基于深度学习的印尼贯穿流反演与预测研究

IOCAS-IR > 海洋环流与波动重点实验室

	基于深度学习的印尼贯穿流反演与预测研究
	辛林超
学位类型	硕士
导师	胡石建
	2023-05
学位授予单位	中国科学院大学
学位授予地点	中国科学院海洋研究所
学位名称	理学硕士
关键词	印尼贯穿流，海表面高度，神经网络，深度学习，卷积神经网络
摘要	印尼贯穿流(Indonesian Throughflow, ITF)是全球海洋中唯一的热带间海洋通道，连接了太平洋低纬度西边界流和印度洋环流系，可以大量地将热带太平洋暖水和盐向印度洋输送，将热带太平洋和印度洋之间的热和盐进行重新的分配。ITF 也影响着两个大洋、尤其是印度洋的流场变化，而这些变化又会作用于海洋和气候系统，在区域和全球气候系统中发挥着重要作用。大洋环流模式和耦合全球气候模式的模拟表明，ITF 的存在对全球海洋环流和气候具有深远的影响。在全球变暖的背景下，包括 ITF 在内的海洋环流预计将发生显著变化。ITF 的变化可能导致印太交流的波动，并影响区域和全球气候。因此，研究印尼贯穿流的长期变化趋势，对于探讨全球变暖背景下太平洋和印度洋两大暖池结构的相互影响，预测全球未来气候变化趋势具有重要意义。然而，由于缺乏长期和连续的 ITF 时间序列，很难对其有更好地理解。前人的研究表明，印度-太平洋压力梯度是 ITF 的主要驱动因素，这意味着可以用印度-太平洋海表面高度来反演 ITF 的输运流量。在本文中，我们主要使用深度学习方法中的卷积神经网络模型来重构 ITF 流量。卷积神经网络模型使用“第六次国际耦合模式比较计划” (Coupled Model Intercomparison Project Phase 6, CMIP6)模拟的 1850-1974 年数据进行训练，并使用 CMIP6 模拟的 1974-2014 年数据进行验证、测试。对训练结果的测试表明，使用海表面高度数据训练的卷积神经网络模型能够重现约 90%的 ITF 流量总方差，反演技能为 0.69。将使用CMIP6 训练的卷积神经网络模型迁移学习至 SODA 数据集中，研究发现迁移学习后的神经网络模型能够再现 SODA 中约 80%的 ITF 流量总方差，反演技能为 0.54。然后通过该模型利用 1993-2021 年期间的卫星观测数据，生成了这段时间的 ITF 流量时间序列，该序列得到了多个观测 ITF 测量计划的验证。深度学习通过卫星数据反演的 ITF 输运流量捕捉到了 ITF 的一般性，并且卷积神经网络通过遥感数据反演的ITF 流量信息与 INSTANT 计划、MITF、IX1 阵列观测到的 ITF 流量信息的相关系数分别为 0.57、0.52、0.43，在 98%的置信水平上显著。生成的 ITF 流量时间序列与 Nino3.4 指数的相关系数为 0.62,并提前 ITF 三个月左右的时间。卷积神经网络模型也能够提前 7 个月对 ITF 流量做出有效的预测，这意味着使用海表面高度数据的深度学习方法预测 ITF 流量是可行的。相比较其他神经网络模型，卷积神经网络具有更高的准确性、预测时长以及更加轻量化的模型参数。
其他摘要	Indonesian Throughflow (ITF) is the only intertropical ocean channel in the global ocean, connecting the western boundary current of the Pacific Ocean at low latitude and the Indian Ocean circulation system, which can transfer a large amount of warm water and salt from the tropical Pacific Ocean to the Indian Ocean, and redistribute heat and salt between the tropical Pacific and the Indian Ocean. It also influences changes in the flow fields of the two oceans, especially the Indian Ocean, which in turn act on the ocean and climate systems, playing an important role in the regional and global climate systems. Simulations of ocean circulation models and coupled global climate models suggest that the existence of ITF also has profound effects on global ocean circulation and climate In the context of global warming, ocean circulation, including the ITF, is expected to change significantly. Changes in the ITF can lead to fluctuations in Indo-Pacific exchanges and affect regional and global climate. Therefore, it is of great significance to study the long-term variation trend of the penetrating Indonesian current in order to discuss the interaction of the two warm pool structures in the Pacific Ocean and the Indian Ocean under the background of global warming, and to predict the future trend of global climate change. However, the lack of long-term and continuous ITF time series makes it difficult to gain a better understanding. Previous studies have shown that the Indo-Pacific pressure gradient is the main driver of the ITF, which means that the ITF transport can be inverted using the surface height of the Indo-Pacific Sea. In this paper, we mainly use a convolutional neural network model in deep learning method to reconstruct ITF transport. The convolutional neural network model was trained using Coupled Model Intercomparison Project Phase 6 (CMIP6) simulated data from 1850-1974. The data from 1974 to 2014 simulated by CMIP6 were used for verification and testing. Tests of the training results showed that the convolutional neural network model trained using sea surface height data was able to reproduce about 90% of the total variance of ITF and the skill was 0.69. The convolutional neural network model trained by CMIP6 was transferred and learned into SODA data set, and it was found that the neural network model after transfer learning could reproduce about 80% of the total variance of ITF transport in SODA and the skill was 0.54. The model then uses satellite observation data from 1993 to 2021 to generate a time series of ITF transport during this period, which has been verified by several observational ITF measurement programs. Deep learning captures the generality of ITF through ITF transport traffic retrieved from satellite data, and the correlation coefficients between ITF traffic information retrieved from convolutional neural network through remote sensing data and ITF transport information observed by INSTANT plan, MITF and IX1 arrays are 0.57, 0.52 and 0.43, which are significant at the 98% confidence level. The generated ITF transport time series has a correlation of 0.62 with the Nino3.4 index and is about three months ahead of it. The convolutional neural network model can also effectively predict ITF transport 7 months in advance, which means that it is feasible to predict ITF transport using the deep learning method of sea surface height data. Compared with other neural network models, convolutional neural network has higher accuracy, prediction time and lighter model parameters.
学科门类	理学::海洋科学
语种	中文
文献类型	学位论文
条目标识符	http://ir.qdio.ac.cn/handle/337002/181233
专题	海洋环流与波动重点实验室
推荐引用方式 GB/T 7714	辛林超. 基于深度学习的印尼贯穿流反演与预测研究[D]. 中国科学院海洋研究所. 中国科学院大学,2023.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
辛林超-基于深度学习的印尼贯穿流反演与预（3884KB）	学位论文		延迟开放	CC BY-NC-SA	浏览 2025-7-1后可获取