application of the soil and water assessment tool for six watersheds in lake erie model parameterization and calibration

After initial model parameterization, the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand SWAT models were calibrated 1998–2001 and conﬁrmed, or validated 2002–2005, individually fo

Trang 1

Application of the Soil and Water Assessment Tool for six watersheds of Lake Erie: Model parameterization and calibration

Nathan S Boscha,⁎, J David Allanb,1, David M Dolanc,2, Haejin Hand,3, R Peter Richardse,4

a Department of Science and Mathematics, Grace College, Winona Lake, IN 46590, USA

b School of Natural Resources and Environment, University of Michigan, Ann Arbor, MI 48109, USA

c Natural and Applied Sciences, University of Wisconsin — Green Bay, Green Bay, WI 54311, USA

d Korea Adaptation Center for Climate Change, Korea Environmental Institute, Seoul 122-706, Republic of Korea

e National Center for Water Quality Research, Heidelberg University, Tifﬁn, OH 44883, USA

a b s t r a c t

a r t i c l e i n f o

Article history:

Received 30 June 2010

Accepted 15 February 2011

Available online 9 April 2011

Communicated by Joseph DePinto

Index words:

SWAT

Great Lakes

Nutrients

Sediments

Catchment

Model conﬁrmation

The Soil and Water Assessment Tool (SWAT), a physically-based watershed-scale model, holds promise as a means to predict tributary sediment and nutrient loads to the Laurentian Great Lakes In the present study, model performance is compared across six watersheds draining into Lake Erie to determine the applicability

of SWAT to watersheds of differing characteristics After initial model parameterization, the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand SWAT models were calibrated (1998–2001) and confirmed, or validated (2002–2005), individually for stream water discharge, sediment loads, and nutrient loads (total P, soluble reactive P, total N, and nitrate) based on available datasets SWAT effectively predicted hydrology and sediments across a range of watershed characteristics SWAT estimation of nutrient loads was weaker although still satisfactory at least two-thirds of the time across all nutrient parameters and watersheds SWAT model performance was most satisfactory in agricultural and forested watersheds, and was less so in urbanized settings Model performance was influenced by the availability of observational data with high sampling frequency and long duration for calibration and confirmation evaluation In some instances, it appeared that parameter adjustments that improved calibration of hydrology negatively affected subsequent sediment and nutrient calibration, suggesting trade-offs in calibrating for hydrologic vs water quality model performance Despite these considerations, SWAT accurately predicted average stream discharge, sediment loads, and nutrient loads for the Raisin, Maumee, Sandusky, and Grand watersheds such that future use of these SWAT models for various scenario testing is reasonable and warranted

Introduction

Nutrient delivery to the Laurentian Great Lakes through tributary

loading has long been identiﬁed as a major contributor to the trophic

status of the lakes Eutrophication in the Great Lakes ﬁrst attracted

attention in the 1960s and 70s as Lake Erie experienced benthic

anoxia and other water quality problems associated with nutrient

enrichment (Boyce et al., 1987; Rosa and Burns, 1987) The 1972 Great

Lakes Water Quality Agreement brought about reductions in point

source loadings of phosphorus (P) that dramatically lowered annual P loads from tributaries and reversed many of the effects of cultural eutrophication in the lakes (DePinto et al., 1986) Despite this major reduction, point-source control programs alone were not adequate to reduce tributary nutrient loading, and attention shifted to nonpoint sources (DePinto et al., 1986; Dolan, 1993; Richards, 1985)

Much research in the Great Lakes basin has provided insight into the relationship between riverine nutrient export and characteristics of watersheds, including land use and nutrients inputs (Baker and Richards, 2002; Bosch and Allan, 2008; Dolan and McGunagle, 2005; Han and Allan, 2008; Han et al., 2011; Robertson, 1997).Dolan and McGunagle (2005)estimated that nonpoint sources often account for more than 70% of total tributary loadings in Lake Erie Watershed-scale nutrient budgets have confirmed this observation and refined it to show specifically that fertilizer application is the largest input of nitrogen (N) and P in agricultural watersheds, while atmospheric deposition of N is

an important input to forested watersheds (Baker and Richards, 2002; Bosch and Allan, 2008; Han and Allan, 2008; Han et al., 2011) In urban watersheds, by contrast, point source P inputs may continue to be signiﬁcant (Nemery et al., 2005)

⁎ Corresponding author Tel.: +1 574 372 5100x6447.

E-mail addresses: boschns@grace.edu (N.S Bosch), dallan@umich.edu (J.D Allan),

doland@uwgb.edu (D.M Dolan), hanhj@kei.re.kr (H Han), prichard@heidelberg.edu

(R.P Richards).

1 Tel.: +1 734 764 6553.

2 Tel.: +1 920 465 2986.

3 Tel.: +82 2 6922 7803.

4 Tel.: +1 419 448 2240.

doi: 10.1016/j.jglr.2011.03.004

Contents lists available atScienceDirect

Journal of Great Lakes Research

j o u r n a l h o m e p a g e : w w w e l s e v i e r c o m / l o c a t e / j g l r

Trang 2

Watershed characteristics and source inputs determine riverine

nutrient export For 18 well monitored tributaries of Lake Michigan

and Lake Superior, topography and surﬁcial deposits were top

predictors of P and sediment yields, followed by land use (Robertson,

1997) Cropland has larger P losses per unit area than pasture or

forest (Alvarez-Cobelas et al., 2009; Castillo et al., 2000; Cooke and

Prepas, 1998; Pieterse et al., 2003) Crop land use is associated with P

fertilization and increased soil erosion caused by tillage practices

(Sharpley, 1999) Mass balance approaches have demonstrated

strong relationships between source inputs to the watershed's

landscape and river export for both N and P (Boyer et al., 2002;

Han and Allan, 2008; Han et al., 2011) Clearly, it is important to

incorporate point and nonpoint sources of N and P into models as

well as other watershed characteristics such as surﬁcial geology and

stream gradients, in order to accurately predict tributary export of

nutrients to the Great Lakes

The Soil and Water Assessment Tool (SWAT), a semi-empirical

hydrologic and water quality model that is calibrated to the

conditions of individual watersheds, may be used to predict changes

in tributary nutrient loads based on nutrient application and

management choices SWAT is a continuous-time, watershed-scale

model that runs at a daily time step (Gassman et al., 2007) SWAT

models are internally organized in a nested spatial hierarchy,

including Hydrologic Response Units (HRUs) within subwatersheds

within watersheds Watersheds are deﬁned by a main outlet point

for the drainage area of interest A variable number of subwatersheds

are delineated in SWAT based on interior outlet points located on the

stream channel and spatially linked to each other through the

adjoining stream channel network Within each subwatershed a

variable number of HRUs are deﬁned as areas with unique

combinations of land cover, soil type, and slope It is important to

note that these HRUs are not spatially referenced except that they

are within a speciﬁc subwatershed Processes modeled within each

HRU are aggregated up to the subwatershed scale by a weighted

average based on land area Thus as a mechanistic model, SWAT

includes spatially explicit parameterization at the subwatershed

spatial scale and partially lumped parameterization within each

subwatershed

SWAT hydrology, sediment, and nutrient processes are modeled in

both upland and water-routing phases (Neitsch et al., 2005) Various

forms of N and P, including mineral P (soluble reactive P (SRP)), other

P (total P - SRP), organic N, nitrate, nitrite, and ammonia are modeled

explicitly with a thorough representation of transport, uptake, loss,

and transformation mechanisms (more detail on SWAT is provided in

Appendix A)

The SWAT model was applied to six watersheds draining into

Lake Erie to model scenarios of riverine nutrient export as part of a

larger investigation of the recent resurgence of anoxic waters in Lake

Erie (Burns et al., 2005) and the contributions of tributary runoff We

include six of the largest U.S watersheds of Lake Erie, which receive

some of the highest nutrient inputs of any in the Great Lakes due

largely to intensive agriculture in the region Previous application of

SWAT for two of the watersheds utilized some similarities in

methodology (Bosch, 2008), but the approach has been improved

and a newer version of the SWAT model has been employed In the

present study, model performance was compared across these

diverse watersheds to determine the performance of SWAT within

watersheds of differing characteristics Care was taken to apply

SWAT similarly to all six watersheds to allow comparison of model

performance There is growing interest in restoring the Great Lakes,

in which nonpoint source nutrient loads are widely considered one

of the most important threats It is our hope that this study will

inform future studies of the suitability of SWAT to predict tributary

sediment and nutrient loads under scenarios of agricultural best

management practices, nutrient source reductions, and future

climates

Methods Study area The Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand watersheds drain into the western and central basins of Lake Erie (Fig 1) from southeastern Michigan, northeastern Indiana, and northern Ohio Primary differences among watersheds are reﬂected

in land cover (Table 1), with the Raisin, Maumee, and Sandusky being predominantly agricultural The Huron and Cuyahoga are the most urbanized watersheds The Grand watershed is mostly forested with relatively little agriculture and urban land Average precipitation increases slightly from west to east due to more lake-effect precipitation in the Cuyahoga and Grand watersheds

Model input data sources The Geographic Information System (GIS) interface created for SWAT, called ArcSWAT (version 2.1.5), was used to develop inputs for the six watershed models Elevation, stream network, land cover, soil type, weather, point source discharges, impoundment (reservoir, lake,

or pond) characteristics, atmospheric N deposition, and land management practices were included Speciﬁc data descriptions, sources, and scales are included in Appendix B

Model setup and parameterization For each of the six watershed models we ﬁrst delineated subwatersheds and distributed HRUs within subwatersheds (Appendix C) The delineation process resulted in 31, 32, 203, 39, 23, and 22 subwatersheds for the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand, respectively, which was the desired level of spatial detail for the study This was an average subwatershed size of 85 km2across all six watersheds There were 441, 468, 2341, 567, 302, and 297 HRUs distributed in the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand watershed models, respectively, based on unique combinations

of land cover and soil type and leading to an average HRU size of 7 km2 Measured data were used for all weather parameters including daily rainfall, minimum and maximum air temperature, windspeed, relative humidity, and solar radiation Weather data were collected for the time period January 1, 1995 through December 31, 2005 Data from the nearest station were used to ﬁll in missing data whenever data records were incomplete

Loading data for point source dischargers were entered into the models as constant average daily loadings For parameters that were not measured by individual municipal wastewater treatments plants,

we used average estimated values (Table 2) obtained from measure-ments during 1995–2005 at eight Midwestern U.S wastewater treatment plants where all parameters were measured For larger wastewater treatment plants, parameters in addition to discharge and

TP were often reported, such that directly measured data could be used Discharge loading information was entered for 36, 23, 83, 13, 25, and 7 dischargers in the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand, respectively

Tile drainage was implemented in the six watershed models following the approach ofGreen et al (2006; Appendix C) Average denitriﬁcation rates in upland crop land over various soil types and N application rates are near 15 kg N ha− 1y− 1(Hofstra and Bouwman,

2005), while rates in well-drained, clay loam, forested soils average

18 kg N ha− 1y− 1(Groffman et al., 1992) Since these six watersheds are dominated with agriculture and forest land cover (Table 1), model parameters were adjusted to approximate these denitriﬁcation rates (Appendix C)

Agricultural land management practices were generalized for each

of the six watersheds based on most common management practices

in each watershed, and used to deﬁne operation schedules in SWAT

Trang 3

for both row-crop and hay agricultural lands Most land classiﬁed as

hay in these watersheds was managed for hay production or used for

grazing and was modeled by SWAT accordingly Row-crop agriculture

in these six watersheds consisted of various rotational schedules and

combinations utilizing mostly corn, soybean, and wheat crop

production SWAT operation schedules were deﬁned according to

these rotation patterns for each watershed individually as well as for

corresponding practices such as tillage, fertilizer application, crop

planting, manure application, and crop harvesting The operation

schedules were adjusted to accurately represent fertilizer and manure

application rates based on county estimates (Ruddy et al., 2006) In

addition, each of the multi-year operation schedules differed based on

the starting crop in the rotation pattern This ensures that the single

year of wheat production in the 6-year rotation of the Maumee

watershed occurs in a staggered manner across the watershed rather

than all subwatersheds producing wheat in the same year These land management schedules were distributed among subwatersheds such that each version was equally represented and applied uniformly across the watershed

Other land cover types (residential, industrial, range, forest, wetlands, and water) were given general operation schedules based

on most common management practices across the six watersheds as

a whole High, medium, and low density residential as well as

Fig 1 The Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand watersheds draining into western and central Lake Erie as delineated in SWAT models.

Table 1

Characteristics of the Huron, Raisin, Maumee, Sandusky, Cuyahoga, and Grand

watersheds for the modeled areas, determined by the watershed outlet location.

Watershed Precipitation Landcover (%)

Size (km 2 ) (mm/y) Agriculture Urban Forested

Table 2 Point source discharge input data by parameter Average estimated concentrations were based on measured data during 1995–2005 at eight Midwestern WWTPs CBOD refers to chemical/biological oxygen demand.

Parameter Units Notes Water ﬂow m 3 Directly measured data used Sediment Mg Based on average total suspended solid concentration of

7.2 mg/L Organic

nitrogen

kg Based on average organic nitrogen concentration of 2.2 mg N/L

Organic phosphorus

kg Assumed to be 30% of directly measured total phosphorus concentration

Nitrate kg Based on average nitrate concentration of 11.3 mg N/L Ammonia kg Based on average ammonia concentration of 3.2 mg N/L Nitrite kg Based on average nitrite concentration of 0.6 mg N/L Mineral

phosphorus kg Assumed to be 70% of directly measured totalphosphorus concentration CBOD kg Based on average CBOD concentration of 4.9 mg/L Dissolved

oxygen

kg Based on average dissolved oxygen concentration of 6.8 mg/L

Trang 4

industrial land covers were set to a plant type of Kentucky Bluegrass

and scheduled to grow from May 1 to September 29 Fertilizer

application equal to the watershed-speciﬁc, non-farm fertilizer

estimates (Ruddy et al., 2006) was applied to low, medium, and

high density residential in four equal applications scheduled on April

1, June 1, August 15, and September 30 Most land classiﬁed as range

in these watersheds is left fallow as conservation reserve areas, so the

plant type was set to Tall Fescue with a growing season from April 1 to

October 31 with no fertilizer or manure application operations

Operations for forest and wetland land covers were scheduled to

correspond with leaf-out in the spring and leaf senescence in the fall,

with the growing season beginning on May 1 and ending on October

10 of each year The water land cover was kept at its default settings

Model calibration and conﬁrmation

The six SWAT models were calibrated and conﬁrmed individually for

stream ﬂow discharge, sediment loads, and nutrient loads (total P (TP),

SRP, total N (TN), and nitrate) based on available datasets Daily mean

stream discharge measurements were available for 1995–2005 from

USGS gage stations near the river mouth of each watershed, including

the Huron River at Ann Arbor, MI (04174500), the River Raisin at

Monroe, MI (04176500), the Maumee River at Waterville, OH

(04193500), the Sandusky River at Fremont, OH (04198000), the

Cuyahoga River at Independence, OH (04208000), and the Grand River

at Painesville, OH (04212100) For the Raisin, Maumee, Sandusky,

Cuyahoga, and Grand rivers, near-daily sediment, TP, SRP, TN, and

nitrate loads were available for sampling sites located near the USGS

gage stations provided by the National Center for Water Quality

Research at Heidelberg College for the entire time period of 1995–

2005 In order to produce observed monthly loads for later calibration

and conﬁrmation tests, missing values for daily loads were determined

using the USGS Estimator protocol (Richards, 1998) For the Huron

River, fewer sediment and nutrient data were available Approximately

biweekly TP, SRP, TN, and nitrate concentration data were collected

during 2003–2005 near the mouth of the river (Bosch, 2007) No

sediment data were available for the Huron watershed As with the

other ﬁve watersheds, missing daily values for TP, SRP, TN, and nitrate

loads were generated for the Huron River using the USGS Estimator

protocol in order to produce observed monthly loads

Observed data were then compared to simulated SWAT output The

ﬁrst 3 years (1995–1997) were used for model spin-up, in order to

minimize the impact of initial model parameter values which may be

suspect, and this model output was not used in calibration or

conﬁrmation The next 4 years of observed data were used for

calibration (1998–2001), and the remaining 4 years (2002–2005) for

model conﬁrmation Model conﬁrmation (using the terminology of

Reckhow and Chapra, 1983 and Oreskes et al., 1994; others have used

the term “validation”) consisted of comparing model predictions with

observations using a data set for years and conditions distinct from those

represented by the calibration data For the Huron watershed only, 2003

and 2004 observed loads were used for calibration, and 2005 loads were

used for conﬁrmation To evaluate model performance during

calibra-tion and conﬁrmacalibra-tion, statistical measures at the monthly time-step

were used as well as visual graphical comparison at the daily time-step

The monthly statistical measures used for calibration and

conﬁrmation evaluation for stream discharge, sediment, TP, SRP, TN,

and nitrate included: coefﬁcient of determination (R2), Nash–Sutcliffe

simulation efﬁciency (NSE), percent bias (PBIAS), and the ratio of the

root mean square error to the standard deviation of the observations

(RSR).Moriasi et al (2007)thoroughly explain the utility of each of

these 4 evaluation statistics, how they are calculated, and what range

of values are satisfactory for watershed hydrology and water quality

modeling with models such as SWAT (Appendix D)

Model calibration included several sequential stages for each

individual watershed, including hydrology sensitivity analysis, hydrology

manual calibration, hydrology autocalibration, sediment manual calibra-tion, and nutrient manual calibration First, an automated sensitivity analysis (Van Griensven et al., 2006) is carried out through the ArcSWAT interface with hydrologic model parameters in order to identify the parameters to be adjusted during the autocalibration procedure as well as

to give some insight into parameters to adjust during manual hydrology calibration The sensitivity analysis procedure uses the Latin hypercube one factor at a time design, and identifies the top 15 most sensitive parameters Next, the hydrology was roughly calibrated manually by changing sensitive hydrologic parameters as described inSanthi et al (2001) Once simulated stream discharge roughly fit the observed discharge for the calibration time period, a second sensitivity analysis was performed The top 15 parameters of both the pre- and post-manual calibration sensitivity analysis runs were then used for the hydrology autocalibration The hydrology autocalibration employed the PARASOL calibration procedure included in the ArcSWAT interface (Van Griensven and Meixner, 2007) The PARASOL method applied a shuffled complex evolution optimization scheme to select the optimal parameter value set for the 15 sensitive hydrologic parameters after several thousand model runs The calibration was based on monthly observed USGS daily mean stream discharge data (1998–2001) and SWAT simulated stream discharge Hydrologic model parameter values were then adjusted to reflect the optimal value set chosen by the autocalibration process After model hydrology calibration was completed, manual cali-bration of sediments and nutrients was completed, followed by model confirmation SWAT sediment parameters were calibrated following the procedure ofSanthi et al (2001) based on monthly observed sediment loads from 1998 to 2001 Since no observed sediment data were available for the Huron River, the optimized sediment parameter values from the adjacent Raisin watershed were used for the Huron SWAT model as well After sediment calibration was completed, TP, SRP, TN, and nitrate calibration were done based on monthly observed nutrient data (Santhi et al., 2001) SWAT output included mineral P and other P, so SWAT mineral P was compared to observed SRP data and observed TP was compared to the sum of SWAT mineral P and other P After model nutrient parameters were optimized for all six watershed models, calibration was complete and model confirmation was initiated During model confirmation, evaluation statistics were calculated for stream discharge, sediments, TP, SRP, TN, and nitrate and for each of the six SWAT models, but no further model parameter changes were made

Results and discussion Stream discharge SWAT predicted measured stream discharge well for all six watersheds during both the calibration and conﬁrmation time periods (Table 3) In fact, NSE, PBIAS, and RSR statistic values were mostly

“very good” with a few categorized as “good” according toMoriasi et

al (2007) Somewhat surprisingly, SWAT prediction of discharge was slightly more accurate during the conﬁrmation period than the calibration time period overall Though all SWAT models performed well, across the six watersheds, stream discharge was best predicted for the Maumee, and least well predicted for the Huron

Successful modeling of stream discharge is expected given the extensive calibration data sets available from stations located near the mouth of each watershed, and the effectiveness of the PARASOL autocalibration method This success indicates that the hydrologic mechanisms included in the model are ﬁt uniquely and well to each of these watersheds and are reliable for future model simulations Stream discharge calibration is critical for subsequent water quality calibration success

The less satisfactory hydrologic performance of the Huron model likely is due to speciﬁc characteristics of the Huron watershed First, the baseﬂow fraction of the Huron River is higher (0.74) than all other

Trang 5

watersheds (0.39–0.60) included in this study (Arnold and Allen, 1999).

Second, the Huron River has far more impoundments than the other ﬁve

rivers combined, storing large volumes of water in its middle and lower

sections Third, and related to the previous two characteristics, the

Huron River has a much less ﬂashy hydrograph than other rivers of this

study and a slow return of discharge to baseﬂow conditions (Fig 2)

While the relatively high urban land cover of the Huron (Table 1) might

be expected to result in ﬂashy stream discharges due to impervious

surfaces, it seems that impoundments effectively dampen peak ﬂows by

temporarily retaining excess water volumes in the river system

Sediment

Sediment calibration was acceptable for the Raisin, Maumee,

Sandusky, and Grand (Table 4) In fact, most statistic values for the

conﬁrmation time period were “very good” for these four rivers

according to the ranges given byMoriasi et al (2007) For the Huron

River, no observed sediment data suitable for calibration or

conﬁrmation was available, so sediment load prediction for the

Huron could not be evaluated Sediment model output for the

Cuyahoga River, despite a strong data set for observed sediments,

was marginal to poor Three of the evaluation statistics had

satisfactory values, while the remaining ﬁve were unsatisfactory

This weak model performance for sediment loads for the Cuyahoga

watershed may be related to its hydrology calibration process as well

as the watershed's landcover This was the only model for which the

observed-simulated sediment R2decreased as a result of the stream

discharge calibration process In the other four models for which sediment calibration was possible, hydrology calibration improved sediment prediction prior to undertaking sediment calibration Parameter adjustments that improved calibration of hydrology in the Cuyahoga model apparently had a negative inﬂuence on subsequent sediment and nutrient calibration, suggesting trade-offs

in calibrating for hydrologic vs water quality model performance In addition, the landcover of the Cuyahoga watershed is the least agricultural and most urban of the six study systems Despite the low extent of agriculture, the sediment load near the mouth of the Cuyahoga River is second only to the much larger and more agricultural Maumee River In fact, the Cuyahoga River had the highest single observed daily sediment load of all the watersheds, including the Maumee River This unique characteristic of the Cuyahoga was not adequately simulated within the sediment transport mechanisms included in SWAT

Total phosphorus Model performance for TP ranged from unsatisfactory to satisfactory across the six study watersheds (Table 5) The Maumee and Sandusky models both showed satisfactory to strong performance across multiple parameters.Fig 3depicts the daily TP load for the Maumee, which is the largest individual TP input to Lake Erie Simulated TP loads for the Raisin and Grand watersheds ranged from unsatisfactory to satisfactory, indicating marginal model performance Simulated TP for the Huron and Cuyahoga was mostly unsatisfactory for calibration and

Table 3

Calibration and conﬁrmation results for monthly stream discharge (m 3 /s) for the six

modeled watersheds Coefﬁcient of determination (R 2 ), Nash–Sutcliffe simulation

efﬁciency (NSE), percent bias (PBIAS), and the ratio of the root mean square error to

observations standard deviation (RSR) are used as evaluators of model performance.

Statistics in bold type are categorized as satisfactory or better ( Moriasi et al., 2007 ).

Observed mean

(m 3 /s)

Simulated mean (m 3 /s)

R 2 NSE PBIAS

(%) RSR (a) Calibration

Maumee 151 162 0.93 0.92 −7 0.28

Sandusky 25 26 0.88 0.87 −2 0.35

Cuyahoga 21 21 0.89 0.88 3 0.34

(b) Conﬁrmation

Raisin 19 21 0.88 0.87 −11 0.36

Maumee 174 176 0.95 0.95 −1 0.22

Sandusky 39 38 0.90 0.90 2 0.32

Cuyahoga 34 30 0.90 0.87 11 0.36

Fig 2 Daily plot of observed and simulated mean stream discharge (m 3 /s) for the Huron River over the calibration (1998–2001) and conﬁrmation (2002–2005) time periods.

Table 4 Calibration and confirmation results for monthly sediment loads (Mg) for the six modeled watersheds Coefficient of determination (R 2 ), Nash–Sutcliffe simulation efficiency (NSE), percent bias (PBIAS), and the ratio of the root mean square error to observations standard deviation (RSR) are used as evaluators of model performance Statistics in bold type are categorized as satisfactory or better ( Moriasi et al., 2007 ) No sediment evaluation was completed for the Huron River due to the lack of observed sediment data.

Observed mean (g)

Simulated mean (Mg)

R 2 NSE PBIAS

Raisin 5203 5008 0.65 0.63 4 0.62 Maumee 63475 63332 0.68 0.67 0 0.58 Sandusky 9572 10490 0.74 0.73 −10 0.53 Cuyahoga 10087 10110 0.46 0.25 0 0.87 Grand 5739 5718 0.85 0.83 0 0.42 (b) Conﬁrmation

Raisin 4104 4865 0.80 0.79 −19 0.47 Maumee 77095 68423 0.92 0.89 11 0.34 Sandusky 20082 17579 0.91 0.85 12 0.39 Cuyahoga 29832 15745 0.76 0.45 47 0.75 Grand 12789 9693 0.75 0.64 24 0.61

Trang 6

conﬁrmation time periods, with particularly poor NSE values for the

Huron model

Poor performance of modeled TP for the Huron and Cuyahoga may

be related to calibration methodology, land cover, and point source

nutrient inputs Relatively sparse observed data are likely to blame for

the particularly poor TP prediction for the Huron TP loads depend

largely on sediment transport, and so the lack of sediment data and

sediment calibration likely contributed to the poor prediction of TP

loads In addition, only 3 years of biweekly TP loads were available for

Huron TP calibration and conﬁrmation, compared to near daily TP

loads for 8 years for the other watersheds, making it difﬁcult to fully

parameterize or evaluate model performance for the Huron model A

different explanation is necessary for the poor performance of the

Cuyahoga model, since ample sediment data were available As noted

previously with respect to sediment prediction, calibration to improve

hydrologic performance of the Cuyahoga model also decreased R2

values between observed and simulated TP loads for both the Huron

and Cuyahoga SWAT models Furthermore, these two watersheds

have the most urban land use and least agriculture of the six study

watersheds, indicating SWAT may better predict TP loads in

agricultural than urban watersheds Both the Huron and Cuyahoga

Rivers receive relatively high point source TP loads relative to total

observed river loads For the Huron River, point source inputs of TP to

the river are estimated (and thus modeled by SWAT) as 29 kg P/d

while the observed average stream load near the Huron River mouth

is 36 kg P/d Similarly, estimated point source inputs for the Cuyahoga

River are 233 kg P/d, roughly one-third of the total stream load of

703 kg P/d

SWAT estimates of TP loads for the Grand and Raisin models can be described as marginally satisfactory, likely due to characteristics of these two watersheds Compared to the Maumee and Sandusky watersheds, which produced strong models, the Grand watershed is much less agricultural and more forested (Table 1), which likely hampers SWAT's ability to predict TP loads Apparently, the high percentage of urban land

in the Huron and Cuyahoga, and correspondingly high point source inputs, account for even poorer model performance for those two watersheds In the case of the Raisin, model TP estimation barely missed the satisfactory threshold for NSE and RSR, possibly due to the presence

of impoundments in the upper Raisin watershed Prior work (Bosch, 2008; Bosch et al., 2009) has shown that impoundments alter the timing and magnitude of TP loads, and this may not have been adequately portrayed in the current Raisin SWAT model

Despite these mixed performance evaluations, it is important to note that PBIAS values were largely categorized as “very good” across all watersheds, indicating that model predictions of TP loads were not generally higher or lower than observed data Thus, SWAT predicted

TP loads accurately on average

Soluble reactive phosphorus SRP estimation was the weakest output from SWAT models for all six watersheds (Table 6) This was apparent during both calibration and confirmation time periods, and evaluation statistics were unsatisfactory just over half of the time across all watersheds and for all statistical measures for SRP (Moriasi et al., 2007) Only the Raisin model consistently performed well in calibration and confirmation Model performance for the Maumee, Sandusky, and Grand was marginal, and more satisfactory during the calibration than the confirmation time period, as expected PBIAS was satisfactory for the Raisin, Maumee, Sandusky, and Grand, once again indicating strong model performance for these four watersheds As observed for TP, the Huron and Cuyahoga models were largely unsatisfactory in their estimation of SRP over time and as a monthly average

Weak prediction of SRP loads for the Maumee and Sandusky is unexplained Inspection of observed and simulated means (Table 6) shows that observed SRP loads increased dramatically from the calibration time period (1998–2001) to the conﬁrmation time period (2002–2005) This same increase is not mirrored in simulated SRP loads Although observed stream discharge increases as well (Table 3), this is insufﬁcient to explain the SRP increase as SWAT should capture this effect Monitoring data show an increase in SRP as a fraction of TP over recent years in these two watersheds (Richards, 2006; Richards,

2007), with no conﬁrmed explanation for this trend It is apparent

Table 5

Calibration and conﬁrmation results for monthly total phosphorus loads (Mg P) for the

six modeled watersheds Coefﬁcient of determination (R 2 ), Nash–Sutcliffe simulation

Observed mean

(Mg P)

Simulated mean (Mg P)

R 2 NSE PBIAS

Maumee 149 149 0.72 0.72 0 0.53

Sandusky 23 24 0.66 0.64 −1 0.60

Cuyahoga 14 14 0.41 0.11 1 0.95

(b) Conﬁrmation

Huron 1 3 0.16 −46.12 −220 56.51

Maumee 182 137 0.86 0.74 25 0.51

Sandusky 43 30 0.87 0.77 29 0.49

Cuyahoga 29 16 0.68 0.39 44 0.79

Fig 3 Daily plot of observed and simulated total phosphorus (TP) loads (kg P/d) for the Maumee River over the calibration (1998–2001) and conﬁrmation (2002–2005) time periods.

Trang 7

that SWAT is not adequately representing the increase in SRP export

regardless of the mechanism causing this shift

Total nitrogen

Model evaluation statistics indicate mixed success in estimating

TN across the six study watersheds (Table 7) However, PBIAS for all

models was less than 20% for the calibration and conﬁrmation stages,

indicating “very good” average estimation of monthly loads (Moriasi

et al., 2007) Model performance for the three agricultural watersheds

received all “good” and “very good” evaluation statistics for

calibration and conﬁrmation.Fig 4illustrates daily TN loads for the

Maumee, the largest single tributary TN load to Lake Erie TN

estimation for the forested Grand watershed was satisfactory or

nearly so for all evaluation statistics Models for the Huron and

Cuyahoga were once again the weakest, with several unsatisfactory

evaluation statistics for both calibration and conﬁrmation time

periods

Similar to TP model results, TN prediction by SWAT was strongest

for agricultural watersheds and weakest for highly urban watersheds

The urbanized Huron and Cuyahoga once again had much higher

point source inputs relative to observed stream loads, in comparison

with other study watersheds For the Cuyahoga, point source inputs

modeled by SWAT were 5851 kg N/d, while observed stream TN loads

were only 6814 kg N/d In the Huron watershed, point source inputs

were 1460 kg N/d and the river's observed load was 1638 kg N/d

Assuming additional nitrogen inputs from terrestrial lands from fertilizer, manure, soil, and atmospheric deposition, a great amount of

N must be removed in the stream channels of these two watersheds SWAT only includes N settling as a removal mechanism in stream channels, despite strong evidence that denitriﬁcation in stream

2005; Smith et al., 2006) Due to this shortcoming, SWAT is unable to strongly predict the timing of TN loading, though it does predict average loads with reasonable accuracy as seen by the PBIAS statistic (Table 7)

Nitrate Nitrate is the largest fraction of TN in river export from these watersheds and thus is modeled by SWAT with similar accuracy as was observed for TN (Table 8) Models for the Raisin, Maumee, and Sandusky performed acceptably across all four evaluation statistics during both the calibration and conﬁrmation times periods The Huron, Cuyahoga, and Grand models performed inconsistently All six SWAT watershed models, however, had PBIAS statistic values considered to be “very good”

It is evident that SWAT model performance is consistently better for the three agricultural watersheds than it is for the more urban and forested watersheds SWAT was developed and optimized for agricultural watersheds (Gassman et al., 2007), and thus incorporates mechanisms best suited for highly agricultural systems In the case of

Table 6

Calibration and conﬁrmation results for monthly soluble reactive phosphorus loads (Mg P)

for the six modeled watersheds Coefﬁcient of determination (R 2 ), Nash–Sutcliffe

simulation efﬁciency (NSE), percent bias (PBIAS), and the ratio of the root mean square

error to observations standard deviation (RSR) are used as evaluators of model

performance Statistics in bold type are categorized as satisfactory or better ( Moriasi

et al., 2007 ).

Observed mean

(Mg P)

Simulated mean (Mg P)

R 2 NSE PBIAS

Huron 0.2 0.2 0.37 −1.12 −17 1.49

Raisin 2.3 2.2 0.81 0.79 3 0.47

Maumee 31.5 32.2 0.65 0.61 −2 0.63

Sandusky 5.2 5.1 0.43 0.44 1 0.75

Cuyahoga 2.8 8.0 0.05 −38.26 −182 6.33

Grand 0.6 0.6 0.53 0.50 10 0.72

(b) Conﬁrmation

Huron 0.1 0.9 0.75 −182.45 −709 41.87

Raisin 2.1 2.1 0.70 0.70 2 0.55

Maumee 56.9 29.8 0.69 0.29 48 0.85

Sandusky 11.2 7.1 0.45 0.35 37 0.82

Cuyahoga 3.9 8.6 0.11 −34.31 −119 6.00

Grand 1.4 0.8 0.63 0.40 42 0.78

Table 7 Calibration and confirmation results for monthly total nitrogen loads (Mg N) for the six modeled watersheds Coefficient of determination (R 2 ), Nash–Sutcliffe simulation efficiency (NSE), percent bias (PBIAS), and the ratio of the root mean square error to observations standard deviation (RSR) are used as evaluators of model performance Statistics in bold type are categorized as satisfactory or better ( Moriasi et al., 2007 ).

Observed mean (Mg N)

Simulated mean (Mg N)

R 2 NSE PBIAS

Raisin 380 383 0.77 0.77 −1 0.49 Maumee 3199 3189 0.86 0.81 0 0.44 Sandusky 578 604 0.70 0.66 −5 0.59 Cuyahoga 172 190 0.40 0.24 −11 0.88

(b) Conﬁrmation

Raisin 332 338 0.81 0.80 −2 0.45 Maumee 3806 3186 0.80 0.67 16 0.58 Sandusky 867 810 0.80 0.76 6 0.50 Cuyahoga 243 196 0.43 0.27 19 0.86

Fig 4 Daily plot of observed and simulated total nitrogen (TN) loads (kg N/d) for the Maumee River over the calibration (1998–2001) and conﬁrmation (2002–2005) time periods.

Trang 8

nitrate and denitriﬁcation in terrestrial soils, SWAT potentially

predicts this removal mechanism more appropriately in agricultural

soils compared to forest soils In the case of dominant nitrate sources

in the watersheds, SWAT may more accurately simulate nitrate

fertilizer application on agricultural lands than nitrate deposition on

forested lands

Related to nitrate dynamics and land cover is the urban inﬂuence

seen in the Cuyahoga watershed where Cuyahoga N point source

inputs are over twice as high as the stream loads Since the majority of

the Cuyahoga River stream TN load is nitrate, a plot showing daily

nitrate loads illustrates these effects (Fig 5) The most striking

observation is the relatively high baseline of plotted nitrate daily loads

(provided by the high point source inputs) with numerous peaks in

both positive and negative directions These peaks stretching below

the baseline are likely periods when the nitrate concentrations in the

stream were diluted by moderately increased stream discharge SWAT

is not able to adequately predict these downward peaks, which offers

another explanation for weaker SWAT performance in predicting the

Cuyahoga nitrate load

Implications

Model performance statistics and graphical plots reveal that SWAT

was effective in capturing system dynamics and estimating nutrient

and sediment export for the three agricultural watersheds SWAT also

performed reasonably well for the highly forested Grand watershed

Sediment and nutrient results for the more urban Cuyahoga and

Huron watersheds, however, suggest that SWAT should be applied

with caution when agriculture is not the dominant land use Although SWAT accurately simulated the hydrology of these urbanized watersheds, water quality simulation was disappointing, evidently because SWAT does not incorporate adequate mechanisms to remove high point source inputs from the channels

This research also indicates that trade-offs may exist in calibrating SWAT parameters to best represent hydrologic response variables versus water quality variables This appears to be the case for the Cuyahoga sediment and TP calibration as well as Huron TP calibration;

in each case, ﬂow calibration resulted in a decreased ﬁt as measured

by the R2between observed and simulated water quality data — 0.44

to 0.40, 0.52 to 0.30, and 0.44 to 0.27, respectively Usually calibration

of hydrology results in better sediment calibration which, in turn, leads to better P and N calibration success In the two urban watersheds, however, it appears that hydrology calibration resulted

in some parameters and processes important for sediment and TP simulation being altered negatively It is uncertain which parameters were responsible for this effect, but two hydrologic parameter values were notably different for the Huron and Cuyahoga compared to the other four watersheds Groundwater delay (GW_DELAY) was set to less than 1 day for both the Huron and Cuyahoga watersheds, versus

30 to 31 days for the other four watersheds, and the snowpack temperature lag factor (TIMP) was higher for the Huron (0.74) and Cuyahoga (0.23) compared to other watersheds (0.06–0.13) This may

be further cause for caution when applying SWAT to watersheds with substantial urban area

It is apparent that extensive empirical data of high quality are critical for SWAT model applications In the case of the Huron watershed, the limited availability of water quality data for calibration and confirmation was a likely contributor to weak model performance Three years of biweekly water quality data for the Huron model was insufficient to separate the data set into meaningful calibration and confirmation time periods The 8 years of water quality data that were available for the other five watersheds clearly enhanced model performance Even longer time series might be useful, but typically the land cover data comes from one time period (2001 in this study), so lack of correspondence between the time periods of hydrologic and water quality data with the land cover data may then limit the realism of scenarios Sampling frequency also is important, as biweekly sampling does not allow meaningful calibration with the model's daily time step, and even modeling at the monthly time step requires filling in many gaps using some estimation approach (such as the Beale Ratio Estimator

or USGS Estimator) In such cases, the monthly calibration is done based

on another largely simulated data set rather than observed data The ﬁnal implication of this research is that despite weak performance results for some watersheds and certain water quality parameters, SWAT predicted loads, on average, that were not seriously biased This is evidenced by the relatively low PBIAS statistics calculated throughout the calibration and conﬁrmation process While SWAT predictions of the timing or magnitude of

Table 8

Calibration and conﬁrmation results for monthly nitrate loads (Mg N) for the six

modeled watersheds Coefﬁcient of determination (R 2 ), Nash–Sutcliffe simulation

Observed mean

(Mg N)

Simulated mean (Mg N)

R 2 NSE PBIAS

Raisin 306 303 0.61 0.61 1 0.63

Maumee 2613 2632 0.65 0.62 −1 0.62

Sandusky 515 530 0.58 0.53 −3 0.69

Cuyahoga 115 130 0.10 −0.21 −13 1.11

(b) Conﬁrmation

Raisin 277 259 0.72 0.69 6 0.57

Maumee 2958 2659 0.69 0.61 10 0.63

Sandusky 728 717 0.59 0.59 2 0.65

Cuyahoga 135 130 0.16 0.11 4 0.95

Fig 5 Daily plot of observed and simulated nitrate loads (kg N/d) for the Cuyahoga River over the calibration (1998–2001) and conﬁrmation (2002–2005) time periods.

Trang 9

certain loading events or durations cannot be assumed to be accurate

and should be evaluated using appropriate statistics, SWAT accurately

predicted overall average loads for nearly all parameters for all six

watersheds during both the calibration and conﬁrmation time

periods Thus, when accompanied by acknowledgment of model

uncertainty, we have conﬁdence that these six models will be readily

useful for future studies in this region to predict average sediment or

nutrient load changes in response to land use and climate alteration

scenarios

Conclusion

The present study demonstrates the applicability of SWAT in the

Great Lakes basin and also identiﬁes certain considerations for future

SWAT application in general SWAT effectively predicted hydrology and

sediments at daily time steps across a range of watershed characteristics

SWAT estimation of daily nutrient loads was weaker, although still

satisfactory at least two-thirds of the time across all nutrient parameters

and watersheds Furthermore, the PBIAS statistic consistently showed

monthly average loads to be estimated without serious bias Agricultural

and forested watersheds lend themselves particularly well to SWAT

modeling of hydrology, sediments, and nutrients This study also

emphasizes the importance of the availability of observed data with

high sampling frequency and long duration for calibration and

conﬁrmation evaluation and effectiveness Despite these considerations,

SWAT accurately predicted average stream discharge, sediment loads,

and nutrient loads for the Raisin, Maumee, Sandusky, and Grand

watersheds such that future use of these SWAT models for various

scenario testing is reasonable and warranted

Supplementary materials related to this article can be found online

atdoi:10.1016/j.jglr.2011.03.004

Acknowledgments

We are grateful to the following people for their insight related to

this work: Steve Davis, Tim Hunter, Les Ober, Jim Selegean, Michael

Winchell, and Tom VanWagner Comments by reviewers on an earlier

draft substantially improved the manuscript This is publication 10-003

of the EcoFore Lake Erie project, funded by the NOAA Center for

Sponsored Coastal Ocean Research under award NA07OAR4320006

This work was performed under the authority of Section 516(e) of the

Water Resources Development Act of 1996 through additional funding

by the United States Army Corps of Engineers

References

Alvarez-Cobelas, M., Sanchez-Carrillo, S., Angeler, D.G., Sanchez-Andres, R., 2009.

Phosphorus export from catchments: a global view J N Am Benthol Soc 28,

805–820.

Arnold, J.G., Allen, P.M., 1999 Automated methods for estimating baseﬂow and ground

water recharge from streamﬂow records J Amer Water Resour Assoc 35, 411–424.

Baker, D.B., Richards, R.P., 2002 Phosphorus budgets and riverine phosphorus export in

northwestern Ohio watersheds J Environ Qual 31, 96–108.

Bosch NS The inﬂuence of impoundments on the transport and fate of nitrogen and

phosphorus in two southeastern Michigan river catchments Ph.D dissertation,

University of Michigan, Ann Arbor, MI 2007

Bosch, N.S., 2008 The inﬂuence of impoundments on riverine nutrient transport: an

evaluation using the Soil and Water Assessment Tool J Hydrol 355, 131–147.

Bosch, N.S., Allan, J.D., 2008 The inﬂuence of impoundments on nutrient budgets in two

catchments of southeastern Michigan Biogeochem 87, 325–338.

Bosch, N.S., Johengen, T.H., Allan, J.D., Kling, G.W., 2009 Nutrient ﬂuxes across reaches

and impoundments in two southestern Michigan watersheds Lake Res Manag 25,

389–400.

Boyce, F.M., Charlton, M.N., Rathke, D., Mortimer, C.H., Bennett, J., 1987 Lake Erie

research: recent results, remaining gaps J Great Lakes Res 13, 826–840.

Boyer, E.W., Goodale, C.L., Jaworsk, N.A., Howarth, R.W., 2002 Anthropogenic nitrogen

sources and relationships to riverine nitrogen export in the northeastern USA.

Biogeochem 57, 137–169.

Burns, N.M., Rockwell, D.C., Bertram, P.E., Dolan, D.M., Ciborowski, J.J.H., 2005 Trends in temperature, secchi depth, and dissolved oxygen depletion rates in the central basin of Lake Erie, 1983–2002 J Great Lakes Res 31, 35–49.

Castillo, M.M., Allan, J.D., Brunzell, S., 2000 Nutrient concentrations and discharges in a midwestern agricultural catchment J Environ Qual 29, 1142–1151.

Cooke, S.E., Prepas, E.E., 1998 Stream phosphorus and nitrogen export from agricultural and forested watersheds on the Boreal Plain Can J Fish Aquat Sci 55, 2292–2299 DePinto, J.V., Young, T.C., McIlroy, L.M., 1986 Great Lakes water quality improvement — the strategy of phosphosrus discharge control is evaluated Environ Sci Technol.

20, 752–759.

Dolan, D.M., 1993 Point-source loadings of phosphorus to Lake Erie: 1986–1990 J Great Lakes Res 19, 212–223.

Dolan, D.M., McGunagle, K.P., 2005 Lake Erie total phosphorus loading analysis and update: 1996–2002 J Great Lakes Res 31, 11–22.

Gassman, P.W., Reyes, M.R., Green, C.H., Arnold, J.G., 2007 The Soil and Water Assessment Tool: historical development, applications, and future research directions Trans ASABE 50, 1211–1250.

Green, C.H., Tomer, M.D., Di Luzio, M., Arnold, J.G., 2006 Hydrologic evaluation of the Soil and Water Assessment Tool for a large tile-drained watershed in Iowa Trans ASABE 49, 413–422.

Groffman, P.M., Tiedje, J.M., Mokma, D.L., Simkins, S., 1992 Regional scale analysis of denitriﬁcation in north temperate forest soils Landsc Ecol 7, 45–53.

Han, H.J., Allan, J.D., 2008 Estimation of nitrogen inputs to catchments: comparison of methods and consequences for riverine export prediction Biogeochem 91, 177–199.

Han, H., Bosch, N., Allan, J.D., 2011 Spatial and temporal variation in phosphorus budgets for

24 watersheds in the Lake Erie and Lake Michigan basins Biogeochemistry 102, 45–58 Hofstra, N., Bouwman, A.F., 2005 Denitriﬁcation in agricultural soils: summarizing published data and estimating global annual rates Nutr Cycl Agroecosyst 72, 267–278.

Inwood, S.E., Tank, J.L., Bernot, M.J., 2005 Patterns of denitriﬁcation associated with land use in 9 Midwestern headwater streams J N Am Benthol Soc 24, 227–245 Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L., 2007 Model evaluation guidelines for systematic quantiﬁcation of accuracy in watershed simulations Trans ASABE 50, 885–900.

Neitsch, S.L., Arnold, J.G., Kiniry, J.R., Williams, J.R., 2005 Soil and Water Assessment Tool Theoretical Documentation: Version 2005 Agricultural Research Service Temple, TX.

Nemery, J., Garnier, J., Morel, C., 2005 Phosphorus budget in the Marne watershed (France): urban vs diffuse sources, dissolved vs particulate forms Biogeochem 72, 35–66.

Oreskes, N., Shrader-Frechette, K., Belitz, K., 1994 Veriﬁcation, validation, and conﬁrmation of numerical models in the earth sciences Science 263, 641–646 Pieterse, N.M., Bleuten, W., Jorgensen, S.E., 2003 Contribution of point sources and diffuse sources to nitrogen and phosphorus loads in lowland river tributaries J Hydrol 271, 213–225.

Reckhow, K.H., Chapra, S.C., 1983 Conﬁrmation of the water quality models Ecol Model 20, 113–133.

Richards, R.P., 1985 Esimating the extent of reduction needed to statistically demonstrate reduced non-point phosphorus loading to Lake Erie J Great Lakes Res 11, 110–116.

Richards, R.P., 1998 Estimation of Pollutant Loads in Rivers and Streams: A Guidance Document for NPS Programs U.S Environmental Protection Agency, Denver, CO.

108 pp.

Richards, R.P., 2006 Trends in sediment and nutrients in major Lake Erie tributaries, 1975–2004 Section 10.10 in Lake Erie LaMP, 2006 update.

Richards, R.P., 2007 Phosphorus loads and concentrations from the Maumee River In: Hartig, J.H., Zarull, M.A., Ciborowski, J.H., Gannon, J.E., Wilke, E., Norwood, G., Vincent, A (Eds.), State of the Strait: Status and Trends of Key Indicators University

of Windsor, Windsor, pp 68–74.

Robertson, D.M., 1997 Regionalized loads of sediment and phosphorus to Lakes Michigan and Superior — high ﬂow and long-term average J Great Lakes Res 23, 416–439.

Rosa, F., Burns, N.M., 1987 Lake Erie central basin oxygen depletion changes from 1929–1980 J Great Lakes Res 13, 684–696.

Ruddy, B.C., Lorenz, D.L., Mueller, D.K., 2006 County-level estimates of nutrient inputs

to the land surface of the conterminous United States, 1982–2001 National Water Quality Assessment Program U.S Geological Survey.

Santhi, C., Arnold, J.G., Williams, J.R., Dugas, W.A., Srinivasan, R., Hauck, L.M., 2001 Validation of the SWAT model on a large river basin with point and nonpoint sources J Amer Water Resour Assoc 37, 1169–1188.

Sharpley, A.N., 1999 Global issues of phosphorus in terrestrial ecosystems Phosphorus Biogeochem Subtropical Ecosyst 15–46.

Smith, L.K., Voytek, M.A., Bohlke, J.K., Harvey, J.W., 2006 Denitriﬁcation in nitrate-rich streams: application of N-2: Ar and N-15-tracer methods in intact cores Ecol Appl.

16, 2191–2207.

van Griensven, A., Meixner, T., 2007 A global and efﬁcient multi-objective auto-calibration and uncertainty estimation method for water quality catchment models J Hydroinformatics 9, 277–291.

van Griensven, A., Meixner, T., Grunwald, S., Bishop, T., Diluzio, A., Srinivasan, R., 2006 A global sensitivity analysis tool for the parameters of multi-variable catchment models J Hydrol 324, 10–23.

Định dạng
Số trang	9
Dung lượng	743,03 KB