Intelligent Inversion of Magnetic Resonance Sounding Data Using Convolutional Neural Network

Magnetic resonance sounding (MRS) is a geophysical method for direct detection of groundwater and quantitative assessment of the hydrological characteristics of aquifers, which is mainly applied to regional hydrogeological investigations and prediction of groundwater-related geological disasters. Currently, the main inversion method of MRS data is the Q-time (QT) inversion method based on the regularization, which has achieved certain application results. However, it is highly dependent on the initial model which is usually difficult to obtain when dealing with the field data. Besides, there are other limitations such as time consuming and high computational power requirements. To solve these problems, we propose a novel MRS inversion method based on deep learning using convolutional neural networks (CNN). The network is designed to establish a mapping relationship between the MRS data and the parameters of the underground aquifer. Diverse and large datasets are generated from MRS forward modeling methods, so as to train an inversion network with strong generalization ability. Compared with the inversion result of the QT inversion method, the proposed method can get more accurate performance and take less than 1 s each inversion. This paper also discusses the generalization ability analysis of the methods by applying the trained network on the datasets with different subsurface aquifer parameters and varied noise levels.