Deep Learning for Subsurface Conductivity Characterization and Contamination Assessment: a Colorado Wetland Case Study

Acid Mine Drainage (AMD) is a persistent environmental issue in the Western United States (WUS), causing subsurface contamination that makes land and groundwater unfit for human use. Advances in computer processing and geophysical data have enabled Machine Learning (ML)-based inversion algorithms to automatically reconstruct ground electrical properties, such as Relative Dielectric Permittivity, Resistivity, and Conductivity, for contamination assessment. This study has used a newly developed DeepLabv3+ architecture-based Deep Convolutional Neural Network (DCNN) for finding the subsurface distribution of electrical conductivity from Frequency Domain Electromagnetic (FDEM) Induction data acquired along and around the AMD-impacted Cement Creek California Gulch near Silverton, Colorado. Initially, a total 20000 synthetic datasets were prepared by taking a range of conductivity from 1–100 mS/m and a 12 layered earth model with a depth of 6 meter. Each dataset contains an apparent conductivity model and the corresponding true conductivity model. Out of 20000 datasets, 70% (14000) datasets were used for training the DCNN. Out of remaining 30% datasets, 15% (3000) were used for the validation and the rest 15% (3000) for the testing of trained DCNN. Training, validation, and testing processes turned out to be very successful with accuracies of around 99% in each case. Finally, the trained DCNN was applied on the field FDEM induction data for obtaining the subsurface distribution of electrical conductivity to assess the AMD-induced contamination. The findings were found to be in good agreement with the results of published literature. In conclusion, it can be said that the developed Deep Learning (DL) approach can be efficiently used for the rapid and the accurate assessment of subsurface contamination caused by factors that alter the ground electrical properties.