| 基于广义回归神经网络的全球表层海水1°×1°二氧化碳分压数据推演 |
| |
| 引用本文： | 钟国荣,李学刚,曲宝晓,王彦俊,袁华茂,宋金明.基于广义回归神经网络的全球表层海水1°×1°二氧化碳分压数据推演[J].海洋学报,2020,42(10):70-79. |
| |
| 作者姓名： | 钟国荣 李学刚 曲宝晓 王彦俊 袁华茂 宋金明 |
| |
| 作者单位： | 1.中国科学院海洋研究所 海洋生态与环境科学重点实验室，山东 青岛 266071 |
| |
| 基金项目： | 国家重点研发计划（2017YFA0603204）；国家自然科学基金（91958103）；中国科学院战略性先导科技专项（XDA19060401）。 |
| |
| 摘 要： | 表层海水二氧化碳分压是评估海洋碳源汇强度的关键参数，但其实测数据较少、时空分布极不均匀，导致二氧化碳交换通量的估算有很大的不确定性，海洋源汇特征就不能确切获取。为了解决这个难题，在收集的表层大洋二氧化碳地图(Surface Ocean CO2 Atlas，SOCAT)实测数据集基础上，运用广义回归神经网络建立二氧化碳分压与经纬度、时间、温度、盐度和叶绿素浓度间的非线性关系，构建了1998?2018年间全球1°×1°经纬度的表层海水二氧化碳分压格点数据，其标准误差为16.93 μatm，平均相对误差为2.97%，优于现有研究中的前反馈神经网络、自组织映射神经网络和机器学习算法等方法。根据构建的数据所绘制的全球表层海水二氧化碳分压的分布与现有研究有较好的一致性。
|
| 关 键 词： | 广义回归神经网络 表层海水二氧化碳分压 全球大洋格点数据 |
| 收稿时间： | 2019/12/29 0:00:00 |
| 修稿时间： | 2020/3/23 0:00:00 |
A general regression neural network approach to reconstruct global 1°×1° resolution sea surface pCO2 |
| |
| Zhong Guorong,Li Xuegang,Qu Baoxiao,Wang Yanjun,Yuan Huamao,Song Jinming.A general regression neural network approach to reconstruct global 1°×1° resolution sea surface pCO2[J].Acta Oceanologica Sinica (in Chinese),2020,42(10):70-79. |
| |
| Authors: | Zhong Guorong Li Xuegang Qu Baoxiao Wang Yanjun Yuan Huamao Song Jinming |
| |
| Institution: | 1.Key Laboratory of Marine Ecology and Environmental Sciences, Institute of Oceanology, Chinese Academy of Sciences, Qingdao 266071, China2.University of Chinese Academy of Sciences, Beijing 100049, China3.Marine Ecology and Environmental Science Laboratory, Pilot National Laboratory for Marine Science and Technology (Qingdao), Qingdao 266237, China4.Center for Ocean Mega-Science, Chinese Academy of Sciences, Qingdao 266071, China |
| |
| Abstract: | Sea surface partial pressure of carbon dioxide (pCO2) is a crucial parameter for estimating ocean carbon source and sink term, but its sparse and uneven in situ measurements in space and time lead to large uncertainty in the estimate of sea-air CO2 flux and characteristics of ocean carbon source and sink. To eliminate this uncertainty, a general regression neural network approach using the Surface Ocean CO2 Atlas (SOCAT) dataset, based on the non-liner regression of pCO2 and longitude, latitude, time, temperature, salinity and concentration of chlorophyll, was successfully used in the reconstruction of global 1°×1° resolution monthly sea surface pCO2 from 1998 to 2018, with a root mean square error (RMSE) of 16.93 μatm and a mean relative error (MRE) of 2.97%, lower than existing feed-forward neural network (FFNN), self-organizing neural network (SOM) and machine learning approaches. The global distribution of pCO2 obtained by this approach agrees well with existing researches. |
| |
| Keywords: | general regression neural network sea surface pCO2 global ocean grid data |
| 本文献已被 万方数据 等数据库收录！ |
| 点击此处可从《海洋学报》浏览原始摘要信息 |
| 点击此处可从《海洋学报》下载免费的PDF全文 |
|
