Prediction of the Indian Ocean Dipole using Deep Learning Method
-
摘要: 印度洋偶极子(IOD)是热带印度洋秋季最强的年际变率,它会通过大气遥相关来影响世界许多地区的气候。目前耦合气候模式对IOD预报技巧仍非常有限,远低于热带太平洋的厄尔尼诺事件的预报技巧。鉴于深度学习具备高效的数据处理能力,本文使用深度学习中的卷积神经网络(CNN)与人工神经网络中的多层感知机(MLP)处理再分析观测资料,从而进行IOD预报。由于当预报初始时刻为北半球冬春季时,对IOD事件的预报技巧较低。因此,为探索CNN的预报能力,本文仅使用三种(1~3月、2~4月、3~5月)初始时刻的海表温度异常(SSTA)作为CNN的输入数据,来预报其后续七个月的印度洋偶极子指数(DMI)、东极子指数(EIOI)和西极子指数(WIOI)。结果表明:CNN对DMI、EIOI和WIOI的有效预测时效均超过了6个月。与现在耦合动力模式相比,CNN模型能够显著提升DMI和EIOI的预报技巧,但对WIOI的预报技巧提升有限。当预报提前时间为7个月时,CNN模型能够比较准确地预报1994、1997与2019年的IOD事件。由于CNN模型能够更好地抓住印度洋海温的空间结构特征,它对IOD事件的预报技巧比传统神经网络MLP高。Abstract: In autumn, the Indian Ocean Dipole (IOD) has the strongest interannual variability in the tropical Indian Ocean. It will influence the climate in many parts of the world due to atmospheric teleconnection. The current coupled climate model has very limited IOD forecasting skills, which are far inferior to those of El Niño events in the tropical Pacific. The authors used the convolutional neural network (CNN) of the deep learning and the multi-layer perceptron (MLP) of the artificial neural network, respectively, to perform IOD prediction due to the super capability of deep learning in processing data. In order to explore the forecasting capabilities of CNN, this article only uses three initial conditions in the boreal spring with low prediction skill to forecast the Indian Ocean Dipole Mode Index (DMI), East Pole Index (EIOI), and West Pole Index (WIOI) for the next seven months. The results show that the CNN model can make useful predictions for the DMI, EIOI, and WIOI at least six months in advance. When compared with the current state-of-the-art general coupled model, the CNN model significantly improves the prediction skills of the DMI and EIOI while only slightly improving WIOI prediction skills. The CNN model could predict the strong IOD events in 1994, 1997, and 2019 well for the lead time longer than seven months. In general, CNN outperforms the traditional neural network MLP for the IOD prediction due to its strong capability to capture the spatial structure characteristics of the Indian Ocean SST.
-
Key words:
- Indian Ocean Dipole /
- Deep learning /
- Convolutional neural network /
- Climate prediction
-
图 2 1990~2019年以JFM、FMA、MAM为起始态,分别使用CNN模型(实线)与MLP模型(虚线)预报的DMI与观测值的(a)相关系数和(b)均方根误差(RMSE)
Figure 2. (a) Correlation coefficients and (b) RMSE (root mean square errors) between the forecasted and observed DMI during 1990–2019 using CNN (solid lines) and MLP (dashed lines) models, respectively, with JFM (January–March), FMA (February–April), and MAM (March–May) as the initial conditions
图 4 1990~2019年以JFM、FMA、MAM为起始态,分别使用CNN模型(实线)与MLP模型(虚线)预报的EIO指数与观测值的(a)相关系数和(b)RMSE
Figure 4. (a) Correlation coefficients and (b) RMSE between the forecasted and observed EIOI (East Pole Index for Indian Ocean) during 1990–2019 using CNN (solid lines) and MLP (dashed lines) models, respectively, with JFM, FMA, and MAM as the initial conditions
图 6 1990~2019年以JFM、FMA、MAM为起始态,分别使用CNN模型(实线)与MLP模型(虚线)预报的WIO指数与观测值的(a)相关系数和(b)RMSE
Figure 6. (a) Correlation coefficients and (b) RMSE between the forecasted and observed WIOI (West Pole Index for Indian Ocean) during 1990–2019 using CNN and MLP models, respectively, with JFM, FMA, and MAM as the initial conditions
表 1 CNN模型、MLP模型的输入(SSTA)与输出(DMI)
Table 1. Input (Sea Surface Temperature Anomaly, SSTA) and output (Indian Ocean Dipole Mode Index, DMI) of CNN model and MLP (multi-layer perceptron) model
输入(SSTA) 输出 提前1个月 提前2个月 提前3个月 提前4个月 提前5个月 提前6个月 提前7个月 JFM 4月DMI 5月DMI 6月DMI 7月DMI 8月DMI 9月DMI 10月DMI FMA 5月DMI 6月DMI 7月DMI 8月DMI 9月DMI 10月DMI 11月DMI MAM 6月DMI 7月DMI 8月DMI 9月DMI 10月DMI 11月DMI 12月DMI -
[1] Ashok K, Guan Zhaoyong, Yamagata T. 2001. Impact of the Indian Ocean Dipole on the relationship between the Indian monsoon rainfall and ENSO [J]. Geophys. Res. Lett., 28(23): 4499−4502. doi: 10.1029/2001GL013294 [2] Ashok K, Guan Zhaoyong, Yamagata T. 2003. Influence of the Indian Ocean Dipole on the Australian winter rainfall [J]. Geophys. Res. Lett., 30(15): 1821. doi: 10.1029/2003GL017926 [3] Becker E, van den Dool H, Zhang Qin. 2014. Predictability and forecast skill in NMME [J]. J. Climate, 27(15): 5891−5906. doi: 10.1175/JCLI-D-13-00597.1 [4] Behera S, Ratnam J V, Masumoto Y, et al. 2013. Origin of extreme summers in Europe: The Indo-Pacific connection [J]. Climate Dyn., 41: 663−676. doi: 10.1007/s00382-012-1524-8 [5] Feng Rong, Duan Wansuo, Mu Mu. 2014. The “winter predictability barrier” for IOD events and its error growth dynamics: Results from a fully coupled GCM [J]. J. Geophys. Res.: Ocean, 119(12): 8688−8708. doi: 10.1002/2014JC010473 [6] Ham Y G, Kim J H, Luo Jingjia. 2019. Deep learning for multi-year ENSO forecasts [J]. Nature, 573(7775): 568−572. doi: 10.1038/s41586-019-1559-7 [7] Huang Boyin, Thorne P W, Banzon V F, et al. 2017. Extended reconstructed sea surface temperature, version 5 (ERSSTv5): Upgrades, validations, and intercomparisons [J]. J. Climate, 30(20): 8179−8205. doi: 10.1175/JCLI-D-16-0836.1 [8] Kirtman B P, Min D, Infanti J M, et al. 2014. The North American Multi-model Ensemble: Phase-1 seasonal-to-interannual prediction; Phase-2 toward developing intraseasonal prediction [J]. Bull. Amer. Meteor. Soc., 95(4): 585−601. doi: 10.1175/BAMS-D-12-00050.1 [9] Klein S A, Soden B J, Lau N C. 1999. Remote sea surface temperature variations during ENSO: Evidence for a tropical atmospheric bridge [J]. J. Climate, 12(4): 917−932. doi:10.1175/1520-0442(1999)012<0917:RSSTVD>2.0.CO;2 [10] Krizhevsky A, Sutskever I, Hinton G E. 2017. ImageNet classification with deep convolutional neural networks [J]. Communications of the ACM, 60(6): 84−90. doi: 10.1145/3065386 [11] LeCun Y, Bengio Y, Hinton G. 2015. Deep learning [J]. Nature, 521(7553): 436−444. doi: 10.1038/nature14539 [12] Lee J, Kim C G, Lee J E, et al. 2018. Application of artificial neural networks to rainfall forecasting in the Geum river basin, Korea [J]. Water, 10(10): 1448. doi: 10.3390/w10101448 [13] 雷蕾, 伍艳玲, 唐佑民. 2020. 印度洋偶极子预报技巧在多模式中的对比研究 [J]. 海洋学报, 42(7): 51−63. doi: 10.3969/j.issn.0253-4193.2020.07.005Lei Lei, Wu Yanling, Tang Youmin. 2020. A comparison of Indian Ocean dipole prediction skill in a multi-model ensemble [J]. Haiyang Xuebao (in Chinese), 42(7): 51−63. doi: 10.3969/j.issn.0253-4193.2020.07.005 [14] Li Chongyin, Mu Mingquan. 2001. The influence of the Indian Ocean dipole on atmospheric circulation and climate [J]. Advances in Atmospheric Sciences, 18(5): 831−843. doi: 10.1007/BF03403506 [15] Liu Huafeng, Tang Youmin, Chen Dake, et al. 2017. Predictability of the Indian Ocean Dipole in the coupled models [J]. Climate Dyn., 48(5−6): 2005−2024. doi: 10.1007/s00382-016-3187-3 [16] Liu Da, Duan Wansuo, Feng Rong, et al. 2018. Summer predictability barrier of Indian Ocean Dipole events and corresponding error growth dynamics [J]. J. Geophys. Res. :Oceans, 123(5): 3635−3650. doi: 10.1029/2017JC013739 [17] Luo Jingjia, Masson S, Behera S, et al. 2005. Seasonal climate predictability in a coupled OAGCM using a different approach for ensemble forecasts [J]. J. Climate, 18(21): 4474−4497. doi: 10.1175/JCLI3526.1 [18] Luo Jingjia, Masson S, Behera S, et al. 2007. Experimental forecasts of the Indian Ocean Dipole using a coupled OAGCM [J]. J. Climate, 20(10): 2178−2190. doi: 10.1175/JCLI4132.1 [19] McKenna S, Santoso A, Gupta A S, et al. 2020. Indian Ocean Dipole in CMIP5 and CMIP6: Characteristics, biases, and links to ENSO [J]. Sci. Rep., 10: 11500. doi: 10.1038/s41598-020-68268-9 [20] Rasp S, Pritchard M S, Gentine P. 2018. Deep learning to represent subgrid processes in climate models [J]. Proceedings of the National Academy of Sciences of the United States of America, 115(39): 9684−9689. doi: 10.1073/pnas.1810286115 [21] Ratnam J V, Dijkstra H A, Behera S K. 2020. A machine learning based prediction system for the Indian Ocean Dipole [J]. Sci. Rep., 10: 284. doi: 10.1038/s41598-019-57162-8 [22] Sahai A K, Soman M K, Satyan V. 2000. All India summer monsoon rainfall prediction using an artificial neural network [J]. Climate Dyn., 16: 291−302. doi: 10.1007/s003820050328 [23] Saji N H, Yamagata T. 2003. Possible impacts of Indian Ocean Dipole mode events on global climate [J]. Climate Res., 25: 151−169. doi: 10.3354/cr025151 [24] Saji N H, Goswami B N, Vinayachandran P N, et al. 1999. A Dipole mode in the tropical Indian Ocean [J]. Nature, 401(6751): 360−363. doi: 10.1038/43854 [25] Shi Li, Hendon H H, Alves O, et al. 2012. How predictable is the Indian Ocean dipole? [J]. Mon. Wea. Rev., 140(12): 3867−3884. doi: 10.1175/MWR-D-12-00001.1 [26] Venzke S, Latif M, Villwock A. 2000. The coupled GCM ECHO-2. Part II: Indian Ocean response to ENSO [J]. J. Climate, 13(8): 1371−1383. doi:10.1175/1520-0442(2000)013<1371:TCGE>2.0.CO;2 [27] Wang Guojian, Cai Wenju, Santoso A. 2017. Assessing the impact of model biases on the projected increase in frequency of extreme positive Indian Ocean Dipole events [J]. J. Climate, 30(8): 2757−2767. doi: 10.1175/JCLI-D-16-0509.1 [28] Weller E, Cai Wenju. 2013. Realism of the Indian Ocean Dipole in CMIP5 models: The implications for climate projections [J]. J. Climate, 26(17): 6649−6659. doi: 10.1175/JCLI-D-12-00807.1 [29] Wu Yanling, Tang Youmin. 2019. Seasonal predictability of the tropical Indian Ocean SST in the North American multi-model ensemble [J]. Climate Dyn., 53(5−6): 3361−3372. doi: 10.1007/s00382-019-04709-0 [30] Zeiler M D, Fergus R. 2014. Visualizing and understanding convolutional networks [C]//European Conference on Computer Vision. Zurich, Switzerland: Springer, 818–833. doi: 10.1007/978-3-319-10590-1_53 [31] Zhao Mei, Hendon H H. 2009. Representation and prediction of the Indian Ocean dipole in the POAMA seasonal forecast model [J]. Quart. J. Roy. Meteor. Soc., 135(639): 337−352. doi: 10.1002/qj.370 -