Deep Learning Improves GFS Sea Surface Wind Field Forecast Accuracy in the Northwest Pacific Ocean

Abstract

Sea surface winds influence shipping, fisheries, and coastal projects. However, the current sea surface wind forecast exhibits noticeable biases. This study introduces a deep learning (DL)‐based bias correction model, WindNet, to improve the Global Forecast System (GFS) sea surface wind field forecast in the Northwest Pacific Ocean (NWPO). WindNet reduces the Root Mean Squared Errors (RMSEs) of wind speed at lead times of 24, 48, and 72 hr from 1.41–1.95 to 1.11–1.55 m s−1, achieving percentage reductions of 20.51%–21.28%. Simultaneously, the RMSEs of wind direction are reduced from 29.67–41.45° to 25.38–36.81°, demonstrating percentage reductions of 11.19%–14.46%. During typhoon passages, the RMSEs of wind speed and direction at three forecast lead times after using WindNet are reduced from 1.57–2.42 to 1.24–1.95 m s−1 and from 30.31–42.35° to 25.88–37.64°, demonstrating percentage reductions of 19.42%–21.02% and 11.12%–14.62%. By integrating a Squeeze‐and‐Excitation Network into WindNet, we find that utilizing information from the circulation field, apart from the zonal and meridional wind components at 10 m height, is crucial for the correction of the sea surface wind speed. WindNet can effectively capture the non‐linear relationship between other low‐level‐circulation‐related variables and sea surface wind speed. Therefore, WindNet remarkably enhances sea surface wind field forecast accuracy in NWPO.