Magnetotelluric data denoising with recurrent neural network
Hang Chen, Rongwen Guo, Jianxin Liu, Yongfei Wang, Rongpei Lin · 2020
For the magnetotelluric (MT) method, the weak signal obtained can be easily affected by various kinds of noise. How to obtain the unbiased impedance estimation from measured data with complex noise is one focus of current research. In this paper, we design a recurrent neural network (RNN) algorithm with Long Short-Term Memory architecture to denoise MT data. We use synthetic data with errors generated from different stable distributions for training and testing. The network performs well in both the validation dataset and test dataset. Besides, the results show that the impedance estimates from the MT data denoised by RNN is much better than those without application of RNN using the traditional robust and least square methods, especially for MT data with strong noise.