| 摘要: |
| 近岸潮位观测是海洋工程应用、海岸防灾减灾、海岸带管理以及海洋有关科研工作中最基础的工作之一。文章基于视频图像深度学习的方法，使用YOLOv5目标检测算法从安装在近岸的固定摄像机拍摄的视频帧中提取潮汐水位特征进行潮位分析。研究采用厦门高崎码头的分辨率为1920×1080的高清摄像头2023年2月的影像数据作为训练集和验证集，2023年3月的影像数据作为测试集，利用岸边验潮井逐时潮位数据进行标注，采用YOLOv5目标检测算法来训练。计算结果显示，通过视频观测潮位在训练集和测试集上的误差分别为3.9 cm和5.3 cm。视频中1个像素点代表3.8 cm，因此潮位观测的平均误差为像素级。研究表明在近岸通过高清摄像头基于图像深度学习进行潮位观测的方法是可行的，观测精度取决于图像目标物的分辨率。 |
| 关键词: 深度学习 潮位观测 厦门高崎码头 高清视频 |
| DOI：10.20016/j.cnki.hykfygl.20240302.002 |
| 投稿时间：2023-05-09修订日期：2023-11-28 |
| 基金项目:自然资源部第三海洋研究所基本科研业务费专项资金资助项目“不同物源砾石海滩沉积与动力地貌差异性研究”（No.2019029）；厦门市海洋与渔业发展专项资金青年科技创新项目资助“基于高清视频识别技术的近岸潮汐、海流、波浪观测技术研究”（23ZHZB050QCB40）. |
|
| Research on Tidal Level Technology Based on Deep Learning Video Observation： A case study of Xiamen Gaoqi Wharf |
| ZENG Zhi,CHEN Zhijie,SUN Quan |
| Third Institute of Oceanography, MNR;Fujian Provincial Key Laboratory of Marine Physical and Geological Processed |
| Abstract: |
| Nearshore tidal observation is one of the most fundamental tasks in ocean engineering applications, coastal disaster mitigation, coastal zone management, and ocean-related scientific research. In this paper, a method based on video image deep learning is proposed to extract tidal level features from the video frames captured by a fixed camera installed near the shore, using YOLOv5 visual AI model for tidal analysis. The study used the high-definition camera of Xiamen Gaoqi Wharf with a resolution of 1920×1080 as the training and validation dataset for the February 2023, and the test dataset for March 2023. The hourly tide data of the coastal tide verification well is used for annotation, and the YOLOv5 object detection model is used for training. The calculation results show that the errors of tidal observation through video on the training set and the test set are 3.9cm and 5.3cm, respectively. One pixel in the video represents 3.8cm, so the average error of the tidal observation is at the pixel level. The study shows that the method of using highdefinition cameras based on image deep learning for tidal observation near the shore is feasible, and the observation accuracy depends on the resolution of the target object in the image. |
| Key words: Deep learning,Tidal level observation,Xiamen Gaoqi Wharf, High definition video |
