Three-stage ocean wave elements extraction using deep learning based on in-situ monocular videos from offshore infrastructure

Abstract

Ocean wave elements refer to statistical descriptions of waves in ocean engineering, serving to evaluate the fatigue and damage of offshore infrastructures and expand the operational time window for offshore construction activities. In-situ measurement of waves is an effective approach to acquire wave elements. However, traditional non-contact equipment is costly and unsuitable for continuous wave elements acquisition near offshore structures. Therefore, this study utilizes monocular vision to extract marine information, focusing on the accurate recognition of significant wave height and peak frequency near infrastructures. A three-stage wave elements extraction method is proposed: preprocessing with infrared enhancement and highlight removal, Convolutional Neural Network (CNN) feature extraction phase, and convolutional Long Short-Term Memory (convLSTM) regression network based on a self-attention memory mechanism. This approach accurately regresses significant wave heights and peak frequencies from in-situ monocular videos of the ocean, achieving mean absolute percentage errors (MAPE) of 3.01% for significant wave height and 4.07% for peak frequency. Additionally, ablation experiments are conducted at each stage to validate the effectiveness of the method.