High-Precision Deep Learning-Based Extraction of Internal Wave Signatures From SWOT Data Over the Sulu Sea

Abstract

Internal waves (IWs) are a critical dynamic process in the density-stratified ocean. Surface water and ocean topography (SWOT) provides high-resolution, wide-swath, and precise sea-surface observations, creating new opportunities for accurate internal wave detection. In this study, we propose VSSI-Net, a deep learning-based method for intelligent extraction of IW crest lines from SWOT data. This method is specifically designed for the characteristics of SWOT data. By integrating the omnidirectional image scan (ODIS) mechanism and the implicit higher order feature transformation block (IHFTB), it effectively enhances model robustness in complex backgrounds. Furthermore, a specially tailored loss function, termed FocalMCCLoss, is designed to alleviate the class-imbalance problem. Experimental results demonstrate that VSSI-Net achieves high accuracy and robustness in IW crest-line segmentation over the Sulu Sea, attaining a precision of 92.32% and an $F1$ -score of 91.51%, and outperforming the competing models in overall performance. SWOT observations indicate that IW detections in the Sulu Sea are concentrated primarily in the western region, with relatively enhanced observed activity during the rainy season. This apparent spatiotemporal pattern may be modulated by regional bathymetry and propagation conditions while also being influenced by SWOT ground-track coverage and overpass-time sampling. In summary, VSSI-Net provides a novel approach for high-accuracy automated extraction of IW crest lines from SWOT data and shows promising application prospects.