中文摘要：

利用1961-2002年ERA-40逐日再分析资料和江淮流域56个台站逐日观测降水量资料,引入基于自组织映射神经网络（Self-Organizing Maps,简称SOM）的统计降尺度方法,对江淮流域夏季（6-8月）逐日降水量进行统计建模与验证,以考察SOM对中国东部季风降水和极端降水的统计降尺度模拟能力。结果表明,SOM通过建立主要天气型与局地降水的条件转换关系,能够再现与观测一致的日降水量概率分布特征,所有台站基于概率分布函数的Brier评分（Brier Score）均近似为0,显著性评分（Significance Score）全部在0.8以上;模拟的多年平均降水日数、中雨日数、夏季总降水量、日降水强度、极端降水阈值和极端降水贡献率区域平均的偏差都低于11%;并且能够在一定程度上模拟出江淮流域夏季降水的时间变率。进一步将SOM降尺度模型应用到BCCCSM1.1（m）模式当前气候情景下,评估其对耦合模式模拟结果的改善能力。发现降尺度显著改善了模式对极端降水模拟偏弱的缺陷,对不同降水指数的模拟较BCC-CSM1.1（m）模式显著提高,降尺度后所有台站6个降水指数的相对误差百分率基本在20%以内,偏差比降尺度前减小了40%-60%;降尺度后6个降水指数气候场的空间相关系数提高到0.9,相对标准差均接近1.0,并且均方根误差在0.5以下。表明SOM降尺度方法显著提高日降水概率分布,特别是概率分布曲线尾部特征的模拟能力,极大改善了模式对极端降水场的模拟能力,为提高未来预估能力提供了基础。

英文摘要：

Based on the ERA-40 daily reanalysis data from 1961–2002 and observed daily precipitation data at 56 meteorological stations located in the Yangtze–Huaihe River basin, this study applies a new downscaling method based on Self-Organizing Maps（SOMs） to produce downscaled summer precipitation estimates at each station. The simulation capability of the statistical downscaling approach for monsoon precipitation and extreme precipitation over East China have been assessed. The downscaling model is then applied to simulate daily precipitation at the same 56 stations for the period 1986–2005 using predictor sets simulated by BCC-CSM1.1（m）（Climate System Model of the Beijing Climate Center）. Results show that the downscaling approach can realistically reproduce the observed probability distribution and temporal variability of precipitation. The Brier scores are almost zero and the significance scores are above 0.8 for all stations. Average biases of the downscaled number of days with precipitation greater than 1 mm and 10 mm, the summer total precipitation, the simple daily intensity index, the extreme daily precipitation threshold, and the fraction of total precipitation due to events exceeding the 95 th percentile of the climatological wet-day precipitation distribution all are below 11%. Furthermore, the downscaling approach is, to a certain extent, able to reproduce the temporal variability characteristics of precipitation. Compared with that for the raw outputs of BCC-CSM1.1（m）, the biases of the above indices for the downscaled results reduce by 40% to 60%. The spatial correlation coefficients increase to 0.9, and the root mean square errors are below 0.5. Overall, the downscaling model significantly improves the simulation of the probability distribution of daily precipitation, particularly the simulation of extreme precipitation. Thereby it can be applied for the projection of future precipitation changes.