一种海洋混合层深度的智能识别方法研究
张康(1991—), 男, 安徽省亳州市人, 硕士研究生, 从事物理海洋学研究。E-mail: zhangkang16@mails.ucas.ac.cn |
Copy editor: 殷波
收稿日期: 2018-12-14
要求修回日期: 2019-04-10
网络出版日期: 2019-10-09
基金资助
国家自然科学基金项目(91752108)
国家自然科学基金项目(41476167)
国家自然科学基金项目(41706029)
国家自然科学基金项目(41606010)
广东省自然科学基金(2016A030311042)
广东省自然科学基金(2016A030310114)
广州市科技计划重点项目(201804020056)
中科院战略性先导专项资助项目(XDA11030302)
中国科学院南海生态环境工程创新研究院课题(ISEE2018PY05)
版权
Using artificial intelligence for identifying the depth of upper-ocean mixed layer
Copy editor: YIN Bo
Received date: 2018-12-14
Request revised date: 2019-04-10
Online published: 2019-10-09
Supported by
National Natural Science Foundation of China(91752108)
National Natural Science Foundation of China(41476167)
National Natural Science Foundation of China(41706029)
National Natural Science Foundation of China(41606010)
Natural Science Foundation of Guangdong Province(2016A030311042)
Natural Science Foundation of Guangdong Province(2016A030310114)
Guangzhou Science and Technology Program Key Project(201804020056)
Strategic Priority Research Program of Chinese Academy of Sciences(XDA11030302)
Institution of South China Sea Ecology and Environmental Engineering, Chinese Academy of Sciences(ISEE2018PY05)
Copyright
文章提出了一种识别混合层深度的人工智能方法。该方法在温度(密度)与压强(或深度)间建立线性模型, 并且将其系数和方差做成一组表征廓线特征的统计量。初始时为模型设定一个主观的先验分布, 在一个自海表向下移动的窗口内通过贝叶斯链式法则和最小描述长度原理学习新数据, 得到系数均值的最大后验概率估计。用F-检验识别系数发生突变的位置, 以此确定混合层的存在性及其深度。通过2017年2月太平洋海域的地转海洋学实时观测阵(Array for Real-time Geostrophic Oceanography, ARGO)数据进行测试, 并且以质量因子(Quality Index, QI)值作为判断识别混合层深度结果准确性的依据, 发现该方法相比于梯度法、阈值法、混合法、相对变化法、最大角度法和最优线性插值法在识别结果上具备更大的QI值。表明该方法能够准确识别混合层深度。
张康 , 郭双喜 , 黄鹏起 , 屈玲 , 鲁远征 , 岑显荣 , 于璐莎 , 周伟东 , 周生启 . 一种海洋混合层深度的智能识别方法研究[J]. 热带海洋学报, 2019 , 38(5) : 32 -41 . DOI: 10.11978/2018137
An artificial intelligence (AI) method for identifying upper-ocean mixed layer depth (MLD) is proposed. In this method, a linear model, whose coefficient and variance are made into a set of statistics to characterize the profile, is established between temperature (density) and pressure (or depth). A subjective priori distribution is set in an initial window. The maximum posterior probability estimate of the mean coefficient value is obtained when the window is moved down from the sea surface, by learning the new data through the Bayesian chain rule and the minimum description length principle. The existence and depth of the mixed layer are determined when the jump of the coefficient is found by using the F-distribution. Using the Argo buoy data measured in the Pacific Ocean in February 2017, and taking the value of the quality index (QI) to estimate the accuracy of the MLD results, we find that this AI method is superior to the gradient method, the threshold method, the Hybrid method, and the relative-variant method.
图1 数据学习过程示意图图中粉色方框表示窗口 Fig. 1 An example of the learning process. The blue line is the temperature profile. The magenta box represents the current window. The green, gray, and red dotted lines represent the fitting curves in the priori, the updating, and the posterior process, respectively |
图5 突变检验示意图a中实线为温度-深度廓线, 三角标注为突变点(混合层下边界); b中实线表示各点发生突变的概率, 虚线表示突变发生的临界概率, 突变发生概率首次超过临界值的点为突变点 Fig. 5 An example for checking the jump. The solid line in (a) denotes the temperature-depth profile, and the triangle is the depth of the jump (the bottom of the mixed layer). The solid line in (b) indicates the probability of mutation at each point, and the dotted line represents the critical probability of the jump occurring. The jump will occur when the probability exceeds the critical value for the first time |
图7 站位23 (28°57'43"N, 166°38'35"E)的温度(a)和密度(b)均匀混合层廓线和对应混合层深度图中三角形表示混合层下边界位置 Fig. 7 Temperature (density) uniform mixed layer and recognition results. The solid line represents the temperature profile (a) and potential density profile (b). The triangle is the lower boundary of the mixed layer |
图8 站位177 (57°22'55"S, 150°8'7"E)的温度(a)和密度(b)存在弱层结的混合层廓线和对应混合层深度图中三角形表示混合层下边界位置 Fig. 8 The temperature (density) profile with gradual pycnocline and recognition results. The solid line represents the temperature profile (a) and potential density profile (b). The triangle represents the lower boundary of the mixed layer |
表1 各方法从温度廓线廓线计算的QI值的结果Tab. 1 QI values calculated from the temperature profile by each method |
方法 | QImean | QI0.25 | QI0.5 | QI0.75 | QIstd |
---|---|---|---|---|---|
阈值法 | 0.795 | 0.708 | 0.899 | 0.962 | 0.239 |
梯度法 | 0.683 | 0.488 | 0.800 | 0.948 | 0.316 |
混合法 | 0.713 | 0.549 | 0.865 | 0.963 | 0.329 |
相对变化法 | 0.900 | 0.870 | 0.945 | 0.982 | 0.129 |
最大角度法 | 0.710 | 0.556 | 0.820 | 0.956 | 0.307 |
最优线性拟合 | 0.782 | 0.725 | 0.927 | 0.985 | 0.314 |
AI法 | 0.936 | 0.922 | 0.969 | 0.990 | 0.104 |
注: QImean为QI值的均值; QI0.25为QI值的1/4分位数; QI0.5为QI值的中值; QI0.75为QI值的3/4分位数; QIstd为QI值的标准差 |
表2 各方法从位势密度廓线中计算的QI值的结果Tab. 2 QI values calculated from the potential density profile by each method |
方法 | QImean | QI0.25 | QI0.5 | QI0.75 | QIstd |
---|---|---|---|---|---|
阈值法 | 0.783 | 0.672 | 0.915 | 0.974 | 0.267 |
梯度法 | 0.727 | 0.596 | 0.808 | 0.926 | 0.251 |
混合法 | 0.734 | 0.594 | 0.872 | 0.962 | 0.306 |
相对变化法 | 0.893 | 0.862 | 0.945 | 0.983 | 0.144 |
最大角度法 | 0.740 | 0.621 | 0.850 | 0.965 | 0.299 |
最优线性拟合 | 0.797 | 0.752 | 0.930 | 0.985 | 0.297 |
AI法 | 0.921 | 0.902 | 0.959 | 0.988 | 0.117 |
注: QImean为QI值的均值; QI0.25为QI值的1/4分位数; QI0.5为QI值的中值; QI0.75为QI值的3/4分位数; QIstd为QI值的标准差 |
[1] |
范聪慧 , 2007. 多因素对海洋上混合层深度影响的数值模拟[D]. 北京: 中国科学院研究生院( 海洋研究所):1-65.
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
[37] |
|
[38] |
|
/
〈 |
|
〉 |