SFAM-Net: A Novel Dual-Branch Network Based on Spectral Feature and Attention Machine for Building Change Detection in Remote Sensing Imagery

Deep learning techniques have significantly advanced change detection in remote sensing imagery. However, building change detection presents challenges due to the varied appearance of buildings and the complexity of scenes in remote sensing images. Current deep learning-based methods encounter three primary issues. Firstly, CNN-based approaches struggle to model crucial global contextual information essential for remote sensing building images analysis. Transformer-based methods may inadvertently degrade local features. Secondly, traditional attention mechanisms fall short in effectively modeling spatial and spectral features. Thirdly, certain channel attention methods extract excessive redundant information. To address these challenges, this study proposes SFAM-Net, a two-branch hybrid architecture. Our approach initially employs orthogonal methods to minimize redundant information extracted from channels and spaces. Subsequently, we leverage the parallel structure of convolutions and visual transformers to enhance images representation, integrating local features and global representations through cross-attention to better coordinate building and background features. In the CNN and Transformer branches, we adopt spatial-spectral feature coordination and spectral multi-head attention coordination strategies to improve performance in complex scenes. Additionally, we introduce a novel loss function combining edge and center guidance, focusing on changing image edges and centers to enhance sensitivity and accuracy in change area detection. Extensive experiments on widely used LEVIR-CD and WHU-CD datasets validate the effectiveness and efficiency of our network.