A posteriori improvement of wave reanalysis data in the Mediterranean Sea using machine learning

Reliable estimation of ocean wave parameters is fundamental for numerous applications. Nevertheless, wave products from widely used datasets such as the ERA5 reanalysis developed by the European Centre for Medium-Range Weather Forecasts are affected by systematic biases, particularly an underestimation of wave heights in semi-enclosed basins such as the Mediterranean Sea. This study addresses these limitations by developing a comprehensive methodology for calibrating and integrating spectral wave parameters in the Mediterranean Sea, taking the significant wave height as an example. The proposed approach synchronizes ERA5 reanalysis data with wave buoy in situ measurements, to train machine learning functions that map ERA5 to the wave buoy data using multi-layered perceptron. The system learns the intricate nonlinear relationships between ERA5 wave data and concurrent and collocated observed wave buoy data, as evident in the time series, annual, and seasonal trends. The results demonstrate significant improvements in the accuracy of the calibrated ERA5 significant wave height, resulting in substantial reductions of error metrics. In addition, the Nash-Sutcliffe Efficiency coefficient consistently improved to over 0.85, confirming superior predictive capability and alignment with wave buoy measurements across eight Mediterranean locations. Furthermore, the trained models demonstrate strong spatial generalizability, enabling accurate calibration in data-sparse regions and providing robust historical and future wave data.