Objectives The calculation of spatial similarity relationship of road networks has wide applications in the fields of spatial data matching and querying, spatial data multi-scale expression and evaluation, and quality evaluation of cartographic generalization. However, the existing similarity calculation methods of road networks have the problems of insufficient utilization of spatial information and subjective setting of the weights of feature factors.

Methods This paper proposes a road network similarity calculation model based on graph convolutional autoencoder (GCAE). The proposed model employs a self-supervised training strategy to achieve end-to-end learning and strives to reconstruct the original input graph as accurately as possible. As a result, it effectively mitigates the limitations of conventional methods and further improves the accuracy of similarity computation. First, a dual graph of the road network is constructed by converting intersections to edges and edges to nodes. Following the principle of whole-part structural relationships, spatial feature information of the road network is assigned to the dual graph nodes from three aspects, including global, local, and connectivity features, thereby obtaining a quantitative representation of the road network graph structure. Then, GCAE is employed to aggregate and update the node features and structural information of the road network graph, forming a deep understanding of the road network. This process yields a feature encoding of the spatial information, achieving a quantifiable representation of the road network structure. Finally, the complex high-dimensional feature space is mapped by average pooling to a low-dimensional, easily measurable space, generating a set of feature vectors. Cosine similarity is then applied to compute the similarity between road networks.

Results The experimental results demonstrate that the similarity values generated by the proposed model exhibit high sensitivity and consistency, align closely with actual road network changes, and correspond well with human spatial cognitive patterns.

Conclusions The proposed model offers a high degree of automation and yields intuitive, well-founded results, effectively mitigating the need for handcrafted features and rule-based designs in road network similarity calculation.