基于海洋遥感影像的中尺度涡自动识别及与渔场动态关系研究

发布时间：2018-05-16 23:32

本文选题：海洋遥感影像 + 中尺度涡　；参考：《上海海洋大学》2017年博士论文

【摘要】：现阶段,人类已进入大规模开发利用海洋时期,海洋成为世界各国竞相争夺的主要战场。依赖于高新海洋科学技术,不断加深对海洋的全面认识,是实现合理管控、高效开发、可持续发展海洋的核心。中尺度涡作为重要的海洋现象广泛存在于世界大洋和边缘海中,携带了海洋中超过90%的动能,以非规则螺旋状结构持续高速自转和水平运动,并改变着海洋中能量和物质的垂直与水平分布。海洋中物质和能量的时空动态变化对气候和生态具有深远的影响,因此,实现中尺度涡的动态监测和时空特征分析,不仅有助于海洋气候变化和海洋生态资源分布的研究,同时在深海捕捞和远洋渔业等实际应用中发挥重要作用。中尺度涡的自动识别是实现其动态监测、进行时空动态变化特征分析的重要手段。海洋遥感可远距离、非接触、快速地获取海洋现象和海洋环境要素信息,为中尺度涡自动识别研究提供了不可替代的数据源。基于海洋遥感数据的中尺度涡识别成为研究热点,主要的研究方法可分为基于物理特征、基于流场几何特征以及两种方法的结合。但是,现有基于物理特征识别方法在人工设计特征过程中引入了大量人为主观因素,导致中尺度涡识别精准度低的问题。同时,与陆地遥感相比,海洋遥感影像具有显著的弱特征性,主要表现为光谱低反差性和高动态海洋要素特征表达的不确定性,加剧了现有基于人工设计特征识别方法精准度低的问题。此外,基于流场几何特征识别方法采用专家设定阈值的特点缺乏泛化能力。特别地,中尺度涡由具体海域多种海洋要素相互作用形成,呈现出与空间差异相关的高动态性;并且其几何形状和物理特性在运动过程中随着能量的注入或消散均发生高度的动态变化,基于单一阈值的识别方法无法满足高动态中尺度涡自动识别的需求,极大地限制了中尺度涡研究的进展。论文以中尺度涡自动精准识别为研究目标,针对现有方法对于高动态中尺度涡自动识别的局限性,结合深度学习的思想,提出了基于特征学习的中尺度涡自动识别模型,论文主要研究内容为:(1)构建基于SAR影像的中尺度涡训练数据集。SAR卫星具有全天时、全天候、高分辨率的观测优势,为中尺度涡的精细化研究提供了必要的数据基础。本文采用欧空局提供的2005-2010年,5°N-25°N,108°E-125°E范围海域的ESA-2和Envisat SAR影像,基于人工目视方法,采用外接矩形对中尺度涡进行手工标注,并采用数据扩充方法提升训练数据集规模,增加训练数据集的多样性;(2)研究基于特征学习的中尺度涡自动识别模型。中尺度涡的高层本质特征的获取与表达是实现其自动识别的关键。本文基于SAR影像中尺度涡训练数据集规模小的现况,从模型框架和模型参数初始化两方面入手,构建适合于高动态中尺度涡自动识别的多层网络模型Deep Eddy,通过Deep Eddy的多层网络模型对中尺度涡高层本质特征的逐级抽象与表达,进而实现中尺度涡的自动精准识别。实验表明Deep Eddy模型的最优中尺度涡识别精准度达96.88%。(3)提出形态和尺度鲁棒的中尺度涡自动识别模型。中尺度涡存在严重的几何形变和空间尺度差异,显著影响其自动识别的精准度。其中,多尺度空间特征的获取是降低几何形变和空间尺度差异影响识别精准度的关键,本文基于空间金字塔模型对中尺度涡自动识别的多层网络模型进行改进,记作Deep Eddy+,实现中尺度涡多尺度空间特征的提取与表达。实验表明,在同样模型参数设置情况下,Deep Eddy+模型的中尺度涡识别精准度明显优于Deep Eddy模型,其中,Deep Eddy+模型的中尺度涡最优识别精准度达98.47%。(4)中尺度涡自动识别模型的实证分析。采用Deep Eddy+模型对获取的研究海域SAR影像进行自动识别,依据自动识别结果对中尺度涡的尺寸大小、时空特征进行分析,并与基于SSH数据识别的中尺度涡进行对比,结果表明两种数据源识别的中尺度涡在统计特征上具有明显的差异性;此外,本文探讨了研究海域中尺度涡与金枪鱼的空间相关性分析,结果表明大眼金枪鱼与气旋涡分布呈正相关性,黄鳍金枪鱼与反气旋涡呈现一定的相关性。通过上述内容的研究,论文取得了一定的研究成果,具体有:(1)首次构建了基于SAR影像的中尺度涡训练数据集,为中尺度涡自动识别方法研究提供了数据基础;(2)提出了简单有效的中尺度涡自动识别模型,实现了中尺度涡完全自动化的高精准度识别;(3)为高动态中尺度涡的自动识别提供了新理论新方法。同时,也为其他海洋现象的自动识别提供技术参考。
[Abstract]:At this stage, human beings have entered the period of large-scale exploitation and utilization of the ocean, and the ocean has become the main battle field for all countries in the world. Relying on high and new marine science and technology and deepening the comprehensive understanding of the ocean, it is the core of realizing rational control, efficient development and sustainable development of the ocean. Mesoscale vortices exist widely as important oceanic phenomena. In the ocean and the marginal sea of the world, more than 90% of the kinetic energy of the ocean is carried, and the irregular spiral structure continues to rotate at high speed and horizontal movement, and changes the vertical and horizontal distribution of energy and matter in the ocean. The dynamic changes in the space and time of the matter and energy in the ocean have a profound influence on the climate and ecology. Therefore, the mesoscale is realized. The dynamic monitoring and spatio-temporal characteristics analysis of the vortex not only contribute to the study of ocean climate change and the distribution of marine ecological resources, but also play an important role in the practical application of deep-sea fishing and ocean fishing. The automatic recognition of mesoscale vortices is an important means to realize dynamic monitoring and analysis of spatiotemporal dynamic changes. The remote, non contact and rapid acquisition of marine phenomena and information of marine environment elements provides an irreplaceable data source for the study of mesoscale vortex automatic recognition. Mesoscale eddy recognition based on marine remote sensing data has become a hot spot. The main research methods can be divided into physical characteristics based on the geometric characteristics of flow field and two kinds of methods. However, the existing physical feature recognition method has introduced a large number of subjective factors in the artificial design process, which leads to the low accuracy of the mesoscale vortex recognition. At the same time, compared with the land remote sensing, the ocean remote sensing images have significant weak characteristics, and the main purpose is to show low spectral contrast and high dynamic ocean elements. The uncertainty of characteristic expression intensifies the existing problem of low precision based on the artificial design feature recognition method. In addition, the characteristics of the geometric feature recognition method based on the flow field are not generalized by the characteristics of the expert setting threshold. In particular, the mesoscale vortex is formed by the interaction of various marine elements in the specific sea area, showing the difference with the spatial difference. The high dynamic characteristics of the closed form and its geometric and physical characteristics are highly dynamic with the injection or dissipation of energy during the movement. The recognition method based on a single threshold can not meet the requirement of the high dynamic mesoscale vortex automatic recognition, which greatly restricts the progress of the mesoscale eddy research. Quasi recognition is the research goal. Aiming at the limitation of the existing methods for the automatic recognition of high dynamic mesoscale vortices, combined with the idea of deep learning, a mesoscale vortex automatic recognition model based on feature learning is proposed. The main research contents are as follows: (1) the construction of the mesoscale eddy training data set.SAR satellite based on SAR images is all day and all day The high resolution observational advantage provides the necessary data basis for the refinement of mesoscale vortices. In this paper, the ESA-2 and Envisat SAR images of 2005-2010 years, 5 N-25 degrees N, 108 degree E-125 E are provided by ESA. Based on artificial visual method, the mesoscale vortices are manually annotated with the external rectangle, and the data expansion is used. In order to improve the size of the training data set and increase the diversity of the training dataset, (2) research on the automatic recognition model of mesoscale vortex based on feature learning. The key to the automatic recognition is to obtain and express the high level characteristic of the mesoscale vortex. This paper is based on the present condition of the small scale of the training data set of the scale vortex in the SAR image, from the model frame Starting with the two aspects of the model parameter initialization, the multi-layer network model Deep Eddy suitable for the high dynamic mesoscale vortex automatic recognition is constructed. Through the multi-layer network model of Deep Eddy, the essential characteristics of the mesoscale vortex are abstracted and expressed, and the automatic accurate identification of the mesoscale vortices is realized. The experiment shows the optimum of the Deep Eddy model. The accuracy of scale vortex recognition reaches 96.88%. (3), which proposes a robust mesoscale vortex automatic recognition model for shape and scale. The mesoscale vortices have serious geometric and spatial scale differences, which significantly affect the accuracy of their automatic recognition. The acquisition of multi-scale spatial features is the accuracy of reducing the accuracy of geometric and spatial scale differences. The key of this paper is to improve the multi-layer network model of the mesoscale vortex automatic recognition based on the spatial Pyramid model. It is recorded as Deep Eddy+ to extract and express the multi-scale spatial features of the mesoscale vortex. The experiment shows that the accuracy of the mesoscale vortex recognition of the Deep Eddy+ model is obviously better than the Deep Eddy model under the same model parameters setting. The accuracy of the mesoscale vortex recognition accuracy of the Deep Eddy+ model is an empirical analysis of the 98.47%. (4) mesoscale eddy automatic identification model. The Deep Eddy+ model is used to automatically identify the SAR images obtained in the study area, and the spatial and temporal characteristics of the mesoscale vortices are analyzed according to the automatic recognition results, and the SSH data based on the SSH data are also analyzed. The results show that the mesoscale vortices identified by the two sources have obvious differences in statistical characteristics. In addition, the spatial correlation analysis of the mesoscale vortices and tuna in the study area is discussed. The results show that the large Eye Tuna has a positive correlation with the cyclone distribution, and the yellowfin tuna and the anti gas vortex are in a positive correlation. A certain correlation is presented. Through the study of the above content, some research results have been obtained. (1) the data set of mesoscale vortex training based on SAR image is first constructed, which provides a data basis for the study of the mesoscale eddy automatic recognition method. (2) a simple and effective mesoscale eddy automatic identification model is proposed to realize the middle ruler. The high precision recognition of the complete automation of the degree vortices (3) provides a new theory and new method for the automatic recognition of the high dynamic mesoscale vortices. At the same time, it also provides a technical reference for the automatic identification of other marine phenomena.
【学位授予单位】：上海海洋大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：S951.4

【相似文献】