Deep learning-based downscaling of ocean surface vector wind over the Taiwan Strait and its adjacent seas with Pangu-weather

Abstract

As a critical economic hub for fisheries and maritime transport, the Taiwan Strait and its adjacent seas urgently require high-resolution ocean surface vector wind (OSVW) forecasting with extended lead times. However, accurate OSVW forecasting in this region is particularly challenging due to complex monsoon regimes and unique geography. In this study, we developed UNet-based deep learning (DL) downscaling approaches and enhanced the resolution of the DL-based weather prediction model, Pangu-Weather, from 0.25° to 0.03° for OSVW over the Taiwan Strait and its adjacent seas. The model was trained using high-resolution (0.01°) atmospheric reanalysis data from the High-Resolution China Meteorological Administration Land Data Assimilation System (HRCLDAS). Results show that the Temporal Enhanced UNet (TempE-UNet) slightly improved overall prediction skill compared to the time-agnostic UNet during the 120 h (5 days) forecast window. Additionally, TempE-UNet more accurately captured local wind field characteristics than the standard UNet. Our results demonstrate that incorporating temporal information as an additional predictor enhances downscaling performance, offering a promising paradigm for operational OSVW forecasting.