Single-instant spatial wave height forecast using machine learning: An image-to-image translation approach based on generative adversarial networks

Abstract

An innovative single-instant approach for spatial significant wave height forecast based on a Generative Adversarial Networks (GANs) model is presented. The model's most distinctive feature is its ability to provide spatial wave height forecast with only a single snapshot of wind field input, eliminating the need for continuous wind field sequences while simultaneously maintaining high accuracy and enhanced efficiency. The proposed model architecture comprises an image-to-image translation generator and a patch-style discriminator, working in tandem to forecast high-quality wave height data. The model is trained using a 20-year ECMWF Reanalysis v5 (ERA5) wind and wave dataset. A vast sea area covering 15°-40°N, 105°-130°E in West Pacific is used for this study. The presented model is carefully tested against a 10-year period historical dataset and the results are discussed. The model exhibits outstanding overall performances with Mean Absolute Error (MAE) of 0.16 m, Root Mean Square Error (RMSE) of 0.24 m, and Mean Absolute Percentage Error (MAPE) of 13.4 %. The model proves particularly effective in forecasting extreme wave heights during tropical cyclones, capturing the structure of typhoon-driven waves and providing wave heights with favorable accuracy. It is also proficient at delivering satisfactory statistical wave height forecasts for long-term periods.