Selective Applicability of Deep Learning Models for Imputation of Missing Wave Observation Data: Effects of Variable Physical Characteristics and Missing-Data Patterns

The optimal artificial intelligence model for imputing missing wave observation data may vary according to the physical characteristics of the target variable and the missingness pattern. This study compared a temporal model, SAITS, with spatiotemporal models, GRIN, SPIN, and ImputeFormer, using 30-min interval wave observations from five stations along the southeastern coast of Korea. All models were evaluated under identical data, data-splitting, sequence, and masking conditions. Performance differences and statistical ties between model pairs were assessed using window-level paired bootstrap 95% confidence intervals. In an evaluation in which the final time step at the target station was artificially masked, SAITS yielded the lowest errors for significant wave height and significant wave period, whereas SPIN achieved the lowest angular error for wave direction, a circular variable. In the missingness-pattern sensitivity experiments, performance differences among models were generally small across random missingness rates. In contrast, under contiguous missingness within a window, GRIN achieved the best performance for both significant wave height and significant wave period when the missingness rate was at least 20%, exhibiting a pattern distinct from that observed under random missingness. Notably, under 50% contiguous missingness, GRIN was the only deep learning model that produced lower errors than linear interpolation for both significant wave height and significant wave period. These findings indicate that imputation performance may be more sensitive to the temporal continuity of missing values than to the overall missingness rate. They further suggest that exploiting spatiotemporal dependencies can be effective for reconstructing consecutive missing segments, particularly when the contiguous missingness rate exceeds 20%.