Identification and Segmentation of Internal Solitary Waves in the East China Sea: A TransUNet Approach Using Multi-Source Satellite Imagery

The East China Sea (ECS) is a globally active region for internal solitary waves (ISWs); however, its overall spatiotemporal distribution remains poorly understood. To address this gap, this study proposes a deep learning method based on multi-source remote sensing imagery (MODIS and SAR) for the intelligent identification and pixel-level segmentation of ISWs in the ECS. We adopted the TransUNet model, which combines the global context-capturing capability of Transformers with the fine-grained segmentation advantages of U-Net to effectively handle the large-scale continuous characteristics of ISWs. The model achieved a Dice coefficient of 71.0% and a precision of 72.7% on the test set, significantly outperforming existing models such as FCN, SegNet, DeepLabV3+, and U-Net. Using this automated framework, multi-source satellite data from 2002 to 2024 were processed to generate the first high-resolution spatiotemporal map of ISWs covering the entire ECS. The map reveals two spatial hotspots: a primary one at the shelf break northeast of Taiwan and a secondary one in the waters southwest of Jeju Island. Furthermore, ISWs exhibit a marked seasonal cycle in both occurrence frequency and properties, peaking in summer and minimizing in winter. This seasonal pattern aligns closely with the physics of internal tide generation via body forcing. By providing the first long-term, high-resolution ISW dataset for the entire ECS, this study demonstrates the potential of deep learning techniques for ISW research in complex marginal seas.