Crest extraction of internal solitary waves from synthetic aperture radar images based on an improved segformer model

Accurate detection of multiscale oceanic internal solitary waves (ISWs) is a pivotal challenge in physical oceanography, with important applications for marine energy transmission and engineering. This study proposes MS-SF-ISWNet, a multiscale attention network model based on SegFormer, engineered to identify multiscale internal solitary ISWs in Sentinel-1 Synthetic Aperture Radar (SAR) images. The model enhances feature extraction capabilities for ISWs across diverse scales by integrating a split attention mechanism. In addition, it employs a hybrid loss function that combines Dice Loss and Focal Loss to address sample imbalance, thereby improving segmentation accuracy and robustness. Utilizing Sentinel-1 Interferometric Wide Swath (IW) mode images as the data source, a dataset was assembled from prominent ISW-active regions, including the South China Sea and the Andaman Sea. Validation was conducted across various sea areas, such as the Lombok Strait and Maluku Strait, which were excluded from the training dataset. Results demonstrated that the F1-score and recall reached 88.45% and 90.76%, respectively. Furthermore, the model exhibited exceptional generalization capabilities for data obtained from other SAR sensors, including GF-3 and Envisat. Ultimately, this study successfully detected multiscale ISW features in full-orbit SAR images.