Direct-transfer evaluation of a transformer-based framework for cross-vessel sea-state estimation through wave buoy analogy

Accurate wave parameter identification is essential for marine vehicle operability and safety. The Wave Buoy Analogy (WBA) establishes an inverse mapping between wave-induced ship responses and the surrounding wave field using onboard sensor data, enhancing navigation while complementing satellite and buoy-based observations to improve wave data resolution across oceanic regions. Despite recent advances in deep learning for WBA, existing models suffer from ship-specific dependency, limiting generalization to unseen vessels. This study presents an exploratory investigation of Transformer-based transfer learning for sea state estimation (SSE) via WBA. A dual-line architecture is proposed that combines cross-attention between global motion representations and direction-sensitive roll features with speed-based feature scaling, supported by a vectorization methodology and feature transformation techniques to improve class separability. Using six ships for training and two unseen ships for evaluation, we evaluated in-distribution validation alongside transfer behaviour on mid-range and out-of-distribution vessels. Results demonstrate strong validation accuracy, promising direct-transfer performance for a mid-range unseen vessel, and failure on an out-of-envelope case, thereby delineating the potential and current limits of the approach. In the single-run validation, the Bidirectional model achieved the highest wave-height accuracy of 98.72%, while the L2→L1 configuration achieved the highest wave-direction accuracy of 91.67%; however, repeated-seed analysis identified Bidirectional as the most robust model, with macro-F1 scores of 94.44 ± 4.68% and 84.01 ± 3.88% for wave height and direction, respectively. In direct-transfer evaluation, the model achieved 76.2% wave-height and 71.4% wave-direction accuracy on the mid-range unseen vessel, but only 8% and 40%, respectively, on the out-of-envelope VLCC. Limitations and directions for future work are discussed.