cartographer's experience and habits. On the other hand, existing methods tend to consider individual structures separately in different algorithms, such as strokes, meshes and graph networks, lacking a solution that brings the advantages of these methods together. Aiming at the above problems, this study designs a simplification method using the Mesh-Line Structure Unit (MLSU) to consider polyline and polygon characteristics simultaneously with the support of graph-based deep learning networks. In order to make generalization decisions, a model based on graph convolutional network (GCN) is constructed and trained using real data, thus realizing the road network selective omission. The experimental results indicate that the proposed method effectively achieves automatic road generalization. The proposed method uses graph convolutional neural network techniques to construct a road generalization model, and can effectively combine the advantages of geographic domain knowledge with data-driven methods.A new specific MLSU structure is proposed for the road generalization tasks, which combines a road mesh with the road itself, enabling it to capture more road-related features and substitute the road in deep learning network model for training.The road generalization approach proposed in this paper comprehensively considers the roads themselves, the road network, and the neighbouring mesh polygons, thereby combining the advantages of traditional methods based on graph theory, strokes and mesh merging.