基于多尺度注意力网络的SAR影像水体高精度自动检测  

Multi-scale Attention Network for High-Precision Automatic Detection of Water Bodies in SAR Images

作　　者：梁秀娟 肖红光[1] 陈立福[2] 范习谦 LIANG Xiujuan;XIAO Hongguang;CHEN Lifu;FAN Xiqian(School of Computer and Communication Engineering,Changsha University of Science and Technology,Changsha 410114,China;School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China;Hangzhou Institute of Advanced Research,Chinese Academy of Sciences,Hangzhou 310024,China)

机构地区：[1]长沙理工大学计算机通信与工程学院,长沙410114 [2]长沙理工大学电气与信息工程学院,长沙410114 [3]中国科学院大学杭州高等研究院,杭州310024

出　　处：《航天返回与遥感》2024年第6期124-136,共13页Spacecraft Recovery & Remote Sensing

基　　金：国家自然科学基金项目(41201468);“深空探测省部共建协同创新中心”开放课题资助项目(SKTC202203)。

摘　　要：合成孔径雷达(Synthetic Aperture Radar,SAR)系统具有全天时全天候的成像特点,在水体检测中具有重要应用价值,但在提取多尺度水体信息时,仍存在细小支流水体特征提取困难的问题。文章提出了名为多尺度注意力水体分割网络(Multi-scale Attention LinkNet,MATLinkNet)。该网络主要分为编码器和解码器两部分。在编码器前的初始块阶段,采用多个小卷积替代传统的7×7卷积,可以提取到更加细腻的水体信息。随后,在编码阶段构建了注意力机制的多尺度金字塔(Attentional Multi-Scale Pyramid,AMSP)模块,加强对不同尺度水体特征的学习并关注水体的重要特征。最后,设计跳跃连接来链接编码器和解码器的特征,弥补编码阶段多次下采样造成的空间信息损失,在有效提高水体提取精度的同时减少了训练时间。在自制的“哨兵一号”SAR影像水体数据集上进行实验,独立测试结果表明,水体提取的精度和交并比最高值分别达到了90.73%和81.95%,比原始的LinkNet网络分别提高了6.81和5.27个百分点,验证了该网络在SAR水体影像分割任务中的优异性能。Synthetic Aperture Radar(SAR)system owns all-weather imaging characteristics and important application value in water body detection,but there still exists the problem of difficulty in extracting the features of fine tributary water bodies when extracting the information of multi-scale water bodies.The article proposes a network called Multi-scale Attention LinkNet(MATLinkNet).The network is mainly divided into two parts:encoder and decoder.In the initial block stage before the encoder,multiple small convolutions are used instead of the traditional 7×7 convolution,which can extract more delicate water body information.Subsequently,the Attentional Multi-Scale Pyramid(AMSP)module with the attentional mechanism is constructed in the encoding stage to enhance the learning of water body features at different scales and focus on the important features of the water body.Finally,skip connections are designed to chain the features of the encoder and decoder to compensate for the loss of spatial information caused by multiple downsampling in the encoding stage,which effectively improves the extraction accuracy of the water body and reduces the training time at the same time.Experiments are conducted on the self-made Sentinel-1 SAR image water body dataset,and the independent test results show that the highest values of water body extraction accuracy and intersection and concatenation ratio reach 90.73%and 81.95%,respectively,which are 6.81 and 5.27 percentage points higher than that of the original LinkNet network,verifying the network’s excellent performance in the task of segmentation of SAR water body images.

关 键 词：水体提取 语义分割 合成孔径雷达 注意力机制 遥感应用 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] V443[自动化与计算机技术—控制科学与工程]