Author(s): Lei Ren; Yaqi Wang; Jun Wei; Michael Hartnett
Linked Author(s): Lei Ren, Michael Hartnett
Keywords: Submesoscale eddy High Frequency radar Machine learning Prediction Northern South China Se
Abstract: Submesoscale eddies are specialized circulation structures predominantly active in the upper layers of the ocean. They constitute pivotal elements in the exploration of oceanic circulation, marine ecosystems, and the transport of energy and materials within the ocean. Moreover, they serve as focal points in contemporary physical oceanography research, occupying a prominent position at the forefront of scientific inquiry. At present, the identification and prediction of submesoscale eddies still have certain deficiencies, including the accuracy of the identification algorithm needs to be further improved, and the traditional numerical model method is difficult to accurately predict the nonlinear evolution characteristics of eddies. With this background, based on the comparison and evaluation of the results of the traditional algorithms (Vector Geometry algorithm and Okubo-Weiss algorithm), this study proposes a submesoscale eddy identification algorithm based on information entropy to solve the problems of "one vortex with multiple kernels" and identification overload in the traditional algorithm. Further, this study constructed a submesoscale eddy dataset in the northern South China Sea. And its spatial and temporal distribution characteristics, evolutionary laws, and influencing factors were investigated. Furthermore, machine learning methods were used to construct a prediction model for the remaining lifetime and propagation trajectory of submesoscale eddies, and the model prediction results are evaluated. The main conclusions of this study are as follows: (1) Based on the High Frequency Radar surface current data, submesoscale eddies in the northern South China Sea are extracted using the Okubo-Weiss algorithm and the Vector Geometry algorithm. And the eddy data are compared and analyzed using the manual discrimination results as a control group, which show that there is a significant over-identification of eddies in the Okubo-Weiss algorithm, while there are eddy under-measurements and under or over-estimation of sizes in the Vector Geometry algorithm. To address these issues, an automatic eddy identification method incorporating the information entropy is proposed, aiming to separate the eddy region from the background flow field based on the principles of maximum entropy, standard deviation of entropy value, and adding the direction of rotation of the velocity vector as a constraint. The method has a detection rate of 82% for submesoscale vortices and an excess detection rate of 3%, and it is better than the Okubo-Weiss and Vector Geometry algorithms in the identification of parameters such as the number and size of eddies. (2) The information entropy-based submesoscale eddy identification method is applied to construct a submesoscale eddy dataset in the northern South China Sea. A total of 512 submesoscale eddies occur in the northern South China Seaf from April to June 2021, mainly cyclonic eddies, with a major radius of 13 ~ 20 km and a duration of 0 ~ 120 min. The eddies are mainly generated and extinguished in the northwestern part of the study area, and they mainly show a southwesterly propagation tendency. And mean propagation velocity of eddies decreases with increasing latitude, the mean meridional velocity component increases with increasing longitude, and the mean latitudinal velocity component shows a characteristic pattern of increasing and then decreasing with increasing longitude. Spatially, the submesoscale eddies in the northern South China Sea are significantly affected by the background current field. On the one hand, the eddies obtain energy from the background current field, and on the other hand, the eddies regulate the development of the background currents, and the eddies in the northwestern near-shore area are also affected by the friction of the topography. Temporally, the eddy kinetic energy in the northern South China Sea is characterized by a decrease followed by an increase from April to June, with June showing more pronounced changes, attributed to strong southwest winds, which serve as significant triggers for frequent cyclonic eddy occurrences in June and provide the main energy for eddy development. (3) To study the prediction of the remaining lifetime of submesoscale eddies in the northern South China Sea using a ResNet model, and the prediction of eddy propagation trajectories using a CNN-BiLSTM-Attention model consisting of a Convolutional Neural Network (CNN), Bi-directional Long Short-Term Memory (BiLSTM), and Attention mechanism. The prediction results of the remaining lifetime show that the ResNet model has a better prediction performance, with "R" ^"2" up to 0.83, which is improved by 32% compared with the Random Forest algorithm. Among all characteristic parameters, vorticity and latitude, have more significant effects on the model prediction effect. Meanwhile, during the life cycle of the submesoscale eddies, the more significant the degree of change of the physical characteristic parameters of eddies, the higher the predictability of its remaining life. However, when the variation of eddy physical characteristic parameters exceeds a certain threshold range, it negatively affects the prediction of the remaining lifetime. The prediction results of the propagation motion trajectories show that the CNN-BiLSTM-Attention model has better prediction results compared to the LSTM network model at different initial learning rates. As the input time step increases, the model prediction accuracy keeps improving. The R^2 of the latitude parameter prediction model can reach up to 0.97, while the R^2 of the longitude parameter model can reach 0.93. At the same time, the analysis of the importance of the model input parameters shows that the predicted trajectories of the eddy propagation motions are not only related to the historical trajectories, but also associated with their own physical characteristic parameters. To study the prediction of the remaining lifetime of submesoscale eddies in the northern South China Sea using a ResNet model, and the prediction of eddy propagation trajectories using a CNN-BiLSTM-Attention model consisting of a Convolutional Neural Network (CNN), Bi-directional Long Short-Term Memory.
Year: 2025