Using Discriminant Analysis to Determine Sources of Salinity in

Jul 25, 2014 - several salinity sources (Appalachian Basin brines, road salt, septic ... of salinity in 35% of sampled NYS groundwater wells with >20 ...
0 downloads 0 Views 3MB Size
Download PDF
Article pubs.acs.org/est

Using Discriminant Analysis to Determine Sources of Salinity in Shallow Groundwater Prior to Hydraulic Fracturing Laura K. Lautz,*,† Gregory D. Hoke,† Zunli Lu,† Donald I. Siegel,† Kayla Christian,† John Daniel Kessler,‡ and Natalie G. Teale†,§ †

Department of Earth Sciences, Syracuse University, 204 Heroy Geology Laboratory, Syracuse, New York 13244, United States Department of Earth & Environmental Sciences, University of Rochester, 210 Hutchison Hall, Rochester, New York 14627, United States

‡

S Supporting Information *

ABSTRACT: High-volume hydraulic fracturing (HVHF) gas-drilling operations in the Marcellus Play have raised environmental concerns, including the risk of groundwater contamination. Fingerprinting water impacted by gas-drilling operations is not trivial given other potential sources of contamination. We present a multivariate statistical modeling framework for developing a quantitative, geochemical fingerprinting tool to distinguish sources of high salinity in shallow groundwater. The model was developed using new geochemical data for 204 wells in New York State (NYS), which has a HVHF moratorium and published data for additional wells in NYS and several salinity sources (Appalachian Basin brines, road salt, septic effluent, and animal waste). The model incorporates a stochastic simulation to predict the geochemistry of high salinity (>20 mg/L Cl) groundwater impacted by different salinity sources and then employs linear discriminant analysis to classify samples from different populations. Model results indicate Appalachian Basin brines are the primary source of salinity in 35% of sampled NYS groundwater wells with >20 mg/L Cl. The model provides an effective means for differentiating groundwater impacted by basin brines versus other contaminants. Using this framework, similar discriminatory tools can be derived for other regions from background water quality data.

■

INTRODUCTION High-volume hydraulic fracturing (HVHF) is being used to extract natural gas stored in low permeability shale and has led to expansion of gas-drilling operations in shale plays.1,2 Shale gas basins are found throughout the United States and include the Marcellus Shale of the Appalachian Basin.3,4 The Marcellus Shale is the most expansive shale gas play in the U.S.;2 it is estimated to house over 20% of the total recoverable shale gas in the U.S.5,6 Environmental concerns related to shale gas production are broad, and one of the greatest concerns lies in the risk of groundwater and surface water contamination from HVHF fluids and released gases and naturally occurring dissolved solids found deep within the subsurface.7 These concerns can be amplified in rural and agricultural places where a large number of private wells depend on shallow groundwater for irrigation and domestic water use.8,9 During gas-drilling operations, saline “flowback” and produced water may be introduced into the environment through migration of injection fluids and formation waters to shallow aquifers and/or discharge of the water to the environment during transport and disposal (e.g., surface spills of produced water or leaking impoundment ponds).8,10 Failed well casings and defective cement have also been implicated as causes of brine contamination in shallow aquifers.1 Produced and flowback waters typically include two sources of potential contamination: additives used to create optimal fluid © 2014 American Chemical Society

consistency for well stimulation and the metals, dissolved solids, and radionuclides introduced to flowback water from naturally occurring basin brines.10 Although the potential risks for groundwater and surface water contamination from gasdrilling operations, including HVHF, have been clearly identified and articulated,2,10,11 evidence for contamination of water due to gas drilling in the Marcellus is sparse, controversial and hotly debated.8,12−16 There are several challenges to identifying water impacted by gas-drilling operations, not the least of which is the possibility of other sources of contamination in the human-impacted watersheds in which drilling is taking place.10 Other potential sources of contamination include mine drainage, brines from abandoned, shallow oil and gas wells, road salt, wastewater effluent, agricultural runoff, discharge from coal-fired power plants, and industrial discharge.17 Natural conditions also complicate interpretations of potential contamination of shallow groundwater; in regions of Pennsylvania, water wells located in low elevation valleys have significantly higher dissolved methane than water wells in upland areas, regardless of the proximity to gas extraction wells,18 and Appalachian Received: Revised: Accepted: Published: 9061

May 6, 2014 July 25, 2014 July 25, 2014 July 25, 2014 dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

Basin brine, road salt runoff, septic effluent, and animal waste), resulting in a synthetic population of high salinity groundwater. Linear discriminant analysis (LDA) was used to compare those synthetic populations to the actual measured observations of high salinity groundwater taken in the field to assess the most likely population from which the observed samples were derived. In this way, measured high salinity groundwater samples were assigned to one of four populations, indicating they were most likely impacted by mixing with Appalachian Basin brines, road salt, septic effluent, or animal waste. While other techniques such as classification trees could be used to develop a similar model, LDA has the advantage of generating probabilities of class membership for each unknown, which can be used to evaluate the quality of the classification. LDA generated linear combinations of solute concentrations (the “linear classifiers”) that best separate different populations; here, those populations are high salinity groundwater samples impacted by the four potential sources of salinity. LDA requires a training data set to derive the linear classifier, which can then be applied to categorize unknown samples. Generation of End-Member Synthetic Water Chemistry Data. Measured shallow groundwater data were divided into two groups: low salinity (20 mg/L Cl) (Supporting Information Figure S1) because prior studies suggest groundwater with >20 mg/L Cl may be impacted by anthropogenic sources of salinity or mixing with naturally occurring brines.19,31 Synthetic random samples of low salinity shallow groundwater were generated using a multivariate random number generator, parametrized with the statistical distribution of the observed log-transformed low salinity groundwater data, and statistical methods detailed in the Supporting Information (SI Methods). The random number generation methods allow for synthetically increasing the low salinity data set sample size to n = 3000, while also creating a data set with complete solute information for all variables (Supporting Information Figure S2). Random samples of the various potential salinity sources (Appalachian Basin brines, road salt, septic effluent, and animal waste) were created in a similar fashion using the statistical distribution of solute concentrations reported in the literature for each salinity source and the same random number generation methods provided in the SI. Development of Training Data Set and Discriminant Analysis Model Development. A training data set for LDA was created using the synthetic low salinity shallow groundwater and four saline end-member data sets (each n = 3000) and a simple two-component mixing model. To explore expected chemistry of shallow groundwater in NYS if mixed with Appalachian Basin brines, road salt, septic effluent, or animal waste, variable percentages of each salinity source were introduced, in turn, to the synthetic low salinity groundwater data set. The chemistry of synthetic high salinity groundwater samples (GWhigh,i) were computed as

Basin brines may naturally migrate into shallow groundwater.19,20 Produced water may have a distinct elemental and isotopic composition that facilitates the development of geochemical fingerprinting tools that distinguish contamination caused by flowback or produced waters from other sources of contamination.10,17,21,22 To uniquely differentiate the impact of Marcellus brines from other Appalachian brines even at very low concentrations, the strontium isotopic composition of impacted waters may be effective,17,19 but isotopic measurements can be expensive and require specialized instruments. Also, such measurements may not be available in background water quality databases or run in routine water quality tests. Shale gas flowback water from the Marcellus Shale, and other Appalachian Basin units, has a chemical composition similar to deep groundwater found in other Appalachian Basin formations,23 contains high concentrations of calcium, sodium, and chloride, and has metal concentrations consistent with deep Appalachian Basin brines.11,17,24 Total dissolved solids, bromide, iodine, and ratios of halogens and major solutes have been proposed as potentially distinct tracers of groundwater contamination by produced and flowback waters that are relatively inexpensive to measure and often included in background water quality surveys.9,11,19,25−27 Although these parameters show promise to quantify the proportion of a water sample containing Appalachian Basin brine,20 natural variability in these parameters in shallow groundwater can make definitive assessment of contamination uncertain. Several have argued that baseline measurements before and after gas-drilling operations are initiated are needed to effectively test whether such operations have caused groundwater contamination.14,16,20 New York State (NYS) is in a unique position related to hydraulic fracturing, in particular, because of its imposed moratorium on HVHF until state regulations are finalized.28 The objective of our study was to develop a quantitative, geochemical fingerprinting tool to distinguish between several natural and anthropogenic sources of high salinity in shallow groundwater prior to future gas drilling with HVHF and by using water quality parameters commonly available from databases of groundwater quality. The fingerprinting tool was developed from new and published data on shallow groundwater quality in southern New York State, including data from Project SWIFT (this study). We synthesized and then compared shallow groundwater chemistry to published data for several potential sources of salinity, including Appalachian Basin brines,17,19,23,29,30 road salt runoff,31−37 septic effluent,31,32 and animal waste.31,32,38 Multivariate, linear discriminant analysis was used to produce a quantitative tool for identifying the most likely source of salinity in groundwater samples with >20 mg/L Cl.

■

MODEL RATIONALE AND DEVELOPMENT Model Development Framework. The development of our geochemical fingerprinting tool incorporates a form of Monte Carlo simulation, or stochastic simulation, where virtual experiments are run repeatedly on random samples drawn from a known distribution, to obtain the distribution of an unknown population. Here, synthetic shallow groundwater samples are created by repeated random draws from a simulated population, whose distributional parameters were estimated from measured data. The synthetic samples were then mixed with random proportions of contaminant drawn from simulated populations of one of four saline end-members (Appalachian

GWhigh, i = ni(sourcei) + (1 − ni)(GWlow, i)

(1)

where ni is the percent source (e.g., brine) in sample i, sourcei is a vector of the nine solute concentrations in a random sample, i, of the source (e.g., a random sample of synthetic brine), and GWlow,i is a vector of the nine solute concentrations in a random low salinity groundwater sample i. A separate synthetic, high salinity data set was created for each end-member (Supporting Information Figure S3). The four synthetic high salinity data sets were then combined to create one training 9062

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

data set. The two-component mixing model assumes conservative mixing. Although this is a simplifying assumption, it is one also made in prior studies of shallow groundwater major ion chemistry impacted by mixing with various sources of salinity, including basin brines, and our model builds on that prior work.17,19,20,22,27 Interpreting geochemical mixing of brines in linear space can also be problematic due to the potential mathematical problems associated with applying standard statistical methods to interpreting compositional data.39,40 Given the relatively low salinities simulated in our model training data, we anticipate limited mathematical problems introduced by performing our analysis in linear space. The goal of the two-component mixing model was to create synthetic high salinity groundwater data sets that had a similar range and distribution of chloride concentrations as the observed high salinity groundwater, if possible (Supporting Information Figure S4). For this reason, the distributions of percentages used for the mixing model were different for the different salinity sources (Supporting Information Figure S5). Septic effluent and animal waste required unreasonable mixing percentages to create similar distributions of chloride in the synthetic high salinity data (e.g., >20%) because the salinity of septic effluent and animal waste is relatively low. Thus, mixing percentages of greater than 20% were considered unreasonable and capped at 20%, since well water is unlikely to be comprised of greater than 20% septic effluent or animal waste (Supporting Information Figure S5). High salinity shallow groundwater, by definition, has a minimum Cl concentration of 20 mg/L. To create synthetic high salinity groundwater data with similar limits, any samples in the synthetic high salinity data sets with Cl concentrations 20 mg/L Cl, n = 65) from the Project SWIFT database that were categorized by the linear discriminant analysis model; larger symbols show classifications with >80% probability, and smaller symbols show classifications with ≤80% probability.

water chemistry data in the two databases are not significantly different with regard to mean values and distributions of logtransformed values (Supporting Information Table S2). We considered four potential sources of salinity in prehydraulic fracturing groundwater: Appalachian Basin brines, road salt runoff, septic effluent, and animal waste. The end-member compositions were taken from new and published data for Appalachian Basin brines,17,19,23,29,30 road salt runoff,31−37 septic effluent,31,32 and animal waste31,32,38 (Supporting Information Tables S3−S6). The nine solutes considered relatively conservative and measured in both the homeowner wells and salinity source samples are iodine (I), sodium (Na), potassium (K), magnesium (Mg), calcium (Ca), chloride (Cl), bromide (Br), barium (Ba), and strontium (Sr); these solutes were included in the statistical analysis. Shallow groundwater chemistry data were evenly distributed spatially across the five-county study area, reducing the potential influence of spatial bias in the field sampling (Figure 1).

■

RESULTS AND DISCUSSION Chemistry of Shallow Groundwater and Potential Sources of Salinity. Salinity of shallow groundwater samples in southern New York State (NYS) varied over four orders of magnitude (0.2−664 mg/L Cl). Based on distinctions made in the literature,19,31 shallow groundwater samples were classified as having low salinity (20 mg/L Cl, n = 81), with the high salinity samples likely to be impacted by mixing with brines or anthropogenic sources of chloride.31 In this study, four potential sources of salinity (or “end-members”) were considered: natural migration of Appalachian Basin brine, anthropogenic mixing with road salt runoff, introduction of saline septic effluent, and mixing with animal waste. Prior studies have observed the presence of these sources of salinity in water wells in other states.19,31 Acid mine drainage and effluent from coal-fired power plants were not

■

FIELD METHODS AND MATERIALS Groundwater chemistry data for shallow wells in southern New York was compiled from sampling done by Project SWIFT in 2012−2013 (Supporting Information Table S1, n = 204) and the U.S. Geological Survey’s 305(b) Ambient Groundwater Quality Program between 2003 and 2009 (USGS, n = 21) (Figure 1).41 Statistical tests indicate the low salinity ground9063

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

potential end-members. Br and Cl concentrations predicted by conservative mixing of shallow groundwater with formation water overlap with ranges predicted by conservative mixing with road salt runoff, septic effluent, and animal waste (Figure 2). Additionally, if limited solute concentration data are available for high salinity samples, such that information for typical tracers (e.g., Br) are absent, evaluating multiple solutes simultaneously may uniquely characterize the most likely source of salinity, thereby avoiding potentially incorrect classifications that can result from making classifications based on bivariate mixing alone. Discriminating between Sources of Salinity in Shallow Groundwater. The high salinity training data set created for the LDA model reveals differences between populations of high salinity groundwater created by mixing with Appalachian Basin brines, road salt runoff, septic effluent, or animal waste (Supporting Information Figure S3). Not surprisingly, there is notable overlap in the distributions of solute concentrations in the high salinity training data sets, which are impacted by different end-members, particularly at Cl concentrations near the 20 mg/L cutoff (Supporting Information Figure S3). LDA and sequential feature selection were used to identify the minimum number and combination of solutes that could be used to achieve the minimum classification error for the training data set. Backward feature selection indicates barium (Ba) can be removed from the data set without increasing the classification error. LDA generated three linear classifiers and associated scores for each sample in the training data set; the classifiers are used as the basis for classifying samples as from the Appalachian Basin brine, road salt, septic effluent, or animal waste populations (Figure 3). Score 1 summarizes the combination of (standardized, logtransformed) solute concentrations that best separate the samples mixed with Appalachian Basin brine from the road salt runoff and septic effluent samples. Score 1, described as the “Halogen Score,” separates samples with high Br and I, relative to other solutes; this is reflected in the positive score 1 correlations with I and Br (Table 1). As has been observed in prior studies, Appalachian Basin brines have high relative abundances of Br and I compared to other salinity sources.19,27,30,31 Score 2, termed the “Halite Score,” summarizes the combination of solute concentrations that best

considered, as there are no currently active coal-fired power plants in the study area and mines in the area are only for sand and gravel, sandstones, or other crushed rock. The primary form of agriculture in New York State is livestock and livestock products, including dairy farming. All four potential sources of salinity considered fall on the same linear relationship between Na and Cl (Figure 2);

Figure 2. Bivariate plots showing Br−Cl and Na−Cl for all observed data, including high salinity (>20 mg/L Cl) versus low salinity ( 99% of training samples from this group were classified correctly. The model was slightly less effective for classifying samples impacted by mixing with road salt runoff, septic effluent, or animal waste; and misclassified training samples from the road salt and animal waste groups were most commonly mistaken for septic effluent (19.3% and 15.1% of data, respectively).

a

We compared our original model to a model run with only the solutes available from the USGS database (Na, K, Mg, Ca, Cl, Ba, Sr), only those found in another publically available database, the National Uranium Resource Evaluation, or NURE, program (Na, Mg, Cl, Br), and only halogens (Cl, Br, I). Percentages are the correct classification rates for each model (equal to the diagonal of the confusion matrix).

The other commonly measured solutes were still effective for fingerprinting brines in the majority of samples and the classification rates for septic effluent and animal waste were high, similar to the original model (Table 3). Although the NURE program only reports limited concentration data, Cl and Br data are available; the inclusion of Cl and Br in the LDA model maintained an almost perfect classification rate for samples impacted by brines, at 99.2%, although accuracy of classification of the septic effluent and animal waste groups dropped notably relative to the original model (Table 3). Similarly, a model developed using only the halogens (Cl, Br, I) had the highest correct classification for brines, indicating these

Table 2. Confusion Matrix for the Linear Discriminant Analysis Model, Which Is Computed from the Summation of 10 Confusion Matrices from a Stratified, 10-fold Cross Validationa predicted group known group Appalachian Basin brine road salt septic effluent animal waste

Appalachian Basin brine 99.4% 1.3% 1.4% 2.6%

road salt

(2744) (33) (8) (58)

0.0% 79.2% 9.0% 1.3%

(0) (2022) (51) (29)

septic effluent

animal waste

0.6% 19.4% 86.5% 15.1%

0.0% 0.2% 3.2% 81.0%

(17) (495) (492) (340)

(0) (4) (18) (1818)

a Percentages show the portion of the known group classified as from the predicted groups, and values in parentheses show the number of training samples from each known group classified in each predicted group. The rows do not sum to 3000 because training data with Cl < 20 mg/L were removed. Values on the diagonal indicate correct classifications.

9065

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

Figure 4. Bivariate plots showing observations of unclassified groundwater and classified Project SWIFT samples.

mixing of brines and shallow groundwater were taking place at the proportions required to elevate Br to observed concentrations. Other sources of Br and Cl, such as road salt, have lower Sr and I concentrations and, as such, the solute concentration patterns in high salinity shallow groundwater suggest these sources are more likely contributing salinity to these samples. At high Cl concentrations (e.g., > 100 mg/L Cl), halogens show the greatest potential for discriminating between formation water and other sources of salinity (Figure 4). One advantage of using discriminant analysis to classify unknown samples is that model classification results include estimates of certainty of the classification for individual unknowns. For samples that fall on the boundary between classifications (e.g., sample 1, which has a score 1 value of ∼0, Figure 3), the classification is more uncertain than for samples that fall well within the range of scores for a given population. The model results indicate only a 56% probability that sample 1 is from the Appalachian Basin brine group. The certainty of each sample classification for the high salinity Project SWIFT

are the most informative, commonly measured solutes for brine fingerprinting. The original LDA model was used to identify the most likely source of salinity for all observed high salinity groundwater samples (>20 mg/L Cl) for which the complete set of solute information is available (from Project SWIFT, n = 65). Using the model, high salinity groundwater samples from southern NYS were classified as impacted by Appalachian Basin brines (n = 23), road salt (n = 27), or septic effluent (n = 15); none were classified as impacted by animal waste (Figures 3 and 4). Although the Br−Cl relationship for observed high salinity samples suggest Appalachian Basin brines as a potential source of salinity (Figure 3), not all samples with elevated Br:Cl ratios were classified as impacted by basin brines; some were classified as more likely to be impacted by septic effluent or road salt (Figure 4). This is because although concentrations of Br and Cl are elevated in these samples (which is consistent with mixing with brines), concentrations of other solutes, such as Sr and I, are not as high as would be predicted if conservative 9066

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

basins. The statistical fingerprinting model provides an effective means for differentiating between samples impacted by Appalachian Basin brines, road salt, septic effluent, and animal waste and allows for assessment of certainty associated with classifying samples from those populations. The model takes advantage of relatively simple chemical measurements that are routinely made for background water quality surveys and can be adapted to use different combinations of solutes, depending on data availability and/or assumptions of which solutes behave conservatively. From the model results, it appears Appalachian Basin brines are already a source of salinity in many shallow groundwater wells in NYS, potentially complicating detection of future changes in salinity associate with future gas-drilling operations in the region, including HVHF. Baseline water quality data will be critical for future assessments. Future baseline water quality surveys should include measurements of all halogens (Br, Cl, I) for reliable fingerprinting of mixing with brines, but existing databases that do not include halogens are still informative for differentiating the sources of salinity considered here. While the specific geochemical model presented here may not be directly transferable to other shale gas basins or regions where there are other notable sources of salinity to shallow groundwater, background water quality databases for other regions, as well as chemistry data for other saline end-members, can be used to create similar discriminatory tools using the framework presented.

samples is summarized in Figure 1, where large symbols indicate samples classified with a probability of greater than 80% and small symbols indicate samples classified with a probability of less than or equal to 80%. There is no definitive way to test the accuracy of the model classification of the unknown samples from Project SWIFT because there is no way of knowing what the true source of salinity is for each sample. In an effort to corroborate the model classification, concentrations of other solutes were explored. In prior studies, shallow groundwater samples suspected of being impacted by mixing with Appalachian Basin brine also had elevated concentrations of methane.19 It is also expected that samples impacted by mixing with septic effluent or animal waste would have elevated concentrations of nutrients, such as nitrate, because these end-members have elevated nutrient concentrations. Nitrate and methane concentrations were not used in the model development or application, so these solutes represent an independent test of the reliability of the model classification. Project SWIFT samples classified as impacted by Appalachian Basin brine have significantly higher methane concentrations than samples classified as impacted by road salt or septic effluent (p-value, one-way ANOVA < 0.001; Figure 5). One sample classified as impacted by road salt has

■

ASSOCIATED CONTENT

S Supporting Information *

Details on field sampling methods and statistical methods for generating synthetic data sets. The complete shallow groundwater chemistry data set collected by Project SWIFT and used in this study, as well as the complete water chemistry data for the saline end-members (Appalachian Basin brines, road salt runoff, septic effluent, and animal waste) compiled from the literature. This material is available free of charge via the Internet at http://pubs.acs.org/.

Figure 5. Box plots showing the distribution of methane and nitrate concentrations in shallow groundwater samples classified as being influenced by mixing with Appalachian basin brine (ABB), road salt (RS), or septic effluent (SE).

■

AUTHOR INFORMATION

Corresponding Author

relatively high methane (7.1 mg/L). This is not necessarily surprising because there may be various sources of methane in shallow groundwater (e.g., biogenic versus thermogenic) and this one sample may be misclassified. Regardless, there is clearly a correlation between the Appalachian Basin brine classification and the methane concentration. Project SWIFT samples classified as impacted by septic effluent have the highest median concentrations of nitrate (Figure 5), although samples classified as road salt also have relatively high concentrations of nitrate. Samples classified as impacted by Appalachian Basin brine have significantly lower nitrate concentrations than other groups (p-value, one-way ANOVA = 0.001; Figure 5). High nitrate concentrations in samples classified as impacted by road salt may reflect misclassification by the model; if samples impacted by septic effluent are misclassified, it is most likely to be as impacted by road salt (Table 2) and the certainties for the road salt samples with the two highest nitrate concentrations (12.4 and 19.8 mg/L, Figure 5) are low (sample 2 73% and sample 3 59%, respectively, Figure 3). Environmental Implications. Discriminant analysis is a presented as an alternative method for identifying the most probable sources of salinity in shallow groundwater in shale gas

*Phone: 315-443-1196. E-mail: [email protected]. Present Address §

N.G.T.: Department of Geography, Texas A&M University, College Station, Texas . Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS Sunshyne Hummel, Egan Waggoner, Max Gade, Karolina Lubecka, Anthony Carranciejie, and Michael Young assisted with sampling and analysis for Project SWIFT. We thank Broome, Chemung, Chenango, Steuben and Tioga Counties for providing property parcel information. We thank Martin Briggs, Ryan Gordon, three anonymous reviewers, and the Associate Editor for providing constructive reviews on the paper. This work was made possible by support from the National Science Foundation’s RAPID program (EAR1313522), the Syracuse University Office of the Provost and Vice Chancellor, and support from Syracuse University Alumni. 9067

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

■

Article

(22) Whittemore, D. O. Geochemical differentiation of oil and gas brine from other saltwater sources contaminating water resources: Case studies from Kansas and Oklahoma. Environmental Geoscience 1995, 2 (1), 15−31. (23) Osborn, S. G.; McIntosh, J. C. Chemical and isotopic tracers of the contribution of microbial gas in Devonian organic-rich shales and reservoir sandstones, northern Appalachian Basin. Appl. Geochem. 2010, 25 (3), 456−471. (24) Siegel, D. I.; Szustakowski, R.; Frape, S. A regional evaluation of brine mixing in the Albion Group (Silurian) sandstones of New York, Pennylvania, and Ohio. Assoc. Pet. Geochem. Explorationists Bull. 1991, 6 (1), 66−77. (25) Kight, M. D.; Siegel, D. I. A protocol to characterize flowback water contamination to shallow waters from shale gas development. Northeastern (46th Annual) and North-Central (45th Annual) Joint Meeting, Pittsburgh, PA, March 20−22, 2011. (26) Wilson, J. M.; Wang, Y.; VanBriesen, J. M. Sources of high total dissolved solids to drinking water supply in southwestern Pennsylvania. J. Environ. Eng. 2013, 140, B4014003. (27) Davis, S. N.; Whittemore, D. O.; Fabryka-Martin, J. Uses of chloride/bromide ratios in studies of potable water. Ground Water 1998, 36, 338−350. (28) Kramer, D. Shale-gas extraction faces growing public and regulatory challenges. Phys. Today 2011, 64 (7), 23−25. (29) Dresel, P. E.; Rose, A. W. Chemistry and origin of oil and gas well brines in western Pennsylvania; Pennsylvania Geol. Surv, Open-File Rep. 2010, 10−01.0, 1−48. (30) Osborn, S. G.; McIntosh, J. C.; Hanor, J. S.; Biddulph, D. Iodine-129, 87-Sr/86-Sr, and trace elemental geochemistry of northern Appalachain basin brines: evidence for basinal-scale fluid migration and clay mineral diagenesis. Am. J. Sci. 2012, 312, 263−287. (31) Panno, S. V.; Hackley, K. C.; Hwang, H. H.; Greenberg, S. E.; Krapac, I. G.; Landsberger, S.; O’Kelly, D. J. Characterization and identification of Na-Cl sources on ground water. Ground Water 2006, 44, 176−187. (32) Panno, S. V.; Hackley, K. C.; Hwang, H. H.; Greenberg, S. E.; Krapac, I. G.; Landsberger, S.; O’Kelly, D. J. Database for the characterization and identification of the sources of sodium and chloride in natural waters of Illinois, Illinois State Geological Survey, 2005; Open File Series 2005-1. (33) Ledford, S. H.; Lautz, L. K. Floodplain connection buffers seasonal changes in urban stream water quality. Hydrol. Process. 2014, DOI: 10.1002/hyp.10210. (34) Granato, G. E. Deicing chemicals as source of constituents of highway runoff. Transportation Research Record 1533; National Research Council: Washington, D.C., 1996; pp 50−58. (35) Howard, K. W. F.; Beck, P. J. Hydrogeochemical Implications of Groundwater Contamination by Road Deicing Chemicals. J. Contam Hydrol 1993, 12 (3), 245−268. (36) Kelly, W. R.; Panno, S. V.; Hackley, K. C.; Hwang, H. H.; Martinsek, A. T.; Markus, M. Using chloride and other ions to trace sewage and road salt in the Illinois Waterway. Appl. Geochem. 2010, 25 (5), 661−673. (37) Foos, A. Spatial distribution of road salt contamination of natural springs and seeps, Cuyahoga Falls, Ohio, USA. Environ. Geol 2003, 44 (1), 14−19. (38) Becker, M. F.; Peter, K. D.; Masoner, J., Possible sources of nitrate in ground water at swine licensed-managed feeding operations in Oklahoma, 2001; U. S. Geological Survey, 2002; Water-Resources Investigations Report 02-4257. (39) Engle, M. A.; Rowan, E. L. Interpretation of Na-Cl-Br Systematics in Sedimentary Basin Brines: Comparison of Concentration, Element Ratio, and Isometric Log-ratio Approaches. Math Geosci 2013, 45 (1), 87−101. (40) Engle, M. A.; Rowan, E. L. Geochemical evolution of produced waters from hydraulic fracturing of the Marcellus Shale, northern Appalachian Basin: A multivariate compositional data analysis approach. Int. J. Coal Geol. 2014, 126, 45−56.

REFERENCES

(1) Vidic, R. D.; Brantley, S. L.; Vandenbossche, J. M.; Yoxtheimer, D.; Abad, J. D. Impact of Shale Gas Development on Regional Water Quality. Science 2013, DOI: 10.1126/science.1235009. (2) Kargbo, D. M.; Wilhelm, R. G.; Campbell, D. J. Natural Gas Plays in the Marcellus Shale: Challenges and Potential Opportunities. Environ. Sci. Technol. 2010, 44 (15), 5679−5684. (3) Soeder, D. J.; Sharma, S.; Pekney, N.; Hopkinson, L.; Dilmore, R.; Kutchko, B.; Stewart, B.; Carter, K.; Hakala, A.; Capo, R. An approach for assessing engineering risk from shale gas wells in the United States. Int. J. Coal Geol 2014, 126, 4−19. (4) Arthur, J. D.; Bohm, B.; Layne, M. Hydraulic fracturing considerations for natural gas wells of the Marcellus Shale. The Ground Water Protection Council Annual Forum, Cincinnati, OH, 2008. (5) U.S. Energy Information Administration, Annual Energy Outlook, 2012; 2012. (6) U.S. Energy Information Administration, Technically recoverable shale oil and shale gas resources: an assessment of 137 shale formations in 41 countries outside the United States; 2013. (7) Kerr, R. A. Natural Gas From Shale Bursts Onto the Scene. Science 2010, 328 (5986), 1624−1626. (8) Osborn, S. G.; Vengosh, A.; Warner, N. R.; Jackson, R. B. Methane contamination of drinking water accompanying gas-well drilling and hydraulic fracturing. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (20), 8172−8176. (9) Boyer, E. W.; Swistock, B. R.; Clark, J.; Madden, M.; Rizzo, D. E. Impact of Marcellus gas drilling on rural drinking water supplies; The Center for Rural Pennsylvania, 2011. (10) Entrekin, S.; Evans-White, M.; Johnson, B.; Hagenbuch, E. Rapid expansion of natural gas development poses a threat to surface waters. Front. Ecol. Environ. 2011, 9 (9), 503−511. (11) Soeder, D. J. The Marcellus Shale: Resources and Reservations. Eos 2010, 91 (32), 277−288. (12) Jackson, R. B.; Osborn, S. G.; Vengosh, A.; Warner, N. R. Reply to Davies: Hydraulic fracturing remains a possible mechanism for observed methane contamination of drinking water. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (43), E872−E872. (13) Osborn, S. G.; Vengosh, A.; Warner, N. R.; Jackson, R. B. Reply to Saba and Orzechowski and Schon: Methane contamination of drinking water accompanying gas-well drilling and hydraulic fracturing. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (37), E665−E666. (14) Davies, R. J. Methane contamination of drinking water caused by hydraulic fracturing remains unproven. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (43), E871−E871. (15) Saba, T.; Orzechowski, M. Lack of data to support a relationship between methane contamination of drinking water wells and hydraulic fracturing. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (37), E663−E663. (16) Schon, S. C. Hydraulic fracturing not responsible for methane migration. Proc. Natl. Acad. Sci. U. S. A. 2011, 108 (37), E664−E664. (17) Chapman, E. C.; Capo, R. C.; Stewart, B. W.; Kirby, C. S.; Hammack, R. W.; Schroeder, K. T.; Edenborn, H. M. Geochemical and Strontium Isotope Characterization of Produced Waters from Marcellus Shale Natural Gas Extraction. Environ. Sci. Technol. 2012, 46 (6), 3545−3553. (18) Molofsky, L. J.; Connor, J. A.; Farhat, S. K.; Albert, S.; Wylie, J.; Wagner, T. Methane in Pennsylvania water wells unrelated to Marcellus shale fracturing. Oil Gas J. 2011, 5 (Dec), 54−67. (19) Warner, N. R.; Jackson, R. B.; Darrah, T. H.; Osborn, S. G.; Down, A.; Zhao, K. G.; White, A.; Vengosh, A. Geochemical evidence for possible natural migration of Marcellus Formation brine to shallow aquifers in Pennsylvania. Proc. Natl. Acad. Sci. U. S. A. 2012, 109 (30), 11961−11966. (20) Llewellyn, G. T. Evidence and mechanisms for Appalachian Basin brine migration into shallow aquifers in NE Pennsylvania, USA. Hydrogeol. J. 2014, 22, 1055−1066. (21) Sharma, S.; Mulder, M. L.; Sack, A.; Schroeder, K.; Hammack, R. Isotope approach to assess hydrologic connections during Marcellus Shale drilling. Groundwater 2014, 52, 424−433. 9068

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069

Environmental Science & Technology

Article

(41) Nystrom, E. A.; Reddy, J. 305(b) Ambient Groundwater Quality Monitoring. http://ny.water.usgs.gov/projects/305b/index.html (accessed June 2013). (42) Valentine, D. L.; Kessler, J. D.; Redmond, M. C.; Mendes, S. D.; Heintz, M. B.; Farwell, C.; Hu, L.; Kinnaman, F. S.; Yvon-Lewis, S.; Du, M. R.; Chan, E. W.; Tigreros, F. G.; Villanueva, C. J. Propane Respiration Jump-Starts Microbial Response to a Deep Oil Spill. Science 2010, 330 (6001), 208−211. (43) Whitcomb, A. Effectiveness of a vegetation buffer strip and wetland in dairy wastewater filtering. Undergraduate Thesis, State University of New York, College of Environmental Science & Forestry: Syracuse, NY, 2007. (44) Pigott, T. D. A review of methods for missing data. Educational Research and Evaluation 2001, 7 (4), 353−383. (45) Little, R. J. A.; Rubin, D. B. Statistical Analysis with Missing Data, 2nd ed.; John Wiley & Sons, Inc, 2002.

9069

dx.doi.org/10.1021/es502244v | Environ. Sci. Technol. 2014, 48, 9061−9069