Quantitative video-based wave parameter estimation using a 3D-CNN and two-stage transfer learning framework

Abstract

Accurate and efficient wave monitoring is vital for the safety and long-term operation of large-scale marine infrastructures, yet conventional approaches such as Doppler profilers, buoys, and fixed stations are costly, difficult to maintain, and limited in coverage. Recent advances in computer vision have enabled non-contact, video-based methods for measuring. However, most existing studies rely on single-frame analysis or require large volumes of labeled field data, which limits their robustness and applicability. To overcome these challenges, this study develops a deep learning framework for the quantitative identification of wave parameters. The framework employs a 3D Convolutional Neural Network (3D-CNN) to extract spatio-temporal features from monocular video sequences, combined with a two-stage transfer learning strategy that leverages simulated wave data for pre-training and a small set of in-situ measurements for fine-tuning. Additionally, tailored input feature design and denoising based on Proper Orthogonal Decomposition (POD) are incorporated to enhance performance under complex visual conditions. Extensive experiments demonstrate that the proposed approach achieves high accuracy in predicting significant wave height and peak period, while substantially reducing reliance on costly field measurements. The results highlight the potential of the framework as a cost-effective, robust, and scalable solution for non-contact wave monitoring in marine engineering applications.