Error structural characteristic of observation data is one of the most important factors that affect data assimilation. The error structural characteristics of observation data in GRAPES come from numerical weather prediction system abroad. In order to improve the application effect of aircraft meteorological data relay (AMDAR) data, statistical analysis of the temperature difference between AMDAR data and NCEP reanalysis data is done according to different latitude zones, flight states, and identifications. The AMDAR data is from realtime database in National Meteorological Information Centre (NMIC), and the period of time for the data is May-July 2013. The statistical results show that there is a maximum temperature bias in the midlatitude (20°-50°N) of Northern Hemisphere, where the number of AMDAR data is at most. The temperature bias is about -2--1℃, and the standard deviation of temperature is about 1.2-1.6℃. The temperature standard deviation varies with latitude zones in different flight states. The minimum standard deviation of temperature appears in cruise phase in the equatorial region (20°S-50°N), being about 0.8℃. The maximum standard deviation of temperature appears in cruise phase in the high latitude (50°-90°N) of the Northern Hemisphere, and it is about 1.5-2℃. By analyzing the error character of AMDAR data, we can do better in quality control and bias correction, improving the effect of data assimilation and forecasting.