Answers are grounded in this paper and cite their sources.
Demonstrates improved building matching accuracy through edge-augmented graph modeling.
Integrating heterogeneous building footprints is essential for Digital Earth applications, yet cross-dataset representations often contain positional offsets, shape distortions, and many-to-many correspondences. Conventional geometry-driven methods rely on manually tuned thresholds, whereas learning-based approaches usually classify candidate pairs independently and neglect structural constraints. We formulate building matching as edge-probability inference on a bipartite candidate graph and propose the Edge-Augmented Graph Transformer, EAGT. Candidate pairs are generated by bounding-box overlap, with overlap similarity used as the only geometric feature. EAGT injects standardised overlap similarity into attention scoring and value aggregation, jointly modelling minimal geometry and local graph structure on candidate edges. During inference, Max-Gap adaptive pruning and connected-component decoding identify 1:0, 0:1, 1:1, 1:n, m:1, and m:n relationships in a unified pipeline without a manually tuned global threshold. On Zhengzhou, Hamburg, and Kampala datasets, EAGT obtains overall F1-scores of 97.17%, 99.36%, and 97.66%, and improves m:n F1 by 20.96, 36.39, and 25.97 percentage points over the tree-constrained optimisation baseline. Inference requires only 2.7 s, 4.1 s, and 3.2 s, indicating a weakly parameterised, accurate, and efficient solution for multi-source building matching and spatial data integration.
No takes yet. Share an insight, caveat, or question.
Qiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: