EGAFNet: Edge-Guided Adaptive Fusion Network With Spatial–Frequency Interaction for Remote Sensing Change Detection

Change detection (CD) is a pivotal technique for revealing surface differences within the same area across different time periods, serving as a crucial means for monitoring interactions between humans and nature. However, existing CD methods struggle to balance global semantics and local details across multilevel features, leading to blurred boundaries and missed detections within changed regions. Moreover, insufficient fusion of bitemporal features further results in pseudo changes in background areas. To address the aforementioned issues, this paper proposes the Edge-Guided Adaptive Fusion Network (EGAFNet) based on spatial-frequency interaction. Two parallel modules, namely the Frequency Attention Module (FAM) and the Multiscale Perception Module (MPM) are introduced to provide complementary enhancement of global and local information from the frequency domain and spatial multiscale perspectives, respectively. The Adaptive Decoupled Fusion Module (ADFM) is employed to progressively interact with bitemporal features, fully aligning spatiotemporal information and achieving deep fusion by leveraging multilevel attention. Subsequently, based on the interaction of parallel path features, the Edge Driven Wavelet Reconstruction Module (EWRM) is utilized to achieve fine-grained spatial frequency modeling of edge regions. Comparative experiments on three CD datasets (SYSU-CD, JLYH-CD, and SECOND-CD) demonstrate that the proposed method outperforms existing state-of-the-art methods, achieving F1-scores of 84.61%, 78.66%, and 73.01%, respectively. All codes are available at https://github.com/wpeng0166/EGAFNet.