Title: A Graph Attention Network Method Coupled With Geographical Laws for Land Mass Valuation Under Limited Samples: A Case Study of Shenzhen
ABSTRACT:Constructing continuous urban land value surfaces is critical but consistently challenged by “limited samples.” Existing spatial interpolation and machine learning methodologies often fail due to over-smoothing or insufficient theoretical integration. To address this, we propose Geo-PHS-GAT, a novel Graph Attention Network that systematically integrates the Laws of Geographical Proximity, Heterogeneity, and Similarity. The framework features three key innovations: a composite similarity metric for graph edge definition, a node feature vector fusing parcel attributes with neighborhood contexts, and a Transductive Learning strategy for information propagation from sparse labels. Applied to Shenzhen's residential land market, Geo-PHS-GAT outperformed nine baseline models, achieving the lowest RMSE in both Independent Out-of-Sample Validation (2943.07 RMB/m2 ) and Leave One-Out Cross-Validation (7759.94 RMB/m2 ). Mechanism ablation studies demonstrate that the synergistic coupling of all three geographical laws is essential for preventing value distortion, providing a robust tool for reliable land valuation in data-scarce scenarios.
Keywords: Geo-PHS-GAT | land Price | limited samples | mass valuation of urban land | Shenzhen
DOI: https://doi.org/10.1111/tgis.70462



