Remote sensing image change detection is an essential aspect of remote sensing technology application. Existing change detection algorithms based on deep learning do not distinguish between changed and unchanged areas explicitly, resulting in serious loss of edge detail information during detection. Therefore, a new attentional change detection network based on Siamese U-shaped structure (SUACDNet) is proposed in this paper. In the feature encoding stage, three branches are introduced between the Siamese structure to focus on the global information, difference information and similarity information of bitemporal images respectively. In the feature decoding stage, an U-shaped structure is constructed for upsampling and recovery layer by layer. Multi-scale Convolution Residual Module (MCRM) is a new convolution structure designed for multi-scale feature extraction in the network. In addition, this work also proposes three auxiliary modules to optimize the network, namely Spatial Attention Module (SAM), Feature Fusion Module (FFM) and Cross-scale Global Context Semantic Information Aggregation Module (CGCAM), making the network more sensitive to the changed area while filtering out the background noise. Comparative experiments on three datasets show that our method is superior to the existing methods.

遥感影像变化检测是遥感技术应用的一个重要方面。现有基于深度学习的变化检测算法没有明确区分变化区域和不变区域，导致检测过程中边缘细节信息严重丢失。因此，本文提出了一种新的基于连体U形结构的注意力变化检测网络（SUACDNet）。在特征编码阶段，在Siamese结构之间引入了三个分支，分别关注双时态图像的全局信息、差异信息和相似信息。在特征解码阶段，构建一个U形结构，逐层上采样和恢复。多尺度卷积残差模块（MCRM）是一种新的卷积结构，专为网络中的多尺度特征提取而设计。此外，这项工作还提出了三个辅助模块来优化网络，即空间注意力模块（SAM）、特征融合模块（FFM）和跨尺度全局上下文语义信息聚合模块（CGCAM），使网络对过滤掉背景噪音的同时改变了区域。在三个数据集上的对比实验表明，我们的方法优于现有方法。在滤除背景噪声的同时，使网络对变化的区域更加敏感。在三个数据集上的对比实验表明，我们的方法优于现有方法。在滤除背景噪声的同时，使网络对变化的区域更加敏感。在三个数据集上的对比实验表明，我们的方法优于现有方法。