High-Confidence Pseudolabeling-Enhanced Multiorder Dynamic Sparse Graph Aggregation Network for Semi-Supervised...

In recent years, graph convolutional networks (GCNs) have received extensive interest in semi-supervised hyperspectral image change detection (HSI-CD), which aims to achieve more robust change feature learning and difference discrimination with the help of the constructed graph structure. However, existing methods typically utilize superpixels to reduce the computational complexity of graph structure construction, while neglecting the accuracy of the constructed graph structure and the existence of complex multiorder relationships in hyperspectral image (HSI). In addition, the insufficient use of unlabeled data further damages the model's superiority. To solve these issues, we propose a high-confidence pseudolabeling-enhanced multiorder dynamic sparse graph aggregation network (HP-MDSGAN) for semi-supervised HSI-CD. Specifically, a multiorder dynamic sparse graph (MoDSG) structure is proposed to simultaneously capture the differences across k-hop neighborhoods, reduce the interference of redundant edges, and adaptively optimize the accuracy of the graph structure relationship. To learn representative change features, a change-aware feature enhancement (CAFE) module is further designed to suppress irrelevant representations while reinforcing those associated with change regions. Finally, a high-confidence pseudolabel learning strategy is introduced to improve the reliability of the generated pseudolabel. Extensive experiments on four hyperspectral datasets demonstrate that HP-MDSGAN outperforms existing HSI-CD methods.