DSDGNet: Dual-Scale Dynamic Graph Network for Hyperspectral Image Change Detection

Hyperspectral image (HSI) change detection (CD) aims to identify land-cover and land-use changes from bitemporal hyperspectral observations. Although convolutional neural networks (CNNs) have achieved remarkable success in HSI-CD, and graph convolutional networks (GCNs) have demonstrated strong capability in modeling non-Euclidean relationships, both still suffer from inherent limitations that lead to incomplete feature representation. Specifically, CNNs rely on fixed local receptive fields, which restrict their ability to capture long-range contextual dependences, while GCNs typically construct graphs based on superpixels, inevitably losing fine-grained information due to region aggregation. To address these issues, we propose a dual-scale dynamic graph network (DSDGNet) for HSI-CD, which jointly models global structural relationships and pixel-level details in a unified framework. The proposed DSDGNet mainly consists of three components: a dual-scale dynamic graph (DSDG) module, a multiscale dilated convolution (MSDC) module, and a cross-guided fusion (CGF) module. Specifically, the DSDG module constructs graph representations at two different superpixel scales and dynamically updates graph connections to capture multiscale structural dependences. The MSDC module extracts discriminative pixel-level features using multiscale dilated convolution, effectively enlarging the receptive field while maintaining computational efficiency. Furthermore, the CGF module performs cross-guided interaction between the two branches, enabling complementary feature refinement and enhancing both global context and local discriminative representations. By integrating dual-scale graph modeling and multiscale convolutional feature learning, the proposed method achieves a more comprehensive representation of hyperspectral changes. Experimental results on three benchmark datasets demonstrate that DSDGNet consistently outperforms several state-of-the-art methods.