Sea state prediction based on machine learning using images

Abstract

Safe navigation of ships has always been emphasized along with human history, and for this purpose, rigorous measuring equipment and advanced navigation technologies are developing continuously. However, accidents of ships continue to occur, and among them, accidents due to the judgement from inexperienced ship crews occur meaningfully. In addition, in the case of larger ships, the use of precision measuring instruments and the use of professional personnel to make judgements on sea conditions as accurately as possible are attempted to sail the ship safely.

In this study, a system for predicting the ocean environment using the rapidly developing machine learning method was proposed. In judging sea states, it is the captain’s main duty to determine the size and seriousness of waves, and based on this long experience, the winds and waves are classified as the Beaufort wind scale. The purpose of this study is to choose and apply a suitable machine learning method that can predict sea states with sufficient big data, and to evaluate its effectiveness.

Initially, the applicability of machine learning was reviewed and the characteristics of the learning model were identified through numerical graphical wave fields. Attempts were performed to identify the wave height of a certain point for long-crested and short-crested waves based on convolutional neural network. Various simulations with image pre-processing techniques, neural network changes, hyper-parameter tunings were conducted, while it was difficult to improve the prediction accuracy. Though this, we tried to examine the applicability of machine learning in identifying the sea state rather than the strict measurement.

Learning data was collected form the actual site over a certain period of time. The images were obtained using a camera in the southwestern area of Korea, and the sea states were obtained from the public data and a seabed installed wave height meter. In single snapshot-based learning, simple convolutional neural network-based learning showed limitations. Therefore, the combined model with convolutional neural network and long short-term memory was applied. In this case, when an appropriate data augmentation technique was utilized together, it was confirmed that the short video-clip based prediction can be applied to the real ocean environment. The applicability of this network was confirmed on images applied by artificially simulating the motion of a vessel, and to improve this, correction and re-prediction of the angle of view through image processing techniques were attempted.

Lastly, limitations of the present technique are arranged and possible solutions are introduced. In addition, the initial application to real problems was performed through prediction of images acquired from a ship in operation.