Identifying urban villages using graph neural network and transfer learning in Southern Jiangsu, China

Abstract

China’s rapid urbanization has led to the emergence of urban villages (UVs), whose governance is crucial for urban quality and built environment management. However, conventional identification methods rely on high-resolution remote sensing images and extensive localized annotated data, inadequately supporting rapid surveys over large areas. Therefore, this study developed an efficient framework for UV identification. Initially, point of interest, real estate, and building footprint data were integrated to construct a comprehensive feature system of urban function, building properties, and 2D and 3D morphologies. Subsequently, using street blocks as spatial units, the GraphSAGE model achieved deep socioeconomic and morphological feature aggregation. Finally, transfer learning (TL) enabled efficient cross-regional knowledge transfer through domain adversarial training. Results revealed that the Southern Jiangsu region comprised 705 UVs in 2020, covering 167.49 km2, with obvious variations in number, scale, and distribution across cities. With Nanjing as the source domain, the GraphSAGE model achieved recognition accuracy of 93.0%. In target domains of Zhenjiang, Changzhou, Wuxi, and Suzhou, the fused TL model achieved accuracies of 79.3%, 87.3%, 83.3%, and 85.5%, respectively. Per-feature Wasserstein distance analysis attributed the cross-city performance disparity to morphological heterogeneity, social-sensing feature misalignment, and graph topology differences, with Zhenjiang exhibiting the largest domain shift. The model converged with only 50% labeled samples, reducing training time by approximately 40%. Ablation and comparative experiments further verified the efficacy of integrating social sensing and building footprint data. The proposed framework provides a technical reference for cross-region, small-sample, large-scale UV identification, supporting decision-making in scientific urban management and planning.