- Top of page
- MODEL DEVELOPMENT
- CALIBRATION AND EVALUATION
- SUPPLEMENTAL DATA
- Supporting Information
A complex interplay of factors determines the degree of bioaccumulation of Hg in fish in any particular basin. Although certain watershed characteristics have been associated with higher or lower bioaccumulation rates, the relationships between these characteristics are poorly understood. To add to this understanding, a dynamic model was built to examine these relationships in stream systems. The model follows Hg from the water column, through microbial conversion and subsequent concentration, through the food web to piscivorous fish. The model was calibrated to 7 basins in Kentucky and further evaluated by comparing output to 7 sites in, or proximal to, the Ohio River Valley, an underrepresented region in the bioaccumulation literature. Water quality and basin characteristics were inputs into the model, with tissue concentrations of Hg of generic trophic level 3, 3.5, and 4 fish the output. Regulatory and monitoring data were used to calibrate and evaluate the model. Mean average prediction error for Kentucky sites was 26%, whereas mean error for evaluation sites was 51%. Variability within natural systems can be substantial and was quantified for fish tissue by analysis of the US Geological Survey National Fish Database. This analysis pointed to the need for more systematic sampling of fish tissue. Analysis of model output indicated that parameters that had the greatest impact on bioaccumulation influenced the system at several points. These parameters included forested and wetlands coverage and nutrient levels. Factors that were less sensitive modified the system at only 1 point and included the unfiltered total Hg input and the portion of the basin that is developed. Integr Environ Assess Manag 2012; 8: 709–722. © 2012 SETAC
- Top of page
- MODEL DEVELOPMENT
- CALIBRATION AND EVALUATION
- SUPPLEMENTAL DATA
- Supporting Information
Mercury (Hg) is a potent neurotoxin with the developing fetus at greatest risk for adverse outcomes. In the 1999–2002 National Health and Nutrition Examination Study, 6% of women of childbearing age were found to have blood Hg levels at or above the US Environmental Protection Agency (USEPA) level of safety (CDC 2004). The largest source of exposure to Hg is from fish consumption (USEPA 2009). As of 2010, Hg contamination is responsible for fish consumption advisories on over 1.2 million miles (1.9 million km) of the nation's rivers (USEPA 2011). The problem is both national and global in perspective and without a clear understanding of the relationship between the environmental and physiological disposition of Hg, effective strategies to reduce contamination have not been forthcoming.
The source of Hg contamination in most stream systems is from atmospheric deposition, both from global and more local or regional sources. Atmospheric Hg makes its way into aquatic systems through wet and dry deposition either directly to the waterway or through terrestrial compartments. Organic matter in soil often sequesters Hg, releasing it slowly over time. A whole ecosystem study of changes in loading of Hg to a lake system found an initial rapid response. However, the rapid change came from the Hg that was deposited directly to the lake surface and a much slower response was noted in the loading of Hg from the watershed (Harris et al. 2007). For stream systems, which receive almost all Hg loading from the watershed, reductions in local and regional emissions of Hg may not result in substantial reductions in fish tissue levels for some time. Fish tissue Hg levels can vary widely between basins, and understanding the different characteristics of these basins may help regulators determine which stream systems are most likely to have elevated Hg levels and also envisage the impact proposed changes in a watershed will have on fish tissue levels.
The dynamics of Hg transport and transformation in stream systems have received increased research attention but are still not well understood. The degree of bioaccumulation in a system is dependent on the interaction of multiple variables, and such complex systems often react to changes in unexpected ways. Although the dynamics are complex, several factors have been found to be associated with higher fish tissue Hg concentrations in stream systems. Ward, Nislow, and Folt (2010b) have designated this collection of watershed characteristics “bioaccumulation syndrome.” Characteristics associated with increased bioaccumulation of methylmercury (MeHg) include connectivity to wetlands, low pH, and factors that contribute to lower levels of productivity such as low nutrient levels and light limitation. Every stream system has a unique set of characteristics that interact to either increase or decrease bioaccumulation.
A mass balance, mechanistic model was built that examines the movement of Hg in stream systems from water column to fish tissue. The model is controlled by a series of simulated mechanisms or processes that control the movement of Hg through the system. The model is dynamic in the sense that, even at steady state, the dynamics of Hg can be quantified as it moves through the various stock and flow components. The model was also constructed to be dynamic in the sense of being responsive to changes in Hg loading as a result of potential policy scenarios. If the relevant components and relationships of the system are properly accounted for, the model can be used to credibly project outcomes of “what if?” scenarios.
Basins used to calibrate and evaluate the model were primarily from the Ohio River Valley, a region that has not been well represented in the literature examining Hg bioaccumulation. These sites exhibit considerable variation in land use and land cover and span several physiographic regions, from mountainous, heavily forested basins in the Appalachian Highlands to basins in the Interior Plains that are highly cultivated. Although increased bioaccumulation has been associated with wetlands coverage and has been a focus of research (Scudder et al. 2009; Warner et al. 2005), wetlands coverage is minimal in the Ohio River Valley (Emery and Spaeth 2011; White et al. 2005). This region offers the unique opportunity to examine the relationships between other drivers of Hg bioaccumulation.
The purpose of this project was to create a tool to assist regulatory decision makers in the process of determining where and how resources can be used to protect populations from adverse Hg exposures. Because the intended users are regulators, existing data from readily available regulatory sources were used as inputs. The objective was to gain insight into how the unique characteristics of basins in this region interact to impact the bioaccumulation process. This insight can then be used to determine which basins are likely to have elevated fish tissue levels or shed light on the impact of proposed changes in land cover or other watershed characteristics, or the responsiveness of a basin to reductions in Hg loading.
This study describes the second step in the development of an overall model that will simulate Hg movement from atmospheric sources to biomarkers of exposure of susceptible human populations. The scope of this bioaccumulation model is to predict fish tissue Hg levels in generic trophic level (TL) 3, 3.5, and 4 fish given watershed characteristics and water column Hg levels. The model focuses on the processing of Hg in the water column and its movement through the food web. The modeled fish tissue Hg concentration will be used as an input into the model developed previously to project biomarkers of exposure for specified populations, described in Chan et al. (2011). This article describes the development, calibration and evaluation, strengths, and limitations of the bioaccumulation model and discusses the interactions of the factors that impact bioaccumulation in the modeled basins.
CALIBRATION AND EVALUATION
- Top of page
- MODEL DEVELOPMENT
- CALIBRATION AND EVALUATION
- SUPPLEMENTAL DATA
- Supporting Information
Seven watersheds in Kentucky were chosen to calibrate the model. Seven basins from outside of Kentucky were chosen to evaluate the model and to determine the limits of model application. All basins were selected based on the availability of fish tissue and water quality data. Evaluation basins were further limited to sites with similarities to calibration sites in basin area and wetlands coverage. Locations of calibration and evaluation watersheds are shown in Figure 4, with general site characteristics given in Table 1. More detailed maps showing land cover for calibration and evaluation basins can be found in the Supplemental Data. The Floyds Fork basin, at 70 km2, was substantially smaller than the other calibration watersheds, which had areas between 399 and 930 km2. Evaluations basins similarly had 1 smaller basin, Spring Creek in Georgia with an area of 98 km2. The range of the remaining basins was 220 to 1440 km2.
Figure 4. Locations of calibration and evaluation watersheds. Blue basins were used for model calibration, while those in pink were used to evaluate the model. See Table 1 for abbreviations of site names.
Download figure to PowerPoint
Table 1. Site characteristics
|Site||Site abbr||Area km2||Mean annual Q m3/s||Mean pH||Mean P mg/L||Fraction forested||Fraction developed||Fraction wetlands||Sinuosity|
| Floyds Fork||FF||70||1.0||7.79||0.1920||0.359||0.375||0.006||1.81|
| Mayfield Creek||MF||759||12.2||6.96||0.1972||0.234||0.064||0.042||1.00|
| Mud River||MD||689||11.1||7.37||0.0607||0.497||0.059||0.007||2.00|
| Salt River||ST||451||7.2||8.03||0.2889||0.244||0.088||0.000||2.13|
| Slate Creek||SE||399||5.7||7.85||0.1005||0.343||0.027||0.000||1.44|
| Tygarts Creek||TG||720||8.8||7.63||0.0193||0.688||0.078||0.000||1.60|
| Big Creek, IN||BC||220||2.1||8.06||0.0704||0.099||0.092||0.000||1.48|
| Little Kanawha River, WV||LK||1448||17.0||7.32||0.2085||0.857||0.051||0.000||1.40|
| Middle Island Creek, WV||MI||1282||16.1||7.36||0.2085||0.877||0.046||0.000||1.25|
| Muscatatuck River, IN||MS||733||8.0||8.05||0.1506||0.471||0.050||0.000||1.19|
| North Fork Forked Deer River, TN||FD||448||5.2||6.85||0.2141||0.162||0.077||0.095||1.06|
| Spring Creek, GA||SC||98||1.4||8.07||0.0796||0.477||0.072||0.005||1.21|
| Talking Rock Creek, GA||TR||738||8.1||7.83||0.0441||0.796||0.064||0.003||1.25|
Natural variability in environmental systems is considerable. Examination of fish tissue data showed considerable variability with samples of the same species from the same site. To compare model output to TL means that, for a particular site, an understanding of the natural variability inherent within a stream system was necessary. The National Fish Database contains over 100 000 records of Hg in fish tissue from various sources compiled by the USGS (USGS and NIEHS 2006). This database was analyzed to quantify the natural variability found between fish of the same species of similar size from the same site and collection year. Many of these samples were composites of similarly sized fish of the same species. The data set was analyzed by selecting for the most commonly sampled species that fit within the range of 0.75 and 1.25 times the quality length for that species. Selections were further grouped and limited to samples that had 6 or more of the same species from the same site and collected the same year. This resulted in 1010 groups. The standard deviation was determined for each of these groups, and the mean of these standard deviations was found to be 0.116. This mean was used to calculate a 95% confidence interval (CI) around the fish tissue means for calibration and evaluation sites. If the calculated lower bound of the confidence interval was negative, this lower bound was set to 0.
Fit was evaluated by 2 methods; by calculating prediction error and comparing model output to the 95% CI. Sample means, the bounds of the 95% CI, and prediction error are shown in Table 2. Mean error for calibration sites was 26% and 51% for evaluation sites. The model prediction fell within the 95% CI for all but 1 of the 20 TL-site combinations. The site that was outside of these bounds was the Muscatatuck River in Indiana. The model prediction was much lower than the sample mean. Error for the 2 Indiana sites and TL3 of the Little Kanawha River in West Virginia was more than twice that of other evaluation sites. The model prediction was much lower than the actual samples for all 4 TL-site combinations in Indiana, but higher for TL3 of the Little Kanawha River. Without these 3 sites, the mean error for the remaining evaluation sites was 26%. Mean error was lowest for TL4 and highest for TL3 (Table 3).
Table 2. Fish tissue Hg (mg/kg) of sample mean with 95% CI compared to model output and calculated prediction error
| Mean|| || || || || ||0.26|
| Mean|| || || || || ||0.51|
Table 3. Prediction error by TL
Land cover varied significantly between basins. Forested coverage varied widely whereas wetlands coverage was minimal to nonexistent for all but 2 basins. Most basins were in rural areas with little development with 1 exception. Regression analyses of land cover fractions compared to error were undertaken to evaluate if the model performed more accurately based on land cover attributes. No difference was found for model accuracy based on land cover (data not shown).
To test the robustness of our model a sensitivity analysis was carried out. For sensitivity to Hg loading, both unfiltered total and dissolved Hg were manipulated for 3 systems with a range of land cover uses. Figure 5 shows the response of fish tissue to percent reductions in both unfiltered total and dissolved Hg for Tygarts Creek, the Mud River, and Mayfield Creek. Reductions in unfiltered THg showed a threshold effect in fish tissue levels, whereas reductions in dissolved Hg were linear.
The sensitivity of the model to other input variables was tested by assessing the impact of manipulating these variables 1 at a time within a plausible natural range for 1 of the modeled systems. The Mud River site was chosen for the sensitivity analysis because of ample water quality and fish tissue data and land cover and water quality parameters that were average compared to the other systems. Parameters that were manipulated included the nutrient level, pH, sinuosity, water temperature, and the fractions of land cover that were forested, wetlands, or developed. Model output in response to these changes is shown in the Supplemental Data. When P levels were increased by 300%, model output decreased by 10%, but when levels were halved, model output increased by more than 50%. Similarly, when the fraction of land cover that was forested was halved, a 5% change in output was realized, but when it was doubled, the model predicted fish tissue levels more than 100% greater. An increase of wetlands coverage to 35% resulted in a fish tissue increase of more than 300%. Alternatively, manipulating the water temperature, pH, and the fraction of the watershed that was developed resulted in changes in output of less than 10%. Modifying sinuosity resulted in an output change of less than 15%.
- Top of page
- MODEL DEVELOPMENT
- CALIBRATION AND EVALUATION
- SUPPLEMENTAL DATA
- Supporting Information
The STELLA model simulates the fate of Hg from water column into fish tissue in several steps. The first step controls the process of conversion of IHg into MeHg or its movement downstream. This process is treated as a single mixed-box reactor. Land use determines the physical and chemical factors that determine the portion of IHg that is retained in the system, converted into MeHg, or is lost downstream. From this point forward, the model is a simple, multicompartment bioaccumulation model, with minimal impact from land cover factors.
Others have developed models that simulate the fate and bioaccumulation of Hg in aquatic systems. The Mercury Cycling Model (MCM) simulates the fate and bioaccumulation of Hg in lake systems (EPRI 2009). The spreadsheet-based ecological risk assessment for the fate of mercury (SERAFM) model looks at Hg in aquatic systems with a focus on contaminated sites (Knightes 2008). The Hg module in the Water Quality Analysis Simulation Program (WASP) simulates the fate of Hg in both lentic and lotic water bodies; however, it does not simulate bioaccumulation in the food web (Wool et al. 2001). The MERGANSER model is a Hg model that looks at predictors of fish tissue and piscivorous bird blood Hg levels. This model is specific to the northeast United States and limited to lakes and reservoirs (Simcox et al. 2011). Each of these Hg models has its own strengths and weaknesses and focuses on a specific area of Hg dynamics. The STELLA model described here is unique in that it focuses on how watershed characteristics of stream systems affect bioaccumulation. With the exception of the SERAFM model that looks at contaminated sites, the other Hg models simulate Hg dynamics in lake systems. This effort expands Hg models into stream systems. With more than 1.2 million river miles (1.9 million km) under a fish consumption advisory for Hg, a more complete understanding of the drivers of bioaccumulation in stream systems is needed (USEPA 2011).
Phosphorus is assumed to be the limiting nutrient and is used as a surrogate for potential productivity and to estimate the biomass of TL1. In a review of studies of nutrient limitation in North America, Borchardt (1996) found that nutrient limitation varied by region. The analysis found the northern half of the United States was predominantly P limited, the Southwest and Ozarks N limited, and the Pacific Northwest was colimited. The study sites were located in the region that Borchardt identified as to be most likely P limited. Although the nutrient that limits potential productivity within a region may vary, it is likely that the limiting nutrient was P. If not, then colimitation was the next most likely condition. In forested basins, productivity is more frequently limited by light, not nutrients. Limitation by light was also simulated in the model structure.
Dissolved organic carbon is a heterogeneous group of organic molecules consisting primarily of humic substances. Although IHg is thought to bind to thiol groups on DOC, the relationship between DOC and Hg is complex (Haitzer et al. 2003). Dissolved total Hg and DOC in stream systems are highly correlated (Brigham et al. 2009; Dittman et al. 2009; Tsui and Finlay 2011). Increasing levels of DOC have been found to be associated with both increased MeHg and lowered bioavailability in aquatic systems (Tsui and Finlay 2011). DOC is thought to be a transport mechanism for Hg into stream systems (Balogh et al. 2004). Wetlands have been shown to export increased amounts of DOC and dissolved MeHg during rain events, whereas forested systems show increased unfiltered THg inputs into streams during rain events (Balogh et al. 2004; Brigham et al. 2009). For systems with wetlands or soils with high humic content, Hg levels have been found to be correlated with DOC. However, this relationship breaks down for systems that primarily have Hg loading directly from the atmosphere (Ravichandran 2004).The Kentucky study sites have little, if any, wetland coverage. Although DOC data were not available, data were available for total organic carbon (TOC). A test to determine if adjusting the dissolved IHg input into the model based on TOC as a surrogate for DOC found no improvement in fit. As a result, TOC was not included as an input into the model. Because wetlands coverage in Kentucky is minimal, the relationship between DOC and Hg may not be strong. Systems with a strong relationship between DOC and Hg, such as systems with significant wetlands coverage, may have improved fit with adjustment of the dissolved IHg fraction.
Wetlands have been shown to contribute significant loads of MeHg to stream systems during hydrologic events and connectivity to wetlands is associated with higher fish tissue MeHg whereas a lower pH has been shown to increase methylation and bioaccumulation (Balogh et al. 1997, 2008; Haitzer et al. 2003; Miskimmin et al. 1992; Scudder et al. 2009). The pH of a system influences the relationship between DOC and Hg methylation. Although DOC transports increased amounts of Hg into stream systems, the availability of that Hg is dependent on pH. In more acidic systems, the competition for binding sites from H ion increases and results in more availability of Hg for methylation (Haitzer et al. 2003). The dissolved IHg input was adjusted by taking into account the pH of the system, although this adjustment did not improve overall mean prediction error. Calibration sites went from 28% to 26% and evaluation sites from 50% to 51%. However, improvement was substantial for higher TLs at those few sites with higher wetlands coverage and lower pH. Error for Mayfield Creek decreased from 36% to 0% for TL3.5. For N. Fork Forked Deer River, prediction error for TL4 dropped from 46% to 17%. Error increased, however, for TL3.5 from 13% to 34% at that Tennessee site. Because improvement for the systems with lower pH and higher wetlands coverage was substantial, this adjustment was included in the model as it improved model fit for systems that were not well represented in this study but are likely to be closely looked at by regulators and its inclusion extended the applicability of the model to a broader array of systems.
The minimal variation in wetlands coverage and the small range of pHs between the modeled basins are certainly limitations in the evaluation of the model. Only 2 sites had wetlands coverage greater than 1% and a mean pH below 7.0. Although the 2 systems with higher wetlands coverage and lower pH had minimal error, more testing is needed to build confidence that the model adequately represents the impact of wetlands coverage and pH on bioaccumulation.
Unlike lentic systems, nutrient levels may not be the limiting factor in primary productivity in stream systems. Light is often limiting for smaller streams. Canopy cover in conjunction with stream width may limit productivity in the stream and consequently be a primary driver for bioaccumulation within the biota (Lamberti and Steinman 1997; Vannote et al. 1980). In lower order streams, the canopy may shade the entire stream, and productivity may be minimal after leaf out. As stream order increases, the width of the stream increases and the canopy opens up resulting in higher levels of productivity for intermediate sized streams. Larger rivers are deeper and tend to carry more suspended solids, both of which reduce light penetration to periphyton-bearing substrates, lowering productivity. Watersheds that are largely forested have reduced productivity compared to stream systems with abundant light. A reduction in the biomass of primary producers leads to a higher concentration of MeHg at the base of the food web. This effect magnifies through the food web, leading to higher fish tissue concentrations in more pristine settings. The modeled basins were small to intermediate in size. Tygarts Creek in particular is a smaller, heavily forested basin. The model predicted this basin to have the highest fish tissue concentrations, which was confirmed by the collected sample. Although the sample was beyond the upper bound of the length range and therefore not included in the error analysis, the model predicted this sample, a sauger, to have the highest fish tissue concentration at 0.708 mg/kg. The actual sample was 1.140 mg/kg. Quality length for sauger is 300 mm, and this sample was 432 mm. The model does not adjust for length, but the model prediction supported the expectation that this system would have elevated bioaccumulation.
In the sensitivity analysis, the response to changes in Hg loading differed for total and dissolved Hg. Changes in unfiltered THg input had little effect on reductions in fish tissue for Mayfield Creek and the Mud River until the amount was reduced to levels below the 1 ng/L dissolved Hg default. Tygarts Creek, however, showed greater responsiveness to unfiltered THg reductions. Compared to the other 2 systems, Tygarts Creek is more forested, has substantially less agricultural coverage, and has the lowest mean unfiltered THg concentrations. Despite having lower unfiltered THg loading, fish tissue levels are highest in this system. In all 3 systems, dissolved Hg showed a linear relationship with fish tissue Hg. Balogh et al. (2007) found that in areas with high suspended solids, unfiltered THg does not correlate to Hg that is available for microbial conversion (Balogh et al. 1997). The responses to reductions in loading of both unfiltered THg and dissolved Hg for these 3 systems support the concept that measures of unfiltered THg do not necessarily correlate with Hg available for methylation and that dissolved Hg is a better predictor of methylation potential. Although Hg must be available for bioaccumulation to occur, basin characteristics drive the availability and efficiency with which Hg enters the food web and bioaccumulates. The sensitivity analysis of Hg loading shows the model accurately simulates the differences that basin characteristics impose on Hg availability and bioaccumulation efficiency.
The sensitivity analysis showed that manipulation of pH, and the fraction of land cover that was developed resulted in minimal changes in output. Changes in the nutrient level or the forested or wetlands coverage, however, had a nontrivial impact on fish tissue concentrations. Although a lower pH does make inorganic Hg more bioavailable (Haitzer et al. 2003), this is only 1 factor in determining the mass of Hg that enters the food web. Development in a basin increases the speed of transport of Hg through the system, but does not increase or decrease methylation rates. Although each of these variables had some influence on bioaccumulation, the combination of nutrient levels, and forested and wetlands coverage of the basin seem to be the strongest drivers within the modeled systems. The STELLA model allows the examination of the impact on bioaccumulation by the structure of the stream system and reveals the basis for the greater impact of the most critical variables. These variables impact the structure of the system at multiple points.
Nutrient levels and forested coverage are both related to primary productivity. The relationship between bioaccumulation and productivity is complex. Productivity impacts the model at several points in the system: the calculation of the microbial conversion rate, the estimate of biomass in the system, and the trophic transfer factors. Increased productivity in the form of extensive periphyton mats increases the methylation of Hg, thereby increasing the mass of MeHg in the biota. Countering this, increased productivity decreases bioaccumulation through both bloom dilution and somatic growth dilution. Hill and Larsen (2005) found that light limitation increased the concentration of MeHg in biofilm in artificial streams. Hg uptake was constant under different light regimes, but growth of microalgae was dependent on light. Ward et al. (2010a) studied somatic growth dilution in Atlantic salmon. Fry raised under the same conditions were stocked in various streams and sampled approximately 4 months later. Individual growth rate accounted for 38% of the variation in Hg concentration in fish tissue, building support for somatic growth dilution as being an important factor in bioaccumulation in stream systems.
The largest impact on bioaccumulation came from land cover factors that influence the amount of IHg in the stream that is converted into MeHg. Likewise, Chasar et al. (2009) found that, across wide environmental conditions, the supply of MeHg at the base of the food web was the largest factor to influence bioaccumulation, and that trophic structure and transfer efficiencies were only important predictors when comparing streams with similar land cover and Hg loading. Riva-Murray et al. (2011) similarly found that MeHg at the base of the food web was more important than trophic structure across larger environmental scales. At a smaller scale, they found significant spatial heterogeneity in Hg in biota within basins that had heterogeneous in-stream habitat. This heterogeneity may be reflecting the different mix of functional feeding groups. Without having a thorough survey of the feeding groups in each stream, the proportions for each stream will be unknown. The assumption is made that the variation in the composition of functional feeding groups at the base of the food web is relatively small across the region depicted. Bradley et al. (2011) also observed spatial variation in filtered MeHg concentrations in stream reaches corresponding to connectivity to wetlands or floodplains or shallow, open-water areas. Smaller reaches may have atypical land cover and habitat characteristics compared to the characteristics of the larger basin containing that reach. Therefore, when using of the model, the size of the basin should be considered. The model was developed and calibrated primarily for midsized streams. Use for smaller or larger basins has not been tested and is likely beyond the current scope of model performance.
Variation within natural settings is large. Stream flow, sunshine and rain, temperature, nutrient levels, growth rates, and other variables fluctuate daily and from year to year. Fish of the same species and size sampled from the same location will vary in tissue Hg concentration. A limitation of this model is that input parameters inherently have a high degree of variability, and the use of data collected for regulatory purposes further compounds the issue. Variable inputs into the model are monthly or annual means from several years of monitoring. Although flow and meteorological data represent many years, water quality data were typically means from 8 or 9 years for calibration sites and fewer for evaluation sites. The assumption in these simulations is that these means represent a steady state system. It is not known whether the means from these years are typical for these basins over the long term, whether the samples encompass data that are outliers, or whether there have been recent changes in water quality, weather, or flow. If the assumption of steady state is incorrect, then model output may be biased.
The natural variation between fish tissue samples from the same species and site limits confidence in the ability to accurately predict the concentration of Hg in any particular sample. A mean standard deviation of 0.116 was found when examining groups of 6 or more samples from the National Fish Database. When monitoring for regulatory purposes, the number of samples are often limited due to the cost of analyzing the samples. The combination of considerable variability and low sample size, even when these are pooled samples, results in a wide 95% confidence interval. Because confidence intervals are wide, the accuracy of model output is more difficult to assess. The ultimate purpose of the overall model is to gain insight into the complex problem of Hg in the environment and to understand the factors that control bioaccumulation and its impact on local susceptible populations. For populations that regularly consume fish from the same water body, the natural variation between fish will be averaged, and a mean intake concentration can be considered. Regulatory agencies sample fish tissue to determine which waterways present a risk to populations from fish consumption. At times, agencies have based consumption advisories on as few as 1 sample (E Eisiminger, KDOW, personal communication). More frequent and intensive sampling efforts would ensure that an elevated sample was truly representative of the Hg concentrations in that system.
This model seeks to predict the fish tissue MeHg concentration in a generic fish species based on TL and basin characteristics. Accuracy will vary based on the species and size of fish to which the model output is compared. When examining prediction error by TL at each site, the mean error for predicting fish tissue Hg levels was 26% for sites in Kentucky and 51% for sites in nearby states. This compares well with other models that predict fish tissue concentrations. The National Descriptive Model for Mercury in Fish Tissue developed by the USGS reported a prediction error of 38% (Wente 2004). The model developed by Trudel and Rasmussen (2006) was within 25% of actual value for 70% of predictions.
One difficulty in developing this model was defining Hg loading to the stream system. Hg was simulated from water column to fish tissue. The model was developed and calibrated using data from the Kentucky Division of Water, and monthly means of unfiltered THg in the water column were used for the input into the system. Soluble Hg is considered a more accurate measure of inorganic Hg that is available for conversion to MeHg, but dissolved Hg data were not available. Soluble Hg levels are not always correlated to unfiltered THg levels, especially in systems with high suspended solids. Measures of DOC, which are more closely related to dissolved Hg levels in stream systems, were not available (Brigham et al. 2009). Actual loading is dependent on a number of variables including the amount of precipitation and the concentration of Hg in that precipitation, the pH of the soil in the watershed, land use, and more. Using the monthly mean for Hg loading introduces uncertainty. Spikes of Hg loading, both particulate and dissolved, may occur during precipitation events that may not be reflected in monthly sampling, particularly in basins that may be downwind of emission sources. Simulating the processes that influence loading into the stream would increase model accuracy.
The model was evaluated by comparing model output to fish tissue data for 7 basins in the surrounding region. The model was not as accurate at predicting fish tissue concentrations for evaluation sites. Surprisingly, sites proximal to the state of Kentucky had higher prediction errors than other evaluation sites. In particular, the 2 sites in Indiana had prediction errors ranging from 78% to 90% for the 4 TL-site combinations. These sites are in the southeast portion of the state downwind from a number of coal-fired power plants. A report by the USGS on Hg in precipitation in Indiana found that portion of the state to consistently have the highest Hg wet deposition, suggesting that these 2 watersheds may be located in a Hg hotspot (Risch and Fowler 2008). Hammerschmidt and Fitzgerald (2006) found a positive correlation between wet deposition of Hg and MeHg concentrations in largemouth bass fillets. Gorski et al. (2008) found that Hg in precipitation was more bioavailable than Hg in surface waters. The higher loading of dissolved Hg in precipitation may account for the elevated Hg in fish samples at the Indiana sites compared to model output. Without the Indiana sites, mean prediction error for the evaluation sites was 35%. Adding wet deposition of Hg to the model might improve model performance. However, wet deposition data at a spatial scale that is relevant to modeling basins in the Ohio River Valley are lacking.
In addition to the elevated error at the Indiana sites, TL3 of the Little Kanawha River in West Virginia had the highest error, 94%. This elevated error is misleading. For lower TLs, the magnitude of predicted and actual values are much lower, so differences between those values may result in higher error even though the difference may not be biologically relevant. For the Little Kanawha River, the sample mean was 0.018 mg/kg and the predicted value was 0.035 mg/kg resulting in an error of 94%. If we look at TL3.5 of Spring Creek, the sample mean was 0.262 mg/kg and the predicted outcome was 0.191 mg/kg for an error of 27%. The difference of 0.017 mg/kg found at Little Kanawha is biologically less significant than the difference of 0.071 mg/kg found at Spring Creek. The error calculations do not convey the biological significance or magnitude of the error. This scaling of error values at least partially explains the decreasing error with increasing TL found in Table 3. Mean prediction error for evaluation sites without the Indiana and Little Kanawha sites was 26%, the same error found for calibration sites.
The basins used to both calibrate and evaluate the model represent a wide variety of geomorphic land types and land uses. In general, the more forested basins were located in more mountainous areas. Although there was considerable variability in prediction error between sites, land cover was not associated with error. This diversity of land types and uses build confidence in the ability of the model to be generalized to other basins in humid temperate regions with similar basin size and wetlands coverage.