融合多尺度特征Transformer的高分辨率遥感图像变化检测  

Change detection for high-resolution remote sensing images with multi-scale feature transformer

作　　者：李健慷 张桂欣 祝善友 徐永明 李湘雨 LI Jiankang;ZHANG Guixin;ZHU Shanyou;XU Yongming;LI Xiangyu(School of Remote Sensing&Geomatics Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Geographical Sciences,Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,南京210044 [2]南京信息工程大学地理科学学院,南京210044

出　　处：《遥感学报》2025年第1期266-278,共13页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家自然科学基金(编号:42171101,42271351);高分辨率对地观测系统重大专项(编号:30-Y60B01-9003-22/23)。

摘　　要：为了加强变化检测中深度学习网络的语义信息提取能力,捕获更多高阶多尺度特征细节以及突出影像差异信息,本文提出一种融合孪生结构和多尺度特征Transformer的高分辨率遥感影像变化检测模型MFTSNet(Multi-scale Feature Transformer Siamese Network)。该模型设计了语义特征Transformer模块ST(Semantic feature Transformer module)捕获不同层级特征图的语义信息,引入置入Transformer模块GT(Grounding Transformer module)和映射Transformer模块RT(Rendering Transformer module)加强低层和高层语义信息的获取,发掘高阶多尺度特征细节信息以及不同空间位置和通道间的全局上下文关系,进一步提升变化检测精度,增强地物检测结果的完整性、区域内部以及边缘细节。将MFTSNet与另外8种变化检测模型在4个公开数据集上的变化检测结果进行对比,并通过消融实验、参数分析等手段验证MFTSNet中各模块的有效性。对比实验结果表明MFTSNet网络模型在4个数据集上的F1和交并比IoU分别至少提高了0.465%、0.113%、0.369%、2.13%和0.723%、0.188%、0.304%、2.962%。消融实验表明GT、RT、ST 3个模块共同作用可有效提升网络模型性能。参数分析表明MFTSNet模型中的特征信息长度L与编码器—解码器个数是两个重要的网络结构参数,L在CDD、WHU-CD数据实验中取16,在SYSU-CD、LEVIR-CD数据实验中取8,4个数据集上设置(EN,DN)为(1,2)时,MFTSNet模型的检测结果最优。Wetland is an important ecosystem and plays a vital role in maintaining regional ecological security.Wetland structure changes respond sensitively to natural and human activities,and flood wetlands experience drastic seasonal water and vegetation changes due to intermittent flood inundations.Mapping high-accuracy wetland structures is challenging because of frequent water and vegetation alternations,which cause spectral confusion and misclassification in optical satellite images.Several wetland extraction methods are available today,including object-oriented methods,whose parameters need to be decided subjectively,and machine learning methods,which have relatively low accuracy.With the continuous development of deep learning in image semantic segmentation,a precise and automatic remote sensing image binary classification becomes possible.Recent studies have suggested that deep learning semantic segmentation methods show great potential for mapping wetland changes in high-resolution images.However,the extraction of wetland structures in complex floodplain scenarios places high demands on models in terms of mining deep spatial information.The deformable U-Net(D-UNet)semantic segmentation model is improved to enhance the accuracy of the extraction of floodplain wetland structure.In this study,the Taitema Lake in Xinjiang,China was selected as the study area because it is a typical floodplain wetland in the arid zone.A multiscene and multitemporal wetland sample dataset was collected using Sentinel-2 remote sensing images in the study area.The D-UNet for wetland structure extraction used VGG16 to build the encoding/decoding network and focused on improving the convolutional layer in the network.D-UNet was improved by replacing the convolution block before dimensionality reduction with multiscale dilated convolutions to enhance the network’s receptive field,fuse features of different scales,and avoid loss of detailed information in highresolution remote sensing images.After pretraining D-UNet,we determined that a mu

关 键 词：高分辨率遥感 变化检测 深度学习 孪生网络 多尺度特征 TRANSFORMER 语义信息 消融实验 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]