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

Area of Science:

• Geographic Information Science
• Urban Studies
• Computer Science

Background:

• Urban functional zone (UFZ) classification is crucial for urban planning and management.
• Existing methods often overlook the complex interplay of local and long-range spatial dependencies in UFZ formation.
• There's a need to differentiate and integrate these spatial interactions for more accurate urban analysis.

Purpose of the Study:

• To develop a novel framework, the Graph Adaptive Propagation Network (GAPN), for enhanced UFZ classification.
• To effectively integrate multi-scale features from Points of Interest (POIs) and street-view imagery.
• To address the limitations of existing methods by adaptively balancing local and long-range spatial dependencies.

Main Methods:

• Proposed a dual-branch Graph Adaptive Propagation Network (GAPN) architecture.
• Integrated POI data and street-view imagery by adaptively aggregating multi-scale features on a shared graph.
• Employed a cross-modal gating mechanism for selective modality fusion and a graph-diffusion attribution scheme for interpretability.

Main Results:

• Achieved high accuracy in UFZ classification in Qingdao (0.840 OA, 0.790 Kappa), outperforming baseline methods.
• Demonstrated the framework's robustness and applicability in a different urban context (Beijing, 0.826 OA).
• The graph-diffusion attribution scheme provided interpretable insights into multi-hop contributions.

Conclusions:

• The proposed GAPN framework significantly improves urban functional zone classification by effectively modeling multi-scale spatial interactions.
• The adaptive balancing of local and long-range dependencies and multi-modal fusion are key to the framework's success.
• GAPN offers a promising and generalizable approach for precision urban management and planning.