U-Net网络结合多注意力机制的极化SAR影像地物识别方法  

A polarimetric SAR image ground recognition method based on U-Net network combined with multiple attention mechanisms

作　　者：李云川 李勇发 左小清[1] 李永宁 徐浩翔 张彧然 LI Yunchuan;LI Yongfa;ZUO Xiaoqing;LI Yongning;XU Haoxiang;ZHANG Yuran(Faculty of Land and Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China)

机构地区：[1]昆明理工大学国土资源工程学院,昆明650093

出　　处：《测绘科学》2024年第10期69-76,共8页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(42161067);云南省重大科技专项计划项目(202202AD080010);云南省基础研究计划项目(202401AU070173);云南省教育厅科学研究基金项目(2024J0067)。

摘　　要：针对极化信息难以利用,卷积神经网络(CNN)只关注局部感受野内的信息,无法准确提取出关键特征,从而导致识别任务性能下降的问题,该文提出一种U-Net网络结合多种attention模块方法MA U-Net。该方法通过联合时频分析(JTFA)将极化状态的时间序列转化为频率表示,揭示了信号频率成分,有助于提取有用的信息降低了极化信息的利用难度。并且,使用增加多种attention模块的U-Net网络用于特征提取,ResNet网络用于地物识别任务。通过与传统CNN和U-Net网络对比发现,该文提出的MA U-Net地物识别方法在同一数据集中的识别精度更高,平均识别精度分别提高了6.1%和4.5%,在极化合成孔径雷达(SAR)影像目标识别方面有着明显的优势。Aiming at the problem that polarization information is difficult to use,convolutional neural network(CNN)only focuses on the information in the local receptive field,and cannot accurately extract the key features,which leads to the degradation of the performance of recognition tasks,a method based on U-Net network combined with multiple attention modules(MA U-Net)was proposed in this paper.The time series of polarization state was converted into frequency representation through joint time-frequency analysis(JTFA),the frequency components of the signal was revealed,helps to extract useful information,and reduces the difficulty of using polarization information.In addition,the U-Net network with multiple attention modules was used for feature extraction,and the ResNet network was used for feature recognition.Compared with the traditional CNN and U-Net networks,the proposed method had higher recognition accuracy in the same data set,and the average recognition accuracy is improved by 6.1%and 4.5%respectively,which showed obvious advantages in polarimetric synthetic aperture radar(SAR)image target recognition.

关 键 词：极化SAR 地物识别 特征融合 深度学习 U-Net网络改进 

分 类 号：P237[天文地球—摄影测量与遥感] P642.22[天文地球—测绘科学与技术]