作者
Zhiyuan Yang,Jindong Xu,Mengying Ni,Menghui Su,Jiantao Peng
摘要
Change detection (CD) is a fundamental yet challenging task for remote sensing (RS), with wide applications in land cover monitoring, urban planning, disaster assessment, and environmental management. In recent years, vision foundation models (VFMs) have made breakthroughs in the field of computer vision. With their powerful zero-shot generalization ability, VFMs, represented by the Segment Anything Model (SAM), have recently been applied to remote sensing change detection (RSCD). However, existing FastSAM-based detection methods still suffer from significant shortcomings in edge detection. To address the problems of edge blurring and missed detection of small-scale changes in very high resolution (VHR) remote sensing image (RSI) CD, this paper proposes an edge-enhanced detection framework, FEENet, which improves FastSAM for CD on RSIs. For accurate edge detail extraction, an Edge-Aware Module (EAM) is designed to fuse frequency-domain enhancement with semantic features. Furthermore, a Multi-scale Fuzzy Edge-Guided Fusion Module (MFEFM) is innovatively proposed, where Gaussian fuzzy logic is employed to model edge uncertainty, and dynamic feature injection is adopted to strengthen the association between edge semantics and changed regions. Experiments on five public datasets, LEVIR-CD, WHU-CD, SYSU-CD, CLCD and S2Looking, demonstrate that FEENet achieves better detection performance than existing methods, especially in detecting subtle changes in edge regions such as building contours and road boundaries, thus providing an effective solution for edge-aware CD tasks in VHR RSIs.