Machine learning-based bias correction method for ocean buoy sensor observations
-
摘要: 海洋浮标观测是海洋研究数据的重要获取手段，但受传感器本身基线漂移、海洋生物附着和海水腐蚀等多种因素的影响，浮标观测的直接观测数据必须进行严格的偏差校正，以确保其数据的可靠性。当前针对物理海洋参数浮标数据的质控方案已有较多研究和报道，然而对于更加复杂多变的化学参数尚无完善可行的在浮标传感器端的质控方案。为此，本研究基于对实验室溶解氧、叶绿素、pH值和CO2分压参数为期90 d的传感器监测数据的变化分析，发现监测参数的漂移偏差与电导率、传感器读数电压等基础参数呈现较强的相关性，同时也不同程度地与生物因素相关。在此基础上，建立了基于机器学习拟合漂移偏差与传感器基础参数间非线性关系的漂移偏差校正方法，使浮标传感器化学参数监测数据有效化。应用该方法对不同参数的观测数据进行校正，可有效减小漂移数据与真实值间的偏差，为实现海洋化学参数浮标观测数据的长期、稳定、高质量获取提供了一种新的质控思路。Abstract: Ocean buoy observations serve as a vital means of acquiring data for marine research. However, direct measurements from buoys are subject to significant biases induced by factors such as sensor baseline drift, biofouling, and seawater corrosion, necessitating rigorous bias correction to ensure data reliability. While numerous quality control (QC) schemes for physical oceanographic parameters from buoy data have been extensively studied and reported, robust and practical sensor QC measures for more complex and variable chemical parameters remain lacking. To address this gap, this study analyzed 90-day laboratory monitoring data for parameters including dissolved oxygen, chlorophyll concentration, pH, and partial pressure of CO2 (pCO2). The analysis revealed that the drift bias in these monitored parameters exhibits a strong linear correlation with fundamental sensor parameters such as conductivity and sensor output voltage and with biological factors to varying degrees. Building upon these findings, we developed a drift bias correction method based on machine learning algorithm to fit the nonlinear relationships between drift bias and fundamental sensor parameters. This method effectively validates buoy sensor data for chemical parameters. Application of this method to observational data across different parameters significantly reduces the deviation between drifted data and true values. It thus provides a novel QC approach for achieving sustained, stable, and high-quality acquisition of marine chemical parameter data from buoy observations.
-
Key words:
- monitoring data /
- drift deviation /
- quality control /
- machine learning /
- buoy observation
-
图 1 浮标传感器观测数据漂移偏差测定实验
Fig. 1 Buoy sensor observed data drift bias determination experiment
图 2 前反馈神经网络（FFNN）漂移偏差校正模型与LSTM神经网络模型结构
Fig. 2 Structure of the FFNN drift bias correction model and LSTM model
图 3 原始数据与LSTM法剔除异常值后的标准差对比
Fig. 3 Comparison of standard deviation between raw data and LSTM-based outlier removal
图 4 各参数漂移偏差与基础参数相关性分析
Fig. 4 Correlation analysis between variable drift biases and basic parameters
表 1 漂移偏差校正基础参数选择
Tab. 1 Basic parameter selection for drift bias correction
偏差校正目标参数 校正使用的基础参数 DO漂移偏差 时间、温度、盐度、电导率、浊度、pH值、DO、DO电压 Chl漂移偏差 时间、温度、盐度、电导率、浊度、pH值、DO、Chl、Chl电压、pCO2 pH值漂移偏差 时间、温度、盐度、电导率、浊度、pH值、DO、Chl、DO电压、Chl电压 pCO2分压漂移偏差 时间、温度、盐度、电导率、浊度、pH值、DO、Chl、DO电压、Chl电压、pCO2 
下载: 导出CSV
表 2 不同判断阈值下LSTM法剔除异常值效果
Tab. 2 Outlier removal performance of LSTM at different decision thresholds
LSTM阈值
系数（θ）实验组DO 对照组DO 实验组Chl 对照组Chl 实验组pH值 对照组pH值 实验组pCO2 最大
STD剔除
数据量最大
STD剔除
数据量最大
STD剔除
数据量最大
STD剔除
数据量最大
STD剔除
数据量最大
STD剔除
数据量最大
STD剔除
数据量1 0.042 7588 0.108 5582 0.014 20352 0.079 5093 0.022 6462 0.178 6131 8.103 7160 2 0.072 1049 0.147 2231 0.028 2829 0.122 3389 0.028 892 0.204 2104 11.442 166 3 0.110 180 0.246 1524 0.053 1322 0.156 2776 0.028 180 0.362 784 15.405 40 4 0.176 80 0.246 1158 0.063 601 0.212 2586 0.059 85 0.362 387 17.545 20 5 0.176 59 0.418 959 0.074 265 0.313 2411 0.059 61 0.362 227 17.545 20 6 0.176 54 0.507 793 0.075 134 0.313 2354 0.059 58 0.362 187 23.323 18 7 0.176 41 0.507 659 0.075 87 0.313 2273 0.059 28 0.362 158 23.323 15 8 0.176 40 0.507 520 0.104 63 0.380 2192 0.059 24 0.362 151 23.323 14 9 0.176 34 0.507 436 0.104 56 0.380 2169 0.059 24 0.362 146 23.323 12 10 0.176 22 0.507 317 0.104 50 0.405 2106 0.059 23 0.362 145 23.323 9 原始数据 0.238 0.959 0.232 4.321 0.242 2.245 51.709 注：加粗内容为各组最终采用的LSTM阈值。
下载: 导出CSV
表 3 不同观测频次下的校正模型预测值与漂移偏差观测值间平均误差（MAE）对比
Tab. 3 Comparison of MAE between predicted and observed drift deviation with different training sample counts
偏差观测频次 DO MAE/(mg·L−1) Chl MAE/(μg·L−1) pH值 MAE 训练集 验证集 训练集 验证集 训练集 验证集 每4 h 0.260 0.346 0.609 1.288 0.091 0.138 每1 h 0.256 0.291 0.156 0.337 0.081 0.104 每30 min 0.119 0.158 0.172 0.281 0.077 0.098 每5 min 0.063 0.080 0.070 0.084 0.050 0.062
下载: 导出CSV
-
[1] 王军成. 国内外海洋资料浮标技术现状与发展[J]. 海洋技术, 1998, 17(1): 9−15.Wang Juncheng. On the current situation and trend about ocean data buoy at home and abroad[J]. Ocean Technology, 1998, 17(1): 9−15. [2] Roemmich D, Johnson G C, Riser S, et al. The Argo program: observing the global ocean with profiling floats[J]. Oceanography, 2009, 22(2): 34−43. doi: 10.5670/oceanog.2009.36 [3] 王军成, 厉运周. 我国海洋资料浮标技术的发展与应用[J]. 山东科学, 2019, 32(5): 1−20. doi: 10.3976/j.issn.1002-4026.2019.05.001Wang Juncheng, Li Yunzhou. Development and application of ocean data buoy technology in China[J]. Shandong Science, 2019, 32(5): 1−20. doi: 10.3976/j.issn.1002-4026.2019.05.001 [4] Oka E, Ando K. Stability of temperature and conductivity sensors of Argo profiling floats[J]. Journal of Oceanography, 2004, 60(2): 253−258. doi: 10.1023/B:JOCE.0000038331.10108.79 [5] Hall C, Jensen R E. USACE coastal and hydraulics laboratory quality controlled, consistent measurement archive[J]. Scientific Data, 2022, 9(1): 248. doi: 10.1038/s41597-022-01344-z [6] Stoer A C, Takeshita Y, Maurer T L, et al. A census of quality-controlled Biogeochemical-Argo float measurements[J]. Frontiers in Marine Science, 2023, 10: 1233289. doi: 10.3389/fmars.2023.1233289 [7] Gaillard F, Autret E, Thierry V, et al. Quality control of large Argo datasets[J]. Journal of Atmospheric and Oceanic Technology, 2009, 26(2): 337−351. doi: 10.1175/2008JTECHO552.1 [8] Bliss A C, Hutchings J K, Watkins D M. Sea ice drift tracks from autonomous buoys in the MOSAiC distributed network[J]. Scientific Data, 2023, 10(1): 403. doi: 10.1038/s41597-023-02311-y [9] Jiang Sha, Chen Yonghua, Liu Qingkui. Advancements in buoy wave data processing through the application of the Sage-Husa adaptive Kalman filtering algorithm[J]. Sensors, 2023, 23(16): 7298. doi: 10.3390/s23167298 [10] Wong A P S, Johnson G C, Owens W B. Delayed-mode calibration of autonomous CTD profiling float salinity data by θ-S climatology[J]. Journal of Atmospheric and Oceanic Technology, 2003, 20(2): 308−318. doi: 10.1175/1520-0426(2003)020<0308:DMCOAC>2.0.CO;2 [11] Owens W B, Wong A P S. An improved calibration method for the drift of the conductivity sensor on autonomous CTD profiling floats by θ-S climatology[J]. Deep Sea Research Part I: Oceanographic Research Papers, 2009, 56(3): 450−457. doi: 10.1016/j.dsr.2008.09.008 [12] 雷发美, 万艳, 商少平, 等. 海洋浮标表层环境要素质控流程和方法的研究[J]. 海洋技术学报, 2022, 41(4): 10−25.Lei Famei, Wan Yan, Shang Shaoping, et al. Research on quality control process and method for surface environmental elements of marine buoy[J]. Journal of Ocean Technology, 2022, 41(4): 10−25. [13] Hansen D V, Poulain P M. Quality control and interpolations of WOCE-TOGA drifter data[J]. Journal of Atmospheric and Oceanic Technology, 1996, 13(4): 900−909. doi: 10.1175/1520-0426(1996)013<0900:QCAIOW>2.0.CO;2 [14] Bushnell M. Quality assurance/quality control of real-time oceanographic data[C]//OCEANS 2015–MTS/IEEE Washington. Washington: IEEE, 2015: 1−4. [15] 张斌, 冯立强, 王彦俊, 等. 长江口06号大型综合观测浮标2014–2015年观测数据集[J]. 中国科学数据, 2017, 2(1): 95−102.Zhang Bin, Feng Liqiang, Wang Yanjun, et al. A dataset of No. 6 large-scale integrated observation buoy on the Yangtze estuary(2014–2015)[J]. China Scientific Data, 2017, 2(1): 95−102. [16] 卢勇夺, 王朝阳, 王豹, 等. 我国海洋锚系浮标数据异常值检测方法研究——以QF110和QF306为例[J]. 海洋预报, 2019, 36(6): 37−43.Lu Yongduo, Wang Zhaoyang, Wang Bao, et al. Research on outlier detection method for marine anchor buoys in China, using QF110 and QF306 as an example[J]. Marine Forecasts, 2019, 36(6): 37−43. [17] Maurer T L, Plant J N, Johnson K S. Delayed-mode quality control of oxygen, nitrate, and pH data on SOCCOM biogeochemical profiling floats[J]. Frontiers in Marine Science, 2021, 8: 683207. doi: 10.3389/fmars.2021.683207 [18] Li Shuo, Wang Bin, Deng Zeng’an, et al. Data quality control method of a new drifting observation technology named drifting air-sea interface buoy[J]. Journal of Ocean University of China, 2024, 23(1): 11−22. doi: 10.1007/s11802-024-5426-2 [19] Takeshita Y, Martz T R, Johnson K S, et al. A climatology-based quality control procedure for profiling float oxygen data[J]. Journal of Geophysical Research: Oceans, 2013, 118(10): 5640−5650. doi: 10.1002/jgrc.20399 [20] Udaya Bhaskar T V S, Venkat Shesu R, Boyer T P, et al. Quality control of oceanographic in situ data from Argo floats using climatological convex hulls[J]. MethodsX, 2017, 4: 469−479. doi: 10.1016/j.mex.2017.11.007 [21] 谭哲韬, 张斌, 吴晓芬, 等. 海洋观测数据质量控制技术研究现状及展望[J]. 中国科学: 地球科学, 2022, 52(3): 418−437.Tan Zhetao, Zhang Bin, Wu Xiaofen, et al. Research status and prospects of quality control technology for ocean observation data[J]. Scientia Sinica (Terrae), 2022, 52(3): 418−437. [22] Organelli E, Claustre H, Bricaud A, et al. A novel near-real-time quality-control procedure for radiometric profiles measured by bio-Argo floats: protocols and performances[J]. Journal of Atmospheric and Oceanic Technology, 2016, 33(5): 937−951. doi: 10.1175/JTECH-D-15-0193.1 [23] Petrenko B, Ignatov A, Pryamitsyn V, et al. Towards improved quality control of in situ sea surface temperatures from drifting and moored buoys in the NOAA iQuam system[J]. Applied Sciences, 2023, 13(18): 10205. doi: 10.3390/app131810205 [24] Wimart-Rousseau C, Steinhoff T, Klein B, et al. Technical note: Assessment of float pH data quality control methods—a case study in the subpolar northwest Atlantic Ocean[J]. Biogeosciences, 2024, 21(5): 1191−1211. [25] Song Miaomiao, Gao Saiyu, Liu Shixuan, et al. Intelligent quality control method for marine buoy data based on transformer encoder and BiLSTM[J]. Frontiers in Marine Science, 2025, 12: 1528587. doi: 10.3389/fmars.2025.1528587 -
图(6) / 表(3)
计量
- 文章访问数: 71
- HTML全文浏览量: 47
- PDF下载量: 7
- 被引次数: 0
