Research On Intelligent Identification Method Of Ocean Eddies Based On Fully Convolutional Networ

Ocean eddy identification is an important field in marine science research and is of great significance for understanding ocean circulation,ecosystems,and climate change.Although deep learning-based methods have made some progress in ocean eddy identification research,there are still some issues that need to be addressed.Firstly,the radius of eddies varies greatly,ranging from one kilometer to several hundred kilometers,and their edges present highly irregular shapes,which significantly increases the difficulty of eddy identification.Secondly,existing research mainly focuses on two-dimensional eddy identification,neglecting the rich three-dimensional structural information of eddies.To address these issues,we mainly carries out the following two aspects of work:First,in order to effectively identify eddies of different scales and preserve the edge information of eddies,we propose an edge-enhanced fully convolutional network.This network consists of two branches: the eddy identification branch and the edge extraction branch.Both adopt the encoder-decoder framework and share the encoder with each other.The eddy identification branch simultaneously utilizes the multi-scale convolutional module in the encoder and the skip-layer connections between the encoder and decoder to learn multi-scale features,thus effectively identifying eddies of different scales.The purpose of the edge extraction branch is to learn the edge information of eddies,thereby enhancing the edge recognition ability of the eddy identification branch.To evaluate the identification performance of the edge-enhanced fully convolutional network,this paper conducts extensive experiments on publicly available sea surface height datasets.The experimental results show that the method we proposed can achieve higher performance than existing models.Second,in order to fully consider the three-dimensional structural information of eddies,based on the above work,we further propose a three-dimensional eddy identification model that integrates fully convolutional networks and Transformer.This model still adopts the framework of encoders and decoders.Specifically,the encoder extracts the local spatial features of eddies through a fully convolutional network,while simultaneously using the Transformer to model the interrelationships between different depth layers of three-dimensional eddies,thus effectively solving the identification problem of three-dimensional eddies.In addition,since the data volume of three-dimensional eddies is much larger than that of two-dimensional eddies,the overall computational overhead of the network is significantly increased.Therefore,this paper further streamlines the network.In detail,the linear projection operations and multi-layer perceptron layers with a large number of parameters in the Transformer are replaced with convolutional operations,and the regular convolutional layers in the fully convolutional network are replaced with separable convolutions.To evaluate the performance of the model,extensive experiments were conducted on a three-dimensional temperature-salinity dataset.The experimental results show that the method we proposed outperforms existing methods in terms of performance.