Data missing is frequently encountered in climate variables due to many reasons, such as instrument failures in the observatory, meteorological extremes, and observation recording errors. However, several types of climatic analysis require the availability of data not only covering a long enough period of time, but also forming a complete and homogeneous series. This paper presented a novel imputation method for missing air temperature records by merging their spatio-temporal characteristics. On the basis of extending Kriging model, a nonstationary Kriging method which assumes that the mean is known and varying in study area was proposed. Firstly, the trend of air temperature in each station was attained by analyzing its time series data, and linear interpolation was adopted in this study. Then, geostatistical analysis were performed on the errors between the trend and observed values. Finally, the spatio-temporal information of air temperature was integrated into the proposed Kriging model. Three other imputation methods, including linear interpolation, ordinary Kriging based on DEM (OKD) and normal ratio, were introduced to compare with. The results show that: 1) Besides OKD, the imputation accuracy of the other three methods varies obviously in 12 months. For linear interpolation, its imputation accuracy in May and July-October is much higher than that in the rest of the month. Normal ratio has higher imputation accuracy in April-November. The proposed method has higher imputation accuracy in March-October, with mean absolute error (MAE) less than 0.2℃. 2) Normal ratio has the largest MAE (4.17℃) in December and the least MAE (0.18℃) in October, this means that it has poor robustness. Compared with linear interpolation, the difference between the maximum and minimum MAE values of OKD is much less (0.25℃), thus it has better robustness. With the difference being 0.1℃ only, the proposed method has the strongest robustness. 3) Air temperature contains the temporal and spatial characteristics together. Linear interpolation only considers its temporal characteristics but ignores its spatial characteristics, while OKD only considers its spatial characteristic but ignores its temporal characteristics. Therefore, they don't attain the satisfactory imputation results. With partly taking the spatio-temporal characteristics of air temperature into account, normal ratio can attain higher imputation accuracy in March-November. However, this method has poor robustness. When air temperature in study area varies sharply or fluctuates around 0℃, normal ratio has lower imputation accuracy. As a result, its overall imputation accuracy is still lower. Among these methods, the proposed method has the smallest MAE and root mean square error in each month and produces the best imputation results.