基于卷积神经网络的海洋中尺度涡旋检测算法研究

发布时间：2018-07-29 21:02

【摘要】：中尺度涡又称为海洋“风暴”,对海洋能量和物质运输具有重大作用,具有重大的研究价值。传统的基于流场几何特征和高度异常值的涡旋检测算法不仅复杂度高,而且阂值设置受人为影响比较大,适用范围有限。卷积神经网络(Convolutional Neural Network,CNN)是深度学习的算法之一[1],已经被广泛应用在图像的识别方面。本文致力于将卷积神经网络引入海洋中尺度涡旋检测,以期提高涡旋检测的效率和精确度。本文在基于对现有涡旋检测方法研究的基础上,结合卷积神经网络在图像识别方面的有效应用,将卷积神经网络引入到涡旋检测场景中,实现基于卷积神经网络的涡旋检测算法。研究内容主要分为两部分:一方面,本文实现基于流场几何特征和基于高度异常值的涡旋检测算法。分析海洋涡旋在流场和高度异常中的特征,以特征约束方式实现涡旋检测。对比分析两种算法的准确度以及误检和漏检的原因。结果证明:这两种算法实现较容易,但对计算机的计算性能比较高,阂值比较敏感,容易造成误检或漏检,检测准确率比较低,适合数据量较少的涡旋检测。另一方面,本文实现基于CNN的涡旋检测算法。在分析CNN原理和结构研究的基础上,将卷积神经网络应用到中尺度涡旋检测中。流场再分析数据(基于海洋数值模拟计算)可以精确表征中尺度涡的速度和方向但涡心不清晰,海面高度数据可以准确反映涡心位置但容易误检测。结合两种数据特点,利用高度异常值进行全局检测,刷选疑似涡旋中心点,利用流场几何特征构建检测样本集,对疑似涡心点进行局部检测,实现基于CNN的涡旋检测。最后将三种方法检测结果进行对比和分析,结果表明:基于CNN的涡旋检测不仅准确率高,而且更加适合大数据背景下的涡旋检测。
[Abstract]:Mesoscale vortex, also called ocean "storm", plays an important role in ocean energy and material transport, and has great research value. The traditional vortex detection algorithm based on the geometric characteristics of the flow field and the height outliers not only has a high complexity but also has a large artificial influence on the setting of the threshold value and has a limited range of application. Convolutional Neural Network (Convolutional Neural) is one of the depth learning algorithms, which has been widely used in image recognition. In this paper, convolution neural network is introduced into ocean mesoscale vortex detection in order to improve the efficiency and accuracy of vortex detection. Based on the research of the existing vortex detection methods and the effective application of convolution neural network in image recognition, the convolutional neural network is introduced into the scroll detection scene. The algorithm of vortex detection based on convolution neural network is realized. The research is mainly divided into two parts: on the one hand, a vortex detection algorithm based on the geometric characteristics of the flow field and the height outliers is implemented in this paper. The characteristics of ocean vortices in current field and height anomaly are analyzed, and vortex detection is realized by means of feature constraint. The accuracy of the two algorithms and the causes of false detection and missed detection are compared and analyzed. The results show that these two algorithms are easy to implement, but they have high computing performance and sensitive threshold value, which are easy to cause false detection or miss detection, and are suitable for vortex detection with less data volume. On the other hand, this paper implements a vortex detection algorithm based on CNN. On the basis of analyzing the principle and structure of CNN, the convolution neural network is applied to mesoscale vortex detection. The reanalysis data (based on ocean numerical simulation) can accurately characterize the velocity and direction of mesoscale vortices but the vortex center is not clear. The sea surface height data can accurately reflect the location of vortex center but is easy to be misdetected. Combined with the two kinds of data characteristics, the global detection is carried out by using the height outliers, the suspected vortex center is selected by brushing, the sample set is constructed by using the geometric characteristics of the flow field, the local detection of the suspected vortex center is carried out, and the vortex detection based on CNN is realized. Finally, the results of the three methods are compared and analyzed. The results show that the vortex detection based on CNN is not only accurate, but also more suitable for vortex detection under big data background.
【学位授予单位】：山东科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;P714.1

【参考文献】