Application of 3D-fluorescence/PARAFAC to monitor the performance of managed aquifer recharge facilities

Rapid population growth, unpredictable rainfall patterns, lack of conventional fresh water sources, and uncertainties due to climate change are increasing the pressure on water resources around the world (Drewes 2009). As a response to these phenomena, the development of municipal water recycling to achieve drinking water augmentation and recharge of local groundwater resources has gained widespread interest over the past decades (Henderson et al. 2009; Drewes & Khan 2011). Managed aquifer recharge (MAR) systems are widely used to augment groundwater supplies and are characterized by relatively low environmental impacts and low capital costs (Drewes & Khan 2011). In MAR systems, such as riverbank filtration, soil aquifer treatment, or aquifer recharge and recovery, water is purposefully added to a groundwater system via natural (i.e. lakes, rivers) and/or engineered structures (i.e. injection wells, infiltration basins) (Drewes 2009; Drewes & Khan 2011; Parsekian et al. 2014). MAR systems that provide sufficient soil passage during infiltration act as a natural filter attenuating turbidity, bacteria, viruses, trace organic compounds (e.g. pesticides, industrial chemicals), nitrogen, algae toxins, and other inorganic and organic constituents through a combination of hydro-geochemical and biological processes such as precipitation, biotransformation, sorption, and ion exchange (Ray et al. 2008; Rauch-Williams et al. 2010; Hoppe-Jones et al. 2010). Further advantages of MAR systems are subsurface storage and the ability to recover recharged water at a later time.

To ensure reliability of MAR in terms of water quality improvement using recycled water sources, water quality monitoring tools are urgently required to maintain not only full protection of public health but also to protect the environment from anthropogenic contaminants and meet regulatory requirements (Henderson et al. 2009). Three dimensional (3D)-fluorescence spectroscopy has received increasing interest as a monitoring tool for chromophoric dissolved organic matter (DOM) in a range of applications including the monitoring of drinking water quality (Stedmon et al. 2011), wastewater treatment (Hambly et al. 2010), surface water (Hudson et al. 2007; Henderson et al. 2009), and disinfection byproduct formation potentials in drinking water (Hua et al. 2007). 3D-fluorescence spectroscopy is: (1) capable of characterizing DOM with chromophoric features in different water types (Chen et al. 2003; Weishaar et al. 2003; Murphy et al. 2008; Fellman et al. 2009; Kowalczuk et al. 2009; Hao et al. 2012), (2) cheap compared to other advanced analytical techniques, (3) rapid (i.e. minimal sample preparation is required) (Henderson et al. 2009), and (4) sensitive (i.e. one to three orders of magnitude more sensitive than ultraviolet (UV)-visible spectroscopy) (Hambly et al. 2010). In short, 3D-fluorescence spectroscopy records the fluorescence signals of water samples at different emission and excitation wavelengths. For each excitation wavelength, the excitation spectrum is scanned at the same time as the fluorescence-emission spectrum is recorded and results are arranged in a 3D grid (excitation × emission × intensity) (Henderson et al. 2009; Nir 2013). Visual comparison (‘peak-picking’) of excitation-emission wavelength pairs in unknown mixtures with identified fluorescence regions of analyzed standards (i.e. humic acid) allows for a qualitative assessment of sample composition (Guo et al. 2010; Dahm et al. 2013). However, visual comparison can lead to misinterpretation of excitation-emission matrices (EEMs) since fluorescence spectra might have overlaying features (Stedmon et al. 2003; Guo et al. 2010). Multi-way statistical analysis such as parallel factor analysis (PARAFAC) has been shown to be a powerful tool, decomposing these EEMs into their underlying chemical components to allow a semi-quantitative assessment of individual components (Murphy et al. 2013).

In this study, 3D-fluorescence spectroscopy and PARAFAC were used as a tool to characterize and differentiate chromophoric DOM components of natural and anthropogenic origin, with the aim of monitoring the fate and transport of recycled water recharged to the subsurface at a full-scale MAR site in California. Furthermore, DOM fate and transport was investigated for three different water types (imported surface water, recycled water, and stormwater) that were used for recharge during specified time periods. To our knowledge, such a tracing experiment has not been accomplished at a full-scale site before. We hypothesize that 3D-fluorescence spectroscopy, in combination with multi-way statistical analysis (i.e. PARAFAC), is capable of distinguishing between DOM originating from different sources to monitor the fate and transport of recycled water DOM in groundwater during recharge.

In total, 178 samples were collected at the Chino Basin in California from five different recharge basins over the period of one year (January 2014–January 2015) and analyzed by 3D-fluorescence spectroscopy. Subsequently, a site-specific PARAFAC model was developed to decompose EEM data into their underlying fluorescence spectra.

In March 2014, the Metropolitan Water District of Southern California initiated a project to investigate the possibility of enhancing groundwater recharge using recycled water in the Chino Basin, California, using 3D-fluorescence spectroscopy as a monitoring tool for DOM in the aquifer. As part of the project, the Inland Empire Utilities Agency (IEUA) sampled recharge events with different source water types that were diverted to five selected recharge basins (Declez, Hickory, Turner 1, Turner 4, and RP3). Each of the chosen recharge basins at the Chino Basin comprises several cross-gradient and downstream monitoring wells as well as lysimeters at different depths (1.5, 3.1, 4.6, 7.6, and 10.7 m). Furthermore, several upstream municipal monitoring wells were monitored throughout the course of the study to determine the DOM baseline in background groundwater.

Between January 2014 and January 2015, a total of 178 samples were collected by IEUA field hydrologists during recharge events applying imported water to basins Declez and RP3 (February–March 2014), recycled water to basins Hickory, Turner 4, and RP3 (February–November 2014), and stormwater to basins Declez, RP3, and Turner 4 (January 2015). Recharge basin Turner 1 received a mix of recycled water and stormwater (up to 90% recycled water) throughout the sampling campaign, and was omitted for the source water type specific assessment. In addition, source water grab samples were collected directly from delivery pipelines. A summary of all analyzed surface water and groundwater samples (including their respective ID) is provided in Table S1, Supplementary Information (available with the online version of this paper). Samples were collected and transported on ice for bulk parameter analyses (i.e. total organic carbon, total nitrogen, electrical conductivity) to the IEUA water quality laboratory in Ontario, California. Additionally, split samples for analysis of selected recycled water indicators (i.e. acesulfame-K, carbamazepine, dilantin, meprobamate, primidone, sucralose) were collected and shipped to Eaton Analytical Laboratory in Monrovia, California. Concentrations of these recycled water indicators are used to account for dilution and dispersion in the subsurface during recharge. Samples for 3D-fluorescence spectroscopy analyses were pre-filtered through a 0.45 μm filter at the IEUA laboratory prior to shipment to Colorado School of Mines (CSM) in Golden, Colorado. Samples were shipped on ice and were stored at 5 °C pending respective analyses.

During project start-up, the application of weed and vector control in the vicinity of the recharge basins was identified as a possible source of organic carbon that might influence analytical measurements. To investigate the contribution of anthropogenic chemicals on sample EEMs, three herbicides (diluted with water to concentrations applied at the recharge site) were provided by IEUA for further 3D-fluorescence spectroscopy analysis: DuPont Oust XP, Habitat Herbicide, and Monsanto Roundup Pro Concentrate. In addition, aqueous analytical standards of potential anthropogenic contaminants, the polycyclic aromatic hydrocarbon (PAH) phenanthrene (1.6 mg/L) and the pesticide carbofuran (50 mg/L), were prepared at CSM and analyzed by 3D-fluorescence spectroscopy.

In general, the pre-filtered samples were analyzed within 72 hours after arrival at the CSM laboratory. Samples (10 mL volume) for dissolved organic carbon (DOC) analysis were acidified with phosphoric acid. DOC was quantified using a Sievers 5310 TOC analyzer with autosampler (Ionics Instruments, Boulder, CO) according to Standard Methods (APHA 2012). Samples with DOC concentrations higher than 2 mg/L were diluted with ultrapure water to 2 mg/L for further spectroscopic analysis to minimize quenching effects on the fluorescence signal. Ultrapure water was obtained from an EMD Millipore Synergy UV-R system (Billerica, MA). 3D-fluorescence spectroscopy was carried out on a Horiba Jobin Yvon AquaLog spectrofluorometer (Edison, NJ) using a 1 cm quartz fluorospectrometer cell. UV absorbance at 254 nm (UV_254nm) was obtained from AquaLog absorbance scans at the respective wavelength. UV_254nm data and DOC concentrations were used to calculate the specific UV absorbance (SUVA) for each sample. SUVA is defined as the ratio between UV_254nm absorbance and DOC concentration, and is an indicator for the degree of aromaticity in the water sample (Weishaar et al. 2003).

The detailed analytical procedure for 3D-fluorescence measurements using the AquaLog spectrofluorometer is described elsewhere (Gilmore 2011). In brief, EEMs were obtained with the following settings: 1 s integration time, 230–599 nm excitation spectra at 3 nm steps, 156–933 nm emission spectra, 4 pixel (2.52 nm) emission increment, and medium CCD gain. By subtracting the measured blank EEM (ultrapure water) from the sample EEM, Raman scatter influences were eliminated. Compensation for the inner filter effect, which is the absorption of both excitation and emission light by the sample matrix (Stedmon et al. 2003; Nir 2013), as well as removal of Rayleigh scatter lines were applied to the scans by an algorithm provided by Aqualog software (version 3.6). EEMs were normalized to the daily measured integral of the Raman water peak area and corrected by their respective dilution factors to obtain the fluorescence intensity for the original undiluted samples. Thus, fluorescence intensity was reported in Raman Units (R.U.). Following this normalization procedure allowed for a comparison of observed spectra with EEMs measured in other studies at different instrument settings using various fluorescence spectrometers (Lawaetz & Stedmon 2009). Finally, corrected EEM contour plots were exported for excitation wavelengths 240–450 nm and emission wavelengths 250–600 nm, and the corrected EEM data were exported as ASCII files and processed in Solo 7.9.2 (Eigenvector Research Inc., Wenatchee, WA) for PARAFAC analysis.

PARAFAC is a three-way statistical method with its origin in psychometrics. The method can be applied to decompose fluorescence EEMs into a set of tri-linear terms and a residual array. The PARAFAC model is calculated by an alternating least square algorithm that minimizes the sum of squared residuals. That allows the separation of signals from a complex mixture of compounds such as DOM (Kowalczuk et al. 2009).

If Equation (1) is applied to EEMs, is the intensity of fluorescence for the ith sample at emission wavelength j and excitation wavelength k; is the residual representing the variability not accounted for by the model. F defines the number of components in the model. Consequently, each f corresponds to a PARAFAC component. Each such component has I a-values, J b-values, and K c-values: one for each sample, one for each emission wavelength, and one for each excitation wavelength, respectively (Bro 1997; Stedmon et al. 2003; Baghoth et al. 2011; Murphy et al. 2013).

Furthermore, the parameters have direct chemical interpretation in a valid model. The parameter is directly proportional to the concentration of the fth analyte of the sample i and is defined as scores; and are related to the emission and excitation spectra, respectively, for the fth analyte and are defined as loadings (Baghoth et al. 2011).

The calculation of F_max allows direct quantitative and qualitative comparison of the fluorescence signal of a given component, or the ratio of any two components within a sample set. However, if component A has a higher fluorescence signal than component B, it does not necessarily mean that component A has a higher concentration. Different fluorophores can have different efficiencies at absorbing and converting incident radiation to fluorescence (Murphy et al. 2013).

It is important to mention that components identified by PARAFAC in complex mixtures like surface water may represent overlapping spectra of fluorophores sharing the same or similar fluorescence properties (Andersen & Bro 2003; Murphy et al. 2008; Baghoth et al. 2011). Since interferences and scatter can hinder the modeling process, successfully building a PARAFAC model is sensitive to specifying the correct number of components and the availability of a large number of representative samples (Murphy et al. 2008). While there are many ways to evaluate whether the correct number of components was chosen for a model, it is not always easy to assess the results as different diagnostic tools can give undetermined or even conflicting results (Murphy et al. 2008, 2013). Therefore a combination of several methods should be considered for real datasets.

Core consistency is a diagnostic tool, which evaluates the appropriateness of the model. A PARAFAC model can be represented as the relative difference between a so-called Tucker3-like core array, which is calculated from the data and the PARAFAC loadings, and a super diagonal core of ones (Kompany-Zareh 2012). The core consistency is usually expressed as a percentage, indicating how well the loadings represent variation in the data (Bro 1998; Andersen & Bro 2003). For lower numbers of components, the core consistency tends to start high, near 100%, and is decreasing to zero or even negative values if too many components are selected (Andersen & Bro 2003; Murphy et al. 2013). For real-world non-ideal datasets, Murphy et al. (2013) emphasized that core consistency is not always a reliable diagnostic tool to determine the correct number of PARAFAC components, because it does not provide information if the number of components is chosen correctly; however, it can be used as an indication that the model is not well specified (Bro 1998; Murphy et al. 2008, 2013). Another diagnostic tool is split-half analysis. First, the dataset is divided into two halves. Then, for each half a new PARAFAC model is generated. If the correct number of components was chosen, the excitation and emission spectral loadings of the two halves should be identical according to the uniqueness of PARAFAC. In other words, if the wrong number of components is chosen in the split-half experiment, it is very likely that the two models will not be equal due to the differences in the different samples (Andersen & Bro 2003; Murphy et al. 2013).

In this study, 82 samples were used to develop a site-specific six-component PARAFAC model. The number of components was determined by split-half analysis, core consistency test, and the identification and removal of outliers from the dataset. Models that could not pass the split-half analysis were rejected. A sample was considered as an outlier if it either contained some instrument error or artifact, or if it was properly measured but it was very different from other samples. Besides the above-mentioned tools, visual inspections of the spectral shapes as well as investigation of the excitation and emission spectra of each component were conducted to validate the correct number of components. The shape of the contour plots of the components should be clear without valleys; the emission spectra should further exhibit exactly one distinct maximum (Li et al. 2014).

Concentrations of selected wastewater indicators were used to account for mixing effects of water originating from different recharge events as well as dilution effects with native groundwater during recharge (data not shown). Travel time estimates were provided by IEUA based on long-term conductivity readings. Table 1 provides a summary of estimated dilution and travel times of water in the subsurface for lysimeters at recharge basins RP3, Declez, Hickory, Turner 4, and Turner 1. Based on trace organic chemical data, dilution can be neglected for lysimeters at all basins except Hickory basin, where more data is needed to draw sufficient scientific conclusions. Looking closer at the monitoring wells, the following conclusions can be drawn: at basins Declez and Turner 1, comparable concentrations of indicator chemicals to recharge surface water were found, which means that dilution can be neglected for monitoring wells at these two basins; at basins RP3 and Turner 4, dilution effects could not be determined since travel time of the recharged water has been estimated in excess of one year; at Hickory basin, high background concentrations of artificial sweeteners like acesulfame-K and sucralose were detected in the monitoring wells.

To assess the attenuation of DOM in the subsurface, results were averaged for each recharged water type, independent of recharge basin location. Information provided by IEUA indicated similar performance across recharge basins Declez, RP3, Hickory, and Turner 4 during soil aquifer treatment. Table 2 summarizes the average DOC, UV_254nm, and SUVA values and standard deviations of water samples collected from recharge basins and lysimeters at 1.5, 3.1, 4.6, 7.6, and 10.7 m depth for recharge events applying imported water, recycled water, and stormwater, respectively. Recharge basins exhibited the following DOC concentrations on average: 8.1 ± 4.7 mg/L for imported water, 8.3 ± 4.4 mg/L for recycled water, and 7.9 ± 2.2 mg/L for stormwater. Comparing DOC values from a recharge basin to the values measured at a lysimeter at 7.6 m depth, the DOC is reduced in the subsurface by 71.1 ± 10.7% to 2.3 ± 0.9 mg/L for imported water, 77.3 ± 7.1% to 1.9 ± 0.6 mg/L for recycled water, and 78.0 ± 4.9% to 1.7 ± 0.4 mg/L for stormwater. Average UV_254nm-absorption values were determined for recharge basins to be 18.5 ± 13.3 m^–1 (imported water), 15.4 ± 10.7 m^–1 (recycled water), and 16.8 ± 4.4 m^–1 (stormwater). Likewise DOC concentration, UV_254nm absorbance is reduced in the subsurface; samples from a lysimeter at 7.6 m depth exhibited the following UV_254nm values: 4.8 ± 2.0 m^–1 for imported water (74.1 ± 10.5% reduction), 3.4 ± 1.2 m^–1 for recycled water (77.7 ± 7.6% reduction), and 3.6 ± 1.2 m^–1 (78.8 ± 4.6% reduction) for stormwater. The SUVA values for the recharge basins exhibited distinctly different values for the three different recharge events: 2.5 ± 1.9 L/mg m for imported water, 1.8 ± 0.4 L/mg m for recycled water, and 2.2 ± 0.1 L/mg m for stormwater. At the lysimeter at 10.7 m depth, the following SUVA values were determined: 1.7 ± 0.2 L/mg m (imported water), 1.7 ± 0.3 L/mg m (recycled water), and 2.1 ± 0.1 L/mg m (stormwater), indicating that SUVA (expressing DOM aromaticity) did decrease for the imported water, but change very little for recycled water and stormwater during subsurface travel.

Application of herbicides for weed control in the vicinity of the recharge facility was identified as one of these non-point sources for elevated DOC. As summarized in Table 3, additional organic carbon was introduced to the recharge basins via wet and dry deposition, even taking dilution with recharged water into account. All three herbicides are composed of a carbon structure and although biodegradable, the active ingredient in DuPont's Oust XP, sulfometuron methyl (C₁₅H₁₆N₄O₅S, CAS 74222-97-2), for instance, exhibits a reported half-life in water and soil of around 30 days.

Phenanthrene is generally formed during incomplete combustion of organic matter via natural processes or human activities (Wang et al. 2010). The pesticide carbofuran is marked under the trade name Furadan and is widely used to combat agricultural pests (Alves et al. 2002).

Component 1 had an emission maximum at 407 nm, an excitation spectrum with a maximum at 246 nm, and a shoulder at 324 nm. The fluorescence signal in this region results from the presence of both carbon-carbon double bonds and aromatic carbon bonds and is referred to as fulvic-like organic matter (Henderson et al. 2009). Component 2 revealed an emission maximum at 483 nm, an excitation spectrum with a maximum below 240 nm, and a secondary excitation band at 369 nm. Fluorescence phenomena in this area represent humic-like organic matter (Murphy et al. 2008; Singh et al. 2013). Component 3 had an emission maximum at 354 nm, an excitation maximum below 240 nm, and a shoulder at 294 nm. While this overlaps with the region of amino acids, free or bound in proteins (Murphy et al. 2008; Kowalczuk et al. 2009), and tryptophan-like components (Henderson et al. 2009), it was suspected that component 3 might represent the PAH phenanthrene (Wang et al. 2010). Phenanthrene fluoresced at the same wavelength region λ_ex/em: <250/348–384 nm (Wang et al. 2010 and Figure 2), although the shape of component 3 was rather unique. Component 4 revealed an emission maximum at 368 nm, an excitation maximum below 240 nm, and a secondary excitation band at 279 nm. Henderson et al. (2009) referred to it as tryptophan-like components. Component 5 had an emission maximum at 327 nm, an excitation maximum at 279 nm, and a shoulder below 240 nm. This peak was also found in other studies, and overlaps with the region of amino acids, free or bound in proteins (Murphy et al. 2008; Kowalczuk et al. 2009) and tryptophan-like fluorescence (Fellman et al. 2009; Singh et al. 2013). In addition, JiJi et al. (1999) reported two peaks that belong to the pesticides carbaryl (λ_ex/em: 270–320 nm) and 1-naphthol (λ_ex/em: 282–335 nm).

Although component 5 may relate to a combination of compounds, it was concluded that component 5 represents predominantly the herbicide DuPont Oust XP, as component 5 revealed two clear excitation and emission maxima (Figure 3). DuPont Oust XP exhibited exactly the same maxima with the same spectral shape (see Figure 2) while the protein peaks and the pesticide peaks of carbaryl and 1-naphthol revealed only one maximum. Component 6 had an emission maximum at 293 nm, an excitation maximum at 267 nm, and a shoulder below 240 nm. Fluorescence in this region was found in a broad range of environments and is thought to represent tyrosine-like components (Fellman et al. 2009; Singh et al. 2013) and tryptophan-like components (Fellman et al. 2009). In addition, JiJi et al. (1999) reported that the pesticide carbofuran fluoresced in that region (λ_ex/em: 267–295 nm). The spectral shape of carbofuran (Figure 2) only revealed one maximum, but clearly overlapped with the spectral shape of component 6.

Comparing the F_max values for each component from recharge basin to the monitoring wells, it can be summarized that F_max of each component was attenuated during water recharge (Figure 5). F_max can be reduced either by biodegradation of DOM or dilution with native groundwater (Hoppe-Jones et al. 2010). Since dilution is negligible for most of the lysimeters (Table 1), attenuation of the fluorescence signal is mainly related to microbial metabolism and/or sorption during travel through the subsurface. In general, F_max was higher for fulvic-like component 1 than any of the other components for each recharge event (Figure 5 and Table S1, Supplementary Information). While these findings suggest that all the samples were predominated by fulvic-like components, it has to be considered that fluorescence intensity is not only proportional to concentration but also to quantum yield (Baghoth et al. 2011). The differences in the relative intensities of the components are not only related to differences in concentration but to a combination of concentration and/or quantum efficiencies of the individual fluorophores.

The contribution of the herbicide DuPont Oust XP to the fluorescence signal (PARAFAC component 5) became obvious in samples ID90 (Hickory basin, recycled water recharge) and ID56 (Declez basin, imported water recharge), where a clear peak (peak B) was detected in the EEM that was very similar to the peak of DuPont Oust XP (Figure 2).

In sample ID97 (RP3 basin, recycled water recharge) F_max of component 6 was significantly higher (3.38 R.U). This was consistent with visual investigation of the EEM (Figure 6, peak C), suggesting that the peak in the EEM of sample ID97 was very similar to PARAFAC component 6 and possibly the pesticide carbofuran.

High F_max standard deviations were also identified for component 3 in basin samples during imported water recharge (Declez and RP3). The following imported water recharge samples revealed increased F_max values of component 3 and total fluorescence intensity: samples ID36, ID56 (both Declez basin), and ID50 (RP3 basin). F_max of component 3 in these samples were 9.78, 1.47, and 3.30 R.U., respectively. This was consistent with visual investigation of the EEMs (peak D in Figure 6). However, F_max of component 3 was almost zero in the source water (imported water, Figure 4) but significantly higher in the recharge basin samples (Figure 5), which suggests that component 3 was either already present in the basins or was discharged in some other way into the basin (e.g. surface runoff, dry deposition).

In summary, samples ID18, ID36, ID50, ID56, ID70, and ID97 revealed not only significantly higher F_max values compared to the other samples, but also different fluorescence shapes in the EEMs and/or increased fluorescence intensity, DOC concentrations and UV_254nm absorbance.

Monitoring the fate of chromophoric DOM during water recharge with 3D-fluorescence spectroscopy proved to be a sensitive and powerful tool. PARAFAC analysis was applied not only to identify and isolate different chromophoric DOM components, but also to derive semi-quantitative assessments and illustrate the decrease of chromophoric DOM in the subsurface over time and travel distance during groundwater recharge. In establishing a representative source water specific fingerprint as a baseline for MAR applications, some limitations were encountered, for example significant impact on UV and fluorescence signals of source water unspecific anthropogenic chemicals from non-point sources that are introduced to the recharge facilities (i.e. application of weed and vector control in the vicinity, wet and dry deposition at the basins). However, PARAFAC could identify, extract, and factor out at least one herbicide with chromophoric features from surface and groundwater EEMs.

This study was partially supported by the Metropolitan Water District of Southern California. We also acknowledge co-funding through the National Science Foundation (NSF) Engineering Research Center for Reinventing the Nation's Water Infrastructure (ReNUWIt) under cooperate agreement EEC-1028968.