基于字典学习的声速剖面重构和反演  

Reconstruction and Inversion of Sound Speed Profile Based on Dictionary Learning

作　　者：谢龙 刘琛 梁文宇 XIE Long;LIU Chen;LIANG Wenyu(91388 Troops,Zhanjiang 524002,China;College of Electronics and Information Engineering,Guangdong Ocean University,Zhanjiang 524088,China;Guangzhou seaward ocean Technology Co.,Ltd.,Guangzhou 510200,China)

机构地区：[1]91388部队,广东湛江524002 [2]广东海洋大学电子与信息工程学院,广东湛江524088 [3]广州至远海洋科技有限公司,广东广州510200

出　　处：《海洋技术学报》2023年第2期12-19,共8页Journal of Ocean Technology

基　　金：国防科技重点实验室基金资助项目(JCKY2022207CH10)。

摘　　要：声速剖面(Sound Speed Profile,SSP)是海洋环境观测的重要要素之一,引入遥感参数进行SSP反演可以实时获取声速数据。反演首先需要依托精确的基函数对声速场进行降维表示,本文提出一种利用非线性基函数-学习字典(Learned Dictionaries,LDs)提高降维精度的方案,并使用在多源信息融合上表现良好的自组织竞争型神经网络算法(Self-organizing Map,SOM)对南海海域进行SSP反演。实验结果使用均方根误差(Root Mean Square Error,RMSE)作为精度评估。实验结果显示:LDs较传统基函数-经验正交函数(Empirical Orthogonal Functions,EOFs)的降维精度在使用三阶基函数时提升0.13 m/s,五阶基函数提升0.07 m/s。使用五阶基函数进行反演,LDs的反演精度为3.01 m/s,低于EOFs的反演精度2.47 m/s。其原因为反演误差在机器学习训练基函数时被放大,导致所求得反演系数欠优。LDs基函数较之传统的EOFs能够有效突破正交性的限制,更精确地表示声速的扰动,达到了更高的降维精度,为声学信号处理任务的基函数学习提供一种方案。Sound speed profile(SSP)is one of the important elements of marine environmental observation,and the introduction of remote sensing parameters for SSP inversion can obtain sound speed data in real time.The inversion first relies on an accurate basis function for the reduced dimensional representation of the sound speed field.In this paper,we propose a scheme to improve the dimensionality reduction accuracy by using non-linear Learned Dictionaries(LDs),and use the Self-organizing Map(SOM)algorithm,which performs well in multi-source information fusion,for SSP inversion in the South China Sea.The experimental results were assessed using root mean square error(RMSE)as the accuracy.The experimental results show that the dimensionality reduction accuracy of LDs is improved by 0.13 m/s for third-order basis functions and 0.07 m/s for fifth-order basis functions compared with that of Empirical Orthogonal Functions(EOFs).The inversion accuracy of 3.01 m/s for LDs using fifth order basis functions is lower than that of 2.47 m/s for EOFs.The inversion accuracy of LDs is lower than that of EOFs because the inversion error is amplified during the machine learning process of training the basis functions,resulting in suboptimal inversion coefficients.Compared with the conventional EOFs,using LDs basis functions,the orthogonality limitation can be effectively broken,the perturbation of sound speed is more accurately represented,a basis function learning scheme for acoustic signal processing tasks is provided.

关 键 词：声速剖面 学习字典 经验正交函数 自组织竞争型神经网络 

分 类 号：P733.2[天文地球—物理海洋学]