Spatial-Temporal Feature Interaction and Multiscale Frequency-domain Fusion Network for Remote Sensing Change Detection

Remote sensing change detection is fundamental to monitoring Earth's surface dynamics. Deep learning-based remote sensing change detection approaches have achieved remarkable advancements in recent years. However, accurately identifying change regions under complex backgrounds and various interference factors remains a challenging task. Existing methods fail to effectively explore the interaction between bitemporal features. In addition, the modeling of difference features remains insufficient. To overcome these limitations, we propose a novel network (SIMFNet) based on spatial-temporal feature interaction and multiscale frequency-domain fusion. First, the multilevel bitemporal features are extracted through a Siamese hierarchical backbone. Subsequently, we design a spatial-temporal feature interaction module (STFI) to establish cross-temporal semantic relations between the pre-change and post-change images. Following this, a multiscale frequency-domain fusion module (MFDF) is employed to fuse bitemporal features. This module combines spatial and frequency information to effectively extract and refine difference features. Finally, considering the semantic disparities among different feature levels, the inter-level difference feature aggregation module (IDFA) is proposed to adaptively aggregate inter-level contextual information. To validate the efficacy of the proposed SIMFNet, extensive experiments are conducted on three public RSCD benchmark datasets: LEVIR-CD, WHU-CD, and SYSU-CD. The results consistently show superior performance compared to existing state-of-the-art methods.