Change Detection on Remote Sensing Images Using a Dual-Branch Architecture With State Space Modeling and Graph-Based Local Interaction Network

In recent years, with the rapid development of remote sensing technology, the acquisition capability of high-resolution images has been significantly improved. However, change detection (CD) in complex scenes still faces multiple challenges. Traditional methods are often sensitive to seasonal variations and illumination fluctuations, which limits their ability to model global contextual dependencies and effectively suppress background noise. Moreover, local feature interaction mechanisms frequently fail to preserve subtle bitemporal differences due to inefficient cross-scale coupling, resulting in boundary blurring and false detections in multi-scale target scenarios. In addition, many existing CD models rely on computationally expensive architectures, which hinders their practical deployment. To address these issues, we propose DSGNet, a dual-branch change detection network that jointly exploits state space modeling and graph-based local interaction. The core design philosophy is to decouple global semantic dependency modeling from fine-grained local change representation, and to integrate them efficiently through a guided attention–gated fusion strategy. The global branch introduces a Directional State Space Model (DSSM), which extends Mamba and Change State Space Models (CSSM) to explicitly encode directional change trajectories across bitemporal images via state-driven dynamics, thereby enhancing the representation of large-scale structural variations. Meanwhile, the local branch incorporates a graph-based interaction module that dynamically constructs a learnable topological structure to facilitate explicit bitemporal feature alignment, significantly improving sensitivity to small-scale and subtle changes. In the fusion stage, a Guided Attention-Gated Feature Fusion (GAFF) module adaptively balances global and local representations through a gating mechanism, which reinforces true change responses while effectively suppressing irrelevant background fluctuations. Experimental results demonstrate that DSGNet achieves superior performance on the WHU_CD, SYSU_CD, and DSIFN datasets, with F1-scores of 95.20%, 85.69%, and 77.39%, respectively.