Abstract: Based on daily average temperature, relative humidity, sunshine hours and diurnal temperature range at 662 meteorological stations in China from 2004 to 2013, the method of singular value decomposition is employed to estimate missing meteorological data. To reduce the negative influence of randomness, the above 10 years of daily meteorological data are averaged. Both the relative error of singular value decomposition and the similarity matrix are adopted to verify the approximate low-rankness of meteorological data, and the correlations between different meteorological elements are discussed. After expounding the principal base vectors, three groups of estimation experiments are designed. The first group randomly chooses data at 6 meteorological stations for testing, and the data at those remaining stations are used for training to obtain five best base vectors. For each testing station, 12 observations are stochastically selected and other unknown elements are estimated according to the chosen base vectors. The means of estimation errors of average temperature, relative humidity, sunshine hours and diurnal temperature range are 8.00%, 7.83%, 17.17% and 10.82%, respectively. In the second group of experiments, 70% of the meteorological stations are randomly selected for training, the remaining for validating the estimation performance. The experimental results show that the number of base vectors can be chosen in the range from 5 to 15, and the estimation performance can be improved with the increase of the number of base vectors or the number of observations. The third group of experiments estimates the unobserved meteorological data at 10 stations in the first half of 1952 according to the corresponding observed data in the second half, and the estimation results are reasonable and reliable.