Correcting Biases in Historical Bathythermograph Data Using Artificial Neural Networks
Correcting Biases in Historical Bathythermograph Data Using Artificial Neural Networks
复制标题
DOI：
10.1175/jtech-d-19-0103.1
复制
发表时间：
2020-10-01
影响因子：
2.2
通讯作者：
DeVries, Timothy
中科院分区：
文献类型：
--
作者：
Bagnell, Aaron;DeVries, Timothy
海洋热含量（OHC）的历史估计对于了解地球系统的气候敏感性和跟踪地球能量平衡随时间的变化非常重要。在2004年之前，这些估计数主要依赖于海军船只和机会船大规模部署的机械和消耗性深海温度计（BT）仪器的温度测量值。这些BT温度测量结果会受到有据可查的偏差的影响，但即使是最好的校准方法，与高质量的温度数据集相比，仍然会出现残余偏差。在这里，我们使用一种新的方法来减少历史BT数据中的偏差，将它们合并到一个规则的网格中，例如用于估计OHC。我们的方法包括一个人工神经网络的集合，它可以纠正前10米深度，年份和水温的偏差。一个全球性的校正和校正优化到特定的BT探头类型的顶部1800米。我们的方法与大多数先前的研究不同，在一个单一的校正中考虑了多个误差来源，而不是将偏差分为几个独立的组成部分。这些新的全局和探针特定校正与广泛使用的校准方法在一系列指标上的表现不相上下，这些指标检查了相对于高质量参考数据集的残余温度偏差。然而，当将这些不同的校准方法外推到我们的跨仪器比较中不包括的BT数据时，会出现不同的模式，从而导致最终影响OHC估计的不确定性。
Historical estimates of ocean heat content (OHC) are important for understanding the climate sensitivity of the Earth system and for tracking changes in Earth's energy balance over time. Prior to 2004, these estimates rely primarily on temperature measurements from mechanical and expendable bathythermograph (BT) instruments that were deployed on large scales by naval vessels and ships of opportunity. These BT temperature measurements are subject to well-documented biases, but even the best calibration methods still exhibit residual biases when compared with high-quality temperature datasets. Here, we use a new approach to reduce biases in historical BT data after binning them to a regular grid such as would be used for estimating OHC. Our method consists of an ensemble of artificial neural networks that corrects biases with respect to depth, year, and water temperature in the top 10 m. A global correction and corrections optimized to specific BT probe types are presented for the top 1800 m. Our approach differs from most prior studies by accounting for multiple sources of error in a single correction instead of separating the bias into several independent components. These new global and probe-specific corrections perform on par with widely used calibration methods on a series of metrics that examine the residual temperature biases with respect to a high-quality reference dataset. However, distinct patterns emerge across these various calibration methods when they are extrapolated to BT data that are not included in our cross-instrument comparison, contributing to uncertainty that will ultimately impact estimates of OHC.