A potential technique for SAR ocean surface current speed retrieval based on machine learning

This is the authors' abstract. We don't add key points for this paper.

The purpose of this study was to develop a machine learning-based algorithm for azimuthal ocean surface current retrieval from Sentinel-1 (S-1) synthetic aperture radar (SAR) images. More than 4500 images from 2020 to 2022 of regions where the western boundary current occurs, i.e. the western Pacific and western Atlantic, were acquired in interferometric wide (IW) mode and collocated with ocean surface current data from the HYbrid Coordinate Ocean Model (HYCOM). The wave spectra from these images were inverted using a theoretical algorithm called the Parameterized First-guess Spectrum Method (PFSM). Comparisons between SAR retrievals and measurements from the National Data Buoy Center (NDBC) showed a 0.52 m root mean squared error (RMSE) for significant wave height (SWH) with a 0.78 correlation (COR). The azimuthal cutoff wavelength was simulated using the SAR-derived wave spectrum. We then determined the dependence of sea surface dynamics, i.e. wind speed, SWH, and ocean surface current speed, on the Doppler centroid anomaly (DCA) and the difference between SAR-measured azimuthal cutoff wavelength and the simulated value. The three machine learning methods (eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), convolutional neural networks (CNNs) were trained on 4000 images to construct the relationship between azimuthal ocean surface current speed and SAR-measured variables. The other >500 images were used to validate the algorithms against the HYCOM ocean surface current data and observations from NDBC buoys. We found that the RMSE of the current speed by XGBoost was 0.14 m/s with a 0.57 COR, which is better than other two machine learning methods (>a 0.15 m/s RMSE with a <0.45 COR), i.e. LightGBM and CNNs. Based on these results, we believe that the XGBoost method is a potential technique for SAR ocean surface current speed retrieval without any external information.