Wave Density Spectra Estimation with LSTM from Sentinel-1 SAR in the Baltic Sea

Surface wave spectra estimation from Synthetic Aperture Radar (SAR) data has been a subject of many studies. In the open ocean, where high-resolution Wave Mode data capture predominately swell waves, various inverse methods have been developed to extract 2D wave information. However, these methods face limitations in wind-dominated regional seas such as the Baltic Sea, where swell waves are lacking, and ScanSAR (TOPS mode for Sentinel-1 (S1)) acquisitions with lower resolution are widespread. To overcome the problem, a long-short-term memory (LSTM) recurrent neural network is used to estimate 1D wave density spectra from S1 SAR image spectra. A total of 71,110 Sentinel-1 Interferometric Wide (IW) swath subimages along with the corresponding NORA3 WAM hindcast spectra are processed and used to train the LSTM model. The experiments show that the frequency-to-frequency correlations exceed 0.8 between 0.2 Hz and 0.4 Hz in the test dataset. The peak frequency of the estimated spectra differs from that of the baseline spectra by 0.13 Hz. The peak value of the spectra shows a difference of 0.38 m 2 /Hz. The bulk wave parameters calculated from the predicted spectra show a high correlation (0.78 for H <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</inf> ) and low RMSE values (0.39 m for HS ), which achieves similar performance than previous empirical methods.