基于深度学习的海浪SAR图像分类  

Classification of Ocean Wave for SAR Images Based on Deep Learning

作　　者：刘腾腾 曹川川 韩勇[1,2] 陈戈 Liu Tengteng;Cao Chuanchuan;Han Yong;Chen Ge(College of Information Science and Engineering, Ocean University of China, Qingdao 266100 China;Laboratory for Regional Oceanography and Numerical Modeling, Pilot National Laboratory for Marine Science and Technology (Qingdao), Qingdao 266237, China)

机构地区：[1]中国海洋大学信息科学与工程学院,山东青岛266100 [2]青岛海洋科学与技术试点国家实验室区域海洋与数值模拟实验室,山东青岛266237

出　　处：《中国海洋大学学报（自然科学版）》2021年第8期88-95,共8页Periodical of Ocean University of China

基　　金：国家自然科学基金项目“基于多源卫星遥感的全球涌浪起源与传播路径研究”(41331172)资助。

摘　　要：针对海浪场中合成孔径雷达(SAR)图像的灰度特征混杂、人工目视分类困难的问题,本文利用简化的Inception-ResNet-V2模型与注意力机制相结合的方法,在减少网络层数、加快运算效率的同时,大幅度提升了计算机对SAR图像中海浪条纹清晰度的识别准确率。在利用模型进行图像分类时,本文提出分块识别的方式,对各块分类概率结果进行累加后取概率最大的类别,可提升1.8%的识别准确率,使最终准确率达到了89.6%。最后,本文基于深度学习结合12000幅分类样本实现了总计5万幅ASAR图像的分类研究,获得条纹清晰组11069幅、中间组16560幅和不清晰组22371幅分类图像。Aiming at the problem of the mixed gray features and the difficulty of artificial visual classification of synthetic aperture radar(SAR)images in the ocean wave field,this paper uses the simplified Inception-ResNet-V2 model combined with the attention mechanism to reduce the number of network layers and speed up the calculation efficiency,while greatly improving the accuracy of the computer's recognition of the wave stripe definition in the SAR images.When using the model for image classification,this paper proposes a block recognition method.After accumulating the classification probability results of each block,the category with the highest probability is selected,which can increase the recognition accuracy by 1.8%and make the final accuracy rate reach 89.6%.Finally,based on deep learning combined with 12000 classification samples,this paper has realized the classification research of a total of 50000 ASAR images,and obtained 11069 classification images in the clear stripe group,16560 images in the middle group and 22371 classification images in the unclear group.This research not only tests the reliability of using deep learning to classify ocean wave SAR images,but also contributes to interdisciplinary research between computer technology and oceanography.

关 键 词：合成孔径雷达图像 注意力机制 海浪条纹清晰度 图像分类 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]