刘玉龙,王国松,侯敏,徐珊珊,苗庆生.基于深度学习的海温观测数据质量控制应用研究[J].海洋通报,2021,(3):
基于深度学习的海温观测数据质量控制应用研究
Quality control of sea temperature observation data using deep learning neural networks
投稿时间:2021-01-08  修订日期:2021-03-09
DOI:
中文关键词:  海温  质量控制  深度神经网络  多层感知器
英文关键词:sea temperature  quality control  deep neural network  multi-layer perceptron
基金项目:国家重点研发计划 (2016YFC1401905;2016YFC1402605;2017YFC1405302) ; 全球变化与海气相互作用 (GASI-IPOVAI-04);国家自然科学基金 (41606039;41776004)
作者单位E-mail
刘玉龙 国家海洋信息中心天津 300171 yulong0631@163.com 
王国松 国家海洋信息中心天津 300171 xifengbishu2110@163.com 
侯敏 天津市滨海新区气象局天津 300457  
徐珊珊 国家海洋信息中心天津 300171  
苗庆生 国家海洋信息中心天津 300171  
摘要点击次数: 225
全文下载次数: 105
中文摘要:
      基于国家海洋信息中心质量控制后的西太平洋 10 度方区约 100 万站次温盐实测历史调查资料,对经过 26 种严格质量控制方法的综合海温质量符进行分类分析,首次将深度学习技术应用于海洋数据质量控制多分类 (multiclass classification)算法与应用研究。通过人工合成少数类样本和加权损失函数方法减少多数类的频率来降低数据的不平衡,并构建了多层感知器 (Multi-Layer Perceptron, MLP) 和深度神经网络 (Deep Neural Network, DNN) 两个海温资料质量符分类深度学习模型。分类结果表明本文构建的两个深度分类模型能够较准确快速地识别该海域海温数据质量,在 20 975 条温盐剖面资料测试集中分类准确率分别达到 99.63%和 99.69%。海温资料的分类精度评分有着较好的表现,其中正确数据 (QC1) 和数据缺失(QC9) 的正确识别率均达 100%。MLP 和 DNN 多分类质量控制模型可大幅降低传统质量控制方法的工作量,提升海量数据处理速度和分析能力,为海温观测资料在海洋研究与工程中应用提供参考。
英文摘要:
      Based on one million records of thermohaline measured data in the western Pacific, the deep learning neural networks is applied to the multiclass classification algorithm of sea temperature observation data. The quality flags of sea temperature, which is from the National Marine Data and Information Service (NMDIS), was classification and analysis after 26 kinds of strict sea temperature quality control methods. In order to reduce the data imbalance, the frequency of most classes is reduced by artificial synthesis of minority samples, down-sampling and cluster sampling, and the deep learning neural networks of quality flags classification of sea temperature observation data are constructed respectively by Multi-Layer Perceptron (MLP) and deep neural network (DNN). The results show that, the two deep classification models constructed in this paper can accurately and quickly identify the quality of sea temperature data. The classification accuracy of the 20 975 temperature and salt profile data on test sets is 99.63% and 99.69%, respectively. The classification accuracy rating has a better performance, and the classification accuracy rating of correct data (QC1) and data missing (QC9) is 100%. The MLP and DNN multiclass classification quality control model can greatly reduce the workload of traditional quality control methods, improve the processing speed and analysis capability of massive data, and provide availability reference for ocean temperature data in marine data research and engineering applications
查看全文    下载PDF阅读器
关闭