Change detection (CD) is an increasingly popular research direction in the field of remote sensing (RS). With the rapid development of RS, the resolution of RS images is gradually improving, which also puts forward higher requirements for the generalization ability of CD models. Due to its excellent feature extraction capabilities, convolutional neural networks (CNNs) have achieved great success in CD tasks in the past. However, CNN cannot model remote context, which also limits its performance on CD. In contrast, Transformer does well in modeling remote contextual information. To make an end, in order to make full use of the advantages of both, we fully combine CNN and Transformer and propose multiscale and multisupervised hierarchical network for RS Image CD, called DMSHNet. First, we introduce multiple convolution kernels of different sizes in the Siamese encoder to help DMSHNet perceive information of different scales. Second, we develop the change information enhancement module (CIEM), which mainly generates differential features at different layers. CIEM uses simple absolute value subtraction and atrous convolution to highlight changing information and suppress unchanged information in different ranges. Third, we propose multisupervised and multiscale cascade module (DMSCM), which integrates multilayer differential features generated by CIEM on the basis of fully considering local information, global information, and multiscale information. We have conducted sufficient experiments on four public datasets, and the experimental results show that our DMSHNet achieves excellent performance. Our source code is available at https://github.com/ahlpc/DMSHNet.git .

中文翻译：

---

DMSHNet：用于遥感图像变化检测的多尺度和多监督分层网络


变化检测（CD）是遥感（RS）领域日益流行的研究方向。随着RS的快速发展，RS图像的分辨率逐渐提高，这也对CD模型的泛化能力提出了更高的要求。由于其出色的特征提取能力，卷积神经网络（CNN）过去在CD任务中取得了巨大的成功。然而，CNN 无法对远程上下文进行建模，这也限制了其在 CD 上的性能。相比之下，Transformer 在对远程上下文信息进行建模方面做得很好。为此，为了充分利用两者的优势，我们充分结合 CNN 和 Transformer，提出了用于 RS Image CD 的多尺度、多监督分层网络，称为 DMSHNet。首先，我们在 Siamese 编码器中引入多个不同尺寸的卷积核，以帮助 DMSHNet 感知不同尺度的信息。其次，我们开发了变化信息增强模块（CIEM），它主要生成不同层的差异特征。 CIEM使用简单的绝对值减法和空洞卷积来突出不同范围内变化的信息并抑制不变的信息。第三，我们提出了多监督多尺度级联模块（DMSCM），该模块在充分考虑局部信息、全局信息和多尺度信息的基础上集成了CIEM生成的多层差分特征。我们在四个公共数据集上进行了足够的实验，实验结果表明我们的 DMSHNet 取得了优异的性能。我们的源代码可在 https://github.com/ahlpc/DMSHNet.git 获取。