To evaluate the reliability of the atmospheric temperature profiles retrieved by the microwave radiometer (MWR) and the impact of sounding balloon drift on the temperature differences (δT) between MWR and radiosonde, more than two years of temperature profiles derived from MWR of Xi’an Jinghe Station in Shaanxi Province are tested against those from radiosondes. A method for quantitative assessment of bias in radiosonde temperature caused by the drift of sounding balloon is proposed. It is demonstrated that there are significant spatial and temporal changes in the consistency of the temperatures acquired by MWR (TM) and radiosonde (TS). The correlation coefficients between TM and TS are found to be higher in spring and autumn than in summer and winter, and better correlation is always detected at lower altitudes in all seasons. At a given altitude, the δT is 〖JP2〗significantly negatively correlated to the ambient temperature represented by TS. The higher the temperature, the more obvious the underestimation of the temperature by the MWR, and vice versa. This leads to a phenomenon of “overestimating low temperature and underestimating high temperature”, which is more obvious at high altitudes than at low altitudes, and more obvious in winter and summer than in spring and autumn. The drift of sounding balloon causes the sounding temperature to be higher than the actual temperature right above the radiosonde sites in all seasons, and the biases in radiosonde temperature records are negatively correlated to the ambient temperature. Therefore, sounding balloon drift causes mean deviation (MD) to be overestimated, and the degree of dispersion in δT to be underestimated, resulting in underestimation of the correlation between δT and ambient temperature as well as the severity of “overestimating low temperature and underestimating high temperature” of MWR. Overall, the influence of sounding balloon drift on the standard deviation of δT and root mean square deviation (RMSD) of the microwave radiometer and radiosonde temperature is below 2%, and its contribution to δT is much smaller than that of the retrieval algorithm.