基于神经网络的全球三维温盐场重构技术研究  

Research and application of global three-dimensional thermohaline reconstruction technology based on neural network

作　　者：聂旺琛 王喜冬 陈志强 何子康 范开桂 NIE Wangchen;WANG Xidong;CHEN Zhiqiang;HE Zikang;FAN Kaigui(Key Laboratory of Research on Marine Hazards Forecasting,Ministry of Natural Resources,College of Oceanography,Hohai University,Nanjing 210098,China;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519000,China)

机构地区：[1]河海大学,自然资源部海洋灾害预报技术重点实验室,江苏南京2010098 [2]南方海洋科学与工程广东省实验室(珠海),广东珠海519000

出　　处：《热带海洋学报》2022年第2期1-15,共15页Journal of Tropical Oceanography

基　　金：国家自然科学基金(41776004)。

摘　　要：文章利用果蝇优化广义回归神经网络算法FOAGRNN (fruit fly optimization algorithm, FOA;generalized regression neural network, GRNN)对SODA (simple ocean data assimilation)再分析数据进行训练,构建海表温度、盐度、海面高度与次表层温盐场之间的投影关系模型,并在全球范围使用SODA和卫星遥感数据评估了模型的应用性能。首先,利用独立的2016年SODA海表数据作为模型输入进行理想重构试验,结果显示全球重构温、盐平均均方根误差(MRMSE)分别为0.36℃和0.08‰,与世界海洋图集WOA13资料相比减小约50%和60%。然后,利用卫星观测的海表信息作为模型输入进行实际应用试验,并与Argo观测剖面进行比较评估。试验结果表明,重构模型能有效表征海水温、盐特征,其中重构温、盐MRMSE分别为0.79℃和0.16‰,相比WOA气候态减小27%和11%。误差的垂向分布显示,重构温度RMSE从海表向下迅速增大,至100m达到峰值1.35℃,而后又迅速回落,至250m处为0.81℃,跃层往下不断减小;重构盐度RMSE基本随深度增大而减小,误差峰值位于25m附近,约为0.25‰。此外, Argo浮标跟踪分析和区域水团统计结果也表明模型能够较好地刻画海洋三维温盐场的内部结构特征。We apply the FOAGRNN(fruit fly optimization algorithm, FOA;generalized regression neural network, GRNN)method and SODA(simple ocean data assimilation) reanalysis data to construct a global ocean projection relationship model between sea-surface variables(sea surface height, SSH;sea surface temperature, SST;sea surface salinity, SSS) and subsurface thermohaline field. The remote sensing observations are utilized to evaluate the applicability of this global surface-subsurface reconstruction model. First, an ideal reconstruction test is executed using the independent SODA data in 2016. The idealized reconstruction results show that the global mean root mean square error(MRMSE) values of the reconstructed temperature and salinity are 0.36 ℃ and 0.08‰, which are reduced by about 50% and 60% compared to those of the WOA13(World Ocean Atlas), respectively. Then, the satellite observations(Input field) and Argo profiles(verification field) are inputted to evaluate the practical application performance of the model. The results again indicate that our reconstruction model can reasonably reconstruct the thermohaline structures, and the MRMSE values of the reconstructed temperature and salinity are 0.79 ℃ and 0.16‰, which are 27% and 11% lower than those in the WOA13, respectively.Specifically, the RMSE of temperature is small at the sea surface and in the deep ocean, and the largest value exists in the thermocline layer with a maximum value of 1.35 ℃ at 100 m, and then quickly decreases to 0.81 ℃ at 250 m. The RMSE of salinity mostly decreases as depth increases, and has the largest peak of about 0.25‰ around 25 m. Finally, the analysis of Argo floats’ tracks and the statistics of regional water mass confirm that the reconstructed model can better describe the interior characteristics of the three-dimensional thermohaline field.

关 键 词：果蝇优化广义回归神经网络算法 三维温盐场 重构 卫星观测数据 SODA再分析数据 

分 类 号：P731.1[天文地球—海洋科学]