Noise power estimating is the core of modern radio monitoring systems for solving tasks of spectrum occupancy calculation, detecting and estimating signal parameters. The growth of electronic devices number leads to an increase in overall noise level and its fast fluctuations. These devices often emit pulses or separate carriers. Since radio monitoring equipment must operate under these conditions, it may not be possible to exclude these components from radio noise measurements. It was shown that in some cases an increase in the noise power by 20% of the expected value leads to an increase in the false alarm rate by an order. The aim of this work is to develop and explore an iterative method for estimating the noise power with an unknown occupancy of the analysis frequency band, which will have low computational complexity and estimates independent of spectrum occupancy. The essence of the proposed method consists in two-threshold division of frequency samples into signal and noise by a statistical criterion using the coefficient of variation of spectral estimates. Thresholds are selected for a given false alarm rate. When threshold value of the coefficient of variation is exceeded, it is considered that there are occupied frequency channels in the spectrum, and each frequency sample is compared with the second threshold. Those samples that have exceeded the threshold are considered signal, and the rest -- noise. The described procedure is then repeated for noise samples until all signal samples have been discarded. Also was developed method for calculating the noise power in time domain using the obtained noise power in frequency domain. Algorithm evaluation has shown that it remains robust for spectrum occupancy up to 60%. In this case, the relative error in estimating the noise power does not exceed 5%, and the average number of iterations of the algorithm grows with increasing occupancy and does not exceed 10.