SGNet: A Transformer-Based Semantic-Guided Network for Building Change Detection

Building change detection (BCD) is a widely used method for monitoring human activities. Despite advancements in deep learning (DL) in computer vision, recent DL-based BCD methods still face challenges in extracting discriminative features due to irrelevant noisy changes and the lack of consideration for specific semantic information related to changes. To address this limitation, we propose the semantic-guided network (SGNet), a multiscale pure Transformer-based BCD method that delves into more discriminative class-specific feature representations for changed buildings. Specifically, we first breakdown the change detection process into multiple stages: changed semantics extraction, enhancement, fusion, and decoding. Then, we analyze the nature of building change events in remote sensing images (RSIs) and categorize them into three main groups. To detect positive changes and reduce noise, we developed the various changes extraction module. In addition, we introduce the semantic-guided changed-semantics augmentation module to enhance the semantic information of target changes while suppressing irrelevant regions or objects. We also propose the various changes fusion module to model semantic latency in bitemporal RSIs. Extensive experiments on two building change datasets, i.e., LEVIR-CD and WHU-CD demonstrate that our SGNet(MiT-b0) outperforms the state-of-the-art baselines and achieves F1 score of 90.99% and 92.47%, respectively, with improvements of 1.9% and 7.89%, respectively, highlighting the exceptional sample-efficient learning capability.