Validation and calibration of partitioned integral ocean wave parameters from co-polarized synthetic aperture radar data

Abstract

The inversion of wave spectrum is the embodiment of the fine observation of ocean swells by synthetic aperture radar. Due to the nonlinear distortion from velocity bunching, the operational Sentinel-1 Wave Mode (WV) level-2 (L2) Ocean (OCN) wave products retrieved by quasi-linear inversion are still quite different from the real wave spectrum, requiring further calibration. Limited by the azimuth cut-off, the quasi-linear inversed wave spectrum mostly concentrates on swells with a wavelength longer than 150 m. 2D spectral partitioning separates different swell systems, denoted as partitions in the Sentinel-1 WV L2 OCN product. We compared and calibrated the Sentinel-1 WV L2 OCN products' swell partitions (More specifically partitioned integral ocean wave parameters: partitioned wave height, partitioned wavelength, and partitioned wave direction) with WAVEWATCH III (WW3) in different polarization modes (HH and VV polarization) and different incidence angles (∼23°-WV1 and ∼ 36°-WV2). The accuracy of swell partitions in the OCN products varies with peak wave direction relative to the Sentinel-1 flight direction, and it is difficult to calibrate the partitions derived from quasi-linear inversion unless deleting partitions severely distorted by nonlinearity and large-scale features. A machine learning method is proposed to calibrate the wave partitions. Spectral distances of collocated Sentinel-1 and WW3 partition pairs larger than two are considered “bad” partitions to be filtered. Inputting high-dimensional features (twenty-one features) derived from Sentinel-1 WV L2 OCN products into the extreme gradient boosting (XGBoost) model can effectively reduce the bias (root-mean-square error, RMSE) of significant wave height (SWH) of all the partitions from ∼0.10 m (∼0.70 m) to ∼0.00 m (∼0.46 m) and the bias (RMSE) of wavelengths of all the partitions from ∼20 m (∼35 m) to ∼0.20 m (∼26 m) compared with WW3 swell partitions. This method improves the precision of the wave partition and retains more partitions in the Sentinel-1 WV L2 OCN products than previous studies, especially for wave partitions with moderate wave height (< 4 m), potentially improving the understanding of swell dynamics globally.