As the seasons change, so do the water levels in streams and rivers. As snow thaws, the icy water seeps into the ground at a specific rate, quantified by a measurement called subsurface permeability. Other forms of precipitation, like rain, add to this as well. The subsurface flow then contributes to river and stream flow. Recently, a study by researchers at Pacific Northwest National Laboratory (PNNL), Oak Ridge National Laboratory, and Los Alamos National Laboratory used deep learning to estimate subsurface permeability levels in a watershed.
“Subsurface permeability helps determine how much water a river has and impacts the health of a watershed, but it is difficult and expensive to directly measure,” said Erol Cromwell, Advanced Computing, Mathematics, and Data Division staff member at PNNL and lead author of the study. “We increased the speed and accuracy of subsurface permeability estimations by creating and training a neural network using simulated data.”
Typically, subsurface permeability is approximated through inverse modeling, which uses indirect measurements to infer permeability. The deep neural network (DNN) Cromwell built uses stream flow data for subsurface permeability estimation. Using watershed simulations produced by Earth scientists in the Atmospheric Sciences and Global Change Division, the DNN is trained to map the relationship between subsurface permeability and widely available stream discharge data across several geologic types.
“This neural network directly benefits watershed modelers by improving the accuracy of simulated watershed response. The next step for this project is to expand this for a larger watershed area and confirm that approach will hold under different conditions,” said Cromwell.