Global ocean mesoscale vortex recognition based on DeeplabV3plus model

Abstract

Ocean mesoscale vortices play a very important role in global energy and material transport, and to a large extent affect the distribution of nutrients and phytoplankton. Traditional mesoscale vortex extraction methods have the problems of dependence on threshold and sensitivity to noise. In recent years, the machine learning methods that have emerged in recent years have also made the generalization ability of the model poor due to the limited coverage of the training data set, resulting in most methods only suitable for extraction in specific sea areas. This paper uses daily global sea level data from 2008 to 2017, combined with the py-eddy algorithm, to complete the construction of a label data set covering the entire sea. At the same time, based on the DeeplabV3plus model, by adjusting the loss function, a recognition model conforming to the mesoscale vortex characteristics is realized. Experimental results show that the verification accuracy rate of the model reaches 80.5%, and the Kappa coefficient is 0.758. Compared with the previous extraction method, the accuracy of this model is increased by 13.8%, and the Kappa coefficient is increased by 0.255. Experiments show that the mesoscale vortex recognition method proposed in this paper has good recognition accuracy for different sea areas and different scale vortices.