Wind-sea and swell separation of 1D wave spectrum by deep learning

In this study, three existing methods to separate wind-sea and swell from 1D wave spectra, namely the Pier-son-Moskowitz (PM) method, the steepness method (HP method), and the overshooting method (JP method), were evaluated statistically using data from 38 meteorological buoys from National Data Buoy Center (NDBC) over the period 2010-2020. Among the three methods, the PM method shows the best agreement with the 2D separation method, because it uses the wind speed information as an input term. Using these buoy data, a new 1D wind-sea-swell separation method based on deep learning is proposed. This new method can directly compute the wind-sea or swell significant wave height (SWH) from a 1D wave spectrum with or without wind speed data as input. When the wind speed is used as an input term, the overall root-mean-square error (RMSE) of the method for wind-sea/swell SWH is 0.27/0.36 m compared to the 2D separation method, which outperforms all existing 1D separation methods. When there is no wind speed data, the RMSE of wind-sea/swell SWH can still reach 0.36/ 0.41 m, which is similar to the accuracy of the PM method uses wind speed as input. Both methods are robust and have no significant geographical dependence.

Automated summary

In this study, three existing methods to separate wind-sea and swell from 1D wave spectra were evaluated statistically using data from 38 meteorological buoys. Among the three methods, the PM method shows the best agreement with the 2D separation method. A new 1D wind-sea-swell separation method based on deep learning is proposed, which outperforms all existing 1D separation methods. Both methods are robust and have no significant geographical dependence.