Enhancing Significant Wave Height Simulations Using Distribution‐Aware Deep Learning and Environmental Predictors

Abstract Numerical wave models are widely used for wave forecasting and hindcast data set construction; however, their outputs frequently exhibit systematic biases stemming from uncertainties in wind forcing, physical parameterizations, and bathymetry. To address this issue, a deep learning–based bias correction approach is applied to significant wave height simulations in the western North Pacific, incorporating multisource environmental predictors and a distribution‐aware loss function. A U‐Net–based spatiotemporal architecture was adopted as the prediction model and was found to outperform ConvLSTM and CNN alternatives in preliminary comparisons. To improve performance under diverse sea states, a distribution‐aware loss function derived from the Rayleigh distribution is introduced to increase the relative weight of high‐wave samples during training, leading to a 14.9% reduction in RMSE under high‐wave conditions. In addition, wind, wave, and current variables are combined as 12 input channels, together with a current‐dependent gating mechanism and temporal aggregation of preceding states, which enhances model performance in regions of strong ocean currents. Applied to WAVEWATCH III simulations, the proposed model reduces the overall RMSE from 0.37 to 0.24 m with a mean bias of 0.01 m. Under extreme sea states exceeding 6 m, RMSE and bias are further reduced by 22% and 46%, respectively, relative to the ERA5 reanalysis product. Independent validation using satellite altimeter and drifting buoy observations indicates consistent performance improvements across data sets.