Evaluation of deep learning approaches for long-term gap imputation in a sparse wave buoy network in the Gulf of Mexico

Wave buoy records are essential for ocean engineering applications, yet long-term data gaps often compromise observation continuity in sparsely distributed buoy networks. This study systematically evaluates five advanced spatiotemporal imputation models, namely multivariate imputation by chained equations (MICE), bidirectional recurrent imputation for time series (BRITS), graph recurrent imputation network (GRIN), ImputeFormer, and recurrent generative adversarial imputation nets (rGAIN), for reconstructing missing significant wave height ( H s ) and mean wave period ( T ) observations from nine National Data Buoy Centre (NDBC) buoys in the Gulf of Mexico (GoM) during 2020–2024 with synthetic gaps of 1–3 months. In the buoy-only setting, weak inter-buoy correlations limit all the models’ performance: for H s , the coefficient of determination ( R 2 ) ranges from 0.62 to 0.69, with root-mean-square errors (RMSEs) of 0.37–0.41 m, while for T , R 2 ranges from 0.65 to 0.68, with RMSEs of 0.57–0.59 s. MICE remains competitive with more sophisticated models, indicating that complexity cannot compensate for limited spatial information. Incorporating the four nearest reanalysis nodes from three different reanalysis datasets substantially improves the performance, reducing H s RMSE to 0.17 to 0.24 m and T RMSE to 0.38 to 0.53 s. Sensitivity analysis further shows that increasing the number of reanalysis nodes does not produce monotonic improvement.