深度学习方法在海浪有效波高数据高分辨率处理中的应用  

Application of Deep Learning Method in High Resolution Processing of Significant Wave Height

作　　者：朱晓雯 侯宇 刘玉海 吴克俭 ZHU Xiao-Wen;HOU-Yu;LIU Yu-Hai;WU Ke-Jan(College of Oceanography and Atmospheric Sciences, Ocean University of China, Qingdao 266100, China;College of Mathematical Sciences, Ocean University of China, Qingdao 266100, China;Institute of Oceanology, Chinese Academy of Sciences, Qingdao 266100, China;Dawning Information Industry Co.,Ltd., Qingdao 266100, China)

机构地区：[1]中国海洋大学海洋与大气学院,山东青岛266100 [2]中国海洋大学数学科学学院,山东青岛266100 [3]中国科学院海洋研究所,山东青岛266100 [4]中科曙光国际信息产业有限公司,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2020年第11期22-29,共8页Periodical of Ocean University of China

基　　金：国家重点研究发展计划:重大海洋动力灾害致灾机理、风险评估、应对技术研究及示范应用项目(2016YFC1402000)资助。

摘　　要：本文基于欧洲中期天气预报中心的有效波高数据,运用传统的双三次插值(Bicubic interpolation)、克里金插值(Kriging interpolation)以及两种改进的超分辨率卷积神经网络(Super-Resolution CNN,简称SRCNN)进行高分辨率处理,使用峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)、均方根误差(Root-Mean-Square Error,RMSE)、结构相似度(Structural Similarity,SSIM)、余弦相似性(Cosine Similarity)等评价指标比较了各种方法的高分辨率处理效果,并分析了每种方法的误差分布特点。结果表明,改进的SRCNN方法(SRCNN_2)是一种在整体效果、局部细节和计算效率方面均比较优秀的高分辨率处理方法,是深度学习方法在海洋数据高分辨率处理问题上一次成功的应用,但改进的SRCNN方法在近岸有效波高数据的处理效果方面还有待提高。Based on the significant wave height dataset of the European Center for Medium Range Weather Forecasting,this paper utilized traditional Bicubic interpolation,Kriging interpolation and two improved super-resolution CNN(SRCNN)toaddress the downscalingof ocean data.Several evaluation indicators,including Peak Signal-to-Noise Ratio(PSNR),Root-Mean-Square Error(RMSE),Structural Similarity(SSIM),Cosine similarity,were used to compare the high-resolution processing effect of various methods andthe characteristics of error distribution are also analyzed.The results showed that the improved SRCNN method(SRCNN_2)is an efficient interpolation method in terms of overall effect,detail effect and computational efficiency,which is a successful application of deep learning method in high resolution problems of physical oceanography.However,the SRCNN method still needs to be improved in the processing effect of offshore significant wave height data.

关 键 词：海浪有效波高 双三次插值 克里金插值 超分辨率卷积神经网络 

分 类 号：P731.22[天文地球—海洋科学]