SOM clustering analysis in the discrete wavelet transform domain for filtering noisy magnetotelluric data

Despite the magnetotelluric method (MT) is one of the most prominent geophysical technique for deep subsoil exploration, it is not yet completely reliable when applied in urban or industrialized areas due to the presence of anthropic electromagnetic noise. The latter, indeed, may severely affect the MT recordings and, as a consequence, the impedance tensor estimates, which allow to retrieve the apparent resistivity values describing the underground electrical behaviour. In this work, a new filtering approach for MT data denoising is proposed. The procedure is based on the clustering of the impedance tensor estimates by using the Self-Organizing Map (SOM) neural network model. The clustering is performed in the time-frequency domain by a discrete wavelet transformation of the MT signals. In addition, a criterion for selecting, in each wavelet scale, the clusters that lead to the most reliable apparent resistivity estimates has been suggested. The application of the proposed filtering approach to synthetic MT signals has shown that the SOM clustering is very sensitive to the presence of noise and that it is possible to get consistent apparent resistivity curves.