Buildings are some of the basic spatial elements of a city. Changes in the spatial distributions of buildings are of great significance for urban planning and monitoring illegal construction. Building change detection (CD) with high-resolution remote sensing images based on deep learning can be used to quickly identify large-scale spatial distribution changes, saving many workforce and material resources. However, existing CD networks mainly focus on regional accuracy, ignoring the importance of accurate boundary identification. It is often difficult for CD networks to achieve accurate boundary segmentation, especially for dense and continuously distributed buildings. In addition, due to the inconsistencies among classes and the discontinuities within classes, it is difficult for CD networks to obtain complete change results. In response to the above problems, a novel method called EGDE-Net that focuses on boundary accuracy and change region integrity is proposed in this paper. First, an edge-guided Transformer block is designed to encode dual-branch networks for EGDE-Net; this block combines a hierarchical Transformer with an edge-aware module (EAM) for long-range context modeling and feature refinement. Second, a feature differential enhancement module (FDEM) is developed to learn highly discriminative change feature maps by exploiting the differences between bitemporal features. In addition, feature maps are recovered through multiple upsampling operations and dilated asymmetric modules (DAMs) in the decoding part of the network. Finally, prior information for boundaries and change information are jointly used to implement a supervision process and obtain the optimal model. The proposed EGDE-Net achieves better results based on the WHU building CD dataset and LEVIR-CD dataset than do similar methods. Notably, F1 scores of 93.02% and 90.05% and intersection over union (IoU) scores of 86.96% and 81.91% are obtained for these two datasets.

建筑是城市的一些基本空间元素。建筑物空间分布的变化对城市规划和违法建设监测具有重要意义。基于深度学习的高分辨率遥感影像建筑物变化检测（CD）可用于快速识别大规模空间分布变化，节省大量人力和物力。然而，现有的 CD 网络主要关注区域精度，忽略了准确边界识别的重要性。CD 网络通常难以实现准确的边界分割，尤其是对于密集且连续分布的建筑物。此外，由于类之间的不一致性和类内的不连续性，CD网络很难获得完全的变化结果。针对上述问题，本文提出了一种关注边界精度和改变区域完整性的新方法EGDE-Net。首先，设计了一个边缘引导的 Transformer 块来对 EGDE-Net 的双分支网络进行编码；该模块将分层 Transformer 与边缘感知模块 (EAM) 相结合，用于远程上下文建模和特征细化。其次，开发了一个特征差分增强模块（FDEM），通过利用双时间特征之间的差异来学习具有高度区分性的变化特征图。此外，特征图通过网络解码部分的多个上采样操作和扩张的非对称模块（DAM）来恢复。最后，边界的先验信息和变化信息共同用于实施监督过程并获得最优模型。所提出的 EGDE-Net 基于 WHU 构建 CD 数据集和 LEVIR-CD 数据集取得了比类似方法更好的结果。值得注意的是，这两个数据集获得了 93.02% 和 90.05% 的 F1 分数和 86.96% 和 81.91% 的交集（IoU）分数。