To evaluate the noise during atmosphere observation, the imaginary part of the atmospheric complex Fourier transform spectra are taken into consideration. However, the imaginary spectrally correlated noise introduced by sampling jitters would be added to the random noise inherent to infrared detectors, which elevates the total instrumental noise floor or even exceeds the sensibility threshold. Utilizing principal component analysis technique, this correlative noise could be reconstructed and filtered out. Then the remaining noise is represented as the noise equivalent differential temperature and compared with that from the calibration target radiance. The results show that the random spectra noises from different scenes are consistent with each other, and all meet the sensibility requirments of 0.4 K, 0.7 K and 1.2 K corresponding to three spectral bands.