A Dual-Attention Mechanism Deep Learning Network for Mesoscale Eddy Detection by Mining Spatiotemporal Characteristics

Mesoscale eddies are a mechanism for ocean energy transfer, and identifying them on a global scale provides a means of exploring ocean mass and energy exchange between ocean basins. There are many widely used model-driven methods for detecting mesoscale eddies; however, these methods are not fully robust or generalizable. This study applies a data-driven method and proposes a mesoscale detection network based on the extraction of eddy-related spatiotemporal information from multisource remote sensing data. Focusing on the northwest Pacific, the study first analyzes mesoscale eddy characteristics using a combination of gridded data for the absolute dynamic topography (ADT), sea surface temperature (SST), and absolute geostrophic velocity (UVG). Then, a deep learning network with a dual-attention mechanism and a convolutional long short-term memory module is proposed, which can deeply exploit spatiotemporal feature relevance while encoding and decoding information in the gridded data. Based on the analysis of mesoscale eddy characteristics, ADT and UVG gridded data are selected to be the inputs for the detection network. The experiments show that the accuracy of the proposed network reaches 93.38%, and the weighted mean dice coefficient reaches 0.8918, which is a better score than those achieved by some of the detection networks proposed in previous studies, including U-Net, SymmetricNet, and ResU-Net. Moreover, compared with the model-driven approach used to generate the ground-truth dataset, the network method proposed here demonstrates better performance in detecting mesoscale eddies at smaller scales, partially addressing the problem of ghost eddies.