Application of machine learning in the ensemble projection of regional extreme climate indices over China

doi:10.11755/j.issn.1006-7639-2026-03-0398

Abstract

Abstract:

Under the background of climate warming, changes in extreme climate events have received increasing attention. In this study, three machine learning models, namely Random Forest, Extremely Randomized Trees, and Ridge Regression, were used to evaluate their capability in simulating extreme climate indices over China, including the warm day index (TX90p), cold night index (TN10p), maximum 1-day precipitation (RX1day), and maximum consecutive 5-day precipitation (RX5day). Their performances were compared with that of the traditional globally bias-corrected multi-model ensemble method to determine the optimal model schemes. Furthermore, the spatial distributions and temporal trends of the selected extreme climate indices during 2024-2100 under different emission scenarios, including SSP1-2.6, SSP2-4.5, and SSP5-8.5, were analyzed relative to the baseline period of 1961-1990. The results show that the machine-learning-based schemes can improve the simulation capability for extreme climate events to varying degrees and effectively reduce simulation biases. During 2024-2100, extremely warm events over China increase significantly under all emission scenarios, with the largest increase under the high-emission scenario SSP5-8.5, reaching approximately 52% relative to the baseline period. In contrast, extremely cold events decrease significantly, and the reduction becomes more pronounced with increasing emissions. Spatially, TX90p shows relatively large increases in North China, the middle and lower reaches of the Yangtze River, the Sichuan Basin, and parts of South China, while the increases are relatively smaller over the Qinghai-Xizang Plateau and some high-elevation areas in Northwest China. Extreme precipitation indices, including RX1day and RX5day, show increasing trends under all emission scenarios, with the most pronounced increase under the high-emission scenario. The enhancement of RX5day is generally stronger than that of RX1day. Northeast China and North China show the strongest responses in extreme precipitation, especially for RX5day, indicating higher sensitivity to climate warming.

Key words: machine learning, extreme climate indices, future projection

摘要：

在气候变暖背景下，极端气候事件的变化趋势备受关注。基于3种机器学习（随机森林、极端随机树和岭回归）模型评估其对中国区域极端气候指数[暖昼指数（TX90p）、冷夜指数（TN10p）、日最大降水量（RX1day）及5 d最大降水量（RX5day）]的模拟能力，并与传统全局偏差订正后的多模式集合方法对比，确定最优模型方案；进一步分析所选取的极端气候指数在不同排放情景（SSP1-2.6、SSP2-4.5、SSP5-8.5）下2024—2100年相对于基准期（1961—1990年）的空间分布及其变化趋势。结果表明：基于机器学习的极端气候指数模拟方案能够在不同程度上提升对极端气候事件的模拟能力，有效减小模拟偏差；2024—2100年中国地区极端暖事件在各排放情景下均显著上升，且高排放（SSP5-8.5）下增幅最大（相较基准期上升约52%），极端冷事件显著减少，且随排放增加减少更明显；空间分布上，TX90p在华北、长江中下游、四川盆地及华南部分地区增幅相对明显，而青藏高原及西北部分高海拔地区增幅相对较小；极端降水指数（RX1day、RX5day）在所有排放情景下均呈增加趋势，且高排放情景下增加最明显，其中RX5day的增强幅度整体高于RX1day；东北和华北地区在极端降水（特别RX5day）上的响应最强，对气候变暖的敏感性更高。

关键词: 机器学习, 极端气候指数, 未来预估

CLC Number:

P467

LIU Mingming, XU Ying. Application of machine learning in the ensemble projection of regional extreme climate indices over China[J]. Journal of Arid Meteorology, 2026, 44(3): 398-411.

刘明铭, 徐影. 机器学习在中国区域极端气候指数集合预估中的应用[J]. 干旱气象, 2026, 44(3): 398-411.

Figures/Tables 9

Tab.1 Definitions of the four extreme climate indices recommended by the ETCCDI

指数缩写	指数名称	定义	单位
TX90p	暖昼指数	每月日最高气温大于基准期内90%分位值的天数百分率	%
TN10p	冷夜指数	每月日最低气温小于基准期内10%分位值的天数百分率	%
RX5day	5 d最大降水量	每月最大的连续5 d降水量	mm
RX1day	日最大降水量	每月最大日降水量	mm

Tab.2 Basic information of the 19 CMIP6 global climate models （GCMs）

序号	模式名称	模式单位（国家）	分辨率（经度×纬度）
1	ACCESS-CM2	澳大利亚气候科学研究中心（澳大利亚）	1.25°×1.875°
2	ACCESS-ESM1-5	澳大利亚气候科学研究中心（澳大利亚）	1.25°×1.875°
3	BCC-CSM2-MR	北京气候中心（中国）	1.125°×1.125°
4	CanESM5	加拿大气候模型与分析中心（加拿大）	2.8°×2.8°
5	CNRM-CM6-1	法国国家气象研究中心（法国）	1.4°×1.4°
6	CNRM-ESM2-1	法国国家气象研究中心（法国）	1.4°×1.4°
7	EC-Earth3	EC-Earth联盟（欧洲）	0.7°×0.7°
8	FGOALS-g3	中国科学院大气物理研究所（中国）	2.25°×2.25°
9	GFDL-ESM4	地球物理流体动力学实验室（美国）	1.0°×1.25°
10	HadGEM3-GC31-LL	英国气象局（英国）	1.0°×1.0°
11	INM-CM4-8	俄罗斯数值数学研究所（俄罗斯）	1.5°×1.5°
12	KIOST-ESM	韩国海洋科学技术院（韩国）	2.0°×2.0°
13	MIROC6	日本海洋地球科学技术研究所（日本）	1.0°×1.0°
14	MPI-ESM1-2-HR	马克斯·普朗克气象研究所（德国）	1.5°×1.5°
15	MPI-ESM1-2-LR	马克斯·普朗克气象研究所（德国）	1.5°×1.5°
16	MRI-ESM2-0	气象研究所（日本）	1.125°×1.125°
17	NESM3	国家气候中心（中国）	0.75°×0.75°
18	NorESM2-LM	挪威气候中心（挪威）	1.25°×1.25°
19	UKESM1-0-LL	英国气象局（英国）	0.5°×0.5°

Fig.1 Research framework

Fig.2 Spatial distribution of climate state bias （simulation values minus observed values） of different schemes in various extreme climate index simulations and the corresponding TSS scores during the verification period （2005-2014） in China （The highest TSS score in each index is highlighted in red）

Fig.3 Taylor diagrams comparison of four extreme climate indices during the verification period （a） TX90p，（b） TN10p，（c） RX1day，（d） RX5day （Each subplot shows the distribution of spatial correlation coefficient， standard deviation， and centered root mean square error （CRMSE）（red dashed lines） for a single CMIP6 model （gray markers） and different scenarios （colored markers） compared with observations）

Fig.4 Temporal trends of anomalies of extreme climate indices in China under different SSP scenarios （relative to 1961-1990）（a） TX90p，（b） TN10p，（c） RX1day，（d） RX5day （The extreme temperature indices are an absolute change， while the extreme precipitation indices are a relative change）

Tab.3 Future trends （2024-2100） of extreme climate indices over China estimated by the AM and machine learning under different emission scenarios （relative to 1961-1990）

指数	排放情景	AM	机器学习
TX90p	SSP1-2.6	+1.07	+0.95（Ridge）
	SSP2-4.5	+2.86	+2.62（Ridge）
	SSP5-8.5	+6.00	+5.49（Ridge）
TN10p	SSP1-2.6	-0.20	-0.19（RF）
	SSP2-4.5	-0.42	-0.41（RF）
	SSP5-8.5	-0.56	-0.54（RF）
RX1day	SSP1-2.6	+0.92	+1.01（ET）
	SSP2-4.5	+1.75	+1.63（ET）
	SSP5-8.5	+3.90	+3.76（ET）
RX5day	SSP1-2.6	+0.98	+1.22（ET）
	SSP2-4.5	+1.71	+1.81（ET）
	SSP5-8.5	+3.73	+3.92（ET）

Fig.5 Spatial distribution of anomalies of the TX90p （the top， based on the Ridge model） and the TN10p （the bottom， based on the RF model） over China under SSP1-2.6 （the left）， SSP2-4.5（the middle）， and SSP5-8.5 （the right） scenarios for the late 21st century （2080-2099）， relative to the 1961-1990 baseline （Unit： %）

Fig.6 The spatial distribution of anomalies of the RX1day （the top） and the RX5day （the bottom） over China under SSP1-2.6 （the left）， SSP2-4.5 （the middle）， and SSP5-8.5 （the right） scenarios for the late 21st century （2080-2099）， relative to the 1961-1990 baseline （Unit： %）（based on the ET model）

References 59

[1]	高学杰, 2007. 中国地区极端事件预估研究[J]. 气候变化研究进展, 3(3): 162-166.
[2]	韩振宇, 童尧, 高学杰, 等, 2018. 分位数映射法在RegCM4中国气温模拟订正中的应用[J]. 气候变化研究进展, 14(4): 331-340.
[3]	吉璐莹, 2021. 基于多模式集成和动力降尺度的东亚夏季降水数值预报研究[D]. 南京: 南京信息工程大学.
[4]	鞠琴, 吴金雨, 王兴平, 等, 2024. 机器学习算法在降水和气温多模式集成中的应用[J]. 水资源保护, 40(3): 106-115.
[5]	匡志远, 宋振亚, 董昌明, 2020. 基于机器学习订正模型的未来百年全球海表温度预估研究[J]. 气候变化研究快报, 9(4): 270-284.
[6]	赖雨曈, 徐影, 2024. CMIP6全球气候模式对中国地区干旱模拟能力评估与预估[J]. 大气科学, 48(6): 2 157-2 177.
[7]	刘新, 赵宁, 郭金运, 等, 2020. 基于LSTM神经网络的青藏高原月降水量预测[J]. 地球信息科学学报, 22(8): 1 617-1 629.
[8]	任宏利, 吴捷, 赵崇博, 等, 2015. MJO预报研究进展[J]. 应用气象学报, 26(6): 658-668.
[9]	童尧, 高学杰, 韩振宇, 等, 2017. 基于RegCM4模式的中国区域日尺度降水模拟误差订正[J]. 大气科学, 41(6):1156-1 166.
[10]	王苗, 秦鹏程, 佘晨旭, 等, 2023. 基于CMIP6多模式集合的华中地区气候变化模拟与预估[J]. 气象与环境学报, 39(6): 51-60.
[11]	吴佳, 高学杰, 2013. 一套格点化的中国区域逐日观测资料及与其他资料的对比[J]. 地球物理学报, 56(4): 1 102-1 111.
[12]	吴佳, 周波涛, 徐影, 2015. 中国平均降水和极端降水对气候变暖的响应:CMIP5模式模拟评估和预估[J]. 地球物理学报, 58(9): 3 048-3 060.
[13]	徐影, 2015. 中国未来极端气候事件变化预估图集[M]. 北京: 气象出版社:59-110.
[14]	徐影, 周波涛, 吴婕, 等, 2017. 1.5-4 ℃升温阈值下亚洲地区气候变化预估[J]. 气候变化研究进展, 13(4): 306-315.
[15]	朱欢欢, 2023. 中国气候未来变化的集合预估比较研究[D]. 南京: 南京信息工程大学.
[16]	ABDELMOATY H M, PAPALEXIOU S M, RAJULAPATI C R, et al, 2021. Biases beyond the mean in CMIP6 extreme precipitation: A global investigation[J]. Earth’s Future, 9(10): e2021EF002196. DOI: 10.1029/2021EF002196.
[17]	ACHARYA N, SHRIVASTAVA N A, PANIGRAHI B K, et al, 2014. Development of an artificial neural network based multi-model ensemble to estimate the northeast monsoon rainfall over south peninsular India: An application of extreme learning machine[J]. Climate Dynamics, 43(5): 1 303-1 310. DOI URL
[18]	AHMED K, SACHINDRA D A, SHAHID S, et al, 2020. Multi-model ensemble predictions of precipitation and temperature using machine learning algorithms[J]. Atmospheric Research, 236: 104806. DOI:10.1016/j.atmosres.2019.104806.
[19]	BILBAO-BARRENETXEA N, MARTÍNEZ-ESPAÑA R, JIMENO-SÁEZ P, et al, 2024. Multi-model ensemble machine learning approaches to project climatic scenarios in a river basin in the Pyrenees[J]. Earth Systems and Environment, 8(4): 1 159-1 177. DOI
[20]	CHEN L, FRAUENFELD O W, 2014. A comprehensive evaluation of precipitation simulations over China based on CMIP5 multimodel ensemble projections[J]. Journal of Geophysical Research: Atmospheres, 119(10): 5 767-5 786. DOI URL
[21]	COOK B I, MANKIN J S, MARVEL K, et al, 2020. Twenty-first century drought projections in the CMIP6 forcing scenarios[J]. Earth’s Future, 8(6): e2019EF001461. DOI: 10.1029/2019EF001461.
[22]	CRAWFORD J, VENKATARAMAN K, BOOTH J, 2019. Developing climate model ensembles:A comparative case study[J]. Journal of Hydrology, 568: 160-173. DOI URL
[23]	DARON J, MACADAM I, KANAMARU H, et al, 2018. Providing future climate projections using multiple models and methods:Insights from the Philippines[J]. Climatic Change, 148(1): 187-203. DOI
[24]	DUNNE J P, HEWITT H T, ARBLASTER J M, 2025. An evolving Coupled Model Intercomparison Project phase 7 (CMIP7) and Fast Track in support of future climate assessment[J]. Geoscientific Model Development, 18(19): 6 671-6 700. DOI URL
[25]	EYRING V, BONY S, MEEHL G A, et al, 2016. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization[J]. Geoscientific Model Development, 9(5): 1 937-1 958. DOI URL
[26]	GLECKLER P J, TAYLOR K E, DOUTRIAUX C, 2008. Performance metrics for climate models[J]. Journal of Geophysical Research: Atmospheres, 113(D6): 2007JD008972. DOI: 10.1029/2007JD008972.
[27]	GUO L Y, GAO Q, JIANG Z H, et al, 2018. Bias correction and projection of surface air temperature in LMDZ multiple simulation over central and Eastern China[J]. Advances in Climate Change Research, 9(1): 81-92. DOI URL
[28]	JOSE D M, VINCENT A M, DWARAKISH G S, 2022. Improving multiple model ensemble predictions of daily precipitation and temperature through machine learning techniques[J]. Scientific Reports, 12(1):4678. DOI:10.1038/s41598-022-08786-w. PMID
[29]	KIM Y H, MIN S K, ZHANG X B, et al, 2020a. Evaluation of the CMIP6 multi-model ensemble for climate extreme indices[J]. Weather and Climate Extremes, 29:100269. DOI: 10.1016/j.wace.2020.100269.
[30]	KIM Y, ROCHETA E, EVANS J P, et al, 2020b. Impact of bias correction of regional climate model boundary conditions on the simulation of precipitation extremes[J]. Climate Dynamics, 55(11): 3 507-3 526. DOI
[31]	LENDERINK G, BUISHAND A, VAN DEURSEN W, 2007. Estimates of future discharges of the river Rhine using two scenario methodologies: Direct versus delta approach[J]. Hydrology and Earth System Sciences, 11(3):1145-1 159.
[32]	LI D W, YUAN J C, KOPP R E, 2020. Escalating global exposure to compound heat-humidity extremes with warming[J]. Environmental Research Letters, 15(6):064003. DOI:10.1088/1748-9326/ab7d04.
[33]	LI Z Y, FANG H Y, 2016. Impacts of climate change on water erosion: A review[J]. Earth-Science Reviews, 163: 94-117. DOI URL
[34]	MARAUN D, 2016. Bias correcting climate change simulations: A critical review[J]. Current Climate Change Reports, 2(4): 211-220. DOI URL
[35]	MEAURIO M, ZABALETA A, BOITHIAS L, et al, 2017. Assessing the hydrological response from an ensemble of CMIP5 climate projections in the transition zone of the Atlantic region (Bay of Biscay)[J]. Journal of Hydrology, 548: 46-62. DOI URL
[36]	MENAPACE A, DHAWAN P, DALLA TORRE D, et al, 2025. Review of bias correction methods for climate model outputs in hydrology[J]. Journal of Hydrology,660:133213. DOI: 10.1016/j.jhydrol.2025.133213.
[37]	NGOMA H, WEN W, AYUGI B, et al, 2021. Evaluation of precipitation simulations in CMIP6 models over Uganda[J]. International Journal of Climatology, 41(9): 4 743-4 768. DOI URL
[38]	OH S G, SUH M S, 2017. Comparison of projection skills of deterministic ensemble methods using pseudo-simulation data generated from multivariate Gaussian distribution[J]. Theoretical and Applied Climatology, 129(1): 243-262. DOI URL
[39]	PATEL J, GNANASEELAN C, CHOWDARY J S, et al, 2022. A quantile mapping approach-based bias correction in Coupled Model Intercomparison Project Phase 5 models for decadal temperature predictions over India[J]. International Journal of Climatology, 42(4): 2 455-2 469. DOI URL
[40]	RANASINGHE R, RUANE A C, VAUTARD R, et al, 2021. Climate change information for regional impact and for risk assessment[M]//MASSON-DELMOTTE V,ZHAI P,PIRANI A, et al. Climate Change 2021: The Physical Science Basis. Cambridge: Cambridge University Press: 1 767-1 926.
[41]	REICHLER T, KIM J, 2008. How well do coupled models simulate today’s climate?[J]. Bulletin of the American Meteorological Society, 89(3): 303-312. DOI URL
[42]	SCHILLING J, HERTIG E, TRAMBLAY Y, et al, 2020. Climate change vulnerability, water resources and social implications in North Africa[J]. Regional Environmental Change, 20(1):15. DOI: 10.1007/s10113-020-01597-7.
[43]	SEO G Y, AHN J B, 2023. Comparison of bias correction methods for summertime daily rainfall in South Korea using quantile mapping and machine learning model[J]. Atmosphere, 14(7): 1057. DOI: 10.3390/atmos14071057.
[44]	SILLMANN J, KHARIN V V, ZHANG X, et al, 2013. Climate extremes indices in the CMIP5 multimodel ensemble: Part 1. Model evaluation in the present climate[J]. Journal of Geophysical Research: Atmospheres, 118(4): 1 716-1 733. DOI URL
[45]	SPULER F R, WESSEL J B, JEBEILE J, et al, 2026. Bias adjustment and the question of usable climate information: Methodological assumptions and value judgements[J]. Bulletin of the American Meteorological Society, 107(1): E79-E102.
[46]	TAYLOR K E, 2001. Summarizing multiple aspects of model performance in a single diagram[J]. Journal of Geophysical Research: Atmospheres, 106(D7): 7 183-7 192. DOI URL
[47]	THEMEßL M J, GOBIET A, HEINRICH G, 2012. Empirical-statistical downscaling and error correction of regional climate models and its impact on the climate change signal[J]. Climatic Change, 112(2): 449-468. DOI URL
[48]	TEUTSCHBEIN C, SEIBERT J, 2012. Bias correction of regional climate model simulations for hydrological climate-change impact studies: Review and evaluation of different methods[J]. Journal of Hydrology, 456: 12-29.
[49]	TONG Y, GAO X J, HAN Z Y, et al, 2021. Bias correction of temperature and precipitation over China for RCM simulations using the QM and QDM methods[J]. Climate Dynamics, 57(5): 1 425-1 443. DOI
[50]	WANG B, ZHENG L H, LIU D L, et al, 2018. Using multi-model ensembles of CMIP5 global climate models to reproduce observed monthly rainfall and temperature with machine learning methods in Australia[J]. International Journal of Climatology, 38(13): 4 891-4 902. DOI URL
[51]	WANG D, LIU J H, SHAO W W, et al, 2021. Comparison of CMIP5 and CMIP6 multi-model ensemble for precipitation downscaling results and observational data: The case of Hanjiang River Basin[J]. Atmosphere, 12(7): 867. DOI: 10.3390/atmos12070867.
[52]	WANG D, LIU J H, WANG H, et al, 2022. Performance evaluations of CMIP6 and CMIP5 models for precipitation simulation over the Hanjiang River Basin, China[J]. Journal of Water and Climate Change, 13(5): 2 089-2 106. DOI URL
[53]	XU Y, GAO X J, SHEN Y, et al, 2009. A daily temperature dataset over China and its application in validating a RCM simulation[J]. Advances in Atmospheric Sciences, 26(4): 763-772. DOI URL
[54]	YUE Y L, YAN D, YUE Q, et al, 2021. Future changes in precipitation and temperature over the Yangtze River Basin in China based on CMIP6 GCMs[J]. Atmospheric Research, 264: 105828. DOI: 10.1016/j.atmosres.2021.105828.
[55]	ZHANG S X, HARROP B, LEUNG L R, et al, 2024. A machine learning bias correction on large-scale environment of high-impact weather systems in E3SM atmosphere model[J]. Journal of Advances in Modeling Earth Systems, 16(8): e2023MS004138. DOI: 10.1029/2023MS004138.
[56]	ZHANG X D, REN G Y, YANG Y D, et al, 2022. Extreme historical droughts and floods in the Hanjiang River Basin, China, since 1426[J]. Climate of the Past, 18(8): 1 775-1 796. DOI URL
[57]	ZHAO X, MA X W, CHEN B Y, et al, 2022. Challenges toward carbon neutrality in China: Strategies and countermeasures[J]. Resources,Conservation and Recycling,176: 105959. DOI: 10.1016/j.resconrec.2021.105959.
[58]	ZHENG H X, CHIEW F H S, CHARLES S, et al, 2018. Future climate and runoff projections across South Asia from CMIP5 global climate models and hydrological modelling[J]. Journal of Hydrology: Regional Studies, 18: 92-109. DOI URL
[59]	ZHU H H, YANG J N, 2026. Evaluation of extreme precipitation over East China in CMIP6 models[J]. Atmosphere, 17(2): 136. DOI: 10.3390/atmos17020136.