LAUSR

Abstract This work presents a systematic study for testing the effectiveness of Self-Organizing Map (SOM) neural networks in filtering magnetotelluric (MT) data affected by cultural noise. Although the MT method is widely used for geophysical investigation of the Earth’s interior, it is very sensitive to anthropogenic noise sources (e.g., power lines, electric railways, etc.), which can generate transient artificial electromagnetic fields disturbing the MT records. Thus, when not properly detected, man-made noises could lead to a distortion of the MT impedance tensors and consequently to wrong estimate of the resulting subsoil resistivity distribution. The choice to use SOM networks to filter noisy MT data comes from the expectation that the impedance tensors, estimated by Discrete Wavelet Transform analysis of MT time series, will cluster differently in presence of noise. This expectation is confirmed by the results of our extensive study on synthetic MT signals affected by temporally localized noise, which show that noisy and noise-free impedance tensor values distribute in well separate clusters. Moreover, as the SOM analysis provides a grid of weights (clusters), each one close to a particular subset of the input data, a criterion is proposed for selecting the cluster that gives the most reliable impedance tensor estimate. An application of the proposed SOM-based filtering procedure to actual MT data demonstrates its efficiency in denoising real MT signals.