GDSN-CD: Graph-Guided Diffusion Synergistic Network for Remote Sensing Change Detection

Remote sensing change detection often faces the challenge of accurately capturing discontinuous change features in complex scenes. Traditional methods struggle to balance spatial structure integrity and feature representation accuracy when dealing with irregular changes, such as fragmented urban expansion or disaster damage. To address this issue, this article innovatively combines the advantages of graph convolutional networks and diffusion models. First, a dynamic graph structure is created using a weight-sharing graph convolutional networks feature encoder, transforming discrete changes into topological relationships between nodes. Then, the powerful feature generation capability of the diffusion model is leveraged to optimize the representation of change features during the denoising process. Based on this approach, we propose an end-to-end GDSN-CD framework. To overcome the problem of semantic and noise feature mismatch that may occur during the denoising process of diffusion models, we innovatively design a semantic-noise collaborative module, which uses a grouped channel fusion strategy to achieve precise feature matching. Furthermore, we propose a hierarchical dynamic fusion module that adaptively fuses multilevel features to coordinate the complementary relationship between texture and semantics; and design a dual-feature dynamic selection module that optimizes the fusion of differential and additive features through adaptive weighting, accurately enhancing change signals while suppressing background interference. Experimental results show that GDSN-CD significantly outperforms mainstream methods on three public datasets, especially in fine-grained change detection and anti-interference capability, demonstrating strong potential for practical applications.