Subsurface Permeability Estimation using Deep Neural Networks

Abstract:

Watershed model parameters, such as the subsurface permeability field, are difficult or expensive to measure directly at the spatial extent and resolution required by mechanistic watershed models. Within a watershed, stream discharge measured at stream gauges is usually available with historical record, and thus can be used to infer soil and geologic properties using inverse modeling. However, very few studies have applied inverse modeling assisted by machine learning (ML) to infer subsurface permeability. In this study, we perform ML assisted hydrologic inverse modeling to estimate soil and geologic permeability using observed hydrographs and water table elevations. We train several deep neural network (DNN) model architectures to predict subsurface permeability parameters from simulated stream discharges at the outlet of the Rock Creek watershed. The DNN models are able to successfully predict the permeability using discharges, and the permeability with larger spatial coverage has better predictability. Additionally, we perform a sensitivity study on the models and find permeability parameters with smaller spatial coverage are more sensitive to noise added to the discharge. Furthermore, the DNN models are able to estimate reasonable permeability values for parameters with larger spatial coverage from the observed stream discharge of the Rock Creek watershed. However, the models suffer in their estimations of the permeability with smaller coverage. Finally, we compare the DNN models against the Ensemble Smother (ES) method and find the models outperform the ES permeability estimations.