Aiming at the problem that the hybrid feature extraction and fusion adopted by current deep networks for remote sensing change detection (RSCD) easily leads to blurred boundaries of predicted categories, a Feature Disentangling Representation and Fusion Deep Network (FDRFCD) is proposed. The network uses the Mamba Out for feature extraction, which can effectively capture local and global multi-scale fused features. Meanwhile, a feature decoupling module is designed to distinguish the unique and shared features of each-phase image, improving the detection accuracy of changed and unchanged regions. To avoid information loss during the decoupling process, the decoupled features are reconstructed through the fusion module to ensure the completeness of information. Finally, a Positive Example Threshold Pull (PETP) loss regularization term is introduced to enhance the attention to positive samples. The experimental results show that this network is superior to other models in both quantitative analysis and qualitative analysis.