GSDNet: A deep learning model for downscaling the significant wave height based on NAFNet

Abstract

Finer resolution is one of the development trends in ocean surface wave simulation and forecasting. However, high-resolution numerical models for ocean surface waves have led to an enormous increase in computational complexity, posing a challenge with respect to balancing computational efficiency and timeliness. To meet the demand for refined ocean surface wave forecasting and to address the computational efficiency challenge of high-resolution ocean surface wave models, we propose a downscaling model called the Global location-Specific transformation Downscaling Network (GSDNet) based on the non-autoregressive fusion network (NAFNet). By incorporating global location-Specific transformation and introducing a land–sea distribution indicator, GSDNet can quickly and accurately map low-resolution significant wave heights to high-resolution grids. The results show that, compared with traditional interpolation methods such as the bilinear, inverse distance weight interpolation (IDW), and bicubic methods, the GSDNet model can reduce the global mean absolute error (MAE) by >77%. Compared with those of FourCastNet (FCN), the Koopman neural operator (KNO), the original NAFNet, and residual networks in deep learning from empirical downscaling methods (DL4DS_ResNet), the MAE decreases by >21%. Furthermore, the GSDNet model outperforms the other downscaling methods at the coastal boundary and for identifying the maximum significant wave height. In this work, we provide an effective solution for balancing computational efficiency and timeliness, which is important for improving the accuracy and reliability of wave forecasting.