It is important to suitably estimate the root mean square of observation error i n a given assimilation system. Firstly, by using Bessel fitting method, the root mean square of observation error and root mean square of first guess error are separated from the covariance of the deviation of radiosounding observation an d first guess. Then the root mean square of other observation error for other ob servation types are estimated by deducting the root mean square of first guess e rror from the covariance of the deviation of radiosounding observation and fir st guess. The observation data during 1-31, August 2006 and the first guess from T213 numerical weather prediction system are used for analyzing the root mean s quare of observation error. The results show that the estimated root mean square of observation error for each observation type is reasonable.