Weakly-Supervised Trend Change Detection Network Based on Mamba and SoftMatch Distance

In the field of remote sensing imagery, general change detection (GCD) plays a crucial role by focusing on identifying binary changes (i.e., change or no change). However, such simple binary information often falls short of providing detailed and actionable insights. To address this limitation, researchers have developed semantic change detection (SCD), which extends GCD by providing semantic categories for the detected changes. While SCD offers richer information, it comes with the drawback of high annotation costs. To balance the trade-off between GCD and SCD, a novel approach called trend change detection (TCD) has been introduced. TCD classifies changes detected by GCD into three distinct trends: "Appear," "Disappear," and "Transform." This approach enhances the richness of detected information while significantly reducing annotation costs compared to SCD. In this paper, to further minimize annotation costs, we propose a weakly-supervised TCD method, which performs the TCD task under the binary change label supervision of GCD. The method utilizes an encoder-decoder architecture to extract features and employs SoftMatch distance to effectively distinguish the foreground from the background. Experimental results on two public datasets, SVCD and LEVIR-CD, demonstrate the superior performance and effectiveness of the proposed method in addressing the weakly-supervised TCD task.