SCGNet: Semantic Correlation-Guided Network for Remote Sensing Semantic Change Detection

Semantic change detection (SCD) is an important research task in remote sensing (RS), as it not only identifies changed regions but also characterizes “from-to” semantic transitions, which are of great value in practical applications. However, under multi-task learning architectures, task-coupling interference between semantic segmentation and binary change detection, together with the pronounced intra-class variance and subtle inter-class differences in bi-temporal RS images, poses significant challenges to SCD. This paper proposes a new SCD method called Semantic Correlation Guided Network (SCGNet) to overcome the mentioned limitations. It consists of a semantic branch and a change branch, between which continuous information interaction is established to enable collaborative information utilization and joint optimization. SCGNet integrates three key modules: a Semantic-Guided Fusion (SGF) module exploits semantic relevance cues from the semantic branch to guide adaptive bi-temporal feature fusion in the change branch; a Semantic Consistency Enhancement (SCE) module introduces invariant information extracted from the change branch to stabilize bi-temporal semantic representations and enhance overall semantic consistency; and a Multi-Scale Change Activation (MSCA) module activates and refines change features under semantic constraints across multiple scales, further strengthening the collaboration between semantic information and change responses. By coupling semantic reasoning and change perception through a continuous information interaction mechanism, SCGNet enables more reliable semantic change modeling. Positive experimental results counted on two benchmark datasets show that SCGNet outperforms many existing methods, confirming its effectiveness. Our source code is available at https://github.com/TangXu- Group/Semantic-Change-Detection/tree/main/SCGNet.