【作者】 张旭；

【导师】 何忠杰；

【作者基本信息】 哈尔滨工程大学 ， 仪器科学与技术， 2023， 硕士

【摘要】 海水的温度和盐度是海洋环境最基本的两个要素,在海洋预报、气候研究、渔业养殖、国防军事等领域有极其重要的作用。空间分辨率是温盐观测数据的一个重要指标,但现阶段传统获取温盐观测数据的手段由于在观测方式、成本等方面的限制,在空间分辨率上难以达到较高的水平。在地震海洋学中,声波反射探测方式的特点决定了其探测到的数据具有较高的横向分辨率,通过对观测的声波地震记录进行反演即可得到高分辨率的海水温盐数据。在众多从波阻抗反演温盐的方法中,普遍使用拟合的数学关系来简化反演流程中涉及的海水变量间复杂的非线性问题。由于拟合的单一数学关系无法完全解释海水变量间的关系,所以这类简化问题的方法会对反演结果造成影响。对主流的温盐反演方法进行分析总结,对这些方法中普遍使用的拟合曲线关系结合温盐数据进行分析,探究其对反演结果可能造成的影响,并提出使用更为适合解决海水反演问题的神经网络来进行改善。以正演-反演为研究框架,通过哥白尼海洋环境观测服务提供的全球模式再分析数据和基于褶积模型的地震正演为反演研究提供所需的合成地震记录。使用BP神经网络完成了波阻抗数据下的温盐反演,并通过设置扰动实验和对比实验来探究和验证方法的优越性。最后将改进方法应用在水下滑翔机(Glider)的观测数据上,以验证方法在真实海水情况下的反演效果。实验结果表明,在基于波阻抗反演下的BP神经网络温盐反演方法相较于直接插值温盐观测数据获取的温盐场有更准确的结果;相对于拟合曲线法和迭代法能有效提升0～500m深度的温盐反演精度,其中温盐均方根误差分别降低约:温度40%、盐度8%。基于Glider的观测数据的反演断面能明显展示出主要的海水结构;其中全局反演均方根误差为:温度0.37℃,盐度0.034;500m以下均方根误差为:温度0.30℃,盐度0.031。实验结果也证明了使用的拟合曲线关系来简化温盐反演问题会降低浅层海水的反演结果准确度,而使用BP神经网络反演则可以规避简化反演的问题并提高在浅水层的反演精度。但由于波阻抗反演中的递推波阻抗过程存在误差累积,使得500m以上水深的反演误差逐渐增大,导致增加了全局的均方根误差。考虑到海洋学研究主要集中在500m深度以内,且本文的方法在该深度范围内有较好的表现,所以本文使用BP神经网络进行温盐反演的方法仍具有研究意义。更多还原

【Abstract】 The temperature and salinity of seawater are the two most basic elements of Marine environment,which play an extremely important role in the fields of Marine forecasting,climate research,fishery and breeding,national defense and military.Spatial resolution is an important index of thermohaline observation data,but at present,the traditional means of obtaining thermohaline observation data is difficult to reach a high level of spatial resolution due to the limitations of observation methods and costs.In seismic oceanography,the characteristics of acoustic reflection detection mode determine that the detected data has a high horizontal resolution.By inversion of the observed acoustic seismic record,high-resolution sea temperature and salt data can be obtained.In many methods of thermohaline inversion from wave impedance,fitting mathematical relation is commonly used to simplify the complex nonlinear problem of seawater variables involved in the inversion process.Since the fitted single mathematical relation cannot fully explain the relationship between seawater variables,such simplified problem methods will affect the inversion results.This paper analyzes and summarizes the mainstream thermohaline inversion methods,analyzes the fitting curve relationship commonly used in these methods combined with thermohaline data,explores its possible influence on the inversion results,and proposes to use the neural network more suitable for solving the seawater inversion problem to improve.Using forward inversion as the research framework,the global model reanalysis data provided by Copernicus Marine environmental observation service and seismic forward modeling based on convolution model provide the required synthetic seismic record for inversion research.The BP neural network is used to complete the inversion of temperature and salinity under the wave impedance data,and the advantages of the method are explored and verified by disturbance experiment and comparison experiment.Finally,the improved method is applied to the observation data of underwater Glider to verify the inversion effect of the proposed method in real seawater.The experimental results show that the BP neural network inversion method based on wave impedance inversion has more accurate results than that obtained by direct interpolation of temperature and salinity observation data.Compared with the fitting curve method and the iterative method,the inversion accuracy of temperature and salinity at 0 ～ 500 m depth can be effectively improved,in which the root mean square error of temperature and salinity is reduced by about 40% and 8% respectively.The inversion section based on Glider data can clearly show the main seawater structure.The root mean square error of global inversion is as follows:temperature 0.37℃,salinity 0.034;Under 500 m,the root-mean-square error is 0.30℃ for temperature and 0.031 for salinity.The experimental results also prove that using the fitting curve relationship to simplify the inversion of temperature and salinity will reduce the accuracy of shallow seawater inversion results,while using BP neural network inversion can avoid the problem of simplified inversion and improve the accuracy of shallow water inversion.However,due to the accumulation of errors in the recursive wave impedance process of wave impedance inversion,the inversion errors of water depths above 500 m gradually increase,resulting in an increase in the global root mean square error.Considering that oceanographic research is mainly concentrated within 500 m depth,and the method proposed in this paper has a good performance in this depth range,the method of temperature and salinity inversion using BP neural network in this paper is still of research significance.更多还原