Towards Development of Visual-Range Sea State Image Dataset for Deep Learning Models

Abstract

Wind waves are generated by winds blowing over long stretches of the sea surface. They are considered as one of the important elements of marine weather. A sea state describes prevailing wind wave conditions. Due to its constant presence, it is important to classify the sea state for safety and optimal operations of coastal and offshore structures, maritime traffic, and recreational activities etc. The Beaufort wind force scale provides an empirical solution for sea state classification. Additionally, wave parameters acquired from sea buoys can be used to identify the sea state. However, the deployment and maintenance costs of buoys are high. Recent advancements in deep learning-based image classification can lead toward the development of low-cost sea state classification solutions. However, to train and test such models, required visual-range sea state image dataset is not yet publicly available. Hence, the authors have proposed the development of said dataset, which is currently in its later stages of construction. In this paper, we present general observations, design considerations, and guidelines formulated during the development of the visual-range sea state image dataset. The paper discusses the important factors related to sensor and field observation site selection, data acquisition considerations, data processing, and the manual sea state identification mechanism. The paper also provides guidelines for application specific augmentation policy development and recommends a baseline number of representative instances per class for the dataset. The research community can refer to the presented work for further research in the development of sea state image datasets.