Abstract: Building change detection plays a crucial role in urban planning and environmental monitoring. However, deep learning-based change detection algorithms exhibit limited capabilities in understanding spatiotemporal relationships, alleviating sample imbalance, and identifying the boundaries of changing objects. Given this, this study proposed a change detection network that coupled adaptive attention and edge refinement (A²ERNet), aiming at aggregating multi-layer contextual features to enhance the performance of building change detection. First, this study designed an adaptive attention fusion mechanism that includes self-attention, cross-attention, and anchor-primary attention modules. This mechanism facilitates faster model convergence to optimal feature representations and enhances both local and global relationships of change features. Subsequently, it introduced a plug-and-play, lightweight edge enhancement and refinement module to improve building edge details. Finally, it established an adaptively weighted binary cross-entropy loss function to address sample imbalance issues, ensuring appropriate supervision for the network's intermediate layers and guiding the training process. Through experiments and tests conducted on the LEVIR-CD and WHU building datasets, the A²ERNet achieved F1-scores of 0.908 7 and 0.915 0, respectively, demonstrating significant superiority over comparative methods.