Various schemes for processing magnetotelluric (MT) data have been reported aiming at suppressing the strong effect of artificial electromagnetic noises, especially coherent noise that is correlated between electric and magnetic time series. Many of the recent denoising schemes are based on decomposing MT data into the responses of the natural signals and noises. However, a steady differentiation of the natural signals from noises independent of any empirical choice of parameter setting is critically important. In addition, improper subtraction of values from the separated signals can lead to losing useful values of the natural signals or missing noise-affected values, which may result in failure in deriving the true MT responses. We propose a novel data-processing method that applies frequency-domain independent component analysis (FDICA) to both the local MT data and the reference magnetic data. Among the separated signals, the proposed method can quantitatively distinguish the natural signals from the noise-affected components by calculating the ratio of the cross-power spectrum with the reference data to the auto-power spectrum for each component. When determining which values to subtract from the separated signals, we introduce an evaluation index with respect to two characteristics of the MT response function: stationarity in the time domain and smoothness in the frequency domain. We conducted two types of experiments: with MT time series severely contaminated by synthetic coherent noises and with MT field data interfered with direct-current (DC) railways. Consequently, we confirmed the noise-suppression performance superiority of the proposed method over the conventional methods of MT data processing.