Graph Representation Learning-Guided Diffusion Model for Hyperspectral Change Detection

Due to its capability to monitor subtle changes occurring on the Earth’s surface, hyperspectral images change detection (HSI-CD) has emerged as a focal research area in the field of remote sensing. Recently, diffusion models have demonstrated remarkable performance in the field of HSI-CD. However, vanilla diffusion models are mostly constructed by CNN, which struggles to model global context relationships in complex scenes to result in limited change detection accuracy. In order to overcome the shortcomings about vanilla diffusion models, we innovatively design graph representation learning-guided diffusion model (GDM) and propose the GDM-based HSI-CD network (GDMCD). Specially, we utilize graph convolutional to construct the GDM as the feature extractor, which can adequately extract global difference features of HSIs. Then, we design the difference perception amplification module (DPAM) to increase the distinction between difference features extracted by GDM. Finally, we obtain the change map by classifying difference features which are processed by DPAM. Experiments conducted on three publicly available datasets with $\mathbf {1}\%$ sample size demonstrate that the proposed method outperforms the other state-of-the-art methods in terms of Overall Accuracy (OA), Kappa Coefficient (KC) achieving improvements of approximately $\mathbf {0.006}\%$ , $\mathbf {1.61}\%$ , and $\mathbf {0.34}\%$ , respectively.