Application of Generative Adversarial Networks in Geoelectrical Field Data Interpretation

The geoelectrical survey method generates subsurface cross-sections images based on physical properties, but requires solution of an inverse problem with potential ambiguities in model interpretation and substructure uncertainties. The integration of electrical resistivity tomography (ERT) and machine learning computational approaches is adopted in this study to reduce these ambiguities and uncertainties as well as the labour-intensive nature of the standard computational approach. A data set collected from landfill locations in Nigeria is inverted by the conventional geophysical method of interpretation using the RES2DINV software. The inverted data (true resistivity tomography images) along with the source data (apparent resistivity images) are used as training samples to develop predictor models based on the Pix2Pix conditional generative adversarial networks (Pix2Pix-cGAN). Initial results with a small number of training samples reveal about 89% structural similarity between the true resistivity topography obtained by the standard inversion method and those predicted by the Pix2Pix translator. Our novel approach can be applied to the analysis of seismic data from the lithosphere and mantle.