Deep Learning Approach to 2D Capacitive Resistivity Imaging Inversion

Abstract

This study developed deep learning models to perform the inversion of capacitive resistivity imaging (CRI) data to acquire 2D subsurface resistivity maps. Inversion of resistivity data using deep learning models had been applied in previous works where it addresses the time-consuming process and inaccurate generation of maps obtained from the conventional iterative process of inversion. The present study contributes to state of the art by training and evaluating the deep learning inversion model dedicated for CRI, which previous works were limited for use with DC ERT surveys. A synthetic dataset of several subsurface resistivity models and expected CRI equipment recordings were generated. Then four deep learning architectures, namely U-Net, SegNet, Wavelet U-Net, and Wavelet SegNet, were trained and evaluated using the synthetic dataset. Results revealed that the Wavelet U-Net performed best among the considered architectures as it had the lowest mean square error, second highest R2, and a more accurate resistivity map generated from a sample in the dataset. All deep learning inversion models produced one sample resistivity map, with a survey length of 100 m, within 1 s compared with the Res2dinv software (iterative method), which took 40 s. Overall, the study concludes that the trained deep learning 2D inversion models, especially Wavelet U-Net, can invert CRI data to a high degree of accuracy at a fast computational speed, which becomes essential when requiring real-time and on-site results and dealing with long surveys.