The Ocean Mesoscale Eddy Detection Algorithms Research Based Onconvolutional Neural Network

Mesoscale eddies is also known as sea "storm" and has an important effect on ocean energy and material transportation, so it is of great research value. The traditional eddies detection algorithms based on geometrical characteristics of the flow field and sea surface height anomaly are more complex. Algorithms threshold settings are artificially affected. Algorithm applicable scope is limited. Convolution neural network is one of the deep learning algorithms and has been widely used in image recognition. In this paper, in order to improve the efficiency and accuracy of eddies detection, convolution neural network is applied to ocean mesoscale eddies detection.Based on the study of existing eddies detection methods and combining with the effective application of convolution neural network in image recognition, the convolution neural network is applied to the eddies detection. Finally, a vortex detection algorithm based on convolution neural network is realized. The research content is divided into two parts:On the one hand, the vortex detection algorithms based on the geometrical characteristics of the flow field and sea surface height anomaly are realized. The characteristics of the ocean eddies in the flow field and sea surface height anomaly are analyzed, and the vortex detection algorithms are realized by the characteristic constraint method.Contrast analysis of the accuracy of the two algorithms and the reasons for false detection and missed detection. The results show that the two algorithms are easy to implement, but have a high demand for computer performance.The threshold is sensitive, it is easy to cause false detection or missed detection.So the detection accuracy is relatively low. These two algorithms are suitable for eddies detection with less data.On the other hand, the vortex detection algorithm based on the convolution neural network is realized. Based on the analysis of CNN principle and structure, the convolution neural network is applied to mesoscale eddies detection.Flow field reanalysis data (based on ocean numerical simulation)can accurately characterize the velocity and direction of the mesoscale eddies but the eddy center is not clear, the sea surface height anomaly data can accurately reflect the eddy center position but is easy to missed detection. Combined with two kinds of data characteristics, the use of highly anomalous values for global detection, brushing the suspected eddy center point, the use of geometric characteristics of flow field to build a sample set. The purpose is to detect suspected eddy points for location detection and realize the vortex detection based on CNN. Finally,the detection results of the three methods are compared and analyzed.The results show that the eddy detection algorithm based on CNN is not only high accuracy, but also more suitable for eddies detection in large data background .