In the domains of geographic information systems and remote sensing image analysis, change detection is vital for examining surface variations in high-resolution remote sensing pictures. However, the intricate texture characteristics and rich details found in high-resolution remote sensing photos are difficult for conventional change detection systems to deal with. Target misdetection, missed detections, and edge blurring are further problems with current deep learning-based methods. This research proposes a high-resolution city change detection network based on difference and attention mechanisms under multi-scale feature fusion (MDANet) to address these issues and improve the accuracy of change detection. First, to extract features from dual-temporal remote sensing pictures, we use the Siamese architecture as the encoder network. The Difference Feature Module (DFM) is employed to learn the difference information between the dual-temporal remote sensing images. Second, the extracted difference features are optimized with the Attention Refinement Module (ARM). The Cross-Scale Fusion Module (CSFM) combines and enhances the optimized attention features, effectively capturing subtle differences in remote sensing images and learning the finer details of change targets. Finally, thorough tests on the BTCDD dataset, LEVIR-CD dataset, and CDD dataset show that the MDANet algorithm performs at a cutting-edge level.

中文翻译：

---

MDANet：多尺度特征融合下基于差异和注意力机制的高分辨率城市变化检测网络

在地理信息系统和遥感图像分析领域，变化检测对于检查高分辨率遥感图片中的表面变化至关重要。然而，高分辨率遥感照片中复杂的纹理特征和丰富的细节是传统变化检测系统难以处理的。目标误检、漏检和边缘模糊是当前基于深度学习的方法的进一步问题。本研究提出了一种基于多尺度特征融合下的差异和注意机制的高分辨率城市变化检测网络（MDANet）来解决这些问题并提高变化检测的准确性。首先，为了从双时相遥感图片中提取特征，我们使用 Siamese 架构作为编码器网络。采用差异特征模块（DFM）来学习双时相遥感图像之间的差异信息。其次，使用注意力细化模块（ARM）对提取的差异特征进行优化。跨尺度融合模块（CSFM）结合并增强了优化的注意力特征，有效捕获遥感图像中的细微差异并学习变化目标的更精细细节。最后，对BTCDD数据集、LEVIR-CD数据集和CDD数据集的全面测试表明，MDANet算法的性能处于前沿水平。