Magnetotelluric (MT) signal processing can increase the reliability of data. Traditional MT de-noising methods are usually filtered in entire MT time-series sequence, which result in losing of useful MT signals and the degradation of electromagnetic inversion imaging accuracy. However, targeted MT noise separation will retain the part of data that is not affected by strong noise, and enhance the quality of MT data. Thus, we proposed a novel method which using refined composite multiscale dispersion entropy (RCMDE) and orthogonal matching pursuit (OMP) for MT noise separation. Firstly, we extracted the RCMDE characteristic parameters from each segment of the MT time-series. Then, the characteristic parameters are input to the fuzzy c-mean (FCM) clustering for automatic signal-noise identification. Next, the identified noise segments were independently denoised by using OMP method. Finally, the reconstructed signal consists of the denoised data segments and the identified useful signal segments. We conducted the simulation experiments and algorithm evaluation on the EMTF data, simulated data and measured sites. The results indicate that a variety of entropy without scale factor is not stable enough, and the RCMDE can improve the stability of multiscale dispersion entropy (MDE) and multiscale entropy (MSE) by analyzing the characteristics of the signal samples library, and thus effectively dividing MT signals and noise. In this paper, the existing techniques of the entire time domain de-noising and signal-noise identification method are compared. Only for RCMDE and OMP as characteristic parameter and noise separation method, the proposed method simplified the multi-features fusion method, and improved the accuracy of signal-noise identification, accelerated de-noising efficiency, and improved the MT data quality of low-frequency band.