LGCANet: Local–Global and Change-Aware Network via Segment Anything Model for Remote Sensing Images Change Detection

Change detection (CD) is a very fundamental and challenging task in remote sensing. Many deep learning-based CD methods generally utilize Siamese networks to extract image features. However, the semantic features extracted by these methods are still not fine-grained. In addition, these CD methods ignore the object scale diverse in remote sensing images and the interaction information between bi-temporal images, which leads to the problem that the network is unable to capture more efficient feature embeddings, with ambiguous or erroneous detection results. To alleviate the above issues, we propose Local-Global and Change Aware Network via Fast Segment Anything Model (LGCANet). The Segment Everything Model (SAM) can accurately segment objects in various scene images. In this work, we intend to utilize the powerful recognition capabilities of SAM to refine the CD task. Therefore, LGCANet employs more efficient FastSAM and ResNet as encoders to extract potential feature representations in remote sensing images. FastSAM can effectively extract global contextual information, combined with ResNet’s powerful deep feature extraction capability, which enables the network to comprehensively model features. LGCANet contains three modules: content aware attention module (CAAM), fore-background aware module (FAM), and edge-reinforce hybrid-selection module (EHM). CAAM delivers feature extraction from local to global perception, realizing dynamic attention to various scales of objects. FAM can effectively learn foreground and background representations through feature interaction, which significantly enhances the model’s capability of recognizing changed regions. EHM can utilize direction-awareness to extract edge information and generate fine-grained detection maps by adaptively selecting discriminative features through designed attention mechanisms. Experiments on publicly available CD datasets show that LGCANet achieves superior detection performance compared to other state-of-the-art methods. The code is available at https://github.com/Jscript10/LGCANet.