Deep Learning for X Band Radar Significant Wave Height Estimation

Abstract

Significant wave height (𝐻𝑠) is one of the sea-state parameters on which offshore workability decisions depend. Personnel transfers are commonly restricted around 𝐻𝑠 =1.5 m, motion-compensated lifting around 2.5 m, and many operations cease above 3 m. Near these limits, even a small error in the estimated sea state can change the operational decision. For 𝐻𝑠 estimates to be useful in this setting, they therefore need to be accurate to roughly 0.25 m RMSE across the workability range where such decisions are still being made.

Several methods are already used to estimate or describe offshore wave conditions, but each has shortcomings when considered against the needs of real-time vessel operations. Numerical wave models provide useful regional context, although their resolution is too coarse to capture the local conditions
around a single vessel. Satellite altimetry can support large-scale sea-state observation, but revisit times are too long for workability decisions made from hour to hour. Buoys provide direct measurements, but only at fixed locations, while onboard physics-based radar methods depend on processing assumptions
that may break down in the same conditions where reliable estimates are most needed. None of these sources fully provides a local, real-time, and sufficiently accurate estimate at vessel scale.

This thesis examines whether deep learning can narrow that gap by estimating 𝐻𝑠 directly from operational vessel data. The dataset combines X-band radar imagery, six-degree-of-freedom vessel motion measurements, and reference 𝐻𝑠 values from open-source wave buoys and ERA5 reanalysis, drawn from three operational vessels over several years. A Vision Transformer backbone is applied to preprocessed radar images, with optional vessel-motion fusion. In the sequence variant, the model is given a short series of consecutive radar images rather than a single image, allowing it to use temporal information in the sea surface pattern. Eight model variants are trained across a four-axis ablation covering preprocessing route, backbone initialisation, motion inclusion, and single-image versus sequence-based input.

On the development vessel, the best model comfortably reaches an RMSE lower than the set target under normal wind across the operationally relevant 𝐻𝑠 ≤3 m range, with little systematic bias. The strongest within-vessel configuration is the radar-only sequence model. Adding vessel motion does not consistently improve the estimate and, in the tested configuration, tends to introduce a positive offset at low sea states. For use on the same vessel, the radar-only sequence variant is therefore the preferred model.

Cross-vessel transfer remains the main unresolved part of the problem. Blindly applying successful models from one vessel to another does not meet the operational target, although the reasons are partly identifiable. In one direction, performance is mainly limited by the available sea-state coverage, while in the other it is dominated by a vessel-specific calibration offset. Low wind is the most consistent within-vessel failure mode, which is consistent with the weaker radar wave signature expected under reduced Bragg scattering. Overall, the thesis shows that deep learning can provide vessel-scale 𝐻𝑠 estimates from X-band radar imagery with accuracy well below the defined operational target, while also showing that reliable use across vessels requires either vessel-specific adaptation or broader multi-vessel training data.