The road network similarity calculation based on autoencoders and its application in map generalization quality evaluation

Spatial similarity underpins multi-scale data representation and cartographic quality assessment. However, evaluating the similarity of structurally intricate road networks using conventional approaches often falls short due to inadequate spatial feature extraction and subjective weighting. To address this, we introduce a graph convolutional autoencoder (GCAE) model for road network similarity computation. The model learns deep spatial representations via self-supervised training, combining graph convolutions for feature extraction with autoencoders for graph reconstruction. Similarity is quantified using latent-space cosine similarity and validated against geometric and topological baselines, as well as human perception. Furthermore, we construct a multi-scale road network standard library to mathematically model the relationship between similarity and scale variation, serving as a benchmark for cartographic quality evaluation. Applied to Chengdu and Hefei, the framework assesses map generalization quality via deviation and evaluation-set methods. The results align closely with geographic evolution and human judgment, delivering objective, robust metrics that significantly enhance the intelligence of automated map generalization.