S 2 LDM: Synergizing Local Decomposition and Structural Memory for Remote Sensing Change Detection

Remote sensing change detection (RSCD) is a key technology to analyze the spatio-temporal change information of land surface, which is of great significance in urban expansion monitoring and disaster assessment. However, existing remote sensing feature extraction methods rely on traditional fusion mechanisms, which lack the ability to perceive boundary shifts and dynamic changes of small objects, and are easy to lose important long-term dependencies in the process of structure guidance. To address these limitations, this paper proposes a RSCD change detection network combining local decomposition and structural memory. It integrates the Local Structure Decomposition Convolution (LSDC) module and the Gated Key Pixel Structure Memory (GKSM) module. The former is injected into the residual network, and the large receptive field is used to dynamically guide the local convolution weights to realize the enhanced feature representation of global-local complementarity. Through the selective feature enhancement mechanism of cyclic update, the GKSM model prevents excessive flow or loss of information while maintaining efficient cross-level feature interaction. Extensive experiments demonstrate the effectiveness of the proposed S2LDM. The proposed method achieves state-of-the-art performance on the three datasets, and exceeds the existing comparison methods in the accuracy index.