Three-dimensional thermohaline reconstruction of mesoscale eddies under remote sensing observation: From the perspective of deep learning of layer depth sequences with fusion of physical mechanisms

Abstract

Mesoscale eddies significantly impact the thermohaline structure of the ocean on a global scale. However, current three-dimensional reconstruction techniques for mesoscale eddies, based on multi-source data fusion, tend to focus on the profile while neglecting the depiction of mesoscale eddies in higher dimensions through three-dimensional structures. To address this issue, we first propose a hybrid recognition algorithm for mesoscale eddies. We then extract the dataset based on the recognition results and ocean reanalysis, transforming the mesoscale eddies reconstruction problem into a prediction problem of layer depth sequences, and implement the model construction using deep learning technology. Simultaneously, we incorporate the globally uniform vertical and horizontal structure of mesoscale eddies into the model input module as a binding physical mechanism and add an attention mechanism to enhance the model's output. Experiments demonstrate that the model developed in this paper performs comparably to the base model in deep learning metrics, exhibits specific advantages in measuring the three-dimensional structure of mesoscale eddies from multiple perspectives, and shows robust generalization across different oceanic regions and data sources. Inspired by the work of many researchers, this paper achieves promising results in the three-dimensional reconstruction of mesoscale eddies, offering valuable insights for deep learning research in marine data-related fields.