Air temperature is an important climatological variable and is usually measured in meteorological stations. Accurate mapping of its spatial and temporal distribution is of great interest for various scientific disciplines, but low station density and complexity of the terrain usually lead to significant errors and unrepresentative spatial patterns over large areas. Fortunately the current studies have shown that the regression models can help overcome the problem with the help of time series remote sensing data. However, noise induced by cloud contamination and other atmospheric disturbances variability impedes the application of LST data. An improved Savizky-Golay (SG) algorithm based on the LST background library is used in this paper to reconstruct MODIS LST product. Data statistical analysis included 12 meteorological stations and 120 reconstructed MODIS LST images of the period from 2001 to 2010. The coefficient of correlations (R²) for 80% of the stations was higher than 0.5 (below 0.5 for only 2 stations) which illustrated that there is a considerably close agreement between monthly mean TA (air temperature) and the reconstructed LST in the Lancang River basin. Comparing to the regression model for every month with only LST data, the regression model with LST and NDVI had higher R² and RMSE. Finally, the LST-NDVI regression method was applied as an estimate model to produce distributed maps of air temperature with month intervals and 1 km spatial in the Lancang River basin of 2010.