基于YOLO算法的海洋中尺度涡三维结构构建

3D Strucure Construction of Mesoscale Eddies Based on YOLO Intelligent Algorithm

摘要:

海洋中尺度涡是一种常见的中尺度海洋现象，研究海洋中尺度涡的分布及运动特性对航运、气候、军事等具有重要作用，海洋中尺度涡的识别是海洋学和计算机科学领域的一个热门研究课题。运用深度学习的方法和框架，对中尺度涡的二维识别和三维结构构建展开研究分析。首先，获取全球海洋再分析数据并进行流线可视化，构建涡旋流线数据集；其次，利用 YOLO v5s 卷积神经网络对涡旋流线数据集进行训练， 并对南海区域中尺度涡进行有效检测。实验结果表明，YOLO v5s 训练后得到最优模型经过测试，平均检测精度均值达到了 86.10%；最后，根据涡旋检测结果，对检测出的同时刻不同深度的涡旋判断是否属于同一涡旋， 确定后进行该涡旋的三维结构构建。

Abstract:

Mesoscale eddies are a common marine mesoscale phenomenon. Studying the distribution and motion characteristics of mesoscale eddies plays an important role in shipping，climate，and military affairs. At present，the identification of marine mesoscale eddies is a hot research topic in the field of oceanography and computer science. In this paper，the two-dimensional identification and three-dimensional structure construction of mesoscale eddies are studied and analyzed by using deep learning method and framework. Firstly，the global marine physics reanalysis data are obtained to visualize streamlines to build the dataset of eddy streamlines. Secondly，the YOLO v5s convolutional neural network is used to train the eddy dataset and to effectively detect mesoscale eddies in the South China Sea region. The experimental results show that the mean precision accuracy of the optimal model obtained after YOLO v5s training reaches 86.10%，which could provide an effective basis for the identification of mesoscale eddies and the construction of 3D structures. Finally，based on the eddy detection results at the same time， the detected eddies at different depths are determined whether they belonged to the same eddy，then the 3D structure of the eddy is built.