An urban functional zone classification framework based on graph adaptive propagation network

Urban functional zone (UFZ) classification is fundamental to urban planning and precision management of cities. Urban functions emerge from multi-scale spatial interactions, where both local dependencies and long-range structural associations coexist with varying importance. However, existing studies typically focus on local dependencies, while only a few consider long-range structural associations between UFZs, and even these fail to differentiate the two spatial interactions, thereby missing their heterogeneous influence patterns. To address this, we propose a Graph Adaptive Propagation Network (GAPN) utilizing a dual-branch architecture to integrate POIs and street-view imagery. GAPN adaptively aggregates multiscale features on a shared graph, learning weights to balance local and long-range dependencies, before selectively fusing modalities via a cross-modal gating mechanism. Additionally, a graph-diffusion attribution scheme ensures interpretability by quantifying multi-hop contributions. Evaluations in Qingdao's Shinan District achieved an accuracy of 0.840 (0.790 Kappa), outperforming baselines, while retraining and validation in Beijing's Chaoyang District (OA = 0.826) confirmed the framework's applicability to other urban contexts.