Change Detection Incorporating Hypergraph Learning and Hierarchical Interaction Attention

Addressing the challenges of efficiently extracting higher-order features and inadequately explored interaction relationships in change detection tasks, a novel dynamic hypergraph learning architecture based on hierarchical interaction attention is proposed, transforming the generation and analysis of difference maps into a hypergraph learning problem. Specifically, a Feature Extraction module is first constructed to model the features of remote sensing images comprehensively from spatial, density, and spectral dimensions. Subsequently, a Hierarchical Interaction Hypergraph Learning module is designed, introducing hypergraph attention networks, which aware the dynamic interactions between source vertices and target vertex, further learning higher-order semantic information. Finally, a Multi-Level Aggregation Generation module is customized to complete the global multi-semantic fusion at the spatial, density, and spectral levels, thereby achieving the detection of change areas. Experimental results show that the method achieves impressive quantitative and qualitative analysis results on the two remote sensing datasets, thus verifying its robustness and accuracy in change detection.