Improved frequency-dependent bias correction method for GCM daily precipitation and its application in Yangtze River Basin
YUE Shuxu,1,2, HU Shi1, MO Xingguo,1,2, ZHAN Chesheng1, LIU Suxia1,21. Key Laboratory of Water Cycle and Related Land Surface Processes, Institute of Geographic Sciences and Natural Resources Research, CAS, Beijing 100101, China 2. College of Sino-Danish Center, University of Chinese Academy of Sciences, Beijing 100049, China
Abstract Bias correction of global climate model (GCM) outputs is essential for studies on the impact of climate change. Equiratio cumulative distribution functions matching (ERCDFm) method is a widely used bias correction method, and it has advantages in correcting future projections compared with the traditional Quantile Mapping method by preserving a consistent ratio between the observed and simulated values during reference and projection periods. However, the ability of modifying wet-day frequency would affect the performance of bias correction method. In this study, a newly developed frequency-dependent method was introduced into the ERCDFm to improve the simulation of precipitation days and total precipitation, which was achieved by complementing the insufficiency when the simulated number of precipitation days was underestimated, and adjusting the simulation of future wet-day frequency by preserving the trends in the raw GCM. The method was applied to the daily precipitation simulated by five GCMs from ISIMIP (Inter-Sectoral Impact Model Intercomparison Project) in both historical period and future RCP4.5 emission scenario over the Yangtze River Basin (YRB) using the gridded daily precipitation data (1961-2005) as observations. Results showed that the frequency-dependent ERCDFm correction method significantly improved the simulation with respect to wet-day frequency and mean precipitation. Compared to the original ERCDFm, the spatial correlation coefficients (CORs) between the corrected and observed wet-day frequency increased by 140% in spring, 85% in summer, 19% in autum, and 21% in winter by using the improved frequency-dependent ERCDFm method; the RMSE between the corrected and observed wet-day frequency and total precipitation reduced by 83% and 58%, respectively; and the area percentage of the precipitation biases within 50 mm/a increased from 31% to 49% over the YRB. In particular, the improved ERCDFm could alleviate the underestimation of total precipitation caused by the underestimated wet-day frequency. The bias-corrected GCM projections (2030-2050) of the ensemble mean indicated that the annual precipitation is expected to increase by 6.1% over the YRB under the RCP4.5 scenario relative to 1986-2005, with seasonal precipitation increasing by 8.2% in spring, 6.4% in summer, 4.7% in autumn, and 0.7% in winter. It is worth noting that the contribution of the change in wet-day frequency is of great importance to the total precipitation trend; therefore it is critical to retain the long-term change signal of wet-day frequency in the bias correction of daily precipitation. Keywords:bias correction;ERCDFm;daily precipitation;wet-day frequency;RCP4.5 emission scenario
PDF (5082KB)元数据多维度评价相关文章导出EndNote|Ris|Bibtex收藏本文 本文引用格式 岳书旭, 胡实, 莫兴国, 占车生, 刘苏峡. 一种基于频率的GCM日降水偏差校正方法改进及其在长江流域的应用. 地理研究[J], 2021, 40(5): 1432-1444 doi:10.11821/dlyj020200446 YUE Shuxu, HU Shi, MO Xingguo, ZHAN Chesheng, LIU Suxia. Improved frequency-dependent bias correction method for GCM daily precipitation and its application in Yangtze River Basin. Geographical Research[J], 2021, 40(5): 1432-1444 doi:10.11821/dlyj020200446
偏差校正是基于当前数据得到的统计关系在未来气候变化情景下依然有效的假设,采用统计学的方法降低模拟值与观测值间的偏差。目前常用的降水偏差校正方法可分为两类,一类基于降水量的均值和方差进行校正,另一类主要基于降水量的概率分布。基于降水量均值和方差的校正方法具有简单易行的特点,适用于对平均态的校正,较难应用于日尺度,例如线性缩放[1]、方差比例变换[2]、delta变换[3]等。基于概率分布的校正方法,在校正降水量的均值与方差的同时,还对降雨的累积分布函数进行了修正,逐渐得到了重视与发展。分位数映射法(Quantile Mapping,QM)[4,5,6,7]是典型的概率分布校正方法,该方法假设在长时间序列内的降水会服从一个相对稳定的概率分布,且模拟降水的概率分布应与观测降水一致。概率分布校正的优势,使其能够在特定情景的误差订正中发挥作用,例如作为单变量QM法扩展的温度-降水联合校正可以保留双变量间的相关性[8]、按降水强度分段建立传递函数能显著改善极端降水的校正效果[9]、考虑空间分布特征的BCSA(Bias-Correction and Stochastic Analog method)校正方法可应用于日降水的降尺度中[10]等。
转移累计概率分布法(Cumulative Distribution Function transform,CDF-t)[11,12]和等距分布映射法(Equidistant cumulative distribution functions matching,EDCDFm)[13,14]是在QM方法的基础上发展而来的,弥补了QM在校正未来预估数据中的不足。EDCDFm方法保留了模拟与观测数据间的秩相关关系,使相同分位数下历史模拟数据与未来预估数据之间的偏差一致。Wang等[15]将EDCDFm方法中未来预估数据的偏差计算公式的形式由等距转为等比,解决了降水校正结果存在负值的不足,称为等比分布映射法(Equiratio cumulative distribution functions matching,ERCDFm)。目前,ERCDFm与EDCDFm方法在校正日降水时仍存在降水日数模拟不准确的问题,仅通过设定阈值来剔除过多的小雨日数,无法有效校正GCM降水日数偏少的现象[14],从而影响降水量的校正效果。此外,随着全球升温极端降水加剧,降水日数也有减小趋势[16,17]。考虑到模式捕捉气候自然变率的能力不同,使用固定阈值会导致降水日数模拟值的误差在未来逐渐增大,因此需要一种考虑未来预估期下降水日数动态变化的校正方法。
日降水模式数据来自跨行业影响模式比较计划ISIMIP(Inter-Sectoral Impact Model Intercomparison Project)提供的5个气候模式(表1)逐日降水资料(https://www.isimip.org/protocol/#isimip2b),包括历史气候模拟(1961—2005年)与不同排放情景的未来预估实验(2006—2099年),空间分辨率为0.5°×0.5°。基于模式数据历史期的长度,选取中国气象数据网1961—2005年的格点化日降水资料作为观测数据(http://data.cma.cn/),该数据是基于2472个国家级气象观测站的降水资料,利用薄盘样条法空间插值生成的0.5°×0.5°的日降水格点数据。
Tab. 1 表1 表1GCM模式详细资料 Tab. 1Information on the GCMs used in this study
模式编号
模式
机构
1
GFDL-ESM2M
Geophysical Fluid Dynamics Laboratory
2
HadGEM2-ES
Met Office Hadley Centre
3
IPSL-CM5A-LR
Institute Pierre-Simon Laplace
4
MIROC- ESM-CHEM
Japan Agency for Marine-Earth Science and Technology, Atmosphere and Ocean Research Institute, and National Institute for Environmental Studies
5
NorESM1-M
Norwegian Climate Centre
注:表中的5个模式是ISIMIP为进行气候变化影响研究,从世界耦合模式比较计划第五阶段(Coupled Model Intercomparison Project Phase 5,CMIP5)中选出的全球气候模式,见文献[20]。
(2)概率分布拟合校正的目的是使模拟的降水序列累积分布函数(Cumulative Distribution Function,CDF)与观测数据尽可能接近。通常采用Gamma分布描述降水序列的CDF,如公式(1)所示。本文使用MCMC(Markov Chain Monte Carlo)方法对参数 进行拟合,再分别对历史期(公式(2))、未来预估期(公式(3))两个时段的模式数据进行概率分布拟合校正[13,14,15]:
Fig. 3Spatial distribution of the observed and ensemble averaged annual precipitation and wet-day frequency over Yangtze River Basin during training period (1961-1985) and its relative change from training period to validation period (1986-2005)
注:每个灰色圆圈代表长江流域内的一个格点。 Fig. 4Comparisons of annual precipitation and wet-day frequency during validation period (1986-2005) for observations and the ensemble mean of the original simulations, ERCDFm bias correction, frequency-dependent ERCDFm bias correction
Fig. 6Probability density distribution of the bias of annual precipitation during validation period (1986-2005) for the ensemble mean of the original simulation, ERCDFm bias-correction, and frequency-dependent ERCDFm bias-correction
Fig. 7Spatial correlation coefficients of the original simulated, ERCDFm bias-corrected, frequency-dependent ERCDFm bias-corrected seasonal precipitation and wet-day frequency with observations during validation period (1986-2005)
[ ZhouLi, JiangZhihong. Future changes in precipitation over Hunan province based on CMIP5 simulations using the statistical downscaling method of transform cumulative distribution function Acta Meteor Sinica, 2017,75(2):223-235.] DOI: http://www.dlyj.ac.cn/article/2021/1000-0585/10.11676/qxxb2017.012. URL [本文引用: 1]
LiH, SheffieldJ, Wood EF. Bias correction of monthly precipitation and temperature fields from Intergovernmental Panel on Climate Change AR4 models using equidistant quantile matching Journal of Geophysical Research-Atmospheres, 2010,115(D10101). DOI: http://www.dlyj.ac.cn/article/2021/1000-0585/10.1029/2009JD012882. URL [本文引用: 2]
[ GuanYinghui. Extreme climate change and its trend prediction in the yangtze river basin Yangling: Doctoral Dissertation of Northwest University, 2015: 14.] [本文引用: 1]