In this study,based on the reforecast data of NCEP Global Ensemble Forecast System(GEFS) and China homogeneous grid-point observational data,features of extreme temperature variations in the past 30 years are analyzed,and the performance of the NCEP-GEFS in representing these kinds of features is thoroughly investigated.By estimating the historical climatic percentile of 2 m temperature in the observational and model data,the characteristics of extreme temperature in winter and summer,along with the spatial distribution and multi-year trend of extreme temperature days,are analyzed.The results show that some strong regional features exist in the spatial distributions of winter extreme low temperature(ELT) and summer extreme high temperature(EHT) in China,i.e.there are relatively lower temperatures corresponding to the percentile thresholds of the winter ELT in northeastern China,northern China and the Qinghai-Tibet Plateau,with higher temperatures corresponding to the percentile thresholds of the summer EHT in southern China,northwestern China and the Yangtze River Basin.Both the summer mean temperature and EHT days throughout China show increasing trends in the past 30 years,and the winter mean temperatures are also increasing throughout most of China,yet decreasing in northwestern and northeastern China.Correspondingly,the numbers of days of the winter ELT are decreasing in most areas,and only slightly increasing in small parts of northwestern,northeastern and southern China.The NCEP-GEFS reforecasts are able to accurately reproduce the climatic trends and interannual variations of the seasonal mean temperature and extreme temperature days in the winter and summer of China,yet varying degrees of cold biases exist in the different regions.The biases in winter are significantly larger than those in summer,and as the forecast length increases,these cold biases are gradually strengthened in winter,while gradually weakened in summer.Therefore,it is suggested to adopt the relative definition of extreme temperature based on the percentile threshold,which can automatically correct these systematic biases in the model analysis and prediction products.