基于改进CGAN的海冰SAR-to-Optical影像转换  

Sea ice SAR-to-Optical image translation based on improved CGAN

作　　者：刘翔[1] 王瑞富[1] 孙光 李媛[1] LIU Xiang;WANG Ruifu;SUN Guang;LI Yuan(College of Surveying&Spatial Information,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学测绘与空间信息学院,山东青岛266590

出　　处：《海洋通报》2024年第4期452-462,共11页Marine Science Bulletin

基　　金：山东省自然科学基金(ZR2022MD002)。

摘　　要：遥感海冰监测是当前研究热点,通过条件生成对抗网络(CGAN)将海冰SAR影像转换成光学影像,可获得全天时全天候且形象直观的监测数据,但该方法得到的转换结果存在影像模糊、纹理弱化和颜色失真等问题。本文针对以上问题设计了改进的CGAN网络,综合当前的改进方式,新模型在网络结构上加入了空洞空间金字塔模块并设计了加入交叉特征融合模块的跳跃连接,使用结构相似性和L1范数联合损失函数。本文选取东波弗特海地区5景Sentinel-1影像和7景Sentinel-2影像开展实验,实验结果表明,改进CGAN转换的影像具有更好的视觉效果,峰值信噪比(PSNR)提高了3.4 dB,结构相似性(SSIM)提高了0.11,均方根误差(RMSE)降低了13%,并且经过转换后的影像比SAR影像海冰分类结果准确度提高了7.33%。Remote sensing sea ice monitoring is a current research hotspot.By translating sea ice SAR images into optical images through the Conditional Generative Adversarial Network(CGAN),all-weather and intuitive monitoring data can be obtained.However,the translation results obtained by this method have problems such as blurring of contours,weakening of textures,and color distortion.This article designs an improved CGAN network to address the above issues.Taking into account the current improvement methods,the new model adds a hollow space pyramid module to the network structure and designs a skip connection with a cross feature fusion module.The loss function uses a joint loss function of structural similarity and L1 norm.This article selects 5 Sentinel-1 images and 7 Sentinel-2 images from the East Beaufort Sea area for experiments.The experimental results show that the improved CGAN translated images have better visual effects,with peak signal-to-noise ratio(PSNR)increased by 3.4,structural similarity(SSIM)increased by 0.11,root mean square error(RMSE)decreased by 13%,and the accuracy of the processed images for sea ice classification improved by 7.33%compared to SAR images.

关 键 词：海冰监测 条件生成对抗网络 SAR 光学影像 影像转换 

分 类 号：P731[天文地球—海洋科学] P237