Terahertz frequency resolution is an important factor affecting substance identification, but the presence of the device reflection within the terahertz time domain spectroscopy systems, causes the presence of a plurality of reflection peaks in the reference and measurement signals with low frequency resolution, because of the length truncation of time domain signal. In order to remove the influence of the reflection peak, this paper proposes a method based on empirical mode decomposition to remove the time domain reflection peak, to improve the terahertz frequency resolution. The time domain reflection peaks are positioned by correlation with the real terahertz peak, calculating the reflection peak upper and lower envelope and an average, obtaining intrinsic mode functions, and alternating reflection peaks with intrinsic mode functions, the effective length of the time domain signal is increased to improve the terahertz frequency resolution. Water vapor in the air terahertz transmission spectroscopy results show that this method can self adaptively remove a plurality of reflection peaks and has good repairing effect on the terahertz time-domain signal; the frequency resolution is increased by 12 times; and the useful information of absorption spectrum is not lost; the absorption peak position and the number is consistent with real spectrum; the terahertz spectrum ability to identify is well preserved.