VALIDATING A MULTIVARIATE STATISTICAL MODEL USED TO ASSESS IF FORMATION BRINES ARE THE MOST PROBABLE SOURCE OF HIGH SALINITY IN SHALLOW GROUNDWATER

Development of unconventional gas resources (e.g. high-volume hydraulic fracturing) may pose contamination risks to shallow groundwater. However, assessment of such contamination can prove difficult due in part to other potential sources of groundwater contamination in shale gas basins. In previous work, we developed a statistical model that uses linear discriminant analysis to identify the most probable source of salinity in groundwater samples based on their geochemical fingerprints. The statistical model considers multiple end-members (e.g. formation brines, road salt, septic effluent) and model predictive accuracy was previously assessed using cross-validation methods with training data. While ultimately successful, there was no definitive way to test the model accuracy in prior work because groundwater samples classified for our original study did not have known sources of contamination. Here, we applied the model to a dataset of shallow groundwater with known sources of contamination compiled from two studies of groundwater quality in Illinois: Panno et al., Illinois State Geol. Survey, Open File Series 2005-1 and Hwang et al., Environ. & Eng. Geosci., 11: 75-90 (2015). By predicting the source of salinity in groundwater samples for which sources of contamination are known, we were able to validate model predictive accuracy.

Different combinations of solutes and saline end members were considered. Results show high classification success (>80%) for groundwater samples impacted by formation brines and road salt, with diminishing success for other salinity sources (septic effluent and animal waste). Certain solutes, particularly Br and I, remain relatively important for accurately fingerprinting the source of groundwater salinity. However, we found changes in the relative importance of other solutes for fingerprinting salinity between the original study and the new datasets. For instance, Na and Ba are more strongly correlated with the discriminant analysis scores used to classify samples in this study, relative to the original dataset. These results highlight the importance of measuring certain solutes (such as Br and I) during baseline water quality assessments and indicate the model could be transferred to other shale gas basins and still remain effective.