Polarization-aware multi-scale vision transformer for significant wave height estimation

Document Type

Article

Publication Date

10-30-2026

Abstract

Accurate estimation of significant wave height (Hs[jls-end-space/]) from compact visual sensors becomes increasingly important as marine autonomy develops. While conventional intensity images provide only limited information about wave surface geometry, polarization imaging offers richer cues related to surface slope and orientation. However, existing studies have not fully explored how interactions among polarization channels contribute to Hs estimation. This study proposes polarization-aware multi-scale vision transformer (PAMS-ViT), an end-to-end framework for Hs estimation directly from raw polarimetric images. The architecture centers on a Learnable Stokes Fusion (LSF) module that replaces fixed analytic transforms with jointly optimized polarimetric channel fusion and a Multi-Scale Polarimetric Tokenization (MSPT) module that captures wave-surface structures across fine, intermediate, and coarse spatial scales. Under a leakage-free 12-fold leave-one-wave-train-out (LOWTO) cross-validation protocol, experimental results on an open dataset, with nominal Hs in the centimeter range, demonstrate that PAMS-ViT consistently outperforms the baseline models, achieving the lowest pooled RMSE of 0.134 cm and (Formula presented). Ablation studies further show that learning polarization-channel interactions directly from data leads to more accurate Hs estimation than conventional fixed analytic transforms. This finding suggests a promising alternative to handcrafted polarimetric representations for wave-state estimation.

Publication Title

Ocean Engineering

Link to Full Text

Share

COinS