Multitype Noise Suppression in Magnetic Resonance Sounding Data Based on a Time–Frequency Fully Convolutional Neural Network

Abstract

Magnetic resonance sounding (MRS) is one of the technical applications of nuclear magnetic resonance (NMR) used to directly detect and quantify groundwater content. MRS suffers from a low signal-to-noise ratio (SNR) due to the low amplitude of free induction decay (FID) signals and an inability to shield environmental noise. In this paper, a time-frequency fully convolutional neural network (TFCN) was proposed to suppress random, harmonic, and spike noise from MRS data. The TFCN parameters were trained with the time-frequency spectrum obtained by the short-time Fourier transform (STFT) of the MRS datasets as the input and the noise-free FID signals as the output. Based on the results of synthetic and field data experiments, the TFCN was compared with existing denoising methods. The results showed that the TFCN extracted the envelope of the FID signals from low-SNR random noise with higher accuracy than other methods. Moreover, the TFCN simultaneously suppressed multiple types of noise and exhibited high computational efficiency.