Accurate detection of ocean eddies using satellite remote sensing data is crucial for marine environmental monitoring and climate studies. Although deep learning models have been applied to this task, existing models primarily rely on computationally expensive, large-scale convolutional structures. These models exhibit structural homogeneity and inefficient eddy feature extraction. Therefore, this study presents Ledformer, a lightweight CNN-Transformer model designed for efficient eddy detection. The encoder achieves hybrid eddy feature extraction by integrating Attention Ghost Bottleneck (AGB) modules with Transformer blocks. In the decoder, the Eddy Feature Mixing (EFM) module is proposed, combining it with multiple attention networks to fuse multi-scale eddy features jointly. To improve detection accuracy, a data-fused eddy dataset is constructed using sea level anomaly (SLA) with meridional and zonal velocity data. This approach effectively reduces the misdetection of pseudo-eddies when only relying on SLA data. On the eddy dataset, Ledformer achieves an F1 score of 94.24%, an Accuracy of 98.22%, and a Mean Intersection over Union (mIoU) of 89.34%. Results show that the model maintains high efficiency while achieving a lightweight structure. Its detection performance outperforms existing deep learning models and produces fewer false detections than the physics-based Py-Eddy-Tracker (PET). The proposed Ledformer provides reliable and efficient eddy detection technology for remote sensing applications. • Proposes Ledformer, a CNN-Transformer model for eddy detection. • Constructs a new eddy dataset fusing SLA, Ugos, and Vgos data. • Introduces AGB and EFM modules for efficient eddy feature extraction. • Achieves 89.34% mIoU, outperforming other eddy detection models. • Reduces eddy misdetections compared to the physics-based PET method.
Sun et al. (Sun,) studied this question.