Magnetotelluric Time Series Denoising Using Encoder-Decoder Consisted of LSTM Cells
Sihao Wang, Lanfang He, Xuben Wang, Liang Li · Journal of Physics Conference Series · 2023
Abstract Electromagnetic signals in geophysics are frequently disturbed by various interference in field data acquisition. Denoising for passive electromagnetic methods such as magnetotelluric (MT) or audio magnetotelluric (AMT) data is significant to improve data quality and finally imaging to the geoelectrical structure. Conventionally, most denoising methods are employed in frequency domain and few of them are applied in time domain. However, a great number of irregular noise in the electromagnetic time series prove difficulty to be removed. We propose a denoising method, using Encoder-Decoder consisted of Long Short-Term Memory cells (ED-LSTM), to reduce the effect of the step noise and the random-impulsive noise. Supervised learning and transductive learning are used for the denoising of the step noise and the random-impulsive noise, respectively. Our results indicate that the step and random-impulsive noise could be successfully removed from the raw time series. The result indicate that ED-LSTM could potentially to be wildly used in the electromagnetic time series denoising and then improve data quality.