An Assessment of Drift Correction Alternatives for CMIP5 Decadal Predictions
An Assessment of Drift Correction Alternatives for CMIP5 Decadal Predictions
复制标题
DOI：
10.1002/2017jd026900
复制
发表时间：
2017-10-16
影响因子：
4.4
通讯作者：
Sivakumar, Bellie
中科院分区：
文献类型：
--
作者：
Choudhury, Dipayan;Sen Gupta, Alexander;Sivakumar, Bellie
漂移校正是使用十年预报实验结果之前的一个重要步骤，已经有了大量的研究。然而，大多数漂移校正研究考虑的变量和模型样本相对较少。在这里，我们系统地应用现有的漂移校正策略，从一套五个最先进的气候模式(CanCM4i1、GFDL-CM2.1、HadCM3i2&i3、MIROC5和MPI-ESM-LR)对各种基于海表面温度的指标进行年代际预测。报告了每种度量和模型的最佳漂移校正方法。初步分析表明，没有一种单一的漂移校正方法能够始终如一地表现最佳。基于初始条件的漂移校正为MIROC5和两个HadCM3模型提供了大多数区域中最低的误差，而基于趋势的漂移校正在该区域的最大份额上为CanCM4i1、GFDL-CM2.1和MPI-ESM-LR提供了最低的误差。对于这些漂移校正方法，使用k近邻方法没有优点。此外，在几乎所有情况下，多模型集合的表现都好于单个模型，因此，研究结论建议使用基于多模型平均值的预测。我们还展示了在模型平均之前使用每个模型/度量的最佳校正方法校正漂移可以获得的一些额外好处，并建议这里给出的结果将帮助潜在用户在处理这些实验的输出时明智地花费时间和资源。
Drift correction is an important step before using the outputs of decadal prediction experiments and has seen considerable research. However, most drift correction studies consider a relatively small sample of variables and models. Here, we present a systematic application of the existing drift correction strategies for decadal predictions of various sea surface temperature-based metrics from a suite of five state-of-the-art climate models (CanCM4i1, GFDL-CM2.1, HadCM3i2&i3, MIROC5, and MPI-ESM-LR). The best method of drift correction for each metric and model is reported. Preliminary analysis suggests that there is no single method of drift correction that consistently performs best. Initial condition-based drift correction provides the lowest errors in most regions for MIROC5 and the two HadCM3 models, whereas the trend-based drift correction produces lowest errors for CanCM4i1, GFDL-CM2.1, and MPI-ESM-LR over the largest share of the area. There is no merit in using a k-nearest neighbor approach for these drift correction methods. Further, in almost all cases, the multimodel ensemble outperforms the individual models, and thus, the study conclusively suggests using forecasts based on multimodel averages. We also show some additional benefit to be gained by drift correcting each model/metric using their best correction method prior to model averaging and suggest that the results presented here would help potential users expend time and resources judiciously while dealing with outputs from these experiments.