| 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,但研究印尼贯穿流的长期变化最缺乏的是长时间序列。基于卷积神经网络深度学习方法,利用第六次国际耦合模式比较计划历史模拟数据进行模型训练和测试,反演了印尼贯穿流的体积输运。结果表明,使用海表面温度反演的印尼贯穿流可再现其总方差的80%左右。进而使用该模式结合HadlSST数据,首次构建了1870年至2023年长达154年的印尼贯穿流时间序列。通过与历史潜标观测数据对比,验证了该时间序列的可靠性。使用Grad-CAM与注意力机制研究了该模型中影响印尼贯穿流反演的敏感性区域。 |
| 关键词: 印尼贯穿流 海表面温度 深度学习 卷积神经网络 |
| DOI：10.11693/hyhz20240900195 |
| 分类号: |
| 基金项目: |
|
| Using a deep learning method and sea surface temperatures to construct a time series of Indonesian throughflow transport since the industrial revolution |
|
XIN LINCHAO, hu shi jian
|
|
Institute of Oceanology, CAS
|
| Abstract: |
| The Indonesian Transversal Flow (ITF) connects the tropical Pacific and Indian Oceans and is critical to the regional and global climate systems. However, what is most lacking in the study of long-term changes in the Indonesian Throughflow is a lengthy time series. Here we use deep learning methods of convolutional neural networks (CNN) to infer ITF volume transport. The CNN model was trained and tested with the results of the sixth International Coupled Model Comparison Program (CMIP6). Results show that the CNN model using sea surface temperature (SST) was able to reproduce about 80% of the total variance of ITF transport. A time series of 154 years ITF transport from 1870 to 2023 is inferred with the HadISST dataset for the first time. The time series is validated by historical observations from subsurface moorings. Sensitivity regions in referring the ITF in the model are investigated with the Grad-CAM and attention mechanisms. |
| Key words: Indonesian Throughflow, sea surface temperature, deep learning, convolutional neural network |