Using Machine Learning Techniques to Optimize Subsurface Hydrologic Data Collection

Abstract:

Machine learning (ML) techniques have made a major impact on the field of atmospheric sciences and are gaining a strong foothold in surface hydrology. To date, their use for subsurface hydrologic applications has been limited. In part, this limitation has to do with the relatively scarcity of subsurface data. This same lack of observational data plagues physics-based models, leading to large (often unacceptably large) uncertainties in predictions related to flow and solute transport. Given the cost of monitoring, there have been many efforts to identify valuable subsurface data for collection. More recent efforts have made use of multiple-model analyses and have cast the measurement optimization in a decision context. That is, physics-based model ensembles are used to identify discriminatory observations that are most likely to affect predictions that are used as the basis for decision making. This study represents an initial step toward using ML techniques for this purpose. We rely on a range of algorithms of varying complexity and use both built-in and post-analysis methods to assess data worth through feature importance analyses. Our results show both promise and limitation for this technique, but generally point to measurement network optimization through ML methods as a promising avenue of applied research in subsurface hydrology.