Correlations between in situ sensor measurements and trace organic pollutants in urban streams

Abstract

Quantification of organic and microbial pollutant loading is expensive and labor-intensive because collection and analysis of grab samples are needed. Instruments are available, however, for in situ analysis of basic water quality parameters at high temporal resolution. Throughout the late summer and fall of 2008 a two-node water quality monitoring network was deployed to measure turbidity, specific conductance, pH, depth, temperature, dissolved oxygen, and nitrate at high frequencies in two urban streams in the Minneapolis, MN metropolitan area. Grab samples also were collected at 2 h intervals for 22 h during two dry periods and five rain events and analyzed for organic and microbial pollutants. This study investigated the viability of using in situ near real-time sensors to predict fecal coliforms, prometon (a residential herbicide), atrazine (an agricultural herbicide), and caffeine (a wastewater indicator) concentrations. Such pollutants can be used as indicators of sources that contribute to what is often termed “urban stream syndrome.” At one stream, linear correlations were observed between nitrate and caffeine (R² = 0.66), turbidity and prometon (R² = 0.91), and discharge and prometon (R² = 0.92). At another location, caffeine linearly correlated with specific conductance (R² = 0.64). A lack of correlation with sensed water quality parameters was also observed with some of the pollutants. When one considers that error is estimated to be as high as 200% when using monthly grab samples to estimate pollutant loading in streams, even moderate correlations, such as the ones found in this study, can provide better loading estimates if frequently sensed parameters can be used for load estimation. Therefore, such site-specific relationships can be used to estimate the loading of specific pollutants in near real-time until robust low-cost technologies to analyze these pollutants in situ become available.

---

Environmental impact

Quantification of organic and microbial pollutant loading is expensive and labor-intensive because collection and analysis of grab samples are needed. Instruments are available, however, for in situ analysis of basic water quality parameters at high temporal resolution. If correlations can be generated between these near real-time measurements and more difficult to measure organic and microbial pollutants, understanding of surface water dynamics and annual pollutant loading would be increased, manual labor requirements would be decreased, and resources could be better-managed. These results show that site-specific correlations are possible and that the error associated with these correlations is similar to that expected from analysis of infrequent grab samples.

Introduction

It is well-documented that urbanization has significantly accelerated the deterioration of surface water quality, a process often termed the “urban stream syndrome.”¹ Land development and the corresponding increase in impervious cover have altered the hydrology of watersheds. Furthermore, urbanization leads to increased application of fertilizers and pesticides and human and animal waste discharges (containing pharmaceutically active compounds (PhACs) and fecal indicator bacteria (FIB)) that have increased the overall pollutant load into urban water bodies. For example, increased sediment toxicity has been strongly correlated with increased pesticide application rates within a watershed.² Also, PhACs have been detected at biologically active concentrations within urban streams and effects on non-target organisms have been observed.^3–5 Finally, FIB are prevalent in urban streams and correlate with pathogen occurrence.⁶ As a result of these pollutant inputs, aquatic organisms inhabiting urban waterways and the humans who use those water resources for recreation or as a source of drinking water are potentially subjected to detrimental health effects.^3,4,7

Organic and biological contaminants are discharged into surface waters from both point and non-point sources and the loading of these contaminants can vary with time. Examples of point sources include wastewater treatment plant (WWTP) discharges and industrial wastewater discharges.^3,4,8,9 Stormwater run-off is a primary means by which non-point source organic and biological pollutants are transported into surface waters. Storm events can flush fertilizers, pesticides, and other constituents from agricultural and residential areas, mobilize shallow groundwater contaminants derived from leaky sewer lines and septic systems, as well as cause overflow within combined-sewer systems.¹⁰ Contaminant concentrations during individual storm events vary temporally based on source location, percent imperviousness of the surrounding area, rainfall intensity, as well as the physico-chemical properties of the contaminant.^11–13 In regions characterized by infrequent high-intensity storm events, such as the Midwest, storms often exhibit first-flush phenomena.¹¹ First-flush occurs when the majority of a pollutant load is discharged during the initial onset of a storm, and is controlled by rain intensity, percent imperviousness of the surrounding area, and duration of the antecedent dry period.¹⁰

The loading and transport of trace organic pollutants in the environment are not well understood.^2,13–15 Although it should be possible to determine the discharge of these compounds into surface waters, accurate quantification of dilute organic pollutant loading rates is prohibitively expensive and labor-intensive.¹² In anticipation of current and future total maximum daily load requirements for such compounds, effort is being placed on developing in situ monitoring instruments and methodologies to more accurately assess surface water quality and pollutant loads.¹⁶ Near real-time data could be used to determine the effectiveness of water quality management practices on current and emerging contaminants with the goal of reducing pollutant loading. For example, an extensive multiple-year study took place from 1995–1998 and 1999–2004 in Wichita, KS and the data collected from near real-time sensors were correlated with results from the analyses of periodic discrete samples.^17,18 Although the most accurate correlations were developed for inorganic pollutants (i.e. chloride and sulfate correlated with specific conductance and discharge), predictive relationships were also established for both atrazine (a function of month, discharge, and specific conductance) and fecal coliforms (a function of month and turbidity). In light of the success of previous studies, this study investigated the utility and feasibility of such correlations within a large urban environment. Subsequently, in this study the focus was on developing predictive relationships for urban and domestic wastewater pollutants based on measurements from near real-time in-stream sensors.

The objective of this study was to determine whether fundamental water quality parameters measured in situ could be correlated with trace organic contaminant and fecal indicator bacteria concentrations in urban streams. Thus, we collected and analyzed grab samples for four target analytes (prometon, atrazine, caffeine, and fecal coliforms) from streams equipped with sensors measuring fundamental water quality parameters at high frequency.

Experimental

Site locations

Shingle Creek. Shingle Creek begins in east-central Hennepin County, MN and travels 11.3 miles before discharging into the Mississippi River 3 miles north of downtown Minneapolis, MN. The 44.5 square mile watershed surrounding the creek comprises 9 urban and suburban municipalities, including Minneapolis, and several major transportation routes. No municipal WWTPs discharge into the creek. An unknown, but small percent of residential homes uses septic systems.¹⁹ The land use in the watershed is a mixture of dense residential neighborhoods (65%), commercial areas (20%), industrial areas (5%), and natural areas (10%). The region is highly impervious (>35%) as a result of urbanization, and the non-impervious areas comprise sandy to sandy-loam soils, which drain well.¹⁹ Typical flow rates within the creek range from 75 to 1000 L s⁻¹,¹⁹ however, flow rates ranged from 25 to 2000 L s⁻¹ during this study. Within the reach that was studied, there was a stormwater input that combined run-off from a several-mile stretch of a 6-lane highway (MN-HWY 100) with that from an approximately 5 square mile residential area. Additionally, there were approximately 2 square miles of commercial/industrial activity immediately upstream from the monitoring site. There was no agricultural activity in the area.

Minnehaha Creek. Minnehaha Creek begins in the west central portion of Hennepin County at Lake Minnetonka in Minnetonka, MN and flows 22 miles before discharging into the Mississippi River in Minneapolis, MN. The 181 square mile watershed comprises 29 municipalities, including Minneapolis. Near the headwaters, the landscape consists of single-family homes, parks, and recreational areas. An increase in industrial activity and imperviousness gradually occurs as the creek approaches the Mississippi River.²⁰ Within the watershed neither municipal WWTPs nor septic tanks are present. Flow into the creek from Lake Minnetonka is regulated by a dam and is a function of the current water levels in both the lake and the creek. Typical flow rates in the creek range from 1000 to 6000 L s⁻¹.²⁰ In 2008, low water levels within the lake resulted in the dam being closed for the year on August 26^th. This study was conducted from September 22^nd through October 30^th, thus all flow and pollutant loading within the creek were from localized water sources (i.e. rainfall or groundwater) and not from Lake Minnetonka. Consequently, creek flow rates during this monitoring period ranged from 200 to 1200 L s⁻¹ and averaged 610 L s⁻¹. Within the location studied, run-off from several miles of a 6-lane highway (I-494) and stormwater inputs from residential areas discharged into the creek. Approximately 30% of the watershed is devoted to agriculture, however, there was no agricultural activity within 10 miles of the monitoring location and no industrial activity upstream from the study site.

Target analytes

The target analytes for this study were prometon, atrazine, caffeine, and fecal coliforms. Atrazine is a commonly used agricultural herbicide and has a relatively low K_oc value (93), low Henry's law constant (2.36 × 10⁻⁹ atm m³ mol⁻¹), and moderate half-life in aerobic surface waters (13–60 days).²¹ In this study atrazine served as an indicator of agricultural pollution. Prometon is an herbicide that is found in many household weed-killing agents²² and is commonly mixed into asphalt to prevent plant growth within roads and driveways.²³ Prometon has a relatively low K_oc value (124), low Henry's law constant (9.09 × 10⁻¹⁰ atm m³ mol⁻¹), and long half-life in aerobic surface waters (30–600 days) and in this study served as an indicator of urban stormwater input.²² Caffeine is relatively abundant in raw domestic wastewater (7–73 µg L⁻¹) with a low K_oc value (∼0), low Henry's constant (1.9 × 10⁻¹⁹ atm m³ mol⁻¹), and moderate half-life in aerobic surface waters (50–1000 days).²⁴ Hence, caffeine served as an indicator of wastewater inputs. Compounds with high K_oc values were not monitored in this study.

Network description

A wireless water quality monitoring network was deployed in Shingle Creek (mile 3) from August 8^th, 2008 to September 17^th, 2008 and in Minnehaha Creek (mile 21) from September 19^th, 2008 to October 30^th, 2008. The network comprised two monitoring stations, each equipped with a HydroLab Data Sonde DS5X (Hach Environmental, Loveland, CO), an ISCO Model 6712 automatic sampler (Lincoln, NE), a MicroLab nutrient analyzer (Enviro-Tech, LLC, Chesapeake, VA), a solar panel (Campbell Scientific, Logan, UT), a 12 V marine battery for supplemental power to the MicroLab analyzer (Attwood Marine Products, Lowell, MI), a radio antenna, a datalogger (CR1000 or CR206, Campbell Scientific), and a power supply (PS100, Campbell Scientific). Additionally, one of the stations was equipped with a rain gauge (Texas Electronics, Dallas, TX) and the other was connected to a cellular-phone modem for communication.

The HydroLab Data Sonde DS5X was equipped with highly precise, EPA-certified sensors that measured dissolved oxygen (DO), specific conductance, turbidity, depth, pH, and temperature every minute. Cleaning was performed as necessary (typically once per week) and the sondes were calibrated every four weeks, as per manufacturer recommendation. Prior to each calibration event, prepared standards were analyzed to determine sensor drift. Throughout this study minimal (<1%) drift was observed for each sensor.

The MicroLab is an in situ nutrient analyzer that measures either nitrogen (nitrate plus nitrate) or phosphorus at programmable intervals as frequent as every 20 min (Enviro-Tech, LLC, 2005). In this study, the analyzer was set up to monitor nitrate via the automated cadmium reduction method (4500-NO3-F)²⁵ at 2 h intervals. Although more frequent sampling was possible, due to power supply limitations, a sampling frequency of 2 h was chosen. Power was predominately supplied to the MicroLab by the solar panel; nevertheless, supplementary power was needed and supplied via a 12 V battery that was recharged in the laboratory on a weekly basis. The instrument was automatically recalibrated after every 5 samples.

The ISCO Model 6712 sampler was programmed to collect twelve 900 mL samples at 2 h intervals over a 22 h period into 1 L solvent-washed glass jars. Residual air and water were purged from the 3/8 inch Teflon tubing before and after each sample was collected. The ISCO was connected directly to a 12 V battery for power. Within 30 min after the last sample was collected, the jars were removed from the autosampler, sealed with Teflon-lined caps, and transported to the laboratory on ice for subsequent analysis of FIB and organic contaminants.

Each piece of equipment was wired through the datalogger, which temporarily stored data and was used to initiate and terminate sampling. The radio antennas enabled communication between monitoring stations and the modem enabled communication with, and control over, the network from a remote location. Data were automatically transferred from the network to a server at the University of Minnesota campus every 24 h.

Monitoring strategy

Two monitoring stations were deployed approximately 200 m apart, one upstream and one downstream from a stormwater input. Throughout the duration of the study (except during times of equipment failure), each HydroLab and MicroLab collected data on a near-continuous basis. Grab samples were collected by the ISCO samplers on four occasions at the Shingle Creek site (one dry period and three rain events) and on three occasions at the Minnehaha Creek site (one dry period and two rain events).

Grab sample collection and analysis

Enumeration of fecal coliforms. The membrane filtration technique (Method 9222 D)²⁵ was used to enumerate fecal coliforms. Sterile water blanks were run as controls, as described by the method. The sample jars were removed from the autosampler at the end of the 22 h sampling period and processed within 2 h. Hence, the period of time between collection and analysis varied from 2 h for the final sample to 24 h for the first sample. To evaluate the effect of storage time, the final sample was analyzed with the rest of the samples and again 22 h later. No significant difference (<5%) was observed between these replicate analyses. Aliquots (100 µL to 10 mL) were removed from each sample jar for coliform analysis and the remainder of the sample was used for trace organic analysis.

Caffeine, atrazine, and prometon. The water samples were stored at 4 °C in the dark and were analyzed within one week. A modification of the EPA Method 525.2 was used for analysis of prometon and caffeine. All of the protocols specified were followed except as noted below. Briefly, water samples were brought to room temperature and then filtered through a Whatman GF/F glass fiber filter using a glass vacuum filtration apparatus. The volume of filtrate was determined gravimetrically. Using a Chaney adaptor and a 100 µL glass syringe, 100 µL of a 1 mg L⁻¹ terbutylazine solution in isopropanol (Pestanal reagents, Crescent Chemical Company, Islandia, NY) were added to each water sample prior to extraction to serve as a surrogate (recovery 88 ± 9.1%, n = 160).

The analytes were concentrated using solid phase extraction (SPE). SPE cartridges (Oasis HLB 200 mg, 6 mL, Waters Corporation, Milford, MA) were preconditioned with 3 mL of methanol and then 3 mL of ultra-pure water as per manufacturer recommendation. The filtered samples (∼900 mL) were then passed through the SPE cartridges. After sample loading, the analytes were eluted by passing 3.0 mL of 3 : 1 hexane : isopropanol (by gravity) through the extraction cartridge, through a pipette filter (to remove residual water), and into a tapered glass collection vial. The solution was blown-down to approximately 100 µL and was transferred to a 100 µL gas chromatography (GC) vial inserted within an amber GC vial. A phenanthrene-d₁₀ internal standard solution (10 µL of a 20 mg L⁻¹ solution, Supelco, Park Bellefonte, PA) was added to each extracted sample. The volume of the final extract was determined gravimetrically. The vial was crimped closed and placed in a freezer (to prevent solvent or analyte volatilization) and analyzed via gas chromatography-mass spectrometry (GC-MS) within 7 days.

Extracted samples (1 µL) were injected into a Hewlett Packard (HP) G1800A System GC-MS (Santa Clara CA) by an HP G1512A autosampler controller. The injector was operated in a splitless mode at 250 °C and compound separation was achieved on a 30 m RTX-5 capillary column (Restek, Bellefonte, PA) with an internal diameter of 0.25 mm and a film thickness of 0.25 µm. Additionally, a single tapered gooseneck splitless liner (Restek, Bellefonte, PA) was used, and was replaced as needed (typically every 40 samples). Within the GC, the following gradient program was applied upon sample injection: isothermal at 70 °C for 2 min, ramp up 25 °C min⁻¹ to 180 °C, ramp up 7 °C min⁻¹ to 210 °C, and ramp up 25 °C min⁻¹ to 300 °C, where the temperature was held for 5 min. For increased sensitivity, the mass spectrometer was operated in a single ion monitoring (SIM) mode and only the characteristic m/z ratios associated with prometon (m/z: 210, 183, 225), atrazine (m/z: 200, 173, 138), caffeine (m/z: 194, 109, 82), terbutylazine (m/z: 173, 138, 231), and phenanthrene-d₁₀ (m/z: 188) were quantified at the predetermined elution times. The areas under the curve were calculated using the manual integration option on HP Chem-Station's data analysis package. The method detection limits (MDLs) were 42 ng L⁻¹, 40 ng L⁻¹ and 18 ng L⁻¹ for prometon, atrazine, and caffeine, respectively.

Two sets of standard solutions of prometon, atrazine, terbutylazine, and caffeine (Pestanal reagents, Crescent Chemical Company, Islandia, NY) were prepared and used in the creation of a calibration curve. Using a Chaney adaptor, known additions of phenanthrene-d₁₀ internal standard were added to each solution. Plots of analyte concentration divided by internal standard concentration versus target analyte MS area divided by internal standard MS area were created and used as calibration curves.

Solutions (∼900 mL) containing known concentrations (25 ng L⁻¹ to 1 µg L⁻¹) of prometon, atrazine, and caffeine were analyzed to quantify method recoveries, and blanks containing only the terbutylazine surrogate in ultra-pure water were analyzed to quantify contamination. Recovery studies were performed in both ultra-pure HPLC-grade water (93 ± 8.0%, 97 ± 8.0%, 80 ± 7.0%, and 63 ± 8.0% for prometon, atrazine, terbutylazine, and caffeine, respectively, for 20 samples) and creek water (114 ± 1.0%, 117 ± 12.0%, 88 ± 7.0%, and 72 ± 9.0% for prometon, atrazine, terbutylazine, and caffeine, respectively, for 11 samples). Approximately 20% of all samples analyzed were dedicated to recovery studies. Values were not adjusted for recovery. No target analytes were detected in any blanks, confirming sufficient cleaning procedures. One 3 L environmental sample was split into 3 aliquots and analyzed for prometon, atrazine, and caffeine to determine analytical reproducibility; the mean concentrations (±standard error) were 103 ± 0.70 ng L⁻¹, non-detected, and 183 ± 17 ng L⁻¹, for prometon, atrazine, and caffeine, respectively.

Results and discussion

Caffeine

Caffeine was detected in 100% (n = 102) of the samples collected from Shingle Creek and in 33% (n = 60) of the samples collected from Minnehaha Creek. Within Shingle Creek, concentrations of caffeine ranged from 39 to 1030 ng L⁻¹, and generally increased after the initial rainfall that followed a prolonged dry period (Fig. 1). The consistent detection of caffeine suggests contamination of the creeks by domestic wastewater. Because there are no direct wastewater discharges to the creeks, the wastewater inputs are due to either leaky sewer lines or septic drainage fields. Such subsurface discharges of wastewater are likely transported to the creeks via shallow groundwater. Shallow groundwater contamination is characterized by slow-moving diffuse plumes with increasing concentrations in the direction of the source.²⁶ Infiltrated rainfall accelerates the mobility of the plume and can increase the pollutant loading into surface waters.²⁶ Based on the caffeine concentrations in raw domestic wastewater (7–73 µg L⁻¹)²⁴ and the range of concentrations observed in Shingle Creek, the creek could contain 0.1–14% wastewater.

Fig. 1 Results collected from Shingle Creek on (a) August 11^th to 12^th, (b) August 18^th to 19^th, (c) August 27^th to 28^th, and (d) September 10^th to 12^th. Each vertical grouping represents an individual sampling event. Parameter concentrations are displayed on the primary vertical axis and rainfall is displayed on the secondary vertical axis. Solid lines correspond to the upstream station, dashed lines correspond to the downstream station, and the inverted solid areas correspond to rainfall intensity. Error bars represent the standard deviation of replicate analyses.

At Shingle Creek, increases in caffeine concentrations occurred after small (<5 mm) rain events (Fig. 1c and d). The soils (sandy to sandy-loam) within this drainage area allow for infiltration rates ranging from 0.34–1.09 cm h⁻¹.²⁷ Thus, the transport of caffeine into the creek was likely accelerated by stormwater that infiltrated into the ground above septic system drainage fields or leaky sewer lines. Additionally, the upstream and downstream caffeine concentrations were similar (<10% difference) suggesting that the local stormwater input did not contain any wastewater.

Assuming a consistent, gradual leak of wastewater or treated septage, the size of the plume should increase with time during dry periods. Indeed, longer antecedent dry periods appeared to result in increased caffeine loading. The rain event monitored on August 27^th was preceded by a dry period of 14 days, and resulted in 1300 mg of caffeine being discharged following the initial 4 mm of rain (Fig. 1c). After 10 h, the caffeine concentration returned to base-flow conditions, at which point a high-intensity (∼2.5 cm h⁻¹) rainfall occurred that exceeded the soil infiltration capacity and produced overland flow (visually confirmed). A sharp reduction in caffeine concentrations was observed, again suggesting the transport mechanism for caffeine into the creek was through the groundwater. A 4-day dry period occurred prior to the rain event on September 11^th, and resulted in 630 mg of caffeine being discharged to the creek following the initial 3.5 mm of rain (Fig. 1d). After 6 h of elevated concentrations, caffeine returned to base-flow conditions as additional rainfall (∼3 mm) occurred, suggesting that the groundwater caffeine plume was flushed into the creek with the initial rainfall. The rain event that occurred on August 12^th (3.3 mm), following a 2-day antecedent dry period, did not result in an increase in caffeine concentration (Fig. 1a). Therefore, we speculate that two days provided insufficient time for the caffeine plume to recharge. The on-site rain gauge was not in operation during this event so a rainfall estimate was obtained from the Crystal Airport located <2 miles off-site. The airport readings were within 5–10% of the on-site rain gauge for the other two rain events.

Caffeine concentrations within Minnehaha Creek also increased upon rainfall, but only at the downstream station (Fig. 2). The 9 mm rain event on October 5^th was preceded by a 5-day dry period and resulted in the cumulative discharge of approximately 550 mg of caffeine over 8 h (Fig. 2b). On October 13^th, a 6 mm rain event was preceded by a 7-day dry period, and approximately 760 mg of caffeine were discharged over 8 h (Fig. 2c). Prior to each rain event, caffeine concentrations were below the analytical MDL (18 ng L⁻¹), and slowly decreased to this level again after the rain-induced increase in caffeine concentrations.

Fig. 2 Results collected from Minnehaha Creek on (a) September 29^th to 30^th, (b) October 5^th to 6^th, and (c) October 12^th to 13^th. Each vertical grouping represents an individual sampling event. Parameter concentrations are displayed on the primary vertical axis and rainfall is displayed on the secondary vertical axis. Solid lines correspond to the upstream station, dashed lines correspond to the downstream station, and the inverted solid areas correspond to rainfall intensity. Error bars represent the standard deviation of replicate analyses.

Similar to Shingle Creek, the longer the antecedent dry period at the Minnehaha Creek site, the greater the cumulative discharge of caffeine. Unlike Shingle Creek, however, caffeine was not continuously detected during base-flow conditions (Fig. 2). This suggests there was not a consistent and significant discharge of caffeine into the shallow groundwater near the Minnehaha Creek site. This difference could be accounted for by the absence of septic tanks within the Minnehaha Creek watershed. During rain events, however, stormwater discharge from the retention ponds between the two monitoring stations was a clear source of caffeine (Fig. 2b and c). A leaky sewer line or a combined-sewer overflow may have been present within the drainage area feeding the ponds. Based on the caffeine concentration range in raw domestic wastewater and the observed concentrations in Minnehaha Creek, wastewater could comprise 0.1–1% of the stream flow during peak caffeine concentrations (directly after rain events). Because the retention ponds received stormwater run-off from a several-mile stretch of a 6-lane highway as well as run-off from local residential streets, it is also possible caffeine could be associated with a non-wastewater source (e.g., 350 mL (12 oz) of coffee contains approximately 100 mg of caffeine).

In drainage areas that contain nitrate-contaminated soils and groundwater, nitrate loading into nearby surface waters is characterized by first-flush behavior.^11,28 Indeed, nitrate loading within Shingle Creek increased with initial rainfall and concentrations remained elevated for several hours. The potential for correlations between caffeine and nitrate during rain events was therefore investigated (Fig. 3). These site-specific results suggest nitrate (independent variable) is a good predictor (T-stat > 1.96, P-value < 0.05) of caffeine (Fig. 3). Indeed, within the data obtained from Shingle Creek, nitrate concentrations could be used to estimate caffeine concentrations with an average error of 40% relative to measured caffeine concentrations. The majority (75%) of the caffeine measurements could have been estimated within 56% using nitrate concentration as a surrogate indicator. The minimum and maximum errors between the estimated and measured concentrations were 3 and 150%, respectively.

Fig. 3 Correlation plot between two analytes that demonstrated first-flush behavior at Shingle Creek: caffeine and nitrate. Data are from Shingle Creek from August 27^th to 28^th (red squares, C_o = 104 µg L⁻¹), September 10^th to 12^th (green triangles, C_o = 240 µg L⁻¹), and August 18^th to 19^th (purple circles, C_o = 137 µg L⁻¹).

At the Minnehaha Creek site, caffeine did not correlate with nitrate (R² = 0.13). Because of instrument failure, only one of the two rain events was monitored for nitrate (Fig. 2b). In addition, the upstream MicroLab was not in operation throughout the duration of this study, providing fewer nitrate data points. Nevertheless, the Minnehaha Creek study site was located downstream from a natural area comprising approximately 1 square mile of wetlands and wooded areas. Wetlands in particular reduce nitrate discharge through plant uptake or biological degradation.²⁹ As a result, it would be expected that groundwater contamination of nitrate would be less significant than at the Shingle Creek site and that it may not exhibit first-flush behavior as at Shingle Creek.

A moderate negative correlation (R² = 0.64) between caffeine and specific conductance was observed at the downstream station in Minnehaha Creek (Fig. 4). The ionic activity (specific conductance) of rain is considerably lower than that of surface waters. Because caffeine enters the creek with stormwater from the ponds, caffeine was inversely correlated with specific conductance. Although the correlation is moderate, presumably from a paucity of data, specific conductance measurements could be used to estimate caffeine concentrations at Minnehaha Creek with an average error of 14% relative to measured caffeine concentrations. Additionally, the majority (75%) of the caffeine measurements could have been estimated within 21% error using specific conductance as a surrogate indicator, and the minimum and maximum errors between the estimated and measured concentrations were 0.5 and 51%, respectively.

Fig. 4 Correlation plot between caffeine and specific conductance in Minnehaha Creek where the source of caffeine was a stormwater pond and specific conductance shows the influence of rainwater in stormwater flow. Data are from Minnehaha Creek from October 5^th to 6^th (red squares, C_o = 1109 µS cm⁻¹) and October 12^th to 13^th (green triangles, C_o = 895 µS cm⁻¹).

Prometon

Prometon was detected in 45% (n = 102) of the samples collected from Shingle Creek. It was not detected, however, in any (n = 60) samples from the Minnehaha Creek site. Within the rain events studied at Shingle Creek, the only event that resulted in increased prometon discharge occurred where the rainfall intensity (2.5 cm h⁻¹) exceeded the soil infiltration capacity (0.34–1.09 cm h⁻¹). These results, therefore, suggest that prometon transport was governed by overland flow, which is consistent with previous observations.¹³ During the event on August 28^th (Fig. 1c), prometon concentrations increased from less than the MDL (42 ng L⁻¹) to approximately 550 ng L⁻¹. Prometon linearly correlated with stream discharge (R² = 0.92) (Fig. 5) as well as with turbidity (R² = 0.91) during the storm event that produced overland flow conditions (Fig. 6). Prometon would be expected to accumulate in near-surface soils upon pesticide application, after which it could be flushed to surface waters via overland flow. Overland flow conditions occur following intense (>1 cm h⁻¹) rainfall, which also results in increased stream discharge and suspended sediment. These correlations are therefore expected. The average error between measured and estimated prometon concentrations was 20 and 11% using turbidity and stream discharge as surrogate indicators, respectively. Although the results suggest good predictor relationships (T-stat > 1.96, P-value < 0.05), these correlations are limited by the lack of data, as only one storm event produced overland flow conditions. Further data are needed to determine if this correlation holds for additional high-intensity rain events.

Fig. 5 Correlation plot between prometon and creek discharge during the event when overland flow was observed. Data are from Shingle Creek from September 10^th to 12^th.

Fig. 6 Correlation plot between prometon and turbidity during the event when overland flow was observed. Data are from Shingle Creek from September 10^th to 12^th.

Atrazine

Atrazine was not detected (MDL = 40 ng L⁻¹) in any samples from Shingle Creek (n = 102) or from Minnehaha Creek (n = 60). No agriculture was present within several miles of either sampling location and samples were collected a few months after the period when atrazine is typically applied.¹³

Fecal coliforms

We attempted to correlate fecal coliforms with turbidity, nitrate, and stream discharge; R² values, however, were consistently below 0.1. Similar to previous studies, a suitable surrogate for fecal coliforms remained elusive. The abundance and variety of non-point sources of these bacteria are unparalleled by other water quality parameters, making it extremely difficult to develop predictive correlations. For example, overland flow could result in run-off of fecal coliforms from animal droppings; as a result, turbidity or particle-associated pollutants could be used as potential indicator variables. Fecal coliforms could also be found in septic tank discharge (if not operating properly) or in leaky sewer pipes, in which case concentrations may behave similar to shallow groundwater contaminants (e.g., nitrate, caffeine). In previous studies on larger streams, fecal coliforms correlated with time of year and turbidity.¹⁷ Smaller creeks with low flow rates, like Shingle Creek and Minnehaha Creek, however, are more influenced by temporally variable inputs than larger streams. Because small creeks are often shallow (<1 m deep), the majority of the flow may be subject to solar irradiation during day-time hours.¹⁸ If solar-induced inactivation occurs, incorporation of solar radiation within a multi-parameter regression model could help account for such diurnal variability. Additional data would be needed to develop this kind of multi-parameter site-specific surrogate relationship.

Within the samples collected from both creeks, the chronic Minnesota water quality standard (200 colonies per 100 mL) was exceeded, even under base-flow conditions (Fig. 1 and 2). Additionally, concentrations were as high as 40 000 colonies per 100 mL during high-flow events (Fig. 1c). In general, fecal coliform concentrations increased with flow rate and turbidity (Fig. 1 and 2). The magnitude and duration of elevated bacterial concentrations were variable, however, and as mentioned above, could not be predicted by other parameters measured in this study.

Molecular methods such as BOX-PCR or PCR of mitochondrial DNA followed by a dot-blot assay could be used to determine the sources of fecal bacteria (e.g., human or animal waste),^30,31 which in turn could enable correlations between a particular bacteria source (e.g., humans) and various water quality parameters. Indeed, this would be useful as there would be greater concern over fecal bacteria of a human origin.

Engineering implications

Estimations of bacterial and organic pollutants on a near real-time basis would be useful when considering recreational activities, adjustments in water treatment strategies, or in calculating pollutant loads for current and future regulations. Although the surrogate relationships between easily sensed water quality parameters and specific organic pollutants were site-specific and only moderately strong (R² = 0.64–0.92), our research showed that it is possible to develop site-specific relationships. In fact, despite a seemingly high potential of average error (up to 40%), the errors associated with the derived predictive relationships are within an already accepted error threshold by US regulatory agencies. For example, pollutant loading is commonly estimated from discrete samples collected on a bi-monthly to monthly basis.^32,33 The average loading estimation errors associated with such infrequent sampling range from 19–200% relative to more accurate methods, such as near real-time continuous monitoring.³⁴ Therefore, annual loading estimation errors associated with collection of infrequent discrete samples can exceed the regression errors of resulting surrogate relationships.¹⁷ Additionally, after establishing a surrogate relationship, loading estimations can be derived from in situ sensor data and the need for frequent grab sample collection can decrease, thus reducing manual labor requirements. Field visits could be minimized and limited resources could be more effectively managed. Although compounds with high K_oc values were not monitored in this study, such compounds would be expected to correlate with turbidity or suspended solids.

Previous researchers have suggested that two years of correlative data over 95% of a given stream discharge range are required to develop accurate surrogate relationships.¹⁷ Christensen et al.¹⁷ also reported that additional data collection beyond those two years did not significantly (<10%) improve the correlations. It is therefore suggested that correlations such as the ones developed in this study be established over a longer sampling period (two years) for regulatory use. If hydrological changes or mitigation practices were implemented within a watershed, such as the separation of combined sewers or the addition of impervious surfaces, these relationships would change and supplementary sampling would be required to update the correlations. If such correlations were used for regulatory purposes, the accuracy of the correlation needed would depend on the pollutant in question and the particular stream being regulated.

Conclusions

Predictive models were developed that enabled the estimation of organic pollutants from in situ measurements of basic water quality parameters. Such predictive regression models can be used to estimate organic pollutant concentrations in near real-time when in situ water quality sensors are deployed. This could increase the understanding of surface water dynamics and annual pollutant loading, reduce manual labor requirements associated with surface water monitoring, and increase resource management for surface water and mitigation assessment. There are limitations to the results of this study, the greatest of which was simply a lack of data, particularly throughout multiple seasons. Indeed, a lack of correlation with sensed water quality parameters was also observed with some pollutants, fecal coliforms for example. Additional research is required to collect data over multiple seasons, spanning a wide range of flow rates and pollutant loading rates to develop accurate predictive models. Nevertheless, the use of near real-time sensor data to predict organic contaminant concentrations was shown to be viable, particularly when one considers that error can be as high as 200% when using monthly grab samples to estimate pollutant loading in streams. Even moderate correlations, such as the ones found in this study, can provide better loading estimates if frequently sensed parameters can be used for load estimation. Such data were also useful in identifying sources of pollution (e.g., a particular stormwater pond versus groundwater pollution). This type of work is therefore also useful for developing water resource mitigation strategies.

Acknowledgements

This work was supported by the National Science Foundation (EAR 0607138 and CBET 0414388) and by the National Institutes for Water Resources/US Geological Survey (2006MN187G). Thanks to Miki Hondzo for his input, Kevin Drees for assistance with the GC/MS, and Paul Capel for assistance with analytical method development.

References

1. C. J. Walsh, H. R. Allison, J. W. Feminella, P. D. Cottingham, P. M. Groffman and R. P. Morgan, II, J.N. Am. Benthol. Soc., 2005, 24, 706.
2. J. W. Hunt, B. S. Anderson, B. M. Phillips, R. S. Tjeerdema, N. Richard, V. Connor, K. Worcester, M. Angelo, A. Bern, B. Fulfrost and D. Mulvaney, Environ. Monit. Assess., 2006, 121, 245.
3. B. E. Haggard, J. M. Galloway, G. W. Reed and M. T. Meyer, J. Environ. Qual., 2006, 35, 1078.
4. D. W. Kolpin, M. Skopec, M. T. Meyer, E. T. Furlong and S. D. Zaugg, Sci. Total Environ., 2004, 328, 119.
5. J. P. Seiler, Toxicol. Lett., 2002, 131, 105.
6. S. Bai and W.-S. Lung, J. Environ. Sci. Heal. A, 2006, 41, 1327.
7. M. W. Sandstrom, D. W. Kolpin, M. E. Thurman and S. D. Zaugg, Environ. Toxicol. Chem., 2005, 24, 1029.
8. J. Hur, M. A. Schlautman, T. Karanfil, J. Smink, H. Song and S. J. Klaine, Environ. Monit. Assess., 2007, 132, 171.
9. C. J. Lee and T. J. Rasmussen, Sci. Total Environ., 2006, 371, 258.
10. L. Asaf, R. Nativ, D. Shain, M. Hassan and S. Geyer, J. Hydrol., 2004, 294, 270.
11. P. L. Brezonik and T. H. Stadelmann, Water Res., 2002, 36, 1743.
12. A. W. Fogle, J. L. Taraba and J. S. Dinger, J. Am. Water Resour. As., 2003, 39, 1361.
13. K. E. Hyer, G. M. Hornberger and J. S. Herman, J. Hydrol., 2001, 254, 47.
14. W. E. Asher, W. Luo, K. W. Campo, D. A. Bender, K. W. Robinson, J. S. Zogorski and J. F. Pankow, Environ. Toxicol. Chem., 2007, 26, 1606.
15. C. J. Diblasi, H. Li, A. P. Davis and U. Ghosh, Environ. Sci. Technol., 2009, 43, 494.
16. S. L. MacLeod, E. L. McClure and C. S. Wong, Environ. Toxicol. Chem., 2007, 26, 2517.
17. V. G. Christensen, X. Jian and A. C. Ziegler, Regression analysis and real-time water-quality monitoring to estimate constituent concentrations, loads, and yields in the Little Arkansas River, South-Central Kansas, 1995–1999, U.S. Geological Survey Water-Resources Investigations Report 00-4126, 2000.
18. T. J. Rasmussen, A. C. Ziegler and P. P. Rasmussen, Estimation of constituent concentrations, densities, loads, and yields in Lower Kansas River, Northeast Kansas, using regression models and continuous water-quality monitoring, January 2000 through December 2003, U.S. Geological Survey Scientific Investigations Report 2005-5165, 2008.
19. Wenck & Associates, Inc., Shingle Creek Chloride TMDL Report, 2006, http://www.shinglecreek.org/tmdlimplementation.pdf.
20. Y. Christianson, J. Carlson and U. B. Singh, Minnehaha Creek Watershed District 2006 Hydrologic Data Monitoring Report, 2007, http://www.minnehahacreek.org/documents/Hydrodata_Report_2006_full.pdf.
21. US EPA, Estimation Programs Interface Suite™ for Microsoft® Windows, v. 4.00, United States Environmental Protection Agency, Washington, DC, 2009.
22. S. E. Kegley, B. R. Hill, S. Orme and A. H. Choi, PAN Pesticide Database, Pesticide Action Network, 2009, http://www.pesticideinfo.org/%5D.
23. P. D. Capel, A. H. Spexet and S. J. Larson, Environ. Sci. Technol., 1999, 33, 674.
24. I. J. Buerge, T. Poiger, M. D. Muller and H.-R. Buser, Environ. Sci. Technol., 2003, 37, 691.
25. Standard Methods for the Examination of Water and Wastewater, APHA, AWWA, and WEF, Washington, DC, 19th edn, 1995.
26. K. A. Shepherd, P. A. Ellis and M. O. Rivett, Sci. Total Environ., 2006, 360, 180.
27. L. W. Mays, Water Resources Engineering, 1st edn, John Wiley & Sons, Inc., Hoboken, NJ, 2001.
28. D. O. Hessen, A. Henriksen and A. M. Smelhus, Water Res., 1997, 31, 1813.
29. G. Oberts and A. Plevan, Benefits of Wetland Buffers: A Study of Functions, Values and Size, Minnehaha Creek Watershed District, 2001, http://www.minnehahacreek.org/documents/MCWD_Buffer_Study.pdf.
30. P. E. Dombek, L. K. Johnson, S. T. Zimmerley and M. J. Sadowsky, Appl. Environ. Microbiol., 2000, 66, 2572.
31. R. Kortbaoui, A. Locas, M. Imbeau, P. Payment and R. Villemur, Water Res., 2009, 43, 2002.
32. Minnesota Pollution Control Agency, Minnesota's Water Quality Monitoring Strategy 2004 to 2014: A Report Prepared for the U.S. Environmental Protection Agency, 2004, http://proteus.pca.state.mn.us/water/pubs/wqms-report.html.
33. A. J. Horowitz, K. A. Elrick and J. J. Smith, Results from the City of Atlanta Water-Quantity and Water-Quality Monitoring Program: Suspended Sediment, Trace Element, and Nutrient Fluxes, 2004–2005, in 2007 Georgia Water Resources Conference, University of Georgia, Athens, Georgia, 2007, http://cms.ce.gatech.edu/gwri/uploads/proceedings/2007/6.2.3.pdf.
34. M. B. Henjum, R. M. Hozalski, C. R. Wennen, P. J. Novak and W. A. Arnold, J. Environ. Monitor., 2009 10.1039/b912990a.

---

Footnote

† Part of a themed issue dealing with water and water related issues.

---