Where The Decision Tree Grows: Can Machine Learning Design A Groundwater Monitoring Network?
Presenter: David Murray1
Co-Author(s): -
Advisor(s): Dr. Ty Ferre
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Panapto Presentation Video
Poster PDF
Poster Session 1
Monitoring wells are often the most expensive part of a groundwater investigation. Ideally, wells would be located based on all available hydrogeologic information such that they can provide information to reduce remaining uncertainties. In addition, the well locations should be chosen to provide the most information regarding predictions of interest for scientists and/or decision makers. This investigation examines where simple Machine Learning models (regression trees) can be used to identify informative well locations. Existing information is contained in a basic numerical hydrogeologic model. Uncertainty is introduced through varied hydraulic property values and boundary conditions. The prediction of interest is the streamflow at the downstream outlet of the basin. Since the model generates both water level and stream flow data, the regression trees are trained on the ensemble of model outputs representing a range of plausible hydrogeologic systems. Results show that the built-in feature importance capability of regression trees allows for highly efficient identification of optimal monitoring locations, only requiring a single run of each model in an ensemble. Future work will examine whether wells can be identified to satisfy multiple predictions of interest and whether optimal sets of wells can be identified simultaneously.