基于模糊共生网络的SAR遥感场景分类  

SAR Image Scene Classification Based on ML-FGLCMNet

作　　者：周易[1] 卢延荣 ZHOU Yi;LU Yanrong(Mianyang Vocational and Technical College,Mianyang,Sichuan 621000,China;College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]绵阳职业技术学院,四川绵阳621000 [2]兰州理工大学电气工程与信息工程学院,兰州730050

出　　处：《遥感信息》2023年第6期103-109,共7页Remote Sensing Information

基　　金：国家自然科学基金(62001198);甘肃省青年科技基金计划(21JR7RA246、21JR7RA247)。

摘　　要：合成孔径雷达(synthetic aperture radar,SAR)图像的场景分类被广泛运用于军事和民用领域,而合成孔径雷达图像存在场景复杂、图像分辨率低等特点,使其准确分类成为挑战。由此,文章提出基于模糊共生网络的合成孔径雷达图像场景分类方法。该方法的ML-Net(multi layer convolutional network)模块可提取合成孔径雷达图像的低分辨特征,模糊灰度共生矩阵模块则能多角度融合4组二阶统计量提取合成孔径雷达图像的纹理特征,并输入至多类支持向量机完成场景分类。选用Terra SAR-X和GS-SAR6数据集完成该方法与全频通道注意力网络和多特征融合全局-局部卷积网络的实验。对比结果可知,该方法的准确率与Kappa值较高,可获得更小的初始损失值、更高的训练准确率和更快的收敛速度。The classification of SAR(synthetic aperture radar)image scenes is widely used in military and civilian fields.However,SAR images pose challenges for feature extraction due to their complex scenes and low image resolution,making accurate classification a challenge.To address this issue,a fuzzy co-occurrence network-based method for SAR image scene classification is proposed.The ML-Net(multi-layer convolutional network)module of this method can extract the low-resolution features of SAR images,while the FGLCM(fuzzy convolutional gray level co-occurrence matrix)module can fuse four sets of second-order statistics from multiple angles to extract texture features of SAR images,which are then input into an MSVM(multi-class support vector machine)classifier for scene classification.Experiments comparing this method with full-frequency channel attention networks and multi-feature fusion global-local convolutional networks using Terra SAR-X and GS-SAR6 datasets show that this method has higher accuracy and Kappa values.This method can achieve smaller initial loss values,higher training accuracy,and faster convergence rates.

关 键 词：SAR图像 场景 分类 GLCM SVM 准确率 

分 类 号：TN957[电子电信—信号与信息处理]