An Enhanced Graph Convolutional Network With Multispatial Relationships for Urban Village Identification

Abstract

Precise identification of urban villages is a key link in enhancing urban governance efficiency and an important foundation for optimizing the allocation of spatial resources. However, existing urban village identification methods often overlook the complex spatial relationships between geographical entities during modeling, limiting the model’s ability to understand spatial patterns and interactions. This article proposes an enhanced graph convolutional network with multispatial relationships (GCN_MSR) for urban village identification. Using street blocks as basic units, this method comprehensively integrates multidimensional features such as socioeconomic, land use, physical space, and urban visual characteristics to construct a graph structure that integrates multispatial relationships among geographic entities, including spatial proximity, spatial similarity, and spatial heterogeneity. On this basis, a relationship-specific attention layer and a node-level relationship fusion module are designed, enabling the model to adaptively learn the contributions of different spatial mechanisms to node classification. Experimental results show that the GCN_MSR achieves an $F1$ -score of 0.89, outperforming models using only single or dual relationship combinations. Further interpretability analysis shows that spatial similarity relationships contribute the most to the overall identification, followed by spatial heterogeneity relationships, while spatial proximity relationships are relatively weak. The study results indicate that comprehensively and diversely characterizing spatial relationships in geospatial analysis plays a positive and important role in constructing high-performance, interpretable geographic artificial intelligence models, providing new ideas and empirical support for solving complex geospatial problems using deep learning methods in the future.