Bridging the semantic gap in urban land use change detection with object-centric graph learning

A fundamental challenge in detecting land-use changes from high-resolution remote sensing images lies in the semantic gap between land cover and land use: imagery captures biophysical surface attributes through electromagnetic responses, but does not directly reveal socioeconomic functions. This gap is particularly pronounced in urban environments, where land-cover methods often fail to infer land-use semantics. Moreover, many deep learning methods rely on fixed-granularity image-scene analysis units, which exacerbate geometric and attribute uncertainties, causing imprecise boundaries, mixed-use ambiguity, and degraded performance in fine-grained urban land use change detection. To address these challenges, we propose a semantic-aware multi-spatial domain interaction network (SMDI-Net) that reformulates land-use change detection as a graph learning problem by modeling spatial structural differences of land-cover objects. By treating these objects as graph nodes and their topological relationships as edges, SMDI-Net establishes an explicit mapping between remote sensing imagery and land-use functional semantics, while precisely delineating change boundaries and maintaining semantic consistency within analysis units. Building upon this formulation, SMDI-Net further incorporates a multi-spatial domain interaction encoder to model land-cover dependencies across spatial scales, a hierarchical structural integration module to strengthen the identification of subtle structural transitions, and a discriminative feature enhancement module to characterize spatial structural shifts and semantic variation for refined change localization. Experiments on three datasets validate the superiority of SMDI-Net over seven state-of-the-art methods, achieving F 1 scores of 93.32% for binary and 88.78% for multi-class change detection. These results show that SMDI-Net effectively alleviates geometric uncertainty and mixed-use ambiguity, offering a robust solution for fine-grained urban land-use monitoring. The code is available at https://github.com/Imp83121/SMDI-Net .