Data Reconstruction of Resistivity Method Based on U-Net Convolutional Neural Network

Abstract In the acquisition of field data using the geophysical direct current (DC) resistivity method, interference from complex terrain and natural noise often causes abnormal perturbations in the measured apparent resistivity data at survey points. Considering that the governing equation for the DC resistivity method is essentially a nonlinear Poisson equation, reconstructing the apparent resistivity data through linear interpolation for inversion will reduce the accuracy of the results. To address this problem, this paper develops a data reconstruction algorithm for DC resistivity data based on convolutional neural networks (CNNs). Specifically, a U-Net deep neural network architecture is constructed, and a training dataset is generated using three-dimensional finite-difference forward modeling simulations. After network optimization and training, a data reconstruction model is established. Experiments on synthetic data demonstrate that inverting the apparent resistivity data reconstructed by the neural network produces resistivity models with significantly higher accuracy compared to those obtained from data based on linear interpolation. Field-data results demonstrate that the reconstructed data obtained after processing with the convolutional neural network (CNN) significantly attenuates the various types of noise generated during field acquisition. This confirms that the proposed method effectively preserves the characteristics of the nonlinear field distribution. This method not only provides a new solution for data reconstruction in DC resistivity data, but its technical framework can also be extended to geophysical exploration fields involving multiple physical methods, such as induced polarization and electromagnetic techniques.