基于机器学习方法重建的过去1000年北半球环状模(NAM)指数

RECONSTRUCTION OF THE NORTHERN ANNULAR MODE(NAM)INDEX FOR THE PAST 1000 YEARS BASED ON MACHINE LEARNING METHODS

受限于观测资料的短缺,关于北半球主要大气环流模态(Northern Annular Mode,简称NAM)的演变规律和机理还很不明确。运用树轮、冰芯、沉积物等代用指标重建时间序列更长的NAM指数有助于加深对其演变规律和驱动机制的认识。本文通过评估多种机器学习模型在古气候重建中的适用性,基于PAGES 2k的气候代用指标重建了过去1000年高分辨率(1年)的NAM指数。研究结果表明相比普通线性回归模型和随机森林等模型,Cat Boost、极端随机树和主成分回归模型可以有效地避免过拟合,模型具有更高的稳定性和可靠性,其中Cat Boost模型的重建结果与器测时段内NAM指数的相关系数最高(R=0.93,p<0.01),能够更好地拟合NAM指数的量级和峰谷变化。分析过去1000年NAM指数的变化特征,发现NAM具有显著的百年际周期(167.5年)和多年代际周期(32.3年)波动,1950~2000年NAM由负位相转向正位相的速率在过去1000年中前所未有。进一步探究NAM与温度和海冰的关系发现,1850年之前,暖期对应NAM增强,冷期对应NAM减弱;而在1850年之后NAM的多年代际变化与巴伦支-喀拉海海冰范围的变化趋于一致,1950年之后NAM向正位相快速转变可能是温度和北极海冰异常共同影响的结果。

As the dominant mode of the atmospheric circulation in the Northern Hemisphere(NH),the Northern Annular Mode(NAM)makes important influences on both NH and global climate change.The knowledge of the NAM is limited by the lack of the observed data.The reconstruction of long-term changes of the NAM using proxies such as tree rings,ice cores and sediments can help us understanding the long-term characters and mechanisms of NAM.In this study,we compare eight reconstruction models,including Original Linear Regression(OR),Principle Composite Regression(PCR),Ridge,Support Vector Regression(SVR),Random Forest(RF),Extremely Randomized Trees(ET),Light Gradient Boosting Machine(LightGBM)and CatBoost,to estimate their applicability in NAM reconstruction.And further reconstruct the annual NAM index for the past 1000 years based on the climate proxies obtained from PAGES2k.By removing missing and anomalous data,54 proxies(obtained from 22 tree ring and lake sediment records)can be used for NAM reconstruction were selected after filtering by feature engineering and correlation analysis.Compared with the OR and RF methods,the results show that the CatBoost,ET and PCR model can effectively avoid overfitting problems and better reconstruct the variability of the NAM over the instrumental period(1948~2000).The CatBoost reconstruction has the highest correlation coefficient(R=0.93,p<0.01)with the NAM index over the instrumental period and is able to better fit the magnitude and peak-to-valley variability of the NAM index.Even the reconstruction series of the OR and RF are significantly correlated with the instrumental NAM index from 1948 A.D.to 2000 A.D.,the results of the cross validation show the high correlation in the calibration period(1967~2000)and very low correlation in the validation period(1949~2000),and the Nash-Sutcliffe coefficient of efficiency(NSCE)are both negative in validation period.This suggests there is a high risk of overfitting when use the two models.The reconstructions of OR and Ridge are highly variable in amplitude because of the high sensitivity to noises in proxies,which may produce spurious estimates.The reconstructions of SVR and LightGBM,which their NSCE are lower than 0.1 in validation period,are also less reliable because of the high sensitivity to the noise of proxies or the insufficient amount of data.We recommend the CatBoost and ET as the preferred machine learning models for paleoclimate reconstruction.The ensemble empirical mode decomposition(EEMD)is used to decompose of the NAM series during the past1000 years to obtain its multi-scales characters and nonlinear trends.The results suggest that the NAM has a significant centennial(167.5 a)and multidecadal(32.3 a)oscillations.The positive shift of NAM from 1950 A.D.to 2000 A.D.is unprecedented in the past 1000 years.We further found that prior to 1850 A.D.,the multidecadal variability of NAM was mainly influenced by the global mean temperature,with warm periods corresponding to an enhanced NAM and cold periods corresponding to a weakened NAM.While after 1850 A.D.,the multidecadal changes of NAM tend to coincide with sea ice extent in the Barents-Kara Sea.We infer that rapid shift of the NAM towards a positive phase occurred after 1950 s is probably influenced by the combined effect of the anomalies in both temperature and Arctic sea ice.