基于深度学习方法使用海表面温度构建工业革命以来的印尼贯穿流流量序列  

USING A DEEP LEARNING METHOD AND SEA SURFACE TEMPERATURES TO CONSTRUCT A TIME SERIES OF INDONESIAN THROUGHFLOW TRANSPORT SINCE THE INDUSTRIAL REVOLUTION

作　　者：辛林超 胡石建 XIN Lin-Chao;HU Shi-Jian(Key Laboratory of Ocean Observation and Forecasting,Qingdao 266071,China;Key Laboratory of Ocean Circulation and Waves,Institute of Oceanology,Chinese Academy of Sciences,Qingdao 266071,China;Institute of Oceanology,Chinese Academy of Sciences,Qingdao 266071,China;College of Marine Sciences,University of Chinese Academy of Sciences,Qingdao 266071,China)

机构地区：[1]海洋动力环境观测与预报重点实验室,山东青岛266071 [2]中国科学院海洋环流与波动重点实验室,山东青岛266071 [3]中国科学院海洋研究所,山东青岛266071 [4]中国科学院大学海洋学院,山东青岛266071

出　　处：《海洋与湖沼》2025年第1期42-53,共12页Oceanologia Et Limnologia Sinica

基　　金：崂山国家实验室项目,LSKJ202202702号;中国科学院战略性先导科技专项项目,XDB42010403号。

摘　　要：印尼贯穿流连接热带太平洋和印度洋,对区域和全球气候系统至关重要,但在印尼贯穿流的长期变化研究中非常缺乏长时间序列。利用第六次国际耦合模式比较计划历史模拟数据对卷积神经网络深度学习模型进行训练和测试,基于海表面温度反演了印尼贯穿流的体积输运。结果表明,该模型反演的印尼贯穿流可再现其总方差的80%左右。进而结合数据,首次构建了1870~2023年长达154 a的印尼贯穿流流量时间序列。通过与潜标观测数据对比发现,反演结果与国际努沙登拉层化和输运计划(the international Nusantara stratification and transport,INSTANT)、印尼贯穿流监测项目(monitoring Indonesian throughflow,MITF)观测数据的相关系数分别为0.56与0.69,验证了该时间序列的可靠性。使用梯度加权类激活映射(gradient-weighted class activation mapping,Grad-CAM)与注意力机制研究了该模型中影响印尼贯穿流反演技巧的敏感性区域。The Indonesian throughflow(ITF)connects the tropical Pacific and Indian Oceans and is critical to the regional and global climate systems.However,what is the most lacking in the study of long-term changes in the ITF is the lengthy time series.A deep learning model of convolutional neural networks(CNN)was trained and tested with the model outputs from the Coupled Model Intercomparison Project Phase 6(CMIP6),and the volume transport of ITF is inferred using this model with sea surface temperature(SST).Results show that the CNN model was able to reproduce about 80%of the total variance of ITF transport.A 154-year time series of ITF transport from 1870 to 2023 was constructed for the first time using the HadISST(Hadley Centre Global Sea Ice and Sea Surface Temperature)dataset.The time series was validated by comparing historical observations from subsurface moorings,from which we found that the correlation coefficients between the inversion results and INSTANT and MITF were 0.56 and 0.69,respectively.Sensitivity regions in referring the ITF in the model were investigated with the Grad-CAM(Gradient-weighted Class Activation Mapping)and attention mechanisms.

关 键 词：印尼贯穿流 海洋环流 海表面温度 深度学习 卷积神经网络 

分 类 号：P731[天文地球—海洋科学]