Eddy Detection in SAR Images Based on Internal Wave Transfer Learning

Mesoscale eddies, as a widely occurring phenomenon in global oceans, play a significant role in the transport of energy, materials, and heat. Their observation relies on high-resolution Synthetic Aperture Radar (SAR). However, The Bragg scattering mechanism in SAR imaging is insensitive to vorticity of the eddy flow field, resulting in faint eddy features. This complicates sample collection and poses challenges in meeting deep learning requirements for both sample quantity and quality. Therefore, this study adopts a transfer learning approach for eddy detection. By leveraging the richness of internal wave data and the similarity of shallow-layer features between internal waves and eddies, internal waves are selected as the source domain for training, and the model is fine-tuned to adapt to eddy data. This method effectively addresses the scarcity of eddy data, achieving an accuracy of 71.4 % on the test set. Compared to training without transfer learning, the accuracy increased by 15.6 %, the recall rate improved by 16.2 %, the mAP50 rose by 12.7 %, and the convergence speed was faster.