X-밴드 레이더 이미지를 활용한 3D 합성곱 신경망 기반의 유의파고 추정

Abstract

This paper presents a method for significant wave heights estimation based on 3D convolutional neural network model using X-band radar image sequences. To train the proposed significant wave height estimation model, we utilized image sequences acquired by X-band radar and actual wave values measured by buoys. The dataset was acquired from an X-band radar set installed in Sokcho, South Korea for 74 days from June 1, 2021 to August 13, 2021. It was collected at approximately 1.43 second intervals and contains 72,180 three-dimensional image sequences. The dataset suffers from data imbalance, so we propose a method to address it. We assigned levels based on wave height range and included a certain number (N) of randomly sampled data from each level in the composition of data patches for training and testing. This allows us to balance each data patch to ensure the diversity of data utilized for training and reduce the risk of overfitting. Data from groups with fewer instances can also be used equally for training, ensuring representativeness across the entire range of data. The deep learning network proposed in this study is based on 3D convolutions, which simultaneously learns temporal and spatial information of radar image sequences and extracts respective features from them. With the proposed model, a correlation coefficient of up to 0.97 and an average of 0.96 is obtained between the actual significant wave height measurements and the estimates. This study confirms that it is feasible to estimate wave height from radar image sequences. In addition to significant waves, it seems that it is possible to obtain various oceanographic information from radar images. The extracted marine information can be applied to route improvement and autonomous navigation of ships, contributing to the construction of an efficient navigation system.