Using Machine Learning to Map Redox Conditions in the Mississippi Embayment Regional Aquifer System
Monday, December 4, 2017: 11:50 a.m.
101 C (Music City Center)
Redox conditions in the Mississippi embayment regional aquifer system were mapped at depth zones used for domestic and public supply using machine learning methods. About 3 million people rely on groundwater from this aquifer system for drinking water and redox processes exert important controls on groundwater quality in this system. Explanatory variables including environmental spatial datasets, groundwater-flow model output, and well characteristics were used to extrapolate groundwater quality (specifically redox conditions) to areas of the system without existing water-quality data. Machine learning methods are well suited to hydrologic studies because they allow input of continuous and categorical explanatory variables, can accommodate interactions among explanatory variables, and have performed better than linear regression methods for training data and also when tested on data not used as part of model training. In addition, machine learning methods are not constrained by hypothesis testing assumptions such as linear relations and data normality. Environmental explanatory variables included soil data (texture, drainage class, conductance, and geochemistry), climate (temperature and precipitation) and land use. Groundwater-flow model variables included groundwater flux, groundwater age, flowpath length, groundwater altitude, water use, and hydrogeologic unit texture. Well characteristic variables included hydrologic position relative to major surface-water drainages and well-construction data (length of and depth to top of the screened interval). Although there may be limitations to using environmental variables to explain groundwater quality for confined aquifers with long groundwater residence times (that is, old groundwater), if machine-learning methods can be used with environmental and well-characteristic variables, then groundwater quality can be more accurately predicted in areas lacking numerical groundwater-flow models.