Abstract: Magnetic resonance sounding is currently the only direct geophysical method of groundwater exploration. The instrument of magnetic resonance sounding has high sensitivity for detecting and receiving performance, and can receive nanovolt (1nV= V) NMR signals, but is susceptible to environmental electromagnetic noise. This restrains the accuracy of the magnetic resonance sounding interpretation results which cannot be reversed to the actual underground aquifer structure. This study proposes an all-phase model estimation method to suppress the power frequency harmonic noise with the largest specific gravity in electromagnetic noise. High-precision harmonic modeling is the key to modeling cancellation. This method uses two-stage noisy NMR signals with time-shift relationship to perform all-phase FFT processing. The phase value of the main line is taken as the phase estimation. The phase difference of the main line can be used to obtain high-precision frequency estimation and amplitude estimation. The all-phase FFT has good frequency leakage suppression characteristics compared with the conventional FFT. Through simulation experiments, we find that the all-phase model estimation method can obtain the power frequency harmonic parameters more accurately than the traditional model denoising method, and the precision is more than doubled. Through the test of the measured data, we find that the all-phase model estimation method can better cope with the complex and variable conditions of the actual noise, effectively suppress the interference of the power frequency harmonics, and improve the accuracy of the MRS method inversion interpretation.