Detecting Ocean Eddies with a Lightweight and Efficient Convolutional Network

Abstract

As a ubiquitous mesoscale phenomenon, ocean eddies significantly impact ocean energy and mass exchange. Detecting these eddies accurately and efficiently has become a research focus in ocean remote sensing. Many traditional detection methods, rooted in physical principles, often encounter challenges in practical applications due to their complex parameter settings, while effective, deep learning models can be limited by the high computational demands of their extensive parameters. Therefore, this paper proposes a new approach to eddy detection based on the altimeter data, the Ghost Attention Deeplab Network (GAD-Net), which is a lightweight and efficient semantic segmentation model designed to address these issues. The encoder of GAD-Net consists of a lightweight ECA+GhostNet and an Atrous Spatial Pyramid Pooling (ASPP) module. And the decoder integrates an Efficient Attention Network (EAN) module and an Efficient Ghost Feature Integration (EGFI) module. Experimental results show that GAD-Net outperforms other models in evaluation indices, with a lighter model size and lower computational complexity. It also outperforms other segmentation models in actual detection results in different sea areas. Furthermore, GAD-Net achieves detection results comparable to the Py-Eddy-Tracker (PET) method with a smaller eddy radius and a faster detection speed. The model and the constructed eddy dataset are publicly available.