A Graph-Based Hyperspectral Change Detection Framework Using Difference Augmentation and Progressive Reconstruction With Limited Labels

Identifying land cover changes based on hyperspectral images (HSIs) has been a research hotspot in the field of remote sensing (RS). In recent years, deep-learning (DL)-based change detection (CD) methods have advanced the development of this subject due to their powerful feature representation capabilities. However, it is difficult for these methods to mine changed information between bi-temporal HSIs with limited labels. To overcome this limitation, we propose a graph-based hyperspectral CD framework using difference augmentation and progressive reconstruction (ARCD), which enhances the recognition ability of changes in HSIs with limited labels. This framework consists of three components: 1) a dual-branch multiscale dynamic graph convolutional network (DMGCN) subnetwork, which is developed to emphasize the changed information and learn global features of HSIs at various scales; 2) a difference augmentation feature fusion (DAFF) module, which is designed to fuse spectral-spatial augmentation information and the difference information to accurately capture discriminative features for the changes between bi-temporal HSIs; and 3) a progressive contextual information attention reconstruction (PCAR) module, which is proposed to focus on key information in the context, and progressively reconstruct multilevel features to reduce semantic gaps between different scale features. ARCD not only enhances the representation ability of changed features but also alleviates the demand for HSI labels. We test the performance of ARCD on four hyperspectral datasets. Quantitative and qualitative results reveal that it outperforms some state-of-the-art (SOTA) methods with limited labels.