As a new groundwater exploration method, noninvasive surface nuclear magnetic resonance (SNMR) has the benefits of direct, quantitative and uniqueness estimation of water content and relaxation time (T2*) in the near surface groundwater exploration. In practice, the earth magnetic field is difficult to be determined accurately, due to its inhomogeneity, time-varying and susceptible to ambient noise, which results in off-resonance excitation and serious decrease in accuracy of the inversion result. In this paper, based on the model of surface nuclear magnetic off-resonance (SNMOR) and the expression for the kernel function, the influences of the frequency offset on the amplitude and phase of the free induced decay (FID) signal are discussed, and a complex envelop inversion (CEI) based on automatic matching system phase and involving both the real part and imaginary part of the signal is applied. By comparing the synthetic signals generated from the SNMR and SNMOR models, it can be concluded that the phase of the FID signal significantly changes with the increase of the frequency offset, and the amplitude of the signal can be increased by 65.9% for the synthetic model in this paper. Thus when the frequency offset is greater than 2 Hz, the distribution of water content and T2* from the inversion results using the SNMR kernel will have a serious deviation from the actual model. However, using the SNMOR kernel based on the frequency offset, the inversion results are more accurate, and the maximum error of the water content and T2* are 4.2% and 39.3 ms, respectively. Moreover, synthetic data with different noise levels are inverted by the CEI method and conventional amplitude envelop inversion method (or QTI). The results show CEI obtain better performances in stability and reliability at a high noise level. Finally, a field measurement of SNMOR is conducted in Taipingchi Reservoir near Changchun City, China. The off-resonance FID signals are obviously observed by utilizing the JLMRS instrument and can be used to estimate the frequency offset. The characteristics of the FID signal with the frequency offset confirm the correctness of the SNMOR model. And the inversion result of field data using SNMOR kernel show that the distribution of water content and T2* are consistent with the known geological data from the drillings and other geophysical methods, which is much better than that using the SNMR kernel or conventional amplitude envelop inversion method. Therefore, the validities and accuracies of the SNMOR model and CEI method proposed in this paper are verified, which provides a new idea and technique for groundwater exploration in the near surface.