Abstract
<title>Abstract</title> <p>Current classification of rural settlements usually stays at the administrative-village scale. This makes it difficult to reveal structural differences and functional differentiation among settlement units within the same village. Taking Dangyang as a case study, this study develops a method to identify and classify fine-grained settlement units based on a graph neural network (GNN). Settlement units are treated as network nodes. Spatial links are treated as edges. Multi-source attributes are integrated with spatial topology to represent and embed the non-Euclidean spatial relations of rural settlements. Network metrics are then used to describe the structural position of each settlement. SHAP contribution analysis and feature correlation analysis are used to identify the key drivers of structural differentiation. The results show that: (1) GNN embedding markedly improves the recognition of link structures at different distance levels, reduces class confusion, and strengthens network representation; (2) the embedded settlement network shows an axial pattern, with urban and township areas as cores and major transport corridors as the main extension direction, and nodes with high comprehensive network scores are concentrated in the urban-rural transition zone; (3) network metrics are mainly controlled by settlement size and supporting facilities, population size plays a reinforcing role, and agricultural vitality and ecological-cultural factors mainly affect local clusters and edge structures; and (4) fine-grained settlement units can be mapped into the clustered development type, urban-rural integration type, heritage conservation type, improvement type, and relocation and consolidation type according to the network metrics and their contribution mechanisms. This study breaks the scale constraint of administrative villages. It provides a new method and empirical support for refined classification and differentiated governance of rural settlements.</p>