A Machine-Learning-Based Ocean-Current Velocity Inversion Model Using OCN From Sentinel-1 Observations

Abstract

High-precision ocean-current velocity inversion is crucial for maritime activities. Synthetic aperture radar (SAR) has become a key data source for ocean-current velocity inversion. However, traditional methods, such as the Doppler centroid anomaly (DCA) and along-track interferometry methods, face challenges, such as low inversion accuracy, poor robustness, and limited data sources. This study developed OCN-CIM, a machine-learning-based model that directly derives the radial ocean-current velocity from Sentinel-1 observations. The model is trained using 186 scenes of Sentinel-1 Level 2 ocean data (OCN) collected between 14 July 2020 and 16 May 2024, in regions with strong currents along the East Coast of the United States. The ground truth is obtained from matched high-frequency radar data. Built on a fully connected neural network, the OCN-CIM features a custom loss function focused on high ocean-current velocities. The model achieved a mean absolute error (MAE) of 0.16 m/s, root-mean-square error (RMSE) of 0.20 m/s, and mean deviation (MD) of 0.005 m/s on the test dataset. When applying the OCN-CIM to ten independent cases, the average MAE, RMSE, and MD were 0.13 m/s, 0.16 m/s, and −0.03 m/s, compared with 0.26 m/s, 0.34 m/s, and 0.06 m/s for the traditional DCA method, demonstrating significant improvement in inversion accuracy. In addition, the OCN-CIM exhibits robustness, with reduced sensitivity to local wind and SAR data anomalies, and consistent results across various electromagnetic direction error-correction methods.