A Novel Consistency Fusion Approach on Topological Manifold for Urban Building Clustering

ABSTRACT

Urban building clustering is essential for deciphering spatial structures and functional patterns. Yet, existing methods mainly rely on local or pairwise geometric similarity measures, which fail to capture the latent nonlinear topological structure of building layouts and insufficiently handle structural inconsistencies across multi‐view data, leading to compromised robustness. To address these issues, this study proposes a novel multi‐view clustering model that, for the first time, integrates topological manifold modeling and cross‐view diversity detection in a unified framework. This model first constructs three complementary feature views (centroid distance, outline distance, and non‐spatial attributes) and models latent topological manifolds within each view. A novel diversity detection mechanism is then proposed to identify and suppress inconsistent structural information both within individual views and across different views, producing a pure graph for each view, which is then fused into a consensus graph with a clear clustering structure. Finally, an alternating iterative optimization algorithm is proposed to jointly learn topological correlation, multi‐view consistency, and consensus structure within a unified framework. Extensive experiments on 15 urban communities in Wuhan and Chengdu, China, show that our model consistently outperforms 12 state‐of‐the‐art baselines, achieving up to 32% improvement in Adjusted Rand Index (ARI) and over 20% gains in Accuracy and F ‐score. In particular, linear patterns achieve an average accuracy of over 90%.