Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
A MULTI-CONFIGURATION EVALUATION OF THE SOIL AND WATER
ASSESSMENT TOOL (SWAT) IN A MIXED LAND USE WATERSHED IN THE
CENTRAL U.S.A.
_____________________________________________________
A Thesis presented to the
faculty of the Graduate
School at the University of
Missouri-Columbia
______________________________________________________________
In Partial Fulfillment of the
Requirements for the Degree
Master of Science
________________________________________________________
By
DANIEL P. SCOLLAN
Jason A. Hubbart, Thesis Advisor
May 2011
© Copyright by Daniel P. Scollan 2011
All Rights Reserved
The undersigned, appointed by the dean of the Graduate School, have examined the
thesis entitled
A MULTI-CONFIGURATION EVALUATION OF THE SOIL AND WATER
ASSESSMENT TOOL (SWAT) IN A MIXED LAND USE WATERSHED IN THE
CENTRAL U.S.A.
presented by Daniel P. Scollan, a candidate for the degree of Master of Science, and
hereby certify that, in their opinion, it is worthy of acceptance.
Jason A. Hubbart, Ph. D., Thesis Advisor
Professor Stephen H. Anderson
Professor Richard P. Guyette
ii
ACKNOWLEDGMENTS
First and foremost acknowledgement is due to Dr. Jason A. Hubbart, Ph.D.,
Department of Forestry, and director of the Interdisciplinary Hydrology Laboratory
(IHL), who initiated and propelled this thesis. Also, thanks to current and former IHL
members for their tremendous assistance in field hydroclimatic data collection and post-
processing; without the assistance of these lab members, this thesis would not have been
possible. Second, but no less important, great appreciation is expressed to Mr. Gary L.
Wilson, U.S. Geological Survey (USGS) Hydrologist, at the Missouri Water Science
Center, Surface Water Unit, Rolla, Missouri, who not only helped provided essential
streamflow data for this study but did so per request on a schedule that was compatible
with the on-time completion of this work.
In addition, Missouri State Climatologist Dr. Patrick E. Guinan, Ph.D., provided
valuable assistance obtaining and selecting climate datasets from the University of
Missouri - Agriculture Experiment Station‟s Commercial Agriculture Automated
Weather Station Network. Thanks are extended to Dr. Aleksey Y. Sheshukov, Ph.D.,
Research Associate, Kansas State University, whose constructive and thought-provoking
criticism at the American Geophysical Union (AGU) Fall Meeting in 2009 improved this
thesis. Graduate tuition and stipend support was provided by the State of Missouri
Department of Natural Resources (MDNR) during the third year, and the University of
Missouri Division of Biological Sciences during the first and second years.
iii
Lastly, great thanks are due to each and every faculty member at the University of
Missouri who have provided assistance and teaching. They have aided the development
of this thesis in many ways both direct and indirect. A special appreciation is expressed to
both Dr. Christoph E. Geiss, Ph.D., undergraduate advisor at Trinity College, Hartford,
Connecticut, and Mr. Christopher K. Metcalf, M.S., Fish Biologist at the U.S. Fish and
Wildlife Service, Panama City, Florida, who provided critical guidance and direction.
Their contributions were invaluable.
iv
TABLE OF CONTENTS
ACKNOWLEDGMENTS .................................................................................... ii
TABLE OF CONTENTS .................................................................................... iv
LIST OF FIGURES ............................................................................................ vii
LIST OF TABLES ................................................................................................ x
ABSTRACT ......................................................................................................... xv
CHAPTER I Introduction .................................................................................... 1
Hydrologic / Water Quality Modeling ................................................................ 1
The H/WQ Modeling Process ......................................................................... 5
Description of the Soil and Water Assessment Tool .......................................... 5
A New Framework for Model Development and Evaluation ........................... 12
Study Objectives ............................................................................................... 17
CHAPTER II Methods ....................................................................................... 21
Study Site Description ...................................................................................... 21
Climate .......................................................................................................... 22
Land Use and Land Cover ............................................................................ 22
Soils and Vegetation ..................................................................................... 26
Topography ................................................................................................... 28
Water Quantity and Quality .......................................................................... 29
v
Data Collection ................................................................................................. 30
Nested Watershed Design ............................................................................. 30
Climate Data ................................................................................................. 31
Stage Data ..................................................................................................... 32
Streamflow Measurements............................................................................ 32
Rating Curve Development............................................................................... 39
Initial Rating Curve Development ................................................................ 39
Trial Rating Curve Testing and Selection of Final Rating Curves ............... 40
Computation of the Continuous Streamflow Record .................................... 41
Model Implementation ...................................................................................... 43
SWAT Model Configuration ........................................................................ 43
Evaluation of Uncalibrated Model Runs....................................................... 50
Test of the Built-In Automatic Calibration Method in SWAT ..................... 59
CHAPTER III Results ........................................................................................ 68
Observed Climate.............................................................................................. 68
Developed Stage-Discharge Rating Curves ...................................................... 76
Observed Streamflow........................................................................................ 78
Uncalibrated Model Configurations ................................................................. 83
Automatic Calibration Comparison .................................................................. 98
CHAPTER IV Discussion ................................................................................ 104
Analysis of Observed Climate Data ................................................................ 104
Analysis of Observed Streamflow Data .......................................................... 107
vi
Uncalibrated Model Performance ................................................................... 110
Effect of Watershed Subdivision and Input Dataset Selection on Model Fit . 111
Watershed Subdivision ............................................................................... 112
Soil Data Resolution ................................................................................... 114
Climate Data ............................................................................................... 119
Evaluation of the SWAT Automatic Calibration Procedure ........................... 121
Comparison of H/WQ Model Fit Evaluation Methods ................................... 125
CHAPTER V Conclusion ................................................................................. 129
Findings on Hydroclimate in the Hinkson Creek Watershed ......................... 129
Recommendations for H/WQ Model Configuration....................................... 130
Recommendations for H/WQ Model Calibration ........................................... 131
Recommendations for H/WQ Model Evaluation ............................................ 133
Recommendations for Future Research in H/WQ Modeling .......................... 134
Contributions to Science ................................................................................. 135
LITERATURE CITED .................................................................................... 137
APPENDIX A: Stage-Discharge Rating Curves ............................................ 144
APPENDIX B: Complete Modeling Results ................................................... 149
vii
LIST OF FIGURES
Figure Page
FIGURE 1. DIAGRAM OF STREAMFLOW SIMULATION IN THE SOIL AND WATER ASSESSMENT
TOOL (SWAT) MODEL, VERSION 2005. ....................................................................... 8
FIGURE 2. MAP OF GAUGING STATIONS AND ASSOCIATED SUB-BASINS WITH LAND USE /
LAND COVER CLASSIFICATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI,
U.S.A......................................................................................................................... 25
FIGURE 3. HYDROLOGIC SOIL GROUPS BY SOIL TYPE IN THE HINKSON CREEK WATERSHED,
MISSOURI, U.S.A. SOIL DATA FROM THE SOIL SURVEY GEOGRAPHIC DATABASE
(SSURGO), 2011. ..................................................................................................... 27
FIGURE 4. HYPSOMETRIC CURVES SHOWING DISTRIBUTION OF ELEVATION FOR EACH SUB-
BASIN ASSOCIATED WITH FIVE NESTED GAUGING STATIONS AND THE OUTLET OF THE
HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .................................................... 28
FIGURE 5. PHOTOGRAPH DATED MAY 16, 2009 SHOWING VARIABLE BACKWATER AFFECTED
STREAMFLOW AT THE SITE #5 GAUGING STATION, HINKSON CREEK WATERSHED,
MISSOURI, U.S.A. ...................................................................................................... 34
FIGURE 6. PHOTOGRAPH SHOWING WADING STREAMFLOW MEASUREMENT TECHNIQUE AT
SITE #2, HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ....................................... 36
FIGURE 7. PHOTOGRAPH SHOWING OBLIQUE ANGLE BETWEEN HINKSON CREEK AND THE
TRANSPORTATION BRIDGE AT SITE #2, HINKSON CREEK WATERSHED, MISSOURI,
U.S.A......................................................................................................................... 38
viii
FIGURE 8. COMPARISON OF (1) LOW RESOLUTION SIX SUB-BASIN AND HIGH RESOLUTION 34
SUB-BASIN WATERSHED DISCRETIZATION SCHEMES AND (2) LOW RESOLUTION
STATSGO SOIL DATA AND HIGH RESOLUTION SSURGO SOIL USED IN SWAT
MODELING. ................................................................................................................. 47
FIGURE 9. PHOTOGRAPH OF SITE #1 CLIMATE STATION DATED FEBRUARY 2011, SHOWING
THE SURROUNDING WOODLAND, A POTENTIAL SOURCE OF PRECIPITATION
UNDERCATCH. HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ............................. 49
FIGURE 10. DAILY PRECIPITATION AT EACH OF SEVEN CLIMATE STATIONS USED IN
MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ....................... 70
FIGURE 11. DAILY MAXIMUM AIR TEMPERATURE AT EACH OF SEVEN CLIMATE STATIONS
USED IN MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .......... 71
FIGURE 12. DAILY MINIMUM AIR TEMPERATURE AT EACH OF SEVEN CLIMATE STATIONS
USED IN MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .......... 72
FIGURE 13. DAILY MEAN RELATIVE HUMIDITY AT EACH OF SEVEN CLIMATE STATIONS USED
IN MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ................... 73
FIGURE 14. DAILY MEAN WIND SPEED AT EACH OF SEVEN CLIMATE STATIONS USED IN
MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ....................... 74
FIGURE 15. DAILY TOTAL SOLAR RADIATION AT EACH OF SEVEN CLIMATE STATIONS USED
IN MODELING OF THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ................... 75
FIGURE 16. PLOT OF FINAL RATING CURVES USED FOR CONTINUOUS STREAMFLOW
ESTIMATION AT ALL FIVE GAUGING STATIONS THAT ALSO SHOWS STREAMFLOW
MEASUREMENTS TAKEN AT EACH GAUGING STATION IN THE HINKSON CREEK
ix
WATERSHED, MISSOURI, U.S.A. FOR THE USGS-OPERATED GAUGING STATION, SITE
#4, ONLY AVAILABLE INFORMATION IS SHOWN. ......................................................... 77
FIGURE 17. FLOW DURATION CURVE FOR OBSERVED DAILY MEAN STREAMFLOW AT FIVE
GAUGING STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .......... 79
FIGURE 18. HYDROGRAPH FOR OBSERVED DAILY MEAN STREAMFLOW AT FIVE GAUGING
STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ......................... 80
FIGURE 19. FLOW DURATION CURVE FOR OBSERVED MONTHLY MEAN STREAMFLOW AT FIVE
GAUGING STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .......... 81
FIGURE 20. HYDROGRAPH FOR OBSERVED MONTHLY MEAN STREAMFLOW AT FIVE GAUGING
STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ......................... 82
FIGURE 21. GRAPHICAL DIAGRAM SHOWING DIFFERENCES BETWEEN ABSOLUTE ERROR AND
SQUARED ERROR RELATIVE TO ACTUAL ERRORS BETWEEN MODELED AND OBSERVED
DATA. ....................................................................................................................... 128
FIGURE 22. DETAILED PLOT OF SITE #1 RATING CURVE AND FLOW MEASUREMENTS. ...... 145
FIGURE 23. DETAILED PLOT OF SITE #2 RATING CURVE AND FLOW MEASUREMENTS. ...... 146
FIGURE 24. DETAILED PLOT OF SITE #3 RATING CURVE AND FLOW MEASUREMENTS. ...... 147
FIGURE 25. DETAILED PLOT OF SITE #5 RATING CURVE AND FLOW MEASUREMENTS. ...... 148
x
LIST OF TABLES
Table Page
TABLE 1. TIMELINE OF SWAT DEVELOPMENT. ADAPTED FROM NEITSCH ET AL. (2005). .. 11
TABLE 2: TOTAL SUB-BASIN AREA (HA), LAND-USE AREA (%) FOR EACH OF FIVE GAUGE
SITES AND CUMULATIVE CONTRIBUTING AND LAND-USE AREA (ASSUMING 15 LAND-
USE DIVISIONS) IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A.
CONTRIBUTING AREAS AND LAND-USE COVER CLASSES WERE DETERMINED USING 10 M
DEM DATA AND 30 M LAND-USE COVER DATA. ......................................................... 24
TABLE 3. FIELD INSTRUMENTATION AND VARIABLES MEASURED AT FIVE HYDROCLIMATE
STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. ......................... 32
TABLE 4. LIST OF ALL 20 SWAT MODEL CONFIGURATIONS TESTED, THE ABBREVIATED
NAMES OF EACH MODEL, AND THE DEFINING CHARACTERISTICS OF EACH MODEL.
CONFIGURATIONS ARE SYSTEMATICALLY LISTED IN ORDER OF INCREASING INPUT
DATA RESOLUTION. .................................................................................................... 50
TABLE 5. DESCRIPTIVE TABLE SHOWING KEY CHARACTERISTICS OF GOODNESS-OF-FIT
INDICATORS USED FOR MODEL EVALUATION IN THIS STUDY. ...................................... 57
TABLE 6. SETTINGS USED IN ARCSWAT WHEN RUNNING THE BUILT-IN AUTOMATIC
CALIBRATION METHOD. .............................................................................................. 63
TABLE 7. THE SIX INPUT PARAMETERS IN SWAT SELECTED FOR OPTIMIZATION WITH THE
BUILT-IN AUTOMATIC CALIBRATION PROCEDURE, THEIR INITIAL VALUES, AND THE
VARIATION SETTINGS. ................................................................................................ 65
xi
TABLE 8. SUMMARY OF CLIMATE DATA DURING 2009-2010 FOR FIVE HYDROCLIMATE
STATIONS IN THE HINKSON CREEK WATERSHED, MISSOURI, U.S.A. AND THE MU
AGRICULTURAL EXPERIMENTAL STATION‟S SANBORN FIELD AND SOUTH FARMS
WEATHER STATIONS. VALUES FOR STANDARD DEVIATION ARE SHOWN IN
PARENTHESES............................................................................................................. 69
TABLE 9. DESCRIPTIVE STATISTICS FOR OBSERVED DAILY STREAMFLOW IN 2009-2010,
HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .................................................... 79
TABLE 10. DESCRIPTIVE STATISTICS FOR OBSERVED MONTHLY STREAMFLOW IN 2009-2010,
HINKSON CREEK WATERSHED, MISSOURI, U.S.A. .................................................... 81
TABLE 11. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS FOR ALL TWENTY
UNCALIBRATED MODEL CONFIGURATIONS. OPTIMAL GOODNESS-OF-FIT VALUES ARE IN
BOLD. CONFIGURATIONS ARE SYSTEMATICALLY LISTED IN ORDER OF INCREASING
INPUT DATA RESOLUTION. ERROR MEASURES GIVEN IN THE TABLE ARE THE MEAN OF
THE MEASURES FOR ALL FIVE GAUGING STATIONS. .................................................... 86
TABLE 12. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS COMPARING MODEL RUNS
WITH 6 SUB-BASINS AND MODEL RUNS WITH 34 SUB-BASINS. ERROR MEASURES GIVEN
IN THE TABLE ARE THE MEAN OF THE MEASURES FOR ALL FIVE GAUGING STATIONS.
OPTIMAL GOODNESS-OF-FIT VALUES ARE IN BOLD. .................................................... 88
TABLE 13. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS COMPARING MODELS USING
THE LOW RESOLUTION STATSGO SOIL DATASET AND MODELS USING THE SSURGO
SOIL DATASET. ERROR MEASURES SHOWN ARE THE MEAN OF THE MEASURES FOR ALL
FIVE GAUGING STATIONS. OPTIMAL GOODNESS-OF-FIT VALUES ARE IN BOLD. ........... 89
xii
TABLE 14. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS COMPARING MODELS USING
THE SWAT WEATHER GENERATOR, THE SINGLE CLIMATE DATASET FROM URBAN-
LOCATED SANBORN FIELD, A SINGLE CLIMATE DATASET FROM RURAL-LOCATED
SOUTH FARMS, FOUR CLIMATE DATASETS FROM HCW SITES #2-5, AND FIVE CLIMATE
DATASETS FROM HCW SITES #1-5. ERROR MEASURES ARE THE MEAN OF THE
MEASURES FOR ALL FIVE GAUGING STATIONS. OPTIMAL GOODNESS-OF-FIT VALUES
ARE IN BOLD. .............................................................................................................. 91
TABLE 15. UNCALIBRATED MODEL RESULTS RANKED BY A MEASURE OF MASS BALANCE FIT:
DAILY PERCENT BIAS. THE PERCENT BIAS SHOWN IS THE AVERAGE OF THE ABSOLUTE
VALUES OF EACH PERCENT BIAS MEASURE AT ALL FIVE HCW GAUGING STATIONS. . 92
TABLE 16. UNCALIBRATED MODELS RANKED BY MEASURES OF HYDROGRAPH FIT: DAILY
NASH SUTCLIFFE EFFICIENCY AND MODIFIED NASH SUTCLIFFE EFFICIENCY. ........... 93
TABLE 17. UNCALIBRATED MODELS RANKED BY MEASURES OF FLOW DURATION FIT: DAILY
RANKED NASH SUTCLIFFE EFFICIENCY AND RANKED MODIFIED NASH SUTCLIFFE
EFFICIENCY. ............................................................................................................... 94
TABLE 18. DEFAULT UNCALIBRATED MODEL RUNS RANKED BY A MEASURE OF MASS
BALANCE FIT: MONTHLY PERCENT BIAS. PERCENT BIAS IS THE AVERAGE OF THE
ABSOLUTE VALUES OF EACH PERCENT BIAS MEASURE AT ALL FIVE HCW GAUGING
STATIONS. .................................................................................................................. 95
TABLE 19. DEFAULT UNCALIBRATED MODEL RUNS RANKED BY MEASURES OF HYDROGRAPH
FIT: MONTHLY NASH SUTCLIFFE EFFICIENCY AND MODIFIED NASH SUTCLIFFE
EFFICIENCY. ............................................................................................................... 96
xiii
TABLE 20. DEFAULT UNCALIBRATED MODEL RUNS RANKED BY MEASURES OF FLOW
DURATION CURVE FIT: DAILY RANKED NASH SUTCLIFFE EFFICIENCY AND RANKED
MODIFIED NASH SUTCLIFFE EFFICIENCY. .................................................................. 97
TABLE 21. VALUES FOR THE BEST SET OF SIX INPUT PARAMETERS FOR THE SELECTED
MODEL CONFIGURATION OPTIMIZED FIRST BY MINIMIZING THE SUM OF SQUARED
ERROR IN DAILY STREAMFLOW AT SITE #4 AND SECOND BY MINIMIZING THE SUM OF
SQUARED ERROR AND THE SUM OF SQUARED ERROR AFTER RANKING USING THE BUILT-
IN AUTOMATIC CALIBRATION METHOD IN SWAT. .................................................... 100
TABLE 22. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS FOR SITE #4 COMPARING
THE SELECTED UNCALIBRATED MODEL CONFIGURATION (34 SUB-BASIN, SSURGO
SOIL DATA, AND HCW-4 CLIMATE DATASET) WITH SINGLE OBJECTIVE (SSQ) AND
MULTIPLE OBJECTIVE (SSQ & SSQR) AUTOMATICALLY CALIBRATED RUNS OF THE
SELECTED MODEL. OPTIMAL GOODNESS-OF-FIT VALUES ARE IN BOLD. .................... 102
TABLE 23. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS FOR SITES #1-3, AND 5
COMPARING THE SELECTED UNCALIBRATED MODEL CONFIGURATION (34 SUB-BASIN,
SSURGO SOIL DATA, AND HCW-4 CLIMATE DATASET) SINGLE OBJECTIVE (SSQ) AND
MULTIPLE OBJECTIVE (SSQ & SSQR) AUTOMATICALLY CALIBRATED RUNS OF THE
SELECTED MODEL. OPTIMAL GOODNESS-OF-FIT VALUES ARE IN BOLD. .................... 103
TABLE 24. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS COMPARING SWAT MODEL
CONFIGURATIONS WITH VARYING LEVELS OF WATERSHED SUBDIVISION (KUMAR AND
MERWADE 1999). WATERSHED SUBDIVISIONS ARE DEFINED BY PERCENT
CONTRIBUTING SOURCE AREA (CSA). ERROR MEASURES SHOWN ARE THE MEAN OF
THE MEASURES FOR ALL 24 MODELS. ....................................................................... 114
xiv
TABLE 25. GOODNESS-OF-FIT MODEL EVALUATION STATISTICS COMPARING SWAT MODELS
CONFIGURED WITH LOW RESOLUTION SOIL DATA (STATSGO) AND HIGH RESOLUTION
SOIL DATA (SSURGO) (KUMAR AND MERWADE 2009). ERROR MEASURES SHOWN
ARE THE MEAN OF THE MEASURES FOR ALL 24 MODELS. .......................................... 117
TABLE 26. EQUATIONS FOR RATING CURVES APPLIED IN THE STUDY IN HINKSON CREEK
WATERSHED, MISSOURI, U.S.A. THE „Y‟ VARIABLE IS EQUIVALENT TO DISCHARGE IN
M3/S; THE „X‟ VARIABLE IS EQUIVALENT TO STAGE IN M........................................... 144
TABLE 27. FULL DAILY STREAMFLOW DESCRIPTIVE AND GOODNESS-OF-FIT STATISTICS FOR
ALL UNCALIBRATED SWAT MODEL CONFIGURATIONS. AVERAGE PERCENT BIAS IS
THE AVERAGE OF THE ABSOLUTE VALUES OF EACH PERCENT BIAS MEASURE. ......... 149
TABLE 28. FULL MONTHLY STREAMFLOW DESCRIPTIVE AND GOODNESS-OF-FIT STATISTICS
FOR ALL UNCALIBRATED SWAT MODEL CONFIGURATIONS. AVERAGE PERCENT BIAS
IS THE AVERAGE OF THE ABSOLUTE VALUES OF EACH PERCENT BIAS MEASURE. ..... 160
TABLE 29. FULL DAILY AND MONTHLY STREAMFLOW DESCRIPTIVE AND GOODNESS-OF-FIT
STATISTICS FOR THE SELECTED UNCALIBRATED SWAT MODEL CONFIGURATION, THE
SUM OF SQUARE ERROR (SSQ) OPTIMIZED MODEL, AND THE SUM OF SQUARED ERROR
(SSQ) AND SUM OF SQUARED ERROR AFTER RANKING (SSQR) OPTIMIZED MODEL.
AVERAGE PERCENT BIAS SHOWN IS THE AVERAGE OF THE ABSOLUTE VALUES OF EACH
PERCENT BIAS MEASURE. ........................................................................................ 171
xv
A MULTI-CONFIGURATION EVALUATION OF THE SOIL AND WATER
ASSESSMENT TOOL (SWAT) IN A MIXED LAND USE WATERSHED IN THE
CENTRAL U.S.A.
Daniel P. Scollan
Jason A. Hubbart, Thesis Advisor
ABSTRACT
Distributed watershed hydrologic/water quality (H/WQ) models are ubiquitous tools for
watershed management. Despite advancements, there remain impediments for end-users.
This study presents a practical framework for use of the Soil and Water Assessment Tool
(SWAT). Results show variable accuracy across scales and evaluation methods using 20
model configurations based on two watershed subdivisions, two soil datasets, and five
climate datasets. Nine goodness-of-fit indicators were tested, including four new indices
(R-RMSE, R-MAE, R-NSE, and R-NSE1) designed to quantify model fit with flow
distribution. Sixteen of 20 configurations achieved satisfactory monthly streamflow fit
(NSE > 0.5, PBIAS < 25%) without calibration. Watershed and soil resolution had
negligible impact; climate input had considerable impact. Single climate station input is
best used for applications requiring monthly predictions; distributed climate station input
is needed for daily predictions. SWAT multi-objective auto-calibration better predicted
monthly flow (PBIAS=1%, NSE=0.8) than single-objective calibration (PBIAS=16%,
NSE=0.5). SWAT performs well in Central U.S. urbanizing watersheds. Accuracy can
improve with auto-calibration as presented and continued model development.
1
CHAPTER I
INTRODUCTION
HYDROLOGIC / WATER QUALITY MODELING
Fresh water is a critical and irreplaceable natural resource necessary for human
health, agriculture, industry, recreation, and ecosystem integrity (O‟Neill et al. 2006).
Increasingly, fresh water resources are threatened by human development activities
resulting in water quality problems that include sediment from terrestrial and aquatic
erosion processes, and contamination by chemicals from point and non-point sources
(Borah and Bera 2004). The EPA‟s Wade-able Streams Assessment, a biological
assessment of 1,392 randomly selected wade-able stream sites in the conterminous U.S.,
found that 42% of the nation‟s wade-able stream length is in poor biological condition
relative to reference site conditions (USEPA 2006).
To help address and manage pressing water resource problems, many
Hydrologic/Water Quality (H/WQ) models have been developed. H/WQ models are
generally conceptual (Nash and Sutcliffe 1970) mathematical models capable of
simulating hydrologic processes in the land phase of the hydrologic cycle (Borah and
Bera 2003; Migliaccio and Srivastava 2007; Srivastava et al. 2007). H/WQ models
attempt to represent physical processes (i.e. precipitation, evaporation, transpiration,
infiltration, runoff, streamflow) by means of discrete, analytical, and algebraic
mathematical expressions (Migliaccio and Srivastava 2007). The expressions may be
empirical, featuring a small number of parameters that are typically not physically
2
measureable and that are determined by calibration against a large number of observed
data (Merritt et al. 2003; Migliaccio and Srivastava 2007). H/WQ model expressions may
also be physically based, featuring a larger number of physically measurable parameters
arranged as a solution to fundamental physical laws such as the conservation of mass,
momentum, or energy (Merritt et al. 2003; Migliaccio and Srivastava 2007). The overall
hydrologic cycle is generally physically represented via the water (or mass) balance
equation (Srivastava et al. 2007). As presented by Dingman (2002), the mass balance for
a watershed may be written as:
P + Gin – (Q + ET + Gout) = S
where P is precipitation (rainfall, snow, occult precipitation), Gin is the inflow of
groundwater, Q is streamflow (discharge), ET is evapotranspiration (direct surface
evaporation and plant transpiration), Gout is the outflow of groundwater, and S is the
change in storage of water in the watershed.
In addition to simulating the movement of water through the hydrologic cycle, an
essential component to H/WQ models is their inclusion of mathematical simulations
quantifying processes critical to water quality, i.e. erosion, sediment transport, nutrient
transport, and pesticide transport (Borah and Bera 2003). H/WQ models are generally
designed to function at a particular spatial scale, including profile / horizon / pedon (point
scale), field or field scale, or watershed scale (Srivastava et al. 2007), and at a particular
temporal scale, ranging from single event to annual or decadal (Merritt et al. 2003).
Data that are generally required for input into watershed scale H/WQ models
include, but are not limited to, climate, topography, soil physical properties, and land
3
cover / land use (Engel et al. 2007). When models are run, simulated estimates of water
yield, sediment yield, and chemical (nutrient and pesticide) yields at the watershed outlet
are generated (USACE 1999). The estimates provided by H/WQ models have been
shown to be highly useful in water resource management applications including
predicting the effects of climate change (Mehta et al. 2011; Parajuli 2010; Stone et al.
2001), predicting the effects of land use change (Ghaffari et al. 2010), assessing the use
of Best Management Practices (BMPs) for water quality protection (Arabi et al. 2006;
O‟Donnell et al. 2008, Santhi et al. 2006), predicting soil nutrient (e.g. phosphorus) loss
in agricultural fields (White et al. 2009), and assessing bacterial pollutant (e.g.
Escherichia coli) loading in coastal estuaries (Bougeard et al. 2011).
A key component of H/WQ model applicability is their capability to extend
findings from localized monitoring studies for application in watersheds in other
environments and at broader scales. For example, H/WQ models may be used to evaluate
the effectiveness of Best Management Practices (BMPs). However, monitoring studies
conducted to evaluate Best Management Practices (BMPs) that protect water quantity and
quality in one area may not always be applicable to other areas. The lack of
transferability is normally due to the inherent variability of soil characteristics, vegetation
communities, climate and hydrologic regime, and land management practices across the
landscape (O‟Neill et al. 2006). Since it is impractical to monitor runoff, sediment and
other water quality variables in every watershed, modeled estimates of these variables
based on a set of inputs, including soil characteristics, slope, elevation, climate, and
management practices, may be used to evaluate BMPs in other ungauged watersheds and
develop BMP implementation strategies (O‟Neill et al. 2006). For instance, a Soil and
4
Water Assessment Tool (SWAT) model application by O‟Donnell et al. (2008) in the Le
Moine River Basin (3,494 km2) in Illinois, U.S.A., found that the current locations of
BMPs for sediment erosion prevention in the watershed were located in areas where they
had little effect on reducing sediment loads; the study therefore determined areas to target
future sediment BMPs. White et al. (2009) developed a simplified management tool
based on the SWAT model to combine field-scale quantitative management of soil
phosphorus loss with basin-scale estimation of downstream phosphorus water quality.
These examples illustrate the ability of H/WQ models to apply localized assessment of
management activities on a broad scale incorporating varying physical environmental
characteristics.
Hydrologic/Water Quality models may also be used to develop Total Maximum
Daily Loads (TMDLs) for impaired water bodies (Radcliffe et al. 2009; O‟Neill et al.
2006; Borah and Bera 2003). TMDLs are calculations of the maximum amount of a
pollutant that a water body may receive while meeting established water quality
standards; states are required by the Clean Water Act to develop TMDLs for impaired
water bodies (USEPA 2008). H/WQ models may also be used to provide
recommendations for the type and placement of specific BMPs, and for the prediction of
hydrologic and water quality changes in response to changes in land use, land
management, and climate (Borah and Bera 2004; O‟Neill et al. 2006; USACE 1999).
5
The H/WQ Modeling Process
To provide clarity in the coming text, the following terms, supplied in sequential
outline form, are presented describing the main steps for the end-user in H/WQ modeling
implementation and application process:
1) Implementation
a. Configuration: Selection and preparation of input datasets.
i. Goodness-of-fit evaluation of an uncalibrated model.
b. Calibration: Optimization of input parameters to reduce error between
modeled output and corresponding set of observed data. May be
preceded by a sensitivity analysis to identify parameters with the
greatest effect on model output.
c. Validation: A method of assessing calibrated values for the model
input parameters by the evaluation of the modeled output against an
independent set of observed data.
2) Application
a. Application of the model for predictions in ungauged basins or sub-
basins.
b. Alteration of model input datasets or input parameters to make a
modeled prediction of change due to a management scenario.
DESCRIPTION OF THE SOIL AND WATER ASSESSMENT TOOL
The Soil and Water Assessment Tool (SWAT) is a conceptual, physically-based,
semi-distributed parameter, long-term, daily-time-step Hydrologic/Water Quality
6
(H/WQ) model; SWAT was developed by the U.S. Department of Agriculture –
Agricultural Research Service (USDA-ARS) primarily at the Grassland, Soil, and Water
Research Laboratory in Temple, Texas (Gassman et al. 2007). The model was designed
for basin-scale applications, and has been previously applied to the conterminous U.S.A
(Srinivasan et al. 1998) and the entire Missouri River Basin (Stone et al. 2001; Mehta et
al. 2011) with mixed success. SWAT is currently being used as part of the U.S.
Department of Agriculture (USDA) Conservation Effects Assessment Program (CEAP),
which is intended to quantify the cumulative environmental benefits of the USDA‟s
conservation programs on cultivated, range, and irrigated lands in the U.S.A. (Van Liew
et al. 2007). SWAT is capable of predicting water, sediment, and chemical yields in
ungauged basins (Gassman et al. 2007). Major routines of the SWAT model include
weather generation, hydrology, sediment, crop growth, nutrients, and pesticides (Neitsch
et al. 2005).
The SWAT model, first released in the early 1990s, incorporates components
from several models developed over the last 30 years at USDA-ARS, including the
Chemical, Runoff, and Erosion from Agricultural Management Systems (CREAMS)
model, the Groundwater Loading Effects on Agricultural Management Systems
(GLEAMS) model, the Environmental Productivity Impact Calculator (EPIC) model, the
Simulator for Water Resources in Rural Basins (SWRRB) model, the Routing Outputs to
Outlets (ROTO) model, and the QUAL2E in-stream model (Gassman et al. 2007). The
latest version of the SWAT model is SWAT 2009; because the 2009 version had not yet
been released at the time of this study and because documentation for the 2009 version is
7
yet to be released, this study evaluates the next-most recent version, SWAT 2005. A
timeline of SWAT development milestones is presented in Table 1.
SWAT is capable of dividing the watershed into multiple sub-basins, which then
may be divided into multiple non-spatially-explicit hydrologic response units (HRUs)
that are populated with a lumped (single) set of input parameters that define homogenous
soil, management, and land use characteristics (Gassman et al. 2007) (Figure 1). HRUs
are defined by selecting a minimum, or threshold percentage of the sub-basin area that is
composed of a unique combination of a land use, soil type, and/or topographic slope
(Santhi et al. 2005). A geographic information system (GIS) extension, ArcSWAT, is
available to simplify the development of model input files (Di Luzio et al. 2004).
SWAT features routines to model the entire water balance (Neitsch et al. 2005)
(Figure 1). Surface runoff and infiltration are calculated using the empirical Soil
Conservation Service (SCS) Curve Number method by default, or, optionally, the semi-
theoretical Green-Ampt-Mein-Larson infiltration method. Canopy interception is implicit
in the Curve Number method. Potential evapotranspiration (PET) is calculated by the
semi-theoretical Penman-Monteith equation by default, or, optionally, the Priestly-Taylor
or Hargreaves-Samani method. Soil water redistribution is modeled using a storage
routing method (Neitsch et al. 2005; Gassman et al. 2007).
8
Figure 1. Diagram of streamflow simulation in the Soil and Water Assessment Tool (SWAT) model, version 2005.
Over 100 hydrologic calibration and/or validation studies of SWAT were
published in the literature (Gassman et al. 2007). Judged by the criteria set forth by
Moriasi et al. (2007), the majority of the studies reviewed by Gassman et al. (2007) were
considered to adequately simulate streamflow on a monthly basis. This judgment was
based on the Nash Sutcliffe Efficiency (NSE) (Nash and Sutcliffe 1970), a statistical
goodness-of-fit indicator that ranges from -∞ to 1, with 1 being the optimal value.
Gassman et al. (2007) found that most models obtained satisfactory results (NSE > 0.5),
however, daily estimates of streamflow were generally less accurate as judged by the
NSE. Factors weakening model performance included inadequate representation of
9
rainfall input, lack of model calibration, and relatively short (2 years or less) calibration
and validation periods (Gassman et al. 2007).
A validation study by Ahl et al. (2008) compared uncalibrated and calibrated
model runs of SWAT in a 2,251 ha snow-dominated, mountainous watershed in
Montana, U.S.A. The study used four years of observation data for calibration and
validation. The uncalibrated model produced acceptable results (NSE >0.5) for annual
water yield only. A manually calibrated model produced acceptable results (NSE >0.5)
for annual, monthly, and daily water yield for the snowmelt-runoff season, but not during
the baseflow season. NSE values of 0.90 and 0.76 were obtained for monthly and daily
water yield after calibration and independent comparison to a validation period. Negative
NSE values were obtained when assessing the baseflow season alone.
White and Chaubey (2005) conducted a validation study of the SWAT model in
the 362 km2
Beaver Reservoir watershed in Arkansas, U.S.A. During the two-year
calibration period of the study, the calibrated model obtained a NSE value of 0.89 for
monthly water yield. During the two-year study validation period, the calibrated model
obtained a NSE of 0.85.
Du et al. (2009) achieved satisfactory model results (daily NSE > 0.5) in a
calibration and validation study of SWAT in the 276 km2 Upper Oyster Creek watershed
in Texas, U.S.A., using a much more limited observed dataset. Their observed data
consisted of discontinuous, instantaneous measurements of streamflow at five main-stem
stream sites collected over a 24-month period to calibrate and validate the model. Du et
al. (2009) pointed out that many watersheds lack long-term continuous records of
streamflow. Comparison of modeled and measured flow data yielded an NSE of 0.66
10
during the calibration period (41 days during year 2004) and an NSE of 0.56 during the
validation period (70 days during years 2002-2003) (Du et al. 2009).
SWAT was applied previously in the state of Missouri, U.S.A. For example,
Stone et al. (2001) used a regional climate model in conjunction with SWAT to estimate
the effect of a doubling of current atmospheric carbon dioxide levels on water yield in the
Missouri River basin, which encompasses 310 eight-digit hydrologic unit code (HUC)
watersheds. The simulation indicated large increases in water yield in eight-digit
watersheds in the north and northwestern portions of the basin, some greater than 70
percent, and considerable overall decreases in water yield by generally less than 20
percent (Stone et al. 2001). However, the overall change in water yield for the entire
basin was estimated to be a decrease of 10 to 20 percent (Stone et al. 2001). Benham et
al. (2006) performed a validation study of SWAT in the 367 km2 Shoal Creek watershed
in Missouri, U.S.A. The calibrated model obtained monthly and daily NSE values of 0.21
and 0.63, respectively, for water yield during the study calibration period, and values of
0.54 and 0.66, respectively, during the validation period.
The SWAT model is not confined to use in the U.SA. The SWAT model was
applied recently in many countries, including not only developed countries, but also
many developing nations including South Korea (Bae et al. 2011), China (Li 2010),
Ethiopia (Setegn et al. 2010), and Iran (Ghaffari et al. 2010). In all cases, satisfactory
modeling results for streamflow were achieved (daily and/or monthly NSE > 0.5).
11
Table 1. Timeline of SWAT development. Adapted from Neitsch et al. (2005).
Release Year;
Model Version Major Changes Made to the SWAT Model
2005;
SWAT2005
This update includes improvements to the model's bacteria transport routines, the
added ability to input weather forecast data, and a new generator for sub-daily
precipitation data. In addition, the daily curve number calculations are altered to
allow the retention parameter to be a function of soil water content or plant
evapotranspiration.
2000:
SWAT2000
Additions to the model include bacteria transport routines, the Green-Ampt
infiltration method, and the Muskingum stream routing method. Improvements are
made to the built-in stochastic weather generator. Values for daily solar radiation,
relative humidity, and wind speed are allowed to be inputted or generated. Potential
evapotranspiration values are allowed to be inputted or generated by the model.
Elevation band processes for weather inputs are improved. In addition, the model is
modified to enable simulation of an unlimited number of reservoirs and the model's
dormancy calculations are altered for appropriate simulation in tropical
environments.
1999;
SWAT99.2
Several improvements are made to the model's nutrient cycling routines and
rice/wetland routines. Modeling of settling processes in reservoirs, ponds, and
wetlands are added. In addition, the update adds routines to model stream bank
water storage, metal routing in streams, and incorporates urban build up/wash off
equations from the SWMM model and USGS regression equations for modeling of
urban pollutant loading.
1998;
SWAT98.1
Model update includes improvements to snow melt routines, in-stream water quality
modeling, and nutrient cycling routines. New management options added include
grazing, manure applications, and tile flow drainage. The model is also modified for
used in the Southern hemisphere.
1996;
SWAT96.2
Update includes new management options for auto-fertilization and auto-irrigation.
C02 is incorporated into the crop growth model for modeling of climate change. The
Penman-Monteith equation for potential evapotranspiration is added. Subsurface
lateral flow of water is incorporated using a kinematic storage model. In addition,
in-stream nutrient equations from the QUAL2E model and in-stream pesticide
routing equations are added.
1994:
SWAT94.2
Ability to incorporate multiple hydrologic response units (HRUs) in each sub-
watershed is added.
12
A NEW FRAMEWORK FOR MODEL DEVELOPMENT AND EVALUATION
H/WQ model end-users face a number of questions and limitations when
developing a model such as SWAT for streamflow prediction. First, model end-users
must determine how to configure the model, including making a decision about how
finely to subdivide a given watershed to achieve accurate results. With a greater number
of sub-basins included in the modeled watershed; greater computational time is required.
When high-resolution datasets (i.e. climate, soil, and land use / land cover) are available,
model users must decide whether the additional computational time and input data
preparation (e.g. pre-processing climate datasets or high-resolution soil data for model
input) provides measurable improvement in the predictive accuracy of the model. For
many watersheds however, the only publically available land use / land cover (LULC)
datasets may be considerably outdated; that is, the LULC datasets do not accurately
represent the land use in the watershed for the period of interest (Engel et al. 2007).
An inadequate number of climate stations that include the necessary climate input
parameters (precipitation, solar radiation, relative humidity, maximum and minimum air
temperature, and wind speed) may often generate additional uncertainty in model
predictions (Engel et al. 2007), as spatial heterogeneity in rainfall has been shown to
introduce considerable error in modeled streamflow estimates when a uniform rainfall
distribution is assumed (Van Werkhoven et al. 2008). For highly physically-based H/WQ
models, collecting data for physical parameters that are often directly measured in
research applications, such as soil effective hydraulic conductivity (Rachman et al. 2008),
may be prohibitive due to high cost and labor requirements (Engel et al. 2007).
13
If the model is to be calibrated and validated, users must decide what method, if
any, is to be used to calibrate the model. Traditional methods for H/WQ model
calibration require time-consuming manual effort, in-depth experience in fine-tuning
numerous model parameters, and expertise in hydroclimatic, soil physical, and
biophysical processes (Van Liew et al. 2005; van Griensven et al. 2002, van Griensven
and Bauwens 2003). Van Liew et al. (2005) reported that approximately four to six weeks
of labor was required for calibrating streamflow for a single watershed used in their
study. It is notable that van Griensven et al. (2002) indicated that manual calibration
efforts are fundamentally flawed because they must consider one set of parameter
changes at a time; thus only part of the available information is being used at one time.
They point out that the manual approach risks an accumulation of errors. Furthermore,
model calibration requires a large amount of monitored streamflow data; however,
monitoring data for many watersheds is often unavailable (Borah and Bera 2004; Engel et
al. 2007). It was recommended that three to five years of monitoring data including wet,
dry, and average years for model calibration be used, however, long-term continuous
time-series of monitoring data are often unavailable (Engel et al. 2007).
A promising alternative to traditional, labor-intensive, and subjective manual
calibration methods has recently emerged (Nash and Sutcliffe 1970; Duan et al. (1994);
Eckhardt and Arnold 2001; van Griensven and Bauwens 2003). Automatic calibration
methods are often viewed as expensive in terms of time and computing resources. Kumar
and Merwade (2009) reported that two weeks of computing time would be required using
an automatic calibration method built into SWAT if they used a single desktop computer
alone. Conversely, Van Liew et al. (2005) reported only one day of runtime.
14
Automatic calibration methods use computer algorithms to determine the optimal
set of input parameters based on a single or multiple set of objective error functions. Nash
and Sutcliffe (1970) recommended automatic calibration to remove subjectivity in the
fitting of a H/WQ model to observed data. An automatic multi-objective and multi-site
capable calibration method is included in SWAT and is accessible via the ArcSWAT
graphical user interface. The SWAT automatic calibration tool is not capable of reducing
error on multiple time-scales simultaneously (e.g. daily and monthly). Nor is it capable of
simultaneous calibration of baseflow and surface flow as was shown by Zhang et al.
(2011). Recent scientific literature evaluating the SWAT calibration tool has not
investigated the SWAT multi-objective model optimization capabilities (Van Liew et al.
2005; Van Liew et al. 2007; Kumar and Merwade 2009; Setegn et al. 2010). At the time
of this writing, only one researcher in the scientific literature has evaluated SWAT‟s
multi-site automatic calibration capabilities (Zhang et al. 2008). Only the original
developers of SWAT‟s automatic calibration method have taken advantage of its multiple
streamflow objective functions (Van Griensven and Bauwens 2003). It is on this basis
that, in the following study, a single objective function and a multiple objective function
automatic calibration method for streamflow were compared.
Finally, when evaluating the model (calibration and validation), model users face
a confusing array of statistical and graphical options to consider for use as indicators of
model fit. Several publications offer a variety of alternatives (Moriasi et al. 2007; Legates
and McCabe 1999; Willmott 1981; Willmott et al. 1984, 1985; Nash and Sutcliffe 1970),
however, no clear consensus has emerged regarding the optimal technique(s). The
15
problem of properly assessing the predictions obtained from H/WQ models was
eloquently described by Nash and Sutcliffe (1970), with emphasis added:
“The results obtained [from conceptual hydrologic models] are not
always presented in a manner which makes possible a judgment of the
relative efficiency of these models, nor does there appear to be any
general agreement on the method of developing and testing a model for a
given catchment or group of catchments. It is intended to set out in this
paper, tentatively, as a basis for discussion and amendment, a systematic
approach towards developing, testing and modifying a model for a set of
catchments with the development of a forecasting technique for an
ungauged member of the set as a long term objective. These preliminary
ideas will be modified by experience …. It is hoped to encourage a
discussion of the general principles by which the conceptual model
technique may be put to best use in this difficult but intriguing problem.”
The Nash Sutcliffe Efficiency (NSE), the statistical goodness-of-fit indicator first
introduced by Nash and Sutcliffe in 1970, remains today, 41 years later; the most widely
used measure of H/WQ model fit and often is the sole measure considered (e.g.
Bouegeard 2011). The efficiency introduced by Nash and Sutcliffe (1970) corrects for the
unsuitability of the classic coefficient of determination, which represents the square of
the Pearson‟s product-moment correlation coefficient. Curiously, Nash and Sutcliffe used
the mathematical notation R2 when defining their statistical measure, a notation now
predominantly used to represent the classic coefficient of determination (Mehta et al.
2011; O‟Donnell et al. 2008). As Legates and McCabe (1999) clarified and further
explicated, the coefficient of determination (R2 or r
2) is poorly suited to measuring
goodness-of-fit due to its basis in linear correlation. The coefficient of determination may
give high values when model fit is low due to its misrepresentation of additive and
proportional errors. Only a 1:1 line on a scatter plot between observed and modeled
16
values should be viewed as a perfect fit. However, the coefficient of determination may
give high values for models which exhibit a slope not equal to one (proportional error) or
a y-intercept not equal to zero (additive error) (Legates and McCabe 1999).
Unlike the coefficient of determination, however, the NSE lacks a well-
understood error probability distribution (Legates and McCabe 1999). Thus, unlike the
coefficient of determination, statistical significance testing for the NSE requires complex,
time-consuming bootstrapping methods (Efron and Gong 1983) to establish a probability
distribution (Legates and McCabe 1999; Willmott et al. 1985) for use in statistical
significance testing.
Willmott (1984; 1985) and Legates and McCabe (1999) have tried somewhat
unsuccessfully to encourage scientific adoption of absolute value error functions over the
more common squared error functions, like the Nash-Sutcliffe Efficiency. They have
introduced several goodness-of-fit statistical indicators which were not widely reported.
Notably, the scientific and practical value of absolute value based error measures remains
to be fully understood.
Van Griensven (2002) eloquently summarized the previously stated practical
approach to H/WQ modeling. In Chapter XII: Conclusions and perspectives, van
Griensven (2002) states:
“A large amount of money and efforts are put in the development and
application of water quality models all over the world and this is even still
increasing. All these developments aim at the development of a tool that is
useful for decision making in water basin management. As there exist very
few examples of such applications, models seem rather to be a game for
engineers or scientists than useful tools that can help to improve the river
water quality. However, it is shown that river basin management based on
simple emission based standards for point sources failed and that, more
and more, diffuse pollution is responsible for bad water quality. As these
17
problems are characterised by a high temporal and spatial variability,
understanding and solving them requires dynamic models to point out the
important causes of pollution or to predict the effects of pollution
abatement programmes”.
As van Griensven (2002) affirmed, the intention of the study presented here was to
evaluate the SWAT model as a management tool. Furthermore, in the same chapter, van
Griensven eloquently describes the ideal H/WQ model:
“The ideal tool for river basin water quality management incorporates all
relevant process descriptions, to enable the simulation of the output
variables needed by decision makers of any kind, using readily available
information of the basin (such as GIS layers and climate data) without
requiring calibration. Unfortunately, tests of models without calibrations
are only published when there is a reasonable fit to the observations. It is
a very idealistic thought that models will be able to be applied on
ungauged basins.”
STUDY OBJECTIVES
Given the constraints on model development that have been described, for this
study, rather than develop a single model for a watershed, multiple configurations of the
SWAT model were developed and then assessed according to several different goodness-
of-fit statistical indicators. The overarching goal of the study was therefore to provide
answers to SWAT users who face the aforementioned questions and practical limitations.
These limitations include a lack of accurate data for model configuration and calibration,
high time and labor requirements for model calibration, and a lack of consensus on
appropriate model evaluation methods. However, the study addresses specific questions
which have wider implications for all H/WQ modelers.
18
To determine the effect of watershed discretization resolution, soil data resolution,
and the quantity and quality of climate station input, a set of twenty different model
configurations, were developed. The 20 model configurations represent all the possible
combinations when using (A) two different watershed discretization schemes, one with a
high number of sub-basins and hydrologic response units (HRUs) and one with a small
number, (B) two soil datasets, one with low spatial resolution, the State Soil Geographic
Database (STATSGO) and one with high resolution, the Soil Survey Geographic
Database (SSURGO), and (C) five different combinations of automatically generated,
single station, and multiple climate station data.
The 20 uncalibrated configurations were then ranked according to five different
goodness-of-fit statistical indicators, at both a daily and monthly scale. In addition, in
order to test the built-in automatic calibration method that is included in SWAT, one of
the 20 configurations was selected for parameter optimization (calibration). An attempt
was made to run the automatic calibration in a reasonable and practical amount of time
(less than 24 hours). When running the automatic calibration, the control parameters were
set to optimize the model‟s flow parameters based on two different sets of objective
functions: (1) the sum of the squared error (SSQ) for daily streamflow, and (2) both the
sum of the squared error (SSQ) and the sum of the squared error after ranking (SSQR) for
daily flow.
The basic questions addressed in this study may be summarized as follows:
(1) What is the effect of input data resolution (watershed discretization, soil, and
climate) on modeled streamflow predictions?
19
a. Null Hypothesis: If input data resolution is increased, the accuracy of
modeled streamflow predictions will not change.
b. Alternative Hypothesis: If input data resolution is increased, the
accuracy of modeled streamflow predictions will change.
(2) How well does the model predict streamflow without calibration of the input
parameters on both a daily and monthly scale?
a. Null Hypothesis: To meet published standards for model accuracy
(Moriasi et al. 2007), the SWAT input parameters do not need to be
calibrated.
b. Alternative Hypothesis: To meet published standards for model
accuracy (Moriasi et al. 2007), the SWAT input parameters must be
calibrated.
(3) Which goodness-of-fit statistical indicators should be used for evaluating the
model during the calibration and validation process? How are they different?
a. Null Hypothesis: The choice of goodness-of-fit statistical indicators
does not affect the selection of the most accurate model.
b. Alternative Hypothesis: The choice of goodness-of-fit statistical
indicators does affect the selection of the most accurate model.
(4) Can the automatic calibration method built-in to SWAT be run successfully on
a single desktop computer in a reasonable amount of time (less than 24
hours)?
a. Null Hypothesis: The built-in automatic calibration method in SWAT
cannot be run successfully in less than 24 hours.
20
b. Alternative Hypothesis: The built-in automatic calibration method in
SWAT can be run successfully in less than 24 hours.
(5) How do automatically calibrated model outputs compare when optimizing for
a single objective function versus a set of two objective functions?
a. Null Hypothesis: The accuracy of the modeled streamflow will be the
same when optimized for a single objective functions as when it is
optimized for a set of two different objective functions.
b. Alternative Hypothesis: The accuracy of the modeled streamflow will
be different when optimized for a single objective function as when it is
optimized for a set of two different objective functions.
21
CHAPTER II
METHODS
STUDY SITE DESCRIPTION
The SWAT model was tested using data collected in an urbanizing watershed, the
Hinkson Creek Watershed (HCW) in Boone County, central Missouri, U.S.A. The HCW
features high land cover spatial heterogeneity, rapid population growth, increasing spatial
change in land use and land cover (i.e. urbanization), and ongoing community, political,
and legal debate over watershed management. A recently developed Total Maximum
Daily Load (TMDL) for the watershed used urban stormwater runoff as a surrogate for
unidentified pollutants suspected to be impairing the biological diversity in the stream;
the TMDL process continues to generate political controversy and questions about
appropriate implementation (USEPA 2011).
The Hinkson Creek Watershed (HCW) is located in the Lower Missouri-Moreau
River Basin (LMMRB, HUC 10300102) in central Missouri, U.S.A. Comprising
approximately 231 km2, Hinkson Creek originates northeast of Hallsville, Boone County,
Missouri, and flows approximately 42 kilometers in a southwesterly direction to its
mouth at Perche Creek. Hinkson Creek is classified as a Missouri Ozark border stream
located in the transitional zone between Glaciated Plains and Ozark Natural Divisions
(Thom and Wilson 1980). Streams generally originate on level uplands underlain by
shale and descend into hilly terrain underlain by limestone (Pfleiger 1989).
22
Climate
The transitional climate of Missouri includes influences from winter dominant
continental polar air masses, and summer prevalent maritime and continental tropical air
masses. This translates to broad fluctuations in temperature (12.8 ºC yearly average) and
precipitation (1016 mm/year). The heaviest rainfall typically arrives in late spring and
early summer with 70% of the total precipitation falling in the period from April through
August. The driest period is from November through March. Annual snowfall is
approximately 508 mm (Nigh and Shroeder 2002).
Land Use and Land Cover
The HCW encompasses the city of Columbia and the surrounding urban-rural
interface (Hubbart et al. 2010; Hubbart and Freeman 2010; Hubbart and Gebo 2010),
allowing a distinct opportunity to study the dynamics of land use diversity and change.
The city of Columbia has experienced rapid population growth of 28.4% between 2000
and 2010 (current population, 108,500, U.S. Census 2010). The HCW is fully contained
within Boone County, Missouri, which has had population growth of 20.4% during the
same period (U.S. Census 2010).
Land use in the upper portion of the watershed consists of rural pastureland and
wooded areas, whereas the lower portion of the watershed is within the urbanized section
of the city of Columbia, Missouri (MDNR 2006). Sub-basin areas and land-use cover
classes were determined in ESRI© ArcGIS 9.3 software using 10 m resolution digital
elevation model (DEM) data from the National Elevation Dataset (NED) and 30 m land-
use cover data from the 2001 National Land Cover Dataset (NLCD). Total sub-basin
23
areas in hectares, land-use area in percent, and cumulative contributing and land-use area
for each of the five gauge sites located in the HCW are presented in Table 2 and Figure 2.
Site #1 has a sub-basin area of 7742.3 ha and the dominant land-use class is
pasture/hay (44.9%). Site #2 has a contributing area of 2358.5 ha and the dominant land-
use class is also pasture/hay (36.7%). Site #3 has a contributing area of 1327.4 ha and the
dominant land-use class is deciduous forest (29.6%). Site #4 has a contributing area of
6526.8 ha and the dominant land-use class is pasture/hay (28.0%). Site #5 has a
contributing area of 2630.1 ha and the dominant land-use class is developed, low
intensity (31.2%). The cumulative contributing area for all five gauge sites is 20,585.1 ha
and the dominant land-use class is deciduous forest (32.4%).
24
Table 2: Total sub-basin area (ha), land-use area (%) for each of five gauge sites and cumulative contributing and land-use area (assuming 15 land-use divisions) in the Hinkson Creek Watershed, Missouri, U.S.A. Contributing areas and land-use cover classes were determined using 10 m DEM data and 30 m land-use cover data.
Contributing Areas Open Water Developed,
Open Space
Developed,
Low
Intensity
Developed,
Med.
Intensity
Developed,
High Intensity
Site Area
(ha) Area (%) Area (%) Area (%) Area (%) Area (%)
1 7742.3 0.5 4.2 0.5 0.0 0.0
2 2358.5 0.6 5.0 1.8 1.2 0.4
3 1327.4 0.4 14.4 19.0 13.8 4.8
4 6526.8 0.7 8.1 9.6 5.5 1.6
5 2630.1 1.0 18.7 31.2 12.2 4.8
Cumulative 20585.1 0.6 8.1 8.7 4.3 1.5
Contributing Areas Barren Land Deciduous
Forest
Evergreen
Forest
Mixed
Forest Shrub/Scrub
1 7742.3 0.0 34.3 0.6 1.2 0.5
2 2358.5 1.4 34.7 0.8 2.2 1.1
3 1327.4 0.1 29.6 0.4 1.9 0.6
4 6526.8 0.4 33.9 1.0 1.6 0.3
5 2630.1 0.1 22.4 0.1 0.6 0.2
Cumulative 20585.1 0.3 32.4 0.7 1.4 0.5
Contributing Areas Grassland/
Herbaceous
Pasture/
Hay
Cultivated
Crops
Woody
Wetlands
Emergent
Herbaceous
Wetlands
1 7742.3 1.1 44.9 10.3 1.9 0.1
2 2358.5 0.7 36.7 12.2 1.3 0.0
3 1327.4 0.1 10.2 4.3 0.4 0.0
4 6526.8 0.6 28.0 7.6 0.9 0.0
5 2630.1 0.4 5.3 1.0 2.1 0.0
Cumulative 20585.1 0.8 31.3 8.1 1.4 0.0
25
Figure 2. Map of gauging stations and associated sub-basins with land use / land cover classifications in the Hinkson Creek Watershed, Missouri, U.S.A.
26
Soils and Vegetation
The Lower Missouri-Moreau River Basin (LMMRB) which contains the HCW is
largely comprised of prairie-forest transitional soils. Soils are poor to well-drained but are
easily erodible in part due to steep slopes (Perkins 1995). A map showing the distribution
of hydrologic soil groups using a high resolution soil dataset (SSURGO) is shown in
Figure 3. The HCW is dominated by the high runoff and moderately high runoff
hydrologic soil groups C and D. The soil type within the upper segments of Hinkson
Creek is characterized as loamy till with a well-developed claypan (Chapman et al.
2002). The soil types within the lower segments of HCW are characterized as thin cherty
clay and silty to sandy clay. Vegetation is loosely characterized as a mixed deciduous oak
forest (Rickett 1931). Riparian zones contain a diverse array of willows (Salix spp.),
birches (Betula spp.), cottonwoods (Populus deltoides), and sycamores (Platanus
occidentalis). Alluvial fans are covered with elms (Ulmus spp.), soft maples (Acer spp.),
basswoods (Tilia spp.), and woody shrubs (MDNR 2006).
27
Figure 3. Hydrologic soil groups by soil type in the Hinkson Creek Watershed, Missouri, U.S.A. Soil data from the Soil Survey Geographic Database (SSURGO), 2011.
28
Topography
Elevation ranges from 170 meters at the confluence with Perche Creek to 287
meters above sea level in the headwaters, as indicated by analysis of digital elevation
model data from the National Elevation Dataset. Hypsometric curves (Figure 4) showing
the areal distribution of elevation were developed for the entire HCW including each of
the sub-basin areas corresponding to the five gauging stations located in the watershed
and a sixth subbasin located at the outlet of the watershed.
Figure 4. Hypsometric curves showing distribution of elevation for each sub-basin associated with five nested gauging stations and the outlet of the Hinkson Creek Watershed, Missouri, U.S.A.
29
Water Quantity and Quality
A U.S. Geological Survey gauging station (#06910230) is located on Hinkson
Creek 122 m downstream of Providence Road in the city of Columbia, Missouri. The
gauging station has a total contributing drainage area of approximately 179.5 km2.
Average discharge measured at the intermittently-operated gauging station from October
1966 to September 1981, October 1986 to September 1991, and April 2007 to December
2008 was approximately 1.42 m3/s. Average annual discharge ranged from a low of 0.38
m3/s in 1980 to a high of 3.14 m
3/s in 1973. Average monthly discharge measured from
1967 to 1991 ranged from a low of 0.48 m3/s in August to a high of 2.66 m
3/s in May.
Since 2001, the Missouri Department of Natural Resources (MDNR) conducted
water quality and aquatic biota monitoring on main-stem Hinkson Creek and related
storm drainages. MDNR results documented that the aquatic community was impaired.
Toxicity tests showed that approximately 20% of stormwater discharges were polluted.
Pollution source procedures implicated a wide variety of urban-associated chemical
constituents. Visual sediment surveys documented increased sediment in the impaired
segment of Hinkson Creek. Analyses in 2005-2006 were contrary to early work and did
not indicate toxicity or measure organic chemical constituents above laboratory detection
levels. Thus, results to date have been confounding and attributable to any number of
environmental origins including irregular and poorly defined stormwater inputs,
snowmelt, in-stream processes, and other natural and/or anthropogenic factors (MDNR
2006; USEPA 2011).
30
DATA COLLECTION
Nested Watershed Design
Hydroclimatic data were collected at five fully-equipped co-located streamflow
gauging and climate monitoring stations located at five sites along the main-stem of
Hinkson Creek. Hydroclimate stations were installed in the winter of 2008-2009. Site #1
is located at the bridge at Rogers Road (39° 01.418‟ N, 92° 14.761‟ W) in Columbia,
Missouri; Site #2 at the bridge at Mexico Gravel Road, 6.8 km downstream of Site #1
(39° 58.964‟ N, 92° 16.758‟ W); Site #3 at the bridge at Broadway, 5.5 km downstream
of Site #2 (38° 56.891‟ N, 92° 18.321‟ W); Site #4 at the current U.S. Geological Survey
(USGS) gauge location at the bridge at Old Route K Road, 7.7 km downstream of Site #3
(38° 55.670‟ N, 92° 20.391‟ W), and Site #5 at the Scott Boulevard / Highway TT bridge,
9.6 km downstream of Site #4 (38° 54.847‟ N, 92° 24.011‟ W).
In addition to the hydroclimate stations, two additional publically available
climate stations from the University of Missouri Agriculture Experiment Station‟s
Commercial Agriculture Automated Weather Station Network were used in
implementing the SWAT model. One of these climate stations, Sanborn Field
(38.942471° N, 92.320468° W) is located 2.3 km northeast of Site #4 and 1.4 km
southeast of Site #3. The other additional climate station, South Farms (38.904675° N,
92.273542° W), is located 5.5 km southeast of Site #3, 0.5 km beyond the boundary of
the HCW. Figure 2 shows the locations of each gauging station in the Hinkson Creek
Watershed; also pictured are the sub-basin areas corresponding to each site, and the land
use / land cover (LULC) composition of each sub-basin.
31
Climate Data
A complete climate station, installed at each site, measured rainfall, minimum and
maximum daily air temperature, relative humidity, solar radiation, and wind speed. A
description of all measurement instruments installed at each site is presented in Table 3.
Five-minute data were logged on-site. Post-processing of the data involved averaging or
summing (i.e. reducing) five-minute time-series data to daily values. To correct for
occasional data gaps, linear regression models (R2 ≥ 0.92) were developed to correlate
the climate data at the Sanborn Field climate station (located within HCW sub-basin #4)
with the all other HCW climate stations. The regression model was then applied to the
data gap periods to simulate specific climate values at the other HCW sites.
32
Table 3. Field instrumentation and variables measured at five hydroclimate stations in the Hinkson Creek Watershed, Missouri, U.S.A.
Instrument Measurement
Accubar® Constant Flow Bubble Gauge/Recorder 56-0133 Water stage in mm
Campbell Scientific, Inc. Met One 034B Windset anemometer Horizontal wind speed in m/s and
direction
Campbell Scientific, Inc. Model 107 Temperature Probe Soil temperature in °C
Campbell Scientific, Inc. CS616 Water Content Reflectometer Soil volumetric water content in %
Campbell Scientific, Inc. SR50A Snow Depth Sensor Snow depth in cm
Campbell Scientific, Inc. Model HMP45C Temperature and
Relative Humidity Probe with radiation shield
Air temperature in °C and relative
humidity in %
Campbell Scientific, Inc. LI200X Pyranometer Incoming solar radiation (400 to
1100 nm) in W/m2
Campbell Scientific, Inc. TE525 Tipping Bucket Rain Gauge Precipitation in mm
Stage Data
At each streamflow gauging station (Sites #1-3, and #5) water stage in mm was
monitored at 5 minute intervals using an Accubar® Constant Flow Bubble Gauge and
Recorder. The USGS also measures stage at Site #4 using an Accubar® bubble gauge,
but at 15 minute intervals. Continuous data records of stage were logged using Campbell
Scientific, Inc. CR1000 data loggers remotely powered with 12-volt batteries charged by
solar panels.
Streamflow Measurements
Streamflow measurements (i.e. stream cross-sections) were conducted at all four
sites not operated by the USGS (#1-3, and #5) on a biweekly basis for a full one-year
33
period (2009). In addition, streamflow measurements were taken during storm events to
measure streamflow at medium and high (peak) stages. In 2010, the focus of streamflow
measurements shifted exclusively to measuring medium and high flow events as well as
measuring the discharge during periods of acute variable-backwater-affected streamflow
at Site #5.
Variable backwater, a phenomenon observed at Site #5, occurs when the energy
slope in a stream reach is variable for a given stage, resulting in unsteady flow (USGS
1982) (Figure 5). In most cases, variable backwater is caused by variable stage at a
downstream confluence for a given discharge at the stream gauge site or may be caused
by the operation of gates at a downstream dam (USGS 1982).
When variable backwater is present at a stream gauging site, the discharge is not
simply a function of stage; it is also a function of the energy slope in the reach (USGS
1982). Variable energy slopes in a stream channel may not only be a function of a
downstream variable backwater source, but also a function of changing discharge in a
stream reach (USGS 1982). Variable slopes caused by changing discharge occur when
the slope of the stream is nearly flat and the change in discharge is very fast (USGS
1982).
The variable backwater phenomena observed at Site #5 was primarily observed
during periods of historical record spring flooding on the Missouri River, indicating that
the Missouri River, downstream of Hinkson Creek, may serve as a source of the variable
energy slope at Site #5; variable backwater was also observed as a lagging response in
the recession limb of the hydrograph following large storm events in the HCW, which
indicates the possible presence of variable slope caused by changing discharge.
34
Figure 5. Photograph dated May 16, 2009 showing variable backwater affected streamflow at the Site #5 gauging station, Hinkson Creek Watershed, Missouri, U.S.A.
Streamflow was measured using the hydrologic standard velocity-area method
(USGS 1982). During low flows, water velocity was measured by wading the stream
while using a velocity sensor mounted to a top-setting wading rod (Figure 6). Velocity
measurements were made at the mid-points of 25 equal-width sampling intervals along
the wetted cross-section that is perpendicular to the downstream direction of flow. For
each interval, velocity was measured at the 0.6, 0.2, and 0.8 depths of the water column.
35
During high flows, streamflow was measured from atop the co-located
transportation bridges using a USGS Type A sounding reel. The sounding reel was
mounted on a USGS standard bridge board with either a 15 or 30 pound USGS
Columbus-Type sounding weight attached. The sounding weight was attached to the
hangar at the end of the sounding line, a USGS standard 0.25 cm thick stainless steel
Ellsworth cable. The Marsh-McBirney velocity sensor was mounted to the hangar bar
that is attached to the sounding weight. Velocity measurements were collected at the
downstream side of the bridge.
To reduce the time taken to measure the full stream channel cross-section during
high flows, where stage rapidly changes and therefore contributes considerable error to
streamflow determination (USGS 1982), the number of sampling intervals was reduced to
15 and the velocity measurements were only taken at the 0.6 depth, as recommended in
published USGS guidelines (USGS 1982).
36
Figure 6. Photograph showing wading streamflow measurement technique at Site #2, Hinkson Creek Watershed, Missouri, U.S.A.
To measure flow velocity, a Marsh-McBirney® Flo-Mate™ Model 2000 was
used (Marsh-McBirney, Inc. 1990). The Flo-Mate sensor was mounted to either the top-
setting wading rod or the hangar of the sounding weight. The Marsh-McBirney Flo-Mate
is a non-mechanical electromagnetic flow sensor that measures flow velocity (accuracy,
±2% of reading +0.015 m/s) by measuring changes in the voltage amplitude that is
created by a magnetic field (Marsh-McBirney, Inc. 1990). The Flo-Mate was set to
calculate a 30-second-averaged stream velocity in m3/s. Steps were taken to avoid high
tension on the fragile wiring within the sensor cable, as well as entanglement of the Flo-
37
Mate sensor cable with debris; when the cable was suspended from transportation
bridges, the sensor cable was affixed to the hangar bar and the Ellsworth cable (which
suspends the sounding weight from the sounding reel) with plastic tie-wraps at multiple
points. This method was recommended by Hach Flow® staff (Darby, pers. comm., 2010)
after and only after consultation. To measure discharge from bridges with pedestrian
safety fences greater in height than is usable with the bridge board alone (Sites #2 and
#3), a moveable hydraulic lift was used to raise the sounding reel operator and the bridge
board to a sufficient height.
The streamflow through each of the 25 and/or15 individual intervals was
estimated by the three-point method (USGS 1982) in which the average velocity at the
mean of the 0.2 and 0.8 depths is averaged with the velocity at 0.6 depth. For individual
intervals in which velocity was not recorded at all three depths due to extremely shallow
water depths as well as during bridge measurements, the six-tenths-depth method (USGS
1982) was used. For this method, only the velocity at the 0.6 depth is multiplied by the
cross-sectional area of the interval to estimate streamflow at each interval (USGS 1982).
In rare cases where only the 0.2 depth was possible to record due to shallow water depths,
the streamflow was measured using the two-tenths-depth method (USGS 1982). The 0.2
depth velocity was multiplied by a coefficient, 0.87, taken from an average vertical-
velocity index that was empirically calculated for numerous streams by the USGS (USGS
1982).
At Sites #2 and #3, high flows were measured at transportation bridges (co-
located with the hydroclimate stations) that travel at an oblique angle to the downstream
direction of streamflow (Figure 7). To account for resulting streamflow measurement
38
error, the initial cross-sectional width that was measured along the bridge road surface
was multiplied by a coefficient (0.82 for Site #2 and 0.87 for Site #3). The coefficient
was determined by calculating the tangent of the planform angle between the bridge and
the cross-section that lies perpendicular to the downstream direction of flow. The
planform angles were measured by determining the compass bearing for both the bridge
and the cross-section and then calculating the difference.
Figure 7. Photograph showing oblique angle between Hinkson Creek and the transportation bridge at Site #2, Hinkson Creek Watershed, Missouri, U.S.A.
39
RATING CURVE DEVELOPMENT
Initial Rating Curve Development
There are two general approaches in the literature for rating curve development.
The first approach involves the simple fitting of a smooth monotonic curve that is
typically logarithmic over most of its range, as is predicted by hydraulic theory (e.g.
Chezy formula). The fitted curve may also be a 2nd
or 3rd
order polynomial relationship;
polynomial curves may more closely fit measured stage-discharge values compared to the
logarithmic curves (Sivapragasam and Muttil 2005). The logarithmic approach has the
advantage of maintaining a monotonic trend through its entire range, while the
polynomial approach has the advantage of more closely fitting the measured data, and
vice-versa. Curves are fitted to a set of paired measurements of streamflow and stage
(water level) (USGS 1982), where stage is the level (i.e. elevation) of the water surface
above an arbitrary yet stationary reference point (USGS 1982).
The second approach, advocated by the USGS, creates rating curves that are
adjusted using judgments based on hydraulic principles and/or are adjusted based on
changing stream channel conditions. Changing stream channel conditions may occur
cyclically (seasonal variation such as in-channel vegetation growth in the summer or
freezing of stream water in the winter). Changing stream conditions may also occur as
long-term permanent changes to the geometry of the stream channel (USGS 1982).
The USGS also takes into consideration the presence, absence, or submergence of
section controls; section controls are defined as stable and permanent natural or artificial
structural features in a stream channel. The section controls are capable of maintaining a
40
consistent stage-discharge measurement for the stream gauge over a limited range of flow
conditions (USGS 1982). Observations of section controls are used to adjust the rating
curve (USGS 1982). Additionally, the USGS advocates use of an indirect method of peak
discharge calculation, e.g. the slope-conveyance method, and a field determination of the
gauge height of zero flow (USGS 1982). While these methods serve only as imprecise
estimates, they may be a better alternative to extrapolation of a curved function based on
direct discharge measurements only.
In this study, the first method discussed was followed. Thus, a 3rd
order
polynomial function was the primary tool for development of the rating curves
(Sivapragasam and Muttil 2005). It is surmised here that the first method allows for a
faster and more objective determination of the stage-discharge relationship. As stated
previously, the 3rd
order polynomial allows for a closer fit with the observed stage-
discharge values (Sivapragasam and Muttil 2005). To correct for poor simulation of very
low flows by the 3rd
order polynomial, linear functions were applied to the very low
ranges of stage to achieve a precise alignment with the stage estimated to correspond to
zero flow (USGS 1982). In cases where the polynomial, due to its S-shape (non-
monotonic trend) began to differ from the expected monotonic hydraulic trend in the
upper portion of the stage range, a logarithmic curve was used to extend the rating curve
to the upper reach of the stage range.
Trial Rating Curve Testing and Selection of Final Rating Curves
Trial rating curves for each University-operated site were established first using
the previously described method, and then analyzed for accuracy by comparing observed
41
between-site water yield differences. Based on those results, the rating curve definition
method was adjusted as necessary before selection of the final rating curve.
Initial trials of rating curve fits were tested by calculating the flow for a four
month time-series (March – June 2010) and comparing total water yield between all four
University-operated sites (HCW Sites #1-3, and 5) and the USGS-calculated water yield
at Site #4. For Sites #3 and #5, initial trial rating curves produced unrealistic water yields
relative to other sites. To determine whether the initial trial rating curves were realistic or
not, between-site water yields (over the four-month period) were compared to between-
site discharge differences measured on same days. In these cases, the curve fitting
method was adjusted to more realistically match expected between-site water yield
differences. This analysis resulted in adjustments to the rating curves at Site #3 and Site
#5. For Site #3, the logarithmic portion of the rating curve was re-calculated through
least-squares regression to adjust stage-discharge measurement in the upper portion of the
rating. Specific rating curve functions for each site are shown in Appendix A.
Computation of the Continuous Streamflow Record
The observed streamflow regime at Site #5 exhibited a variable backwater
phenomenon (USGS 1982); this observation was evidenced by the location of numerous
(26 out of 56 total) streamflow measurements that plotted to the right of the general
monotonic trend of the data in the stage-discharge plot (USGS 1982) (Appendix A). The
USGS has published methods to develop direct measures of streamflow estimation at
gauging sites that are affected by variable backwater (USGS 1982). The two methods
described by the USGS are the slope-index method and the velocity-index method. The
42
slope-index method uses continuous measurement of the water surface slope (an estimate
of the energy slope) in conjunction with the stage measurement to determine the
discharge record (USGS 1982). The velocity-index method uses a continuous
measurement of stream velocity at a point or points in the stream channel in conjunction
with the stage measurement to determine the discharge (USGS 1982). An unsuccessful
attempt was made to develop a slope-rating function for Site #5 using the slope-index
method. The USGS gauge at Site #4 served as an auxiliary gauge for determination of
water surface slope.
To account for periods in which the variable backwater effect resulted in
streamflow at high stages but with much lower than expected discharge, a linear
regression model (R2
= 0.95) was developed to correlate the stage at Site #4 with the
stage at Site #5. This linear regression model was then used to simulate stage at Site #5
during periods of extreme difference in the trend time-series, thus resulting in a
hydrograph with corrected backwater-affected streamflow.
To then prepare the observed data for analysis of SWAT-modeled output, stage
data were averaged on both a daily and monthly scale for each gauge site. The final rating
curves were then applied to the stage data to calculate a continuous record of streamflow
during year 2009-2010. Post-processed (i.e. approved) daily mean streamflow for Site #4
was obtained from the USGS.
43
MODEL IMPLEMENTATION
SWAT Model Configuration
In this study, a general framework for SWAT model development and evaluation,
designed with the agency or practitioner in watershed management in mind, is described,
tested, and evaluated. Due to practical limits on time, computer resources, access to
accurate datasets, as well as specialized experience and knowledge in modeling
techniques and hydrologic processes, SWAT model end-users (agencies and individuals)
may not have the ability to incorporate sophisticated research-grade methods to optimize
and evaluate their models.
In consideration of these constraints, several rules guided model development in
this study. Model input datasets for land use / land cover, soils, climate, and topography
were selected from publically and widely available datasets that do not require
considerable pre-processing in a geographic information system (GIS). Default model
input parameters were used when configuring the model (Neitsch et al. 2004). An
automatic calibration method, already built-in to the ArcSWAT version 2.3.4 software
package, was used to optimize the parameters in the model that govern the modeled
output of estimated streamflow. Since automatic calibration methods are typically time-
intensive (Kumar and Merwade (2009) reported two weeks of runtime on a single
computer), the parameters that control the automatic calibration routine were set such that
the time required for successful completion would take less than 24 hours. The automatic
calibration routine was run using a single desktop-class computer equipped with two
44
Intel® Core™ 2 Duo CPUs (E8600) running at 3.33 GHz and with 3.21 GB of RAM
(random access memory).
The 1:250,000 scale State Soil Geographic Database (STATSGO) was selected to
supply a soil dataset to parameterize the SWAT model (Figure 8). STATSGO is a
national-scale geo-referenced soil dataset that is included in the ArcSWAT software
package (Neitsch et al. 2004), and it was assumed to be the soil data reference of choice
for end-users. In addition, a higher resolution national-scale geo-referenced soil dataset,
the 1:24,000 scale Soil Survey Geographic Database (SSURGO), was used in other
model runs because it can be freely downloaded via the internet from U.S government
sources (Figure 8). However, the SSURGO dataset requires pre-processing (appending of
SSURGO soil data to the included national-scale STATSGO soil property database) and
the creation of a lookup text file to associate SSURGO soils with the soil property
records added to the SWAT soil database (Di Luzio et al. 2002). Fortunately, an easy-to-
use, free software program named SWATioTools™ is available for automatic pre-
processing of SSURGO soil data. The software program was developed and is shared via
website by Dr. Alexei Sheshukov of Kanas State University.
The 2001 NLCD used in this study was the most current national-scale publically
available LULC dataset at the time of this study. The 2006 National Land Cover Dataset
was released February 16, 2011. While the state of Missouri Spatial Data Information
Service (MSDIS) has produced a LULC dataset dated 2005, the more widely applicable
NLCD dataset was chosen over the more up-to-date dataset from MSDIS because use of
MSDIS data requires the model end-user to pre-process the data before input into the
model. Pre-processing requires creation of a text lookup file to associate MSDIS land use
45
classifications with the land use classification codes in the SWAT database (Di Luzio et
al. 2002). To automate the process for the end-user, SWAT includes a pre-defined text
lookup file for the 2001 NLCD. The 2001 NLCD text lookup file associates each NLCD
land use / land cover class with a default management file from either the SWAT land
cover / plant growth database (e.g. Forest-Deciduous (FRSD) or Hay (HAY)) or the
SWAT urban database (e.g. Residential – Medium Density (URMD)). Default
management settings were not adjusted. In this study, the most widely-applicable LULC
dataset that involved the most efficient method for the end-user was applied.
To define the number of sub-basins in the SWAT model of the HCW, two
schemes were used: 1) accepting the default value determined in ArcSWAT for the
minimum contributing source area used to define each sub-basin outlet, resulting in 34
sub-basins (based on a 10 m resolution DEM from the National Elevation Dataset); and
2) reducing the number of sub-basins using the built-in tools in ArcSWAT to six, one for
each of the five gauging stations and one below Site #5 at the outlet of Hinkson Creek
(Figure 8). Reducing the number of sub-basins in the SWAT model configuration
increases the computational efficiency of the model by generally reducing the quantity of
hydrologic response units (HRUs) because most SWAT computations occur at the HRU
level (Neitsch et al. 2005).
When defining the number of HRUs, the percent threshold values prescribed in
the SWAT manual (Neitsch et al. 2005) were used to limit the number of HRUs created:
a minimum of 20% of the sub-basin area for each land use class and a minimum of 10%
sub-basin area for each soil type (Neitsch et al. 2005). HRUs were not subdivided by
slope. The default setting in ArcSWAT uses a single slope value for all HRUs in each
46
sub-basin. In accordance with the modeling approach used in this study, the default
and/or recommended settings in the model were chosen, and computational efficiency
was optimized.
47
Figure 8. Comparison of (1) low resolution six sub-basin and high resolution 34 sub-basin watershed discretization schemes and (2) low resolution STATSGO soil data and high resolution SSURGO soil used in SWAT modeling.
48
Five climate input data approaches were selected to use in model configurations.
For the first set (Weather Generator), SWAT‟s built-in weather generator algorithm,
WXGEN (Neitsch et al. 2005), was used to simulate all the climate parameters (daily
values for precipitation, solar radiation, wind speed, relative humidity, minimum and
maximum air temperature). WXGEN by default used multi-year averaged climate data
from nearby Moberly, Missouri, to the north of the watershed, and from nearby Jefferson
City, Missouri, south of the watershed. Two of the five climate datasets used a single
climate station for daily climate input. One of these datasets (Sanborn Field) was
populated with data from the Sanborn Field climate station located in sub-basin #4 in a
small experimental agricultural field surrounded by a highly urbanized (university)
setting. To compare the urban-influenced climate input from the state-university-operated
Sanborn Field climate station, another single station climate input dataset (South Farms)
was taken from the South Farms climate station, located 0.5 km outside the HCW,
approximately 5.5 km southeast of Site #3, in an open, expansive agricultural
experimental area. Finally, two sets of multiple station climate input datasets were
produced using the HCW network of five hydroclimate stations. One of these datasets
(HCW-5 Climate) included all five stations. A second dataset (HCW-4 Climate) was
created including only the climate stations from Sites #2-5. This decision was made to
account for suspected precipitation undercatch at Site #1, possibly due to a fetch problem
from nearby woodland (Figure 9). Site #1 climate data other than precipitation (solar
radiation, wind speed, air temperature, and relative humidity) was not used in order to
avoid any other possible fetch problems with the Site #1 climate data.
49
Figure 9. Photograph of Site #1 climate station dated February 2011, showing the surrounding woodland, a potential source of precipitation undercatch. Hinkson Creek Watershed, Missouri, U.S.A.
ArcSWAT designates an input climate dataset with the nearest sub-basin by
determining in GIS the spatial proximity between the sub-basin and the climate station
(Winchell et al. 2009). All 20 combinations of the two watershed discretization schemes,
the two soil data sets STATSGO and SSURGO, and the five climate datasets were
configured in SWAT and ran for the period from 2001-2010. A “warm-up” period where
the model is run but not analyzed is recommended when running SWAT, allowing the
50
model to equilibrate to ambient watershed hydrologic conditions (e.g. soil antecedent
moisture condition) (Ahl et al. 2008). Only the two-year period of model simulated data
from 2009-2010, in which observed streamflow data was available, was used for analysis.
The specific configuration of each run is listed in Table 4.
Table 4. List of all 20 SWAT model configurations tested, the abbreviated names of each model, and the defining characteristics of each model. Configurations are systematically listed in order of increasing input data resolution.
Model Configuration
No. of
Sub-basins Soil Dataset
No. of
HRUs Climate Dataset
6_ST_WGN 6 STATSGO 22 Weather Generator
6_ST_SAN 6 STATSGO 22 Sanborn Field
6_ST_SFM 6 STATSGO 22 South Farms
6_ST_HCW4 6 STATSGO 22 HCW-4 Climate
6_ST_HCW5 6 STATSGO 22 HCW-5 Climate
6_SS_WGN 6 SSURGO 40 Weather Generator
6_SS_SAN 6 SSURGO 40 Sanborn Field
6_SS_SFM 6 SSURGO 40 South Farms
6_SS_HCW4 6 SSURGO 40 HCW-4 Climate
6_SS_HCW5 6 SSURGO 40 HCW-5 Climate
34_ST_WGN 34 STATSGO 108 Weather Generator
34_ST_SAN 34 STATSGO 108 Sanborn Field
34_ST_SFM 34 STATSGO 108 South Farms
34_ST_HCW4 34 STATSGO 108 HCW-4 Climate
34_ST_HCW5 34 STATSGO 108 HCW-5 Climate
34_SS_WGN 34 SSURGO 206 Weather Generator
34_SS_SAN 34 SSURGO 206 Sanborn Field
34_SS_SFM 34 SSURGO 206 South Farms
34_SS_HCW4 34 SSURGO 206 HCW-4 Climate
34_SS_HCW5 34 SSURGO 206 HCW-5 Climate
Evaluation of Uncalibrated Model Runs
To evaluate the goodness-of-fit of the twenty uncalibrated model configurations,
multiple statistical indicators were used. In addition to calculating previously published
51
goodness-of-fit indicators (Moriasi et al. 2007; ASCE 1993; Legates and McCabe 1999):
the percent bias (PBIAS) (Moriasi et al. 2007), [the root mean squared error (RMSE), the
mean absolute error (MAE), the squared-error based Nash Sutcliffe efficiency (NSE)
(Nash and Sutcliffe 1970), and the absolute-error based modified Nash and Sutcliffe
Efficiency (NSE1) (Legates and McCabe 1999)] for the set of 20 uncalibrated model
configurations, versions of the RMSE, MAE, NSE, and NSE1 that use ranked data series
rather than time series, were developed. The additional goodness-of-fit measures were
designed to measure the fit of modeled flow output with the observed flow duration
curve.
The additional measures are based on the sum of the squared error after ranking
(SSQR) automatic calibration objective function developed by van Griensven and
Bauwens (2003), which was incorporated into the built-in SWAT auto-calibration
method. The SSQR was used in a SWAT automatic calibration study by Van Liew et al.
(2005). In automatic calibration, an objective function, an indicator of the deviation
between a measured and simulated data series, is minimized using a computer algorithm
designed to find an optimal set of input parameter values (van Griensven and Bauwens
2003). The SSQR is given by the following equation:
nj
simulatedjmeasuredj xxSSQR,1
2
,,
(1)
where n represents the total number of paired values in a data series, j, sorted (ranked) by
magnitude, xmeasured is the measured value, and xsimulated is the model-simulated value. The
key difference between the use of SSQR in previous publications (van Griensven and
Bauwens 2003; Van Liew et al. 2005) and its use in this study is that the SSQR was
52
further developed into a tool for comparison of H/WQ models, rather than simply a tool
for parameter optimization in an automatic calibration method. As stated by van
Griensven and Bauwens (2003), the SSQR measures the model fitting of the distribution
of flows through time, and ensures that the full range of flows are represented in an
automatically calibrated model. Van Griensven and Bauwens (2003) further identified a
key advantage of the SSQR function over measures of the error in a time series; daily
H/WQ models in “small” basins may be more appropriately evaluated where errors in the
timing of peak flow events causes high residuals, while flows may be otherwise well-
represented. The new measures based on the SSQR, are identical mathematically to their
original non-ranked counterparts (RMSE, MAE, NSE, and NSE1) except that both the
observed and modeled streamflow data are sorted (ranked) according to the magnitude of
the discharge. The ranked versions of the aforementioned formulas are termed the R-
RMSE, R-MAE, R-NSE, and R-NSE1.
PBIAS, a measure in percentage terms of the overall model bias, i.e. the error in
total mass, is given by the following equation:
ni
measuredi
ni
simulatedimeasuredi
x
xx
PBIAS
,1
,
,1
,,
)100(
(2)
where n represents the total number of paired values (measured and model-simulated
streamflow) in a time-series, i, xmeasured is the measured value, and xsimulated is the model-
simulated value. A value of 0% represents optimal model goodness-of-fit. The RMSE is a
measure of the average error between observed and simulated values in a time-series and
53
is calculated as the square root of the squared errors and presented in the units of the
quantity of interest (e.g. m3/s for streamflow), as follows:
ni
simulatedimeasuredi xxn
RMSE,1
2
,,
1
(3)
where n represents the total number of paired values in a time-series, i, xmeasured is the
measured value, and xsimulated is the model-simulated value. A value of 0 represents
optimal model goodness-of-fit. The MAE, another measure of the average error between
observed and simulated values in a time-series, calculated as the absolute value of the
errors and presented in the units of the quantity of interest (e.g. m3/s for streamflow), is
given by the following equation:
ni
simulatedimeasuredi xxn
MAE,1
,,
1
(4)
where n represents the total number of paired values in a time-series, i, xmeasured is the
measured value, and xsimulated is the model-simulated value. A value of 0 represents
optimal model goodness-of-fit. The NSE is a dimensionless statistic that compares the
relative magnitude of the residual variance to the measured data variance (Moriasi et al.
2007). The NSE is given by the following equation:
ni
meanmeasuredi
ni
simulatedimeasuredi
xx
xx
NSE
,1
2
,
,1
2
,,
1
(5)
where n represents the total number of paired values in a time-series, i, xmeasured is the
measured value, xsimulated is the model-simulated value, and xmean is the mean of the
observed values. A NSE value of 1.0 represents optimal model fit, and a value of ≤0.0
54
indicates that the model simulated values are less accurate than a baseline model. The
baseline model is traditionally represented by the linear function f(x) equal to xmean,
where f(x) is equal to the quantity of interest and xmean is equal to the mean value of the
observed time-series (Nash and Sutcliffe 1970). The NSE1 is a dimensionless statistic
based on the NSE that calculates the absolute values of the errors in the time series rather
than the squared errors (Legates and McCabe 2007). The NSE1 is given by the following
equation:
ni
meanmeasuredi
ni
simulatedimeasuredi
xx
xx
NSE
,1
,
,1
,,
1 1
(6)
where n represents the total number of paired values in a time-series, i, xmeasured is the
measured value, xsimulated is the model-simulated value, and xmean is the mean of the
observed values. Similar to the NSE, A value of 1.0 represents optimal model fit, and a
value of ≤0.0 indicates a poor model-simulated fit to observed (Moriasi et al. 2007). a
value of 1.0 represents optimal model fit, and a value of ≤0.0 indicates that the model
simulated values are less accurate than a baseline model. The baseline model is
represented by the linear function f(x) equal to xmean, where f(x) is equal to the quantity of
interest and xmean is equal to the mean value of the observed time-series (Nash and
Sutcliffe 1970). The R-RMSE, a measure of the average error between observed and
simulated values in a series ranked by magnitude, calculated as the square root of the
squared errors and presented in the units of the quantity of interest (e.g. m3/s for
streamflow), is given by the following equation:
55
nj
simulatedjmeasuredj xxn
RMSER,1
2
,,
1
(7)
where n represents the total number of paired values in a ranked series, j, xmeasured is the
measured value, and xsimulated is the model-simulated value. A value of 0 represents
optimal model goodness-of-fit. The R-MAE, an additional measure of the average error
between observed and simulated values in a ranked series, calculated as the absolute
value of the errors and presented in the units of the quantity of interest (e.g. m3/s for
streamflow), is given by the following equation:
nj
simulatedjmeasuredj xxn
MAER,1
,,
1
(8)
where n represents the total number of paired values in a ranked series, j, xmeasured is the
measured value, and xsimulated is the model-simulated value. A value of 0 represents
optimal model goodness-of-fit. The R-NSE is a dimensionless statistic based on the NSE
that calculates the errors in a ranked series rather than a time series. The R-NSE is given
by the following equation:
nj
meanmeasuredj
nj
simulatedjmeasuredj
xx
xx
NSER
,1
2
,
,1
2
,,
1
(9)
where n represents the total number of paired values in a ranked series, j, xmeasured is the
measured value, xsimulated is the model-simulated value, and xmean is the mean of the
observed values. Similar to the NSE, the value of 1.0 represents optimal model fit, and a
value of ≤0.0 indicates that the mean observed value is a better predictor than the model-
simulated values. Finally, the R-NSE1 is a modified version of the dimensionless statistic
56
R-NSE that calculates the absolute values of the errors in the ranked series rather than the
squared errors. The R-NSE1 is given by the following equation:
nj
meanmeasuredj
nj
simulatedjmeasuredj
xx
xx
NSER
,1
,
,1
,,
1 1
(10)
where n represents the total number of paired values in a ranked series, j, xmeasured is the
measured value, xsimulated is the model-simulated value, and xmean is the mean of the
observed values. Similarly, a value of 1.0 represents optimal model fit, and a value of
≤0.0 indicates that the mean observed value is a better predictor than the model-simulated
values.
To aid in the interpretation of statistical goodness-of-fit indicators used in H/WQ
model evaluation, indicators were associated with a graphical or physical concept that
closely matched the error represented by the statistical indicator. In Table 5, the
aforementioned goodness-of-fit indicators were identified by a suitable conceptual
analog, either the mass balance, the hydrograph, or the flow duration curve. These
analogous associations were previously identified by van Griensven and Bauwens (2003).
While van Griensven and Bauwens (2003) used three sets of objective functions for
automatic streamflow calibration, one each fitting the visual/conceptual analogs
described here, they did not make a direct comparison with the subjective
visual/graphical approach often used in manual calibration (Srinivasan et al. 1998; Santhi
et al. 2001) and in interpretation of model fit by the end-user. Accordingly, use of a
conceptual analog may serve to improve understanding of how to appropriately interpret
these statistics.
57
Table 5. Descriptive table showing key characteristics of goodness-of-fit indicators used for model evaluation in this study.
Model Goodness-of-Fit
Indicator
Visual/Conceptual
Analog
Source of error
measured
Type of
error
measured
Standardized
for
Comparison
Percent Bias (PBIAS) Mass Balance Total Mass (Total
Water Yield)
Overall
Error (Bias)
Yes
Root Mean Square Error
(RMSE)
Hydrograph Event Magnitude
and Timing
Squared
Error
No
Mean Absolute Error
(MAE)
Hydrograph Event Magnitude
and Timing
Absolute
Error
No
Nash Sutcliffe Efficiency
(NSE)
Hydrograph Event Magnitude
and Timing
Squared
Error
Yes
Modified Nash Sutcliffe
Efficiency (NSE1)
Hydrograph Event Magnitude
and Timing
Absolute
Error
Yes
Ranked Root Mean Square
Error (R-RMSE)
Flow Duration
Curve
Percent of Time
Flow Exceeded
Squared
Error
No
Ranked Mean Absolute
Error (R-MAE)
Flow Duration
Curve
Percent of Time
Flow Exceeded
Absolute
Error
No
Ranked Nash Sutcliffe
Efficiency (R-NSE)
Flow Duration
Curve
Percent of Time
Flow Exceeded
Squared
Error
Yes
Ranked Modified Nash
Sutcliffe Efficiency
(R-NSE1)
Flow Duration
Curve
Percent of Time
Flow Exceeded
Absolute
Error
Yes
Indicators such as the Percent Bias (PBIAS) measure the difference between the
sums of the modeled and observed streamflow for the period of record under study; then
divide the difference by the sum of the observed flow. The output of the PBIAS thus
gives an indication of how well the modeled water yield compares with the observed
water yield. A close fit for PBIAS (PBIAS near 0%) indicates that observed and modeled
mass balance are very similar (Moriasi et al. 2007). With respect to hydrograph fit, the
study examines four indicators: two absolute-error based statistics, one which reports in
the units of the quantity of interest (MAE), and one which reports a standardized value
comparable across different gauging stations through normalization (NSE1, Legates and
McCabe 1999); and, also, two squared-error based statistics, the RMSE and the NSE
58
(Nash and Sutcliffe 1970). The NSE and NSE1 range from -∞ to 1, with a score of 0
indicating that the summed error, squared or absolute, between the modeled and observed
data is equal to the summed error between the mean value of the observed data and the
actual observed data (Legates and McCabe 1999). Metrics for evaluating flow duration
curve fit, based on the SSQR objective function of van Griensven and Bauwens (2003),
are introduced and are similar to the above mentioned set of four hydrograph fit
measures. In contrast to the set of hydrograph fit measures, the introduced metrics
compare paired values of observed and modeled streamflow after ranking each dataset in
order of magnitude.
In this study, the SWAT model is evaluated using the three classes of goodness-
of-fit statistical measures: total mass fit, hydrograph fit, and flow duration fit. Van Liew
et al. (2005) reported that there are tradeoffs when implementing the SWAT model
between achieving accurate representation of the total mass balance, hydrograph event
response, and the full range of flows. Accordingly, Van Liew et al. (2005) suggested
pursuing a modeling implementation approach that is best suited to the needs or problems
being addressed by the model; accordingly, a suitable balance in terms of amount, timing,
and distribution of the hydrologic or water quality variable may be attained. Furthermore,
van Griensven and Bauwens (2003) argued that achievement of a satisfactory hydrograph
fit alone does not imply that the hydrological processes, i.e. the division of surface runoff,
lateral flow (interflow), and groundwater input, have been adequately described by the
model. Therefore, while developing the auto-calibration method now made available with
SWAT, van Griensven and Bauwens (2003) incorporated three objective functions to
optimize the modeled streamflow output: one to minimize error in the overall mass
59
balance, one to minimize the errors for each event in the time-series, and one to minimize
error in the ranked streamflow series. In agreement with the finding of the
aforementioned studies and in accordance with the modeling approach described in the
introduction, this study used measures from each of the three classes of goodness-of-fit
indicators.
The normalized (dimensionless) statistics (PBIAS, NSE, NSE1, R-NSE, and R-
NSE1), capable of standardized comparison of model performance across different
gauging stations and watersheds, were used to rank the 20 uncalibrated model
configurations in terms of goodness-of-fit. In addition, direct comparisons were made
between the two watershed discretization schemes, the two soil datasets, and the five
climate datasets by averaging the goodness-of-fit indicators for each model configuration
using, for example, SSURGO soil data, and STATSGO soil data, as was done by Kumar
and Merwade (2009).
Test of the Built-In Automatic Calibration Method in SWAT
The ParaSol (Parameter Solutions) automatic calibration method built into SWAT
and developed by van Griensven (2002) is based on the Shuffled Complex Evolution
algorithm developed by the University of Arizona (SCE-UA). SCE-UA is a genetic
algorithm (Eckhardt and Arnold 2001) that was developed and shown by Duan et al.
(1984) to be highly efficient in finding the optimal set of model parameters. As described
by Eckhardt and Arnold (2001), the SCE-UA method considers a sample of points
(models with varying parameter values) spread throughout the parameter space (limited
by the upper and lower bounds assigned to each parameter). The points are conceived as
60
individual members of a biological population with specific genetic information
(parameter values) that evolve towards optimal fitness (i.e. the minimum of an objective
statistical criterion) through reproduction, mutation, and genetic recombination
(shuffling). An initial sample is sub-divided into multiple sub-samples (complexes) that
generate new sets of points (offspring) via the downhill simplex procedure of Nelder and
Mead (1965). The likelihood of an individual reproducing is based on the fitness of the
individual. The algorithm approaches an optimal set of parameters through mutation (the
simplex procedure), reproduction (replacement of parents by offspring according to
fitness), and genetic recombination (shuffling), a regular re-arrangement of individual
points into new complexes. Through this process, genetic algorithms are able to
efficiently and automatically optimize parameters used in H/WQ models (Duan et al.
1994; Eckhardt et al. 2001).
In this study, the auto-calibration routine was set to optimize the model based on
observed daily streamflow from the Site #4 streamflow gauge for the two-year period
2009-2010. Site #4 was chosen because it is a publically available USGS-operated gauge
and has been in operation since 1967. Accordingly, Site #4 was assumed to have a better
developed rating curve relative to the HCW Sites #1-3 and #5 (in operation since January
2009). A well-developed rating curve is necessary for accurate continuous streamflow
estimation. A publically available gauge also fits with the aforementioned practical
modeling framework aimed at end-users that was presented in the Introduction.
A split-sample method, i.e. the use of separate time periods or spatial locations of
observed flow data, one used for calibration of the parameters with the other serving as
an independent test, or validation, of the calibrated parameters, is typically used for
61
SWAT model validation (Gassman et al. 2007; Neitsch et al. 2005). In most examples of
SWAT model calibration and validation in the scientific literature, the calibration and
validation periods are separated by time (Gassman et al. 2007).
The method used in this study, however, was the more rarely used spatial split-
site calibration and validation method. In this method, data from one streamflow gauge
are used for model calibration and data from other gauges are used for model validation
(Gassman et al. 2007). It is surmised here that split-time sample methods are best suited
to forecasting applications, while split-site sample methods, used in this study, are best
suited for prediction in ungauged basins applications. This conclusion is based on the fact
that Nash and Sutcliffe (1970), in their discussion of the principles of conceptual,
physically based hydrologic modeling, were focused on two broad model applications:
flood event forecasting and forecasting the effects of management works on flow regime.
In these applications, model validation using the split-time method serves to directly test
the forecasting ability of the model. Prediction in ungauged basins or sub-basins,
conversely, a stated goal of the SWAT model (Neitsch et al. 2005) and not mentioned in
Nash and Sutcliffe (1970), is an application of spatial prediction or interpolation rather
than future forecasting and, accordingly, such applications may be best tested using split-
site validation.
The potential disadvantage to automatic calibration methodology is its intensive
demand on computing resources (e.g. two weeks of runtime reported by Kumar and
Merwade (2009)). Thus, the control parameters (Table 6) for the SCE-UA automatic
calibration built-in SWAT were chosen to conserve computational runtime. Based on the
experiment-based recommendations of Duan et al. (1994), the number of complexes in
62
the initial population (NGS) was set equal to the number of parameters being optimized
plus one. This value for NGS was suggested by the experiment-based findings of Duan et
al. (1994). Also, the number of evolution steps taken by each complex before shuffling
(NSPL) was set equal to twice the number of parameters being optimized plus one per the
recommendations of Duan et al. (1994). Parameters particular to the SWAT
implementation of the SCE-UA, KSTOP (set to 1), PCENTO (set to 0.01), and MAXN
(set to 1440) were set to levels that would constrain model runtime with a goal of
successful completion of the calibration procedure in under 24 hours.
63
Table 6. Settings used in ArcSWAT when running the built-in automatic calibration method.
SWAT Auto-Calibration Setting Value
Auto-Calibration Method (ICLB) Parasol only
Objective error function (OFMET) First Auto-Calibration: Sum of the
squared error (SSQ) for daily mean
discharge at Site #4 during year 2009-
2010
Second Auto-Calibration: Sum of the
squared error (SSQ) and the sum of the
squared error after ranking (SSQR) for
daily mean discharge at Site #4 during
year 2009-2010
Maximum number of simulations allowed (MAXN) 1440 (24 hour runtime at 1 minute per
simulation)
Maximum number of shuffling loops allowed following
initial loop in which the global objective criterion does
not change (KSTOP)
1
Percentage by which the global objective criterion must
change (PCENTO)
1%
Number of complexes in the initial population (NGS) 7 (Number of parameters to be
optimized plus one)
Number of evolution steps taken by each complex
before shuffling (NSPL)
13 (Twice the number of parameters to
be optimized plus one)
To further reduce runtime, a set of only six (out of a total of 26 streamflow-related
parameters in SWAT (Neitsch 2005) were selected for optimization (Table 7). These
parameters and their calibration bounds were identical to the set of six parameters used
by Van Liew et al. (2005), which were based on model parameter sensitivity analyses by
van Griensven (2002). This parameterization approach was applied to the entire
watershed as was done by Van Liew et al. (2005). Van Liew et al. (2005) reported
modeling results with automatic calibration that produced satisfactory results (as
measured by PBIAS and NSE) in one day of runtime. Kumar and Merwade (2009), on
64
the other hand, optimized a larger set of 14 flow parameters when using the built-in
SWAT auto-calibration tool and reported that the auto-calibration took two weeks using a
single desktop computer equipped with an Intel® Core 2 Duo™ 2.66 GHz processor with
2.00 GB of RAM. This study used the smaller and more computationally efficient set of
six parameters used by Van Liew et al. (2005). As suggested by the Law of Diminishing
Returns (a basic concept in economics), a smaller set of optimized parameters may be
more computationally efficient in terms of improving model accuracy than a larger set of
parameters, as was demonstrated by the comparison of Van Liew et al. (2005) and
Kumar and Merwade (2009).
The six SWAT input parameters selected for optimization (Table 7) were
ALPHA_BF, the baseflow alpha, or recession, factor; CN2, the initial curve number
condition II value, which governs surface runoff volumes; GW_DELAY, the
groundwater delay factor, which governs the length of time between percolation of water
from the soil profile and entry into the shallow groundwater aquifer; GW_REVAP, the
groundwater revap coefficient, which regulates the rate of “revap,” the movement of
water from the shallow aquifer back to the soil profile through evapotranspiration;
RCHRG_DP, the deep aquifer percolation fraction, which governs the amount of water
that permanently seeps out of the watershed‟s shallow aquifer; and REVAPMN, the
threshold water depth in the shallow aquifer for “revap” to occur (Neitsch et al. 2005).
65
Table 7. The six input parameters in SWAT selected for optimization with the built-in automatic calibration procedure, their initial values, and the variation settings.
SWAT Input
Parameter Input Parameter Description
Variation
Method
Bounds
Lower Upper
ALPHA_BF Baseflow alpha factor (days) Replace
value
0 1
CN2 Initial Curve Number Condition II
value
Multiply
default
parameter
value
-50% 50%
GW_DELAY Groundwater delay factor (days) Replace
value
0 500
GW_REVAP Groundwater revap coefficient Replace
value
0.02 0.2
RCHRG_DP Deep aquifer percolation fraction Replace
value
0 1
REVAPMN Threshold water depth in the shallow
aquifer for revap to occur (mm)
Replace
value
0 500
One model configuration arbitrarily selected from the set of twenty model
configurations was optimized using the automatic calibration routine using two separate
schemes. In the first scheme, the model was optimized for sum of the squared error
(SSQ) for daily streamflow, and in the second scheme was optimized for the sum of the
square error (SSQ) in conjunction with the sum of the squared error after ranking (SSQR)
for daily flow. When using the SWAT auto-calibration tool, other researchers optimized
streamflow for only a single objective function (Van Liew et al. 2005; Van Liew et al.
2007; Kumar and Merwade 2009; Zhang et al. 2009; Setegn et al. 2010); unlike van
Griensven and Bauwens (2003), who optimized streamflow using three different
objective functions: the SSQ, the SSQR, and the total mass controller (TMC). The TMC
objective function is not currently available for use in SWAT.
The previously described SSQ objective function, which measures the hydrograph
fit, is given by the following equation:
66
ni
simulatedimeasuredi xxSSQ,1
2
,,
(1)
where n represents the total number of paired values in the time series, i, xmeasured is the
measured value, and xsimulated is the model-simulated value. The SSQR objective function,
which measures the flow duration fit, is given by the following equation:
nj
simulatedjmeasuredj xxSSQR,1
2
,,
(11)
where n represents the total number of paired values in a data series, j, sorted (ranked) by
magnitude, xmeasured is the measured value, and xsimulated is the model-simulated value.
Finally, the TMC objective function, which measures the total mass fit, is given by the
following equation:
1)100(
,1
,
,1
,
ni
simulatedi
ni
measuredi
x
x
absTMC
(12)
where n represents the total number of paired values in the time series, i, xmeasured is the
measured value, and xsimulated is the model-simulated value.
To analyze the data, the modeled output for the selected uncalibrated model
configuration was ranked with the optimal models identified using the two automatic
calibration (both the single objective (SSQ) and the multiple objective (SSQ and SSQR)
methods) model runs following the same ranking system used to evaluate the
uncalibrated model configurations.
The SWAT auto-calibration method automatically identifies the optimal (best) set
of parameters for use in the model. The arbitrarily selected model chosen was configured
67
with 34 sub-basins, SSURGO soil data, and the HCW-4 climate dataset. The values for
the „best‟ set of parameters are stored in a text file named „bestpar.out‟. The model was
re-run using the „best‟ set of parameters by executing the „ReRun Calibrated Model‟
command in the ArcSWAT 2.3.4 extension for ArcGIS 9. The simulated daily
streamflow output for the model configured with the „best‟ set of parameters was then
obtained from the „output.rch‟ text file.
68
CHAPTER III
RESULTS
OBSERVED CLIMATE
The observed climate dataset indicates a wet two-year (2009-2010) study period
relative to historical values. The total two-year precipitation ranged from 1574 mm at
HCW Site #1 to 2709 mm at both HCW Site #3 and Sanborn Field, with a mean of 1234
mm in 2009 and 1243 mm in 2010 (1239 mm/year average) (Table 8). When excluding
Site #1, two-year precipitation totals ranged from 2481 mm at Site #2 to 2709 mm at both
HCW Site #3 and Sanborn Field, with a mean of 1300 mm in 2009 and 1328 mm in 2010
(1314 mm/year average). As reported earlier, historical mean annual precipitation was
1016 mm per year at the time of this study.
Daily maximum air temperature during the study period ranged from 29.4 °C at
HCW Site #1 to 37.3 °C at Sanborn Field with a mean of 32.1 °C (Table 8). The daily
maximum air temperature reported at Sanborn Field and South Farms, 37.3 °C and 36.2
°C, respectively, was substantially higher than the HCW sites, which ranged from 29.4
°C to 30.9 °C. Daily minimum air temperatures during the study period ranged from -
21.3 °C at South Farms to -15.7 °C at Site #3, with a mean of -18.5 °C. The daily
minimum air temperature reported at Sanborn Field and South Farms, -20.6 °C and 21.3
°C, respectively, was notably lower than the HCW sites, which ranged from -18.8 °C to -
15.7 °C.
69
Daily mean relative humidity ranged from 67.9 % at Sanborn Field to 77.4 % at
HCW Site #1, with a mean for all seven stations of 73.3 % (Table 8). Daily mean wind
speed ranged from 0.6 m/s at Site #1 to 3.1 m/s at South Farms, with a mean of 1.4 m/s.
The daily mean wind speed reported at Sanborn Field and South Farms, 2.0 m/s and 3.1
m/s, respectively, was considerably higher than the HCW sites, which ranged from 0.6
m/s to 1.3 m/s. Daily mean solar radiation ranged from 10.0 MJ/m2 at Site #1 to 13.8
MJ/m2 at South Farms with a mean of 12.4 MJ/m
2. Excluding Site #1, solar radiation
values ranged from 11.3 MJ/m2 at Site #4 to 13.8 MJ/m
2 at South Farms with a mean of
12.8 MJ/m2. Graphical time-series of all climate data are presented in Figures 10 through
15.
Table 8. Summary of climate data during 2009-2010 for five hydroclimate stations in the Hinkson Creek Watershed, Missouri, U.S.A. and the MU Agricultural Experimental Station‟s Sanborn Field and South Farms weather stations. Values for standard deviation are shown in parentheses.
2009 - 2010
Climate Data
Statistic HCW
#1
HCW
#2
HCW
#3
HCW
#4
HCW
#5
Sanborn
Field
South
Farms
Precipitation
(mm)
2009-10
total
1573.7 2481.3 2708.9 2670.9 2498.5 2708.9 2699.6
Maximum Air
Temperature
(°C)
max
daily
29.4 30.6 30.9 30.0 30.6 37.3 36.2
Minimum Air
Temperature
(°C)
min
daily
-18.8 -17.9 -15.7 -17.3 -18.2 -20.6 -21.3
Relative
Humidity (%)
daily
mean
77.4
(10.8)
74.3
(11.1)
72.0
(11.9)
75.2
(10.7)
74.7
(10.9)
67.9
(15.0)
71.5
(13.5)
Wind Speed
(m/s)
daily
mean
0.6
(0.4)
1.2
(0.7)
1.3
(0.6)
0.8
(0.4)
1.0
(0.7)
2.0
(0.7)
3.1
(1.3)
Solar
Radiation
(MJ/m2)
daily
mean
10.0
(5.8)
12.5
(7.7)
13.1
(8.2)
11.3
(7.1)
13.3
(7.9)
12.9
(7.4)
13.8
(7.6)
70
Figure 10. Daily precipitation at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
71
Figure 11. Daily maximum air temperature at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
72
Figure 12. Daily minimum air temperature at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
73
Figure 13. Daily mean relative humidity at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
74
Figure 14. Daily mean wind speed at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
75
Figure 15. Daily total solar radiation at each of seven climate stations used in modeling of the Hinkson Creek Watershed, Missouri, U.S.A.
76
DEVELOPED STAGE-DISCHARGE RATING CURVES
Rating curve determination for the streamflow gauges at Sites #1-3 and #5 was
based on a total of 34 streamflow measurements at Site #1, 33 streamflow measurements
at Site #2, 31 streamflow measurements at Site #3, and 56 streamflow measurements at
Site #5, 23 of which were identified as variable backwater influenced. Rating curve
determination at Site #4 was completed by the U.S. Geological Survey, using standard
operating procedures as described by USGS (1982; 2010).
To evaluate final rating curves, the Nash Sutcliffe Efficiency (NSE) was
calculated for each University-operated streamflow gauge. Because the NSE is very
sensitive to extreme outliers, one extreme outlier streamflow measurement at both Sites
#1 and #3, each likely due to measurement error, was eliminated from the calculation of
the NSE. In addition, flow measurements identified as variable backwater influenced at
Site #5 were excluded from the calculation of the NSE. Figure 16 shows each final rating
curve and its NSE. The mathematical functions for each rating curve are shown in
Appendix A. The NSE for the university-operated gauging stations ranged from 0.92 at
Site #1 to 0.99 at Sites #2 and #3. The NSE for the Site #5 rating curve (excluding
backwater measurements) was 0.95. A NSE, or a similar measures of rating curve
goodness-of-fit, for the USGS-operated Site #4 rating curve is not made available by the
USGS.
77
Figure 16. Plot of final rating curves used for continuous streamflow estimation at all five gauging stations that also shows streamflow measurements taken at each gauging station in the Hinkson Creek Watershed, Missouri, U.S.A. For the USGS-operated gauging station, Site #4, only available information is shown.
78
OBSERVED STREAMFLOW
Observed streamflow for each of the five streamflow gauging stations located in
the HCW are presented in this text using summary statistics, flow duration curves, and
hydrographs (Tables 9-10, Figures 17-20).
During the 2009-2010 study period, daily mean streamflow at Site #1 ranged from
0.03 m3/s to 77.52 m
3/s, with a mean of 1.18 m
3/s (median of 0.20 m
3/s). Site #2 ranged
from 0.12 m3/s to 185.35 m
3/s, with a mean of 2.11 m
3/s (median of 0.18 m
3/s). Site #3
ranged from 0.5 m3/s to 115.2 m
3/s, with a mean of 2.07 m
3/s (median of 0.36 m
3/s). Site
#4 ranged from 0.02 m3/s to 177.8 m
3/s, with a mean of 3.52 m
3/s (median of 0.51 m
3/s).
Streamflow at Site #5 ranged from 0.10 m3/s to 141.5 m
3/s, with a mean of 4.54 m
3/s
(median of 0.71 m3/s).
Monthly mean streamflow for the same periods reported above in daily terms,
between January 2009 and December 2010, ranged at Site #1 from 0.03 m3/s to 3.83
m3/s, with a mean of 1.18 m
3/s (median of 0.84 m
3/s). Site #2 ranged from 0.13 m
3/s to
9.56 m3/s, with a mean of 2.11 m
3/s (median of 1.37 m
3/s). Site #3 ranged from 0.5 m
3/s
to 6.50 m3/s, with a mean of 2.07 m
3/s (median of 0.36 m
3/s). Site #4 ranged from 0.24
m3/s to 10.31 m
3/s, with a mean of 3.52 m
3/s (median of 2.43 m
3/s). Streamflow at Site #5
ranged from 0.30 m3/s to 12.80 m
3/s, with a mean of 4.54 m
3/s (median of 3.35 m
3/s).
79
Table 9. Descriptive statistics for observed daily streamflow in 2009-2010, Hinkson
Creek Watershed, Missouri, U.S.A.
Statistic Observed Daily Streamflow Data
Site #1 Site #2 Site #3 Site #4 Site #5
MIN (m3/s) 0.024 0.115 0.046 0.020 0.099
MEAN (m3/s) 1.184 2.105 2.066 3.516 4.540
STDEV (m3/s) 4.926 10.246 7.630 12.424 14.001
MEDIAN (m3/s) 0.196 0.178 0.362 0.510 0.708
MAX (m3/s) 77.518 185.347 115.218 177.830 141.523
Figure 17. Flow duration curve for observed daily mean streamflow at five gauging stations in the Hinkson Creek Watershed, Missouri, U.S.A.
80
Figure 18. Hydrograph for observed daily mean streamflow at five gauging stations in the Hinkson Creek Watershed, Missouri, U.S.A.
81
Table 10. Descriptive statistics for observed monthly streamflow in 2009-2010, Hinkson Creek Watershed, Missouri, U.S.A.
Statistic Observed Monthly Streamflow Data
Site #1 Site #2 Site #3 Site #4 Site #5
MIN (m3/s) 0.026 0.130 0.119 0.238 0.300
MEAN (m3/s) 1.186 2.105 2.069 3.520 4.538
STDEV (m3/s) 1.165 2.458 1.909 3.159 4.060
MEDIAN (m3/s) 0.836 1.373 1.699 2.430 3.349
MAX (m3/s) 3.833 9.556 6.497 10.309 12.799
Figure 19. Flow duration curve for observed monthly mean streamflow at five gauging stations in the Hinkson Creek Watershed, Missouri, U.S.A.
82
Figure 20. Hydrograph for observed monthly mean streamflow at five gauging stations in the Hinkson Creek Watershed, Missouri, U.S.A.
83
UNCALIBRATED MODEL CONFIGURATIONS
In order to compare goodness-of-fit of the 20 uncalibrated model configurations
tested in this study, each of the twenty model configurations were ranked using five
standardized goodness-of-fit indicators, at both daily and monthly scales. Each of the 20
uncalibrated model configurations used default model parameter values, and only varied
in terms of watershed subdivision (number of sub-basins), resolution of the input soil
dataset, and input climate dataset used to force the model.
The standardized goodness-of-fit indicators used to rank the twenty uncalibrated
model configurations were PBIAS (Moriasi et al. 2007), NSE (Nash and Sutcliffe 1970),
NSE1 (Legates and McCabe 1999), R-NSE, and R-NSE1, averaged across all five
streamflow gauging sites (summary data in Table 11, rankings in Tables 15 through 20).
PBIAS was used to measure model goodness-of-fit in terms of total mass; NSE and NSE1
were used to measure goodness-of-fit in terms of event timing and magnitude
(hydrograph fit), based on each the square of the errors (NSE) and the absolute values of
the errors (NSE1); R-NSE and R-NSE1 were used to measure goodness-of-fit in terms of
the percent of time flows were exceeded (flow duration fit), based on each the square of
the errors (R-NSE) and the absolute values of the errors (R-NSE1)
The use of three different types of goodness-of-fit measures was modeled after
the work of van Griensven and Bauwens (2003), and the use of both squared error and
absolute value based statistics is based on the work of Willmott (1984; 1985) and Legates
84
and McCabe (1999). Full descriptive and goodness-of-fit statistics for each model
configuration are presented in Appendix B.
The 20 uncalibrated model configurations ranged from 35.10% to 12.53%
absolute value of PBIAS for daily flow and 35.31% to 12.35% PBIAS for monthly flow,
from a NSE (NSE1) for daily flow of -0.17 (0.07) to 0.19 (0.30), from NSE (NSE1) for
monthly flow of -0.32 (-0.03) to 0.66 (0.47), from a R-NSE (R-NSE1) for daily flow of
0.52 (0.57) to 0.92 (0.72), and from an R-NSE (R-NSE1) for monthly flow of 0.48 (0.44)
to 0.87 (0.65).
When ranked by the mean absolute value of PBIAS across all five sites for daily
streamflow, the 34 sub-basin, STATSGO soil, and HCW-4 climate dataset (using HCW
Sites #2-5) configuration performed the best (PBIAS = 12.53%). When ranked using
NSE for daily flow, the 6 sub-basin, SSURGO, HCW-4 climate (NSE = 0.19) performed
the best; however, using NSE1, the 6 sub-basin STATGSO, HCW-5 climate configuration
performed the best (NSE1 = 0.30). When ranked by R-NSE for daily flow, the 34 sub-
basin, STATSGO, HCW-4 climate configuration and the 34 sub-basin, STATSGO,
Sanborn Field configurations performed close to the observed data, both with a R-NSE of
0.92 and a R-NSE1 of 0.72. Unsurprisingly, the Weather Generator based models
performed poor relative to observed data across all error measures and greatly
underestimated streamflow during the wet years of 2009-2010.
Rankings according to monthly goodness-of-fit statistics indicated a different set
of best performing models. Based on the mean absolute value of PBIAS across all five
sites for monthly streamflow, the best performing model configuration was the 34 sub-
basin, STATSGO soil, and HCW-4 climate dataset configuration, (PBIAS = 12.35%), the
85
same best configuration according to daily PBIAS. When ranked using NSE and NSE1
for monthly streamflow, the 6 sub-basin, STATSGO, South Farms climate (NSE = 0.66,
NSE1 = 0.47) and the 34 sub-basin, STATGSO, South Farms climate (NSE = 0.66, NSE1
= 0.47) configurations performed the best. When ranked by R-NSE for monthly flow, the
6 sub-basin, STATSGO, HCW-4 climate, which also ranked as best using R-NSE for
daily flow, and the 34 sub-basin, STATSGO, HCW-4 climate configurations performed
the best, both with an R-NSE of 0.87. When ranked by R-NSE1, the 34 sub-basin,
STATSGO, HCW-4 climate configuration performed the best (R-NSE1= 0.65); this same
configuration performed the best in daily R-NSE1.
Table 11. Goodness-of-fit model evaluation statistics for all twenty uncalibrated model configurations. Optimal goodness-of-fit values are in bold. Configurations are systematically listed in order of increasing input data resolution. Error measures given in the table are the mean of the measures for all five gauging stations.
Model
Configuration
Total Mass Fit Hydrograph Fit Flow Duration Fit
PBIAS (%)
Daily Flow
PBIAS (%)
Monthly Flow
NSE (NSE1)
Daily Flow
NSE (NSE1)
Monthly Flow
R-NSE (R-NSE1)
Daily Flow
R-NSE (R-NSE1)
Monthly Flow
6_ST_WGN 35.10 35.31 -0.17 (0.08) -0.32 (-0.03) 0.58 (0.64) 0.50 (0.44)
6_ST_SAN 19.54 19.36 0.09 (0.16) 0.54 (0.39) 0.91 (0.70) 0.83 (0.58)
6_ST_SFM 16.71 16.50 0.05 (0.18) 0.66 (0.47) 0.91 (0.70) 0.81 (0.55)
6_ST_HCW4 13.11 12.88 0.15 (0.20) 0.62 (0.43) 0.91 (0.71) 0.87 (0.64)
6_ST_HCW5 23.32 23.47 0.18 (0.30) 0.60 (0.43) 0.85 (0.68) 0.76 (0.60)
6_SS_WGN 34.41 34.61 -0.13 (0.09) -0.29 (-0.02) 0.52 (0.57) 0.50 (0.45)
6_SS_SAN 19.99 19.78 0.16 (0.16) 0.54 (0.37) 0.90 (0.62) 0.81 (0.54)
6_SS_SFM 17.97 17.73 0.10 (0.17) 0.65 (0.44) 0.91 (0.62) 0.78 (0.51)
6_SS_HCW4 14.73 14.40 0.19 (0.18) 0.60 (0.40) 0.90 (0.61) 0.84 (0.60)
6_SS_HCW5 21.92 22.09 0.17 (0.29) 0.57 (0.42) 0.82 (0.62) 0.76 (0.62)
34_ST_WGN 33.34 33.52 -0.17 (0.07) -0.21 (0.04) 0.57 (0.65) 0.49 (0.44)
34_ST_SAN 18.82 18.64 -0.01 (0.13) 0.55 (0.40) 0.92 (0.72) 0.84 (0.59)
34_ST_SFM 16.23 16.04 -0.06 (0.15) 0.66 (0.47) 0.91 (0.72) 0.82 (0.56)
34_ST_HCW4 12.53 12.35 0.06 (0.17) 0.62 (0.43) 0.92 (0.72) 0.87 (0.65)
34_ST_HCW5 22.90 23.07 0.11 (0.28) 0.61 (0.44) 0.87 (0.70) 0.77 (0.61)
34_SS_WGN 33.63 33.80 -0.12 (0.09) -0.19 (0.04) 0.51 (0.58) 0.48 (0.44)
34_SS_SAN 19.70 19.49 0.11 (0.14) 0.55 (0.38) 0.91 (0.64) 0.81 (0.55)
34_SS_SFM 17.71 17.48 0.04 (0.16) 0.66 (0.45) 0.91 (0.64) 0.78 (0.51)
34_SS_HCW4 13.75 13.41 0.13 (0.17) 0.61 (0.41) 0.90 (0.64) 0.85 (0.61)
34_SS_HCW5 21.26 21.45 0.13 (0.28) 0.59 (0.42) 0.83 (0.64) 0.77 (0.63)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
86
86
87
Direct comparisons were made between 6 sub-basin and 34 sub-basin
configurations by averaging the goodness-of-fit statistics for all 10 configurations
discretized with 6 sub-basins and comparing those statistics with the average goodness-
of-fit statistics for all 10 configurations with 34 sub-basins (Table 12). This procedure,
shown briefly on page 1188 of Kumar and Merwade (2009), aggregates the results of all
models using 6 sub-basins and those using 34 sub-basins, thus isolating the effect of
watershed subdivision.
The 34 sub-basin model configurations outperformed the 6 sub-basin
configurations based on the average absolute value of PBIAS, with PBIAS for daily
(monthly) streamflow of 21.67% (21.59%) and 20.92% (20.84%) for the 6 sub-basin and
34 sub-basin models, respectively, a difference of 0.75 % (0.75%) PBIAS. Using NSE
(NSE1) for daily flow, the 6 sub-basin models outperformed the 34 sub-basin models,
with the 6 sub-basin models obtaining 0.08 (0.18) and the 34 sub-basin models receiving
0.02 (0.08), a difference of 0.06 (0.02). According to monthly NSE results (NSE1),
however, the 34 sub-basin models outperformed the 6 sub-basin models, with the 6 sub-
basin models receiving 0.42 (0.33) and the 34 sub-basin models receiving 0.45 (0.35), a
difference of 0.03 (0.02). Based on R-NSE (R-NSE1) for daily flow, the 34 sub-basin
models outperformed the 6 sub-basin models, with the 6 sub-basin models receiving 0.42
(0.33) and the 34 sub-basin models receiving 0.45 (0.35), a difference of 0.03 (0.02).
Using R-NSE (R-NSE1) for monthly flow, the 34 sub-basin models slightly outperformed
the 6 sub-basin models, with the 6 sub-basin models obtaining R-NSE (R-NSE1) values
of 0.75 (0.55) and the 34 sub-basin models obtaining 0.75 (0.56), a difference of 0.00
(0.01).
88
Table 12. Goodness-of-fit model evaluation statistics comparing model runs with 6 sub-basins and model runs with 34 sub-basins. Error measures given in the table are the mean of the measures for all five gauging stations. Optimal goodness-of-fit values are in bold.
Goodness-of-fit Statistic 6 Sub-basin Models 34 Sub-basin Models
Total Mass Fit Daily PBIAS (%) 21.67 20.92
Monthly PBIAS (%) 21.59 20.84
Hydrograph Fit Daily NSE (NSE1) 0.08 (0.18) 0.02 (0.16)
Monthly NSE (NSE1) 0.42 (0.33) 0.45 (0.35)
Flow Duration Fit Daily R-NSE (R-NSE1) 0.82 (0.65) 0.83 (0.67)
Monthly R-NSE (R-NSE1) 0.75 (0.55) 0.75 (0.56)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
Direct comparisons were also made between models configured with the low-
resolution soil dataset, STATSGO, and models configured with the high-resolution,
SSURGO soil dataset. This comparison was made by averaging the goodness-of-fit
statistics for all 10 configurations configured with STATSGO soil data and comparing
those statistics with the average goodness-of-fit statistics for all 10 configurations with
SSURGO data (Table 13).
The STATSGO-configured models outperformed the SSURGO configurations
based on the average absolute value of daily and monthly PBIAS, with a daily (monthly)
PBIAS of 21.08% (21.01%) and 21.51% (21.42%) for the STATSGO and SSURGO
models, respectively, a difference of 0.43% (0.41%) PBIAS. Using daily NSE (NSE1),
the results differed between squared error and absolute error based measures, with the
STATSGO-configured models receiving 0.02 (0.17) and the SSURGO-configured
models receiving 0.08 (0.17), a difference of 0.06 (0.00). According to monthly NSE
(NSE1), again the results differed between squared error and absolute error based
measures, with the STATSGO models obtaining a value of 0.43 (0.35) and the SSURGO
89
models obtaining a value of 0.43 (0.33), a difference of 0.00 (0.02). For the daily R-NSE
(R-NSE1), the STATSGO models outperformed the SSURGO models, with the
STATSGO models obtaining a value of 0.83 (0.69) and the SSURGO models obtaining a
value of 0.81 (0.62), a difference of 0.02 (0.07). Using monthly R-NSE (R-NSE1), the
STATSGO models outperformed the SSURGO models, with the STATSGO models
obtaining values of 0.75 (0.69) and the SSURGO models receiving 0.73 (0.63), a
difference of 0.02 (0.06).
Table 13. Goodness-of-fit model evaluation statistics comparing models using the low resolution STATSGO soil dataset and models using the SSURGO soil dataset. Error measures shown are the mean of the measures for all five gauging stations. Optimal goodness-of-fit values are in bold.
Goodness-of-fit Statistic STATSGO Models SSURGO Models
Total Mass Fit Daily PBIAS (%) 21.08 21.51
Monthly PBIAS (%) 21.01 21.42
Hydrograph Fit Daily NSE (NSE1) 0.02 (0.17) 0.08 (0.17)
Monthly NSE (NSE1) 0.43 (0.35) 0.43 (0.33)
Flow Duration Fit Daily R-NSE (R-NSE1) 0.83 (0.69) 0.81 (0.62)
Monthly R-NSE (R-NSE1) 0.75 (0.69) 0.73 (0.63)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
A final set of direct comparisons were made between the configurations based on
each climate dataset (Table 14). The averages of the goodness-of-fit statistics for all four
configurations using a) The Weather Generator dataset, b) The single-station Sanborn
Field dataset, c) The single-station South Farms dataset, d) The four-station HCW-4
climate dataset (Sites #2-5), and e) The five-station HCW-5 climate dataset (Sites #1-5)
were compared. In terms of PBIAS for daily and monthly flow, the HCW-4 climate
90
dataset performed the best with a daily (monthly) absolute value of PBIAS of 13.53%
(13.39%), followed by South Farms at 17.16% (16.93%), Sanborn Field at 19.52%
(19.32%), HCW-5 climate at 22.35% (22.52%), and Weather Generator at 34.12%
(34.31%). The Sanborn Field, South Farms, and HCW-4 climate datasets generally
overestimated streamflow in terms of overall model bias, while the Weather Generator
and HCW-5 climate datasets generally underestimated streamflow (Appendix B).
Based on NSE and NSE1 for daily streamflow, however, the HCW-5 climate
dataset performed well relative to the observed dataset with an NSE (NSE1) for daily
flow of 0.15 (0.29), followed by HCW-4 climate at 0.13 (0.18), Sanborn Field at 0.09
(0.15), South Farms at 0.03 (0.13), and Weather Generator at -0.15 (0.08). Based on NSE
and NSE1 for monthly flow, the South Farms climate dataset performed the best with a
NSE (NSE1) for monthly flow of 0.66 (0.46), followed by HCW-4 climate at 0.61 (0.42),
HCW-5 climate at 0.59 (0.43), Sanborn Field at 0.55 (0.38), and Weather Generator at -
0.25 (0.01). According to R-NSE and R-NSE1 for daily streamflow, the Sanborn Field,
South Farms, and HCW-4 climate datasets performed the best, all with a daily R-NSE (R-
NSE1) of 0.91 (0.67), followed by HCW-5 climate at 0.84 (0.66), and Weather Generator
at 0.55 (0.61). Considering the R-NSE and R-NSE1 for monthly streamflow, an
inconclusive comparison resulted, with the HCW-4 and HCW-5 climate datasets
generally performing the best with a monthly R-NSE (R-NSE1) of 0.86 (0.62) and 0.77
(0.62), respectively, followed closely by Sanborn Field at 0.82 (0.56) and South Farms at
0.80 (0.56), and Weather Generator at 0.49 (0.44).
Table 14. Goodness-of-fit model evaluation statistics comparing models using the SWAT weather generator, the single climate dataset from urban-located Sanborn Field, a single climate dataset from rural-located South Farms, four climate datasets from HCW Sites #2-5, and five climate datasets from HCW Sites #1-5. Error measures are the mean of the measures for all five gauging stations. Optimal goodness-of-fit values are in bold.
Goodness-of-fit Statistic Weather Generator Sanborn Field South Farms HCW-4 Climate HCW-5 Climate
Total Mass Fit Daily PBIAS (%) 34.12 19.52 17.16 13.53 22.35
Monthly PBIAS (%) 34.31 19.32 16.93 13.39 22.52
Hydrograph Fit Daily NSE (NSE1) -0.15 (0.08) 0.09 (0.15) 0.03 (0.13) 0.13 (0.18) 0.15 (0.29)
Monthly NSE (NSE1) -0.25 (0.01) 0.55 (0.38) 0.66 (0.46) 0.61 (0.42) 0.59 (0.43)
Flow Duration Fit Daily R-NSE (R-NSE1) 0.55 (0.61) 0.91 (0.67) 0.91 (0.67) 0.91 (0.67) 0.84 (0.66)
Monthly R-NSE (R-NSE1) 0.49 (0.44) 0.82 (0.56) 0.80 (0.53) 0.86 (0.62) 0.77 (0.62)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
91
92
Table 15. Uncalibrated model results ranked by a measure of mass balance fit: daily
Percent Bias. The Percent Bias shown is the average of the absolute values of each Percent Bias measure at all five HCW gauging stations.
Mean Absolute Value of
PBIAS for all Sites (%)
Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
12.53 34 STATSGO HCW-4 Climate
13.11 6 STATSGO HCW-4 Climate
13.75 34 SSURGO HCW-4 Climate
14.73 6 SSURGO HCW-4 Climate
16.23 34 STATSGO South Farms
16.71 6 STATSGO South Farms
17.71 34 SSURGO South Farms
17.97 6 SSURGO South Farms
18.82 34 STATSGO Sanborn Field
19.54 6 STATSGO Sanborn Field
19.70 34 SSURGO Sanborn Field
19.99 6 SSURGO Sanborn Field
21.26 34 SSURGO HCW-5 Climate
21.92 6 SSURGO HCW-5 Climate
22.90 34 STATSGO HCW-5 Climate
23.32 6 STATSGO HCW-5 Climate
33.34 34 STATSGO Weather Generator
33.63 34 SSURGO Weather Generator
34.41 6 SSURGO Weather Generator
35.10 6 STATSGO Weather Generator
93
Table 16. Uncalibrated models ranked by measures of hydrograph fit: daily Nash
Sutcliffe Efficiency and Modified Nash Sutcliffe Efficiency.
Mean Daily NSE for all Sites Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
0.187 6 SSURGO HCW-4 Climate
0.180 6 STATSGO HCW-5 Climate
0.170 6 SSURGO HCW-5 Climate
0.163 6 SSURGO Sanborn Field
0.155 6 STATSGO HCW-4 Climate
0.131 34 SSURGO HCW-4 Climate
0.129 34 SSURGO HCW-5 Climate
0.112 34 STATSGO HCW-5 Climate
0.108 34 SSURGO Sanborn Field
0.103 6 SSURGO South Farms
0.090 6 STATSGO Sanborn Field
0.064 34 STATSGO HCW-4 Climate
0.047 6 STATSGO South Farms
0.042 34 SSURGO South Farms
-0.010 34 STATSGO Sanborn Field
-0.063 34 STATSGO South Farms
-0.123 34 SSURGO Weather Generator
-0.129 6 SSURGO Weather Generator
-0.170 34 STATSGO Weather Generator
-0.173 6 STATSGO Weather Generator
Mean Daily NSE1 for all Sites No. of Sub-basins Soil Dataset Climate Dataset
0.300 6 STATSGO HCW-5 Climate
0.293 6 SSURGO HCW-5 Climate
0.282 34 SSURGO HCW-5 Climate
0.278 34 STATSGO HCW-5 Climate
0.198 6 STATSGO HCW-4 Climate
0.185 6 STATSGO South Farms
0.184 6 SSURGO HCW-4 Climate
0.171 6 SSURGO South Farms
0.170 34 SSURGO HCW-4 Climate
0.167 34 STATSGO HCW-4 Climate
0.164 6 STATSGO Sanborn Field
0.157 6 SSURGO Sanborn Field
0.156 34 SSURGO South Farms
0.151 34 STATSGO South Farms
0.144 34 SSURGO Sanborn Field
0.134 34 STATSGO Sanborn Field
0.094 6 SSURGO Weather Generator
0.093 34 SSURGO Weather Generator
0.078 6 STATSGO Weather Generator
0.068 34 STATSGO Weather Generator
94
Table 17. Uncalibrated models ranked by measures of flow duration fit: daily Ranked Nash Sutcliffe Efficiency and Ranked Modified Nash Sutcliffe Efficiency.
Mean Daily R-NSE for all Sites Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
0.920 34 STATSGO HCW-4 Climate
0.916 34 STATSGO Sanborn Field
0.915 6 STATSGO HCW-4 Climate
0.912 6 STATSGO Sanborn Field
0.912 34 SSURGO South Farms
0.910 34 SSURGO Sanborn Field
0.908 34 STATSGO South Farms
0.907 6 STATSGO South Farms
0.906 6 SSURGO South Farms
0.902 34 SSURGO HCW-4 Climate
0.900 6 SSURGO Sanborn Field
0.896 6 SSURGO HCW-4 Climate
0.867 34 STATSGO HCW-5 Climate
0.850 6 STATSGO HCW-5 Climate
0.833 34 SSURGO HCW-5 Climate
0.821 6 SSURGO HCW-5 Climate
0.580 6 STATSGO Weather Generator
0.574 34 STATSGO Weather Generator
0.525 6 SSURGO Weather Generator
0.510 34 SSURGO Weather Generator
Mean Daily R-NSE1 for all Sites No. of Sub-basins Soil Dataset Climate Dataset
0.720 34 STATSGO Sanborn Field
0.720 34 STATSGO HCW-4 Climate
0.718 34 STATSGO South Farms
0.709 6 STATSGO HCW-4 Climate
0.702 6 STATSGO Sanborn Field
0.700 6 STATSGO South Farms
0.698 34 STATSGO HCW-5 Climate
0.679 6 STATSGO HCW-5 Climate
0.653 34 STATSGO Weather Generator
0.645 34 SSURGO Sanborn Field
0.644 34 SSURGO South Farms
0.643 6 STATSGO Weather Generator
0.641 34 SSURGO HCW-5 Climate
0.636 34 SSURGO HCW-4 Climate
0.625 6 SSURGO HCW-5 Climate
0.624 6 SSURGO South Farms
0.619 6 SSURGO Sanborn Field
0.614 6 SSURGO HCW-4 Climate
0.579 34 SSURGO Weather Generator
0.571 6 SSURGO Weather Generator
95
Table 18. Default uncalibrated model runs ranked by a measure of mass balance fit: monthly Percent Bias. Percent Bias is the average of the absolute values of each Percent Bias measure at all five HCW gauging stations.
Mean Absolute Value of
PBIAS for all Sites (%)
Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
12.35 34 STATSGO HCW-4 Climate
12.88 6 STATSGO HCW-4 Climate
13.41 34 SSURGO HCW-4 Climate
14.40 6 SSURGO HCW-4 Climate
16.04 34 STATSGO South Farms
16.50 6 STATSGO South Farms
17.48 34 SSURGO South Farms
17.73 6 SSURGO South Farms
18.64 34 STATSGO Sanborn Field
19.36 6 STATSGO Sanborn Field
19.49 34 SSURGO Sanborn Field
19.78 6 SSURGO Sanborn Field
21.45 34 SSURGO HCW-5 Climate
22.09 6 SSURGO HCW-5 Climate
23.07 34 STATSGO HCW-5 Climate
23.47 6 STATSGO HCW-5 Climate
33.52 34 STATSGO Weather Generator
33.80 34 SSURGO Weather Generator
34.61 6 SSURGO Weather Generator
35.31 6 STATSGO Weather Generator
96
Table 19. Default uncalibrated model runs ranked by measures of hydrograph fit: monthly Nash Sutcliffe Efficiency and Modified Nash Sutcliffe Efficiency.
Mean NSE for all Sites Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
0.665 34 STATSGO South Farms
0.659 34 SSURGO South Farms
0.658 6 STATSGO South Farms
0.652 6 SSURGO South Farms
0.624 34 STATSGO HCW-4 Climate
0.619 6 STATSGO HCW-4 Climate
0.610 34 STATSGO HCW-5 Climate
0.608 34 SSURGO HCW-4 Climate
0.602 6 STATSGO HCW-5 Climate
0.602 6 SSURGO HCW-4 Climate
0.586 34 SSURGO HCW-5 Climate
0.574 6 SSURGO HCW-5 Climate
0.552 34 SSURGO Sanborn Field
0.546 34 STATSGO Sanborn Field
0.545 6 SSURGO Sanborn Field
0.537 6 STATSGO Sanborn Field
-0.187 34 SSURGO Weather Generator
-0.209 34 STATSGO Weather Generator
-0.286 6 SSURGO Weather Generator
-0.317 6 STATSGO Weather Generator
Mean NSE1 for all Sites No. of Sub-basins Soil Dataset Climate Dataset
0.472 34 STATSGO South Farms
0.466 6 STATSGO South Farms
0.448 34 SSURGO South Farms
0.440 6 SSURGO South Farms
0.437 34 STATSGO HCW-5 Climate
0.434 34 STATSGO HCW-4 Climate
0.434 6 STATSGO HCW-5 Climate
0.429 6 STATSGO HCW-4 Climate
0.422 34 SSURGO HCW-5 Climate
0.416 6 SSURGO HCW-5 Climate
0.408 34 SSURGO HCW-4 Climate
0.400 6 SSURGO HCW-4 Climate
0.397 34 STATSGO Sanborn Field
0.389 6 STATSGO Sanborn Field
0.377 34 SSURGO Sanborn Field
0.370 6 SSURGO Sanborn Field
0.044 34 SSURGO Weather Generator
0.035 34 STATSGO Weather Generator
-0.017 6 SSURGO Weather Generator
-0.028 6 STATSGO Weather Generator
97
Table 20. Default uncalibrated model runs ranked by measures of flow duration curve fit:
daily Ranked Nash Sutcliffe Efficiency and Ranked Modified Nash Sutcliffe Efficiency.
Mean R-NSE for all Sites Uncalibrated Model Runs
No. of Sub-basins Soil Dataset Climate Dataset
0.874 34 STATSGO HCW-4 Climate
0.868 6 STATSGO HCW-4 Climate
0.848 34 SSURGO HCW-4 Climate
0.842 6 SSURGO HCW-4 Climate
0.840 34 STATSGO Sanborn Field
0.834 6 STATSGO Sanborn Field
0.816 34 STATSGO South Farms
0.811 34 SSURGO Sanborn Field
0.811 6 STATSGO South Farms
0.806 6 SSURGO Sanborn Field
0.785 34 SSURGO South Farms
0.779 6 SSURGO South Farms
0.774 34 STATSGO HCW-5 Climate
0.769 34 SSURGO HCW-5 Climate
0.762 6 STATSGO HCW-5 Climate
0.757 6 SSURGO HCW-5 Climate
0.501 6 SSURGO Weather Generator
0.497 6 STATSGO Weather Generator
0.495 34 STATSGO Weather Generator
0.484 34 SSURGO Weather Generator
Mean R-NSE1 for all Sites No. of Sub-basins Soil Dataset Climate Dataset
0.651 34 STATSGO HCW-4 Climate
0.637 6 STATSGO HCW-4 Climate
0.628 34 SSURGO HCW-5 Climate
0.618 6 SSURGO HCW-5 Climate
0.615 34 STATSGO HCW-5 Climate
0.611 34 SSURGO HCW-4 Climate
0.605 6 STATSGO HCW-5 Climate
0.596 6 SSURGO HCW-4 Climate
0.586 34 STATSGO Sanborn Field
0.576 6 STATSGO Sanborn Field
0.555 34 STATSGO South Farms
0.548 34 SSURGO Sanborn Field
0.548 6 STATSGO South Farms
0.541 6 SSURGO Sanborn Field
0.515 34 SSURGO South Farms
0.507 6 SSURGO South Farms
0.453 6 SSURGO Weather Generator
0.444 34 SSURGO Weather Generator
0.439 6 STATSGO Weather Generator
0.439 34 STATSGO Weather Generator
98
AUTOMATIC CALIBRATION COMPARISON
The SCE-UA automatic calibration method built into SWAT was successfully run
in well under 24 hours as hypothesized. Previous studies reported up to two weeks of
runtime on a single computer (Kumar and Merwade 2009); model end-users in the
natural resources profession may find two weeks for automatic calibration impracticable.
With the model configuration selected for calibration, the 34 sub-basin, SSURGO soil,
and HCW-4 climate model, running at 1 minute per simulation, the single objective
automatic calibration (sum of the squared error only) took 691 simulations to successfully
complete while the multiple objective automatic calibration (sum of the squared error and
sum of the squared error after ranking) took 547 simulations to complete. In terms of
shuffled complex evolution (SCE) loops, the single objective calibration ran for 4
shuffling loops following the initial loop while the multiple objective calibration ran for 3
shuffling loops after the initial loop. The automatic calibration routine was run on a
single desktop-class computer with a dual-processor Intel® Core™ 2 Duo CPU rated at
3.33 GHz per core with 3.21 GB of RAM.
The automatic calibration procedure identifies the simulation resulting in the least
goodness-of-fit error. The parameter values used in this simulation are identified as the
„best‟ set of parameters. In this study, the single objective and multiple objective
automatic calibrations resulted in two distinct „best‟ sets of optimized parameters (Table
21).
99
The auto-calibration routine resulted in ALPHA_BF, the baseflow alpha factor,
changing from the default value of 0.048 days to 0.56 days in the single objective
calibration best parameter set and 0.48 days in the multiple objective calibration best
parameter set. The initial Curve Number condition II values were reduced basin-wide by
50% in the single objective calibration, while the Curve Numbers were increased by 7%
in the multiple objective calibration. GW_DELAY, the groundwater delay factor, was
changed from the default value of 31 days to 0.62 days in the single objective calibration
and 211.39 days in the multiple objective calibration. The GW_REVAP parameter, the
groundwater revap (movement of water from the shallow aquifer to the vadose
(unsaturated) zone due to evapotranspiration demand) coefficient, was changed from the
default value of 0.02 to 0.079 in the single objective calibration and 0.163 in the multiple
objective calibration. RCHRG_DP, the deep aquifer percolation fraction, was changed
from the default value of 0.05 to 0.00 in the single objective calibration and 0.44 in the
multiple objective calibration. Finally, REVAPMN, the threshold water depth in the
shallow aquifer for revap to occur, was changed from the default value of 1 mm to 417.4
mm in the single objective calibration and 355.3 mm in the multiple objective calibration.
Results show widely varied selected parameter sets between the single and multiple
objective automatic calibration methods.
100
Table 21. Values for the best set of six input parameters for the selected model configuration optimized first by minimizing the sum of squared error in daily streamflow at Site #4 and second by minimizing the sum of squared error and the sum of squared error after ranking using the built-in automatic calibration method in SWAT.
SWAT Input
Parameter Input Parameter Description
SWAT Default
Value
Best Value
SSQ SSQ & SSQR
ALPHA_BF Baseflow alpha factor (days) 0.048 0.56 0.48
CN2 Initial Curve Number Condition II
value
Varies -50% 7%
GW_DELAY Groundwater delay factor (days) 31 0.62 211.39
GW_REVAP Groundwater revap coefficient 0.02 0.079 0.163
RCHRG_DP Deep aquifer percolation fraction 0.05 0.00 0.44
REVAPMN Threshold water depth in the shallow
aquifer for revap to occur (mm)
1 417.4 355.3
Goodness-of-fit statistics based on streamflow at the USGS-operated Site #4
gauge were calculated to compare the selected uncalibrated configuration, the single
objective optimized model, and the multiple objective optimized model (Table 22).
Statistics for Site #4 were calculated because the SCE-UA automatic calibration
procedure was set to minimize the error in daily streamflow model prediction specifically
at Site #4. The calibration procedure was set to minimize error at Site #4 because it is a
long-term USGS-operated gauge with publically available streamflow data. The USGS
gauge has been in operation for a longer period (beginning in 1967) than all other HCW
streamflow gauges (beginning in 2009). Due to the longer operational period of the
USGS gauge, it was assumed that the USGS-operated Site #4 may serve as the most
suitable reference gauge.
In terms of PBIAS, the multiple objective optimized model performed the best
with an absolute value PBIAS for daily (monthly) flow of 0.50% (0.75%), followed by
101
the uncalibrated model at 11.75% (11.42%), and the single objective optimized model at
16.66% (16.08%). Using NSE and NSE1, the single objective optimized model performed
the best at simulating daily flow, with an NSE (NSE1) of 0.52 (0.35), followed by both
the uncalibrated model at 0.25 (0.24) and the multiple objective optimized model at 0.15
(0.30). Based on NSE and NSE1 for monthly flow, however, the multiple objective
optimized model performed the best with an NSE (NSE1) of 0.81 (0.59), followed by the
uncalibrated model at 0.68 (0.24), and the single objective optimized model at 0.54
(0.40). According to R-NSE and R-NSE1 for daily flow, the multiple objective optimized
model performed the best with an R-NSE (R-NSE1) of 0.97 (0.82), followed by the
uncalibrated model at 0.95 (0.66), and the single objective optimized model at 0.80
(0.63). R-NSE and R-NSE1 values for monthly streamflow indicated the uncalibrated
model and the multiple objective optimized model performed similarly with the
uncalibrated model obtaining an R-NSE (R-NSE1) value of 0.89 (0.67) and the multiple
objective optimized model obtaining a value of 0.89 (0.73). followed by the single
objective optimized model at 0.82 (0.67).
102
Table 22. Goodness-of-fit model evaluation statistics for Site #4 comparing the selected uncalibrated model configuration (34 sub-basin, SSURGO soil data, and HCW-4 climate dataset) with single objective (SSQ) and multiple objective (SSQ & SSQR) automatically calibrated runs of the selected model. Optimal goodness-of-fit values are in bold.
Goodness-of-fit Statistic for Site #4 Uncalibrated
Model
SSQ
Calibrated Model
SSQ & SSQR
Calibrated Model
Total Mass Fit Daily PBIAS (%) 11.75 16.66 0.50
Monthly PBIAS (%) 11.42 16.08 0.75
Hydrograph Fit Daily NSE (NSE1) 0.25 (0.24) 0.52 (0.35) 0.15 (0.30)
Monthly NSE (NSE1) 0.68 (0.45) 0.54 (0.40) 0.81 (0.59)
Flow Duration Fit Daily R-NSE
(R-NSE1) 0.95 (0.66) 0.80 (0.63) 0.97 (0.82)
Monthly R-NSE
(R-NSE1) 0.89 (0.67) 0.82 (0.69) 0.89 (0.73)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
To validate the automatically calibrated model results on a spatially independent
dataset, average goodness-of-fit statistics for Sites #1-3 and #5 were calculated (Table
23). In general, validation statistics were slightly lower than the statistics calculated for
Site #4, the USGS-operated gauge used for optimization in the automatic calibration
procedure.
In terms of PBIAS for daily (monthly) flow, the multiple objective optimized
model was closest to the observed data with an absolute value PBIAS of 11.11%
(10.98%), followed by the uncalibrated model at 14.24% (13.91%), and the single
objective optimized model at 19.37% (18.76%). Based on NSE and NSE1 for daily flow,
the single objective optimized model was closest to the observed data with an NSE
(NSE1) of 0.45 (0.31), followed by both the uncalibrated model at 0.19 (0.15) and the
multiple objective optimized model at 0.16 (0.20). Based on NSE and NSE1 for monthly
flow, on the other hand, the multiple objective optimized model was closest to the
103
observed data with an NSE (NSE1) of 0.70 (0.49), followed by the uncalibrated model at
0.59 (0.40), and the single objective optimized model at 0.44 (0.35). According to R-NSE
and R-NSE1 for daily flow, the multiple objective optimized model was closest to the
observed data with an R-NSE (R-NSE1) of 0.91 (0.77), followed by the uncalibrated
model at 0.89 (0.63), and the single objective optimized model at 0.74 (0.57). R-NSE and
R-NSE1 values for monthly flow indicated similar performance between all three models;
the R-NSE for the uncalibrated model on a monthly basis was (R-NSE1) of 0.84 (0.60),
for the single objective optimized model was 0.79 (0.66), and for the multiple objective
optimized model was 0.86 (0.66).
Table 23. Goodness-of-fit model evaluation statistics for Sites #1-3, and 5 comparing the
selected uncalibrated model configuration (34 sub-basin, SSURGO soil data, and HCW-4 climate dataset) single objective (SSQ) and multiple objective (SSQ & SSQR) automatically calibrated runs of the selected model. Optimal goodness-of-fit values are in bold.
Goodness-of-Fit Statistic for Sites #1-3, 5 Uncalibrated
Model
SSQ Calibrated
Model
SSQ & SSQR
Calibrated Model
Total Mass Fit Daily PBIAS (%) 14.24 19.37 11.11
Monthly PBIAS (%) 13.91 18.76 10.98
Hydrograph Fit Daily NSE (NSE1) 0.19 (0.15) 0.45 (0.31) 0.16 (0.20)
Monthly NSE (NSE1) 0.59 (0.40) 0.44 (0.35) 0.70 (0.49)
Flow Duration Fit Daily R-NSE
(R-NSE1) 0.89 (0.63) 0.74 (0.57) 0.91 (0.77)
Monthly R-NSE
(R-NSE1) 0.84 (0.60) 0.79 (0.66) 0.86 (0.66)
* PBIAS shown is the average of the absolute values of the PBIAS at all five gauging stations.
** Absolute-error based statistics are in parentheses.
104
CHAPTER IV
DISCUSSION
ANALYSIS OF OBSERVED CLIMATE DATA
Well above mean precipitation levels occurred during the study period. Mean
annual precipitation for all seven climate stations used averaged 1239 mm per year (22%
above historical average). When excluding the Site #1 precipitation data that was affected
by undercatch, the other six climate stations used averaged 1314 mm per year (29%
above historical average). Therefore, it is suggested that the modeling results presented in
this study be viewed within this context of well-above mean precipitation. The modeling
results do not reflect potential model performance in years with average precipitation or
performance in unusually dry years.
One clear problem in the climate data that arose was the apparent undercatch of
precipitation at Site #1. This problem was suspected due to the considerable difference in
total precipitation (1574 mm) during the study period measured at the six other climate
stations used in the study (2481 – 2709 mm). It was not surprising that Site #1 would
exhibit lower precipitation total due to fetch since the Site #1 climate station is located
closer to surrounding woodland area than any other site in the study (Figure 8). A simple
linear regression model was developed to correlate precipitation and elevation along a
North-South elevation gradient for six precipitation gauges (Sites #1-5 and a Missouri
Agricultural Experiment Station operated precipitation gauge in Auxvasse, Missouri).
The six gauges ranged in elevation from 172 m to 265 m above sea level. The linear
105
regression, however disqualified (coefficient of determination (R2) = 0.010) the
possibility of orographic effects being the sole cause for lower precipitation totals at Site
#1.
The results of the uncalibrated model configuration rankings using PBIAS, a
measure of overall model bias, additionally support the conclusion that the Site #1
climate station was indeed underestimating precipitation (Appendix B). Specifically, all
four model configurations using the HCW-4 climate dataset (Sites #2-5) performed better
based on PBIAS (absolute value of PBIAS for daily (monthly) streamflow ranging from
12.53% (12.35%) to 14.73% (14.40%)) than models using all other climate datasets
(absolute value of PBIAS for daily (monthly) flow ranging from 16.23% (16.04%) to
35.10% (35.31%)). Due to the multiple stations included in the HCW-4 climate dataset,
the results suggest that the HCW-4 best represented the climate parameters in the HCW
through representation of the spatial heterogeneity of climate in the watershed.
The four model configurations using the HCW-5 climate dataset performed
further from the observed data in terms of PBIAS (absolute value of PBIAS for daily
(monthly) flow ranging from 21.26% (21.45%) to 23.32% (23.47%)) than both the single
station South Farms and Sanborn Field climate datasets (absolute value of PBIAS for
daily (monthly) flow ranging from 16.23% (16.04%) to 19.99% (19.78%)). The poor
performance of models configured with the HCW-5 climate dataset relative to the single
station climate datasets suggests that the incorporation of Site #1 climate data in the
HCW-5 dataset considerably reduced model fit with observed data.
The HCW-5 model configurations only outperformed the stochastic weather
generated climate dataset (absolute value of PBIAS for daily (monthly) flow ranging
106
from 33.34% (33.52%) to 35.10% (35.31%)). Positive PBIAS values for models with the
HCW-5 climate dataset (Appendix B) indicated underestimation of streamflow; thus
suggesting that the HCW-5 climate dataset precipitation total accordingly was an
underestimate. The poor performance (streamflow underestimation) of the HCW-5
climate dataset may be attributed to the aforementioned precipitation undercatch at Site
#1.
Results also indicated key differences between the climate data collected at HCW
Sites #1-5 and the data collected at the Sanborn Field and South Farms climate stations.
The maximum and minimum daily air temperature values are more extreme (higher
maximum (37.3° at Sanborn Field and 36.2° at South Farms) and lower minimum (-20.6°
at Sanborn Field and -21.3° at South Farms)) at the Sanborn Field and South Farms
climate stations relative to HCW Sites #1-5 (29.4° to 30.9° maximum temperature and -
18.8° to -15.7° minimum temperature). Additionally, daily mean wind speeds were
considerably higher at Sanborn Field (2.0 m/s) and South Farms (3.1 m/s) than at HCW
Sites #1-5 (range from 0.6 – 1.3 m/s).
Both the moderation in temperature and reduced wind speeds at the HCW sites
may be at least partially attributed to the topographic position of the climate stations. The
HCW sites, co-located with stream gauging stations, are located in valley bottom riparian
zones, where air temperatures and wind speeds may be moderated (Dingman 2008;
Campbell and Norman 1999). The more extreme maximum and minimum daily air
temperatures at Sanborn Field and South Farms may also be attributed to the position of
the temperature gauges at approximately 1 m above the ground at those sites, making
those gauges more susceptible to surface heating and cooling processes than the HCW
107
temperature gauges, which are positioned approximately 3 m above the ground surface
(Campbell and Norman 1999).
Between May 2009 and July 2009, solar radiation values were relatively low at
HCW Sites #1 and #4 relative to all other sites as shown in Figure 15. The decline in
solar radiation during this period at both sites may be attributed to seasonal growth
patterns of leaf area in trees near the climate stations. At Site #1, multiple trees were
felled towards the end of this period to remove the shading problem. At Site #4, the
climate station was moved to an open field on the opposite side of Hinkson Creek
towards the end of this period to address fetch and canopy shading issues.
ANALYSIS OF OBSERVED STREAMFLOW DATA
Streamflow was perennial rather than intermittent at all five HCW gauging sites
during the two-year study period of this study (2009-2010). The lowest daily mean
streamflow, 0.020 m3/s, was recorded at Site #4. This finding supports the use of the
default value of zero for the CH_K (2) parameter in SWAT. This parameter represents
the effective hydraulic conductivity of the main stream channel in mm/hour (Neitsch et
al. 2005). As stated by Neitsch et al. (2005), the effective hydraulic conductivity for
perennial streams is zero.
Observed streamflow at Site #2 had higher total (mean) streamflow for the 2009-
2010 period (2.105 m3/s) than Site #3 (2.066 m
3/s), located 5.5 km downstream of Site
#2. It is also noted that Site #2 had the highest maximum daily streamflow (185.4 m3/s)
compared to all other sites (daily range 77.5 – 177 m3/s). It is suspected that these
seemingly unrealistic flow values for Site #2 may be due to a large flux or slug of
108
bedload sediment in the sand and gravel size-classes that was observed in the Site #2
reach in the Spring of 2009. The bedload sediment flux resulted in the up to 0.5 meter
burial of the orifice of the Accubar® Constant Flow Bubble Gauge during multiple
(approximately five) storm events. The orifice was uncovered following storm events by
shoveling sediment away from the orifice. The large amount of accumulated bedload
sediment that temporarily buried the bubbler orifice on multiple occasions may have
interfered with proper operation of the bubbler gauge, which measures the gauge pressure
(the absolute pressure subtracted by the atmospheric pressure). Sediment covering the
bubbler orifice may have increased the absolute pressure at the bubbler orifice, resulting
in greater than actual computed water levels.
In July 2009, the bubbler was moved approximately 15 meters downstream to
avoid sedimentation. A second possible explanation for the greater discharge at Site #2
may be that the flow at Site #3 has artificially reduced groundwater input due to
impervious surface area in sub-basin #3. Also, there potentially exists an urban
stormwater path(s) that may route water out of sub-basin #3 into sub-basin #4, thereby
bypassing the Site #3 streamflow gauge. No direct evidence of such surface flow paths
exists; further investigation is warranted.
In addition, flow duration curves suggest that Site #2 observed streamflow is
unusual and/or in error. The Site #2 flow duration curve intersects the Site #1 flow
duration curve for daily flow (Figure 17) less than 1 m3/s, and intersects the Site #3 flow
duration curve for monthly flow (Figure 19) above 6 m3/s. The Site #2 flow duration
curve also exhibits an asymptote between ranked days 300 to 730, suggesting that low
(base) flows are far more consistent in frequency and magnitude than all other sites. This
109
phenomenon observed in the flow duration curves is a function of the close spatial
proximity between Sites #2 and #3; the sub-basin area between the sites measures 1327
ha, only 6.4% of the total cumulative contributing area at Site #5, the most downstream
gauge in the study, and only 11.6% of the total cumulative contributing area to Site #3.
The phenomenon also suggests that further refinement of the rating curves at Sites #2 and
#3 is needed, particularly in the low water portion (less than 0.2 m3/s discharge) of the
rating curve where the Site #2 flow duration curve asymptotes. Further refinement may
be achieved by conducting regular (e.g. monthly or seasonal) cross-sections at Sites #2
and #3 during periods of low flow (less than 0.2 m3/s discharge).
While the Site #5 stage record was post-processed to correct for variable
backwater effects, further investigation of the variable backwater phenomenon at Site #5
is warranted. Investigation into backwater at Site #5 may permit a more direct
determination of the streamflow at Site #5; for this study, periods with which the effect of
variable backwater on streamflow was acute required correction using a linear regression-
model based on the Site #4 streamflow. A more direct determination of streamflow at
Site #5 may be obtained through the use of a velocity-index rating curve as described by
the USGS (1982). A velocity-index rating curve would require installation of a
continuously recording velocity meter(s) at a point in the stream or at multiple points
along a transverse line (USGS 1982). Another method suggested by the USGS (1982), a
slope index rating curve, proved unsuccessful for calculation of Site #5 backwater-
affected streamflow when using the USGS-operated Site #4 gauge as an auxiliary gauge
for continuous measurement of water surface slope. The failure of this attempt to develop
a slope-rating function for Site #5 is likely due to the appreciable longitudinal curvature
110
of the stream reach between Sites #4 and #5; the curvature likely led to the measured
water surface slope to be a poor estimate of the energy slope (USGS 1982). This failure
also may be due to the large proportional difference in cumulative contributing area
between Sites #4 and #5 (14.6%, 2630.1 ha), resulting in markedly different magnitudes
in flow between the two gauging stations.
UNCALIBRATED MODEL PERFORMANCE
Moriasi et al. (2007) proposed standard criteria for judging H/WQ model
performance on a monthly time scale. Based on the proposed Percent Bias (PBIAS)
criteria, 4 of the 20 uncalibrated model configurations may be considered “good” (10% <
PBIAS < 15%), 12 of the 20 may be considered “satisfactory” (15% < PBIAS < 25%),
and 4 of the 20 configurations may be considered “unsatisfactory” (PBIAS > 25%).
These results demonstrate that the SWAT model is capable of achieving
acceptable goodness-of-fit with observed data according to published standard criteria
(Moriasi et al. 2007) without calibration (i.e. parameter optimization) in the Hinkson
Creek Watershed, Missouri, U.S.A., when full-suite observed climate data (i.e.
precipitation, air temperature, solar radiation, wind speed, and relative humidity) are
used. Furthermore, these results may be applicable to other watersheds in Missouri and
the Central U.S.A.
The sole common denominator among the “good” model configurations was the
use of the HCW-4 climate dataset; the sole common denominator among the
“unsatisfactory” model configurations is the use of the Weather Generator climate
dataset. Models using the HCW-4 climate dataset may have performed closest to the
111
observed streamflow values as measured by PBIAS because this dataset potentially better
quantified the variability in climate (e.g. precipitation) across the HCW, particularly at a
daily time-scale. Based on the proposed Nash Sutcliffe Efficiency (NSE) criteria of
Moriasi et al. (2007), 4 of the 20 model configurations may be considered “good” (0.65 <
NSE < 0.75), 12 of the 20 model configuration may be considered “satisfactory” (0.50 <
NSE < 0.65), and 4 of the 20 configurations may be considered “unsatisfactory” (NSE <
0.50).
The sole common denominator among the “good” model configurations based on
NSE is the use of the South Farms climate dataset; the sole common denominator among
the “unsatisfactory” model configurations based on NSE is the use of the Weather
Generator climate dataset. For monthly streamflow simulation, models using the well-
sited (negligible fetch problem) South Farms climate dataset most closely matched the
observed streamflow at a monthly scale. These results indicate that a single well-sited
(negligible fetch) and fully-equipped climate station dataset is capable of outperforming
multiple fully-equipped climate stations that better represent the spatial variability in
climate (e.g. precipitation).
EFFECT OF WATERSHED SUBDIVISION AND INPUT DATASET SELECTION ON MODEL FIT
Goodness-of-fit statistical rankings of the uncalibrated model configurations did
not clearly determine the optimal choice for model watershed subdivision, input soil or
climate dataset because the rankings represent the combined effect of watershed
subdivision, input soil, and climate dataset. To overcome this problem, direct
112
comparisons were made between watershed discretization schemes, soil datasets, and
climate datasets. Direct comparisons were obtained by, for example, by averaging the
NSE for each model configured with STATSGO soil data and each model configured
with SSURGO soil data, as was briefly demonstrated on page 1188 in Kumar and
Merwade (2009).
Watershed Subdivision
The direct comparison between 6 sub-basin and 34 sub-basin model
configurations indicated that the choice between low and high resolution watershed
discretization had a negligible effect on streamflow simulation (Table 12). This finding of
negligible effect is based on the fact that differences between goodness-of-fit statistical
values were minimal and varied with the goodness-of-fit evaluation measure used. The
differences in PBIAS for daily streamflow between 6 sub-basin and 34 sub-basin models
was 0.43% and for monthly streamflow was 0.41%. The differences between NSE
(NSE1) between 6 sub-basin and 34 sub-basin models for daily streamflow was 0.06
(0.00) and for monthly streamflow was 0.00 (0.02). The differences between R-NSE (R-
NSE1) for 6 sub-basin and 34 sub-basin models for daily streamflow was 0.02 (0.07) and
for monthly streamflow was 0.02 (0.06). A similar case emerged in the direct comparison
of the low resolution STATSGO soil model configurations and the high resolution
SSURGO soil model configurations.
Direct comparisons of the effect of watershed subdivision on SWAT modeled
streamflow accuracy were not available. Arabi et al. (2006) report the effects of
watershed subdivision on sediment and nutrient modeled predictions, but not streamflow.
113
Kumar and Merwade (2009) do not provide a direct quantitative comparison of results
between model configurations with varying levels of watershed subdivision. To analyze
the modeling results presented in Kumar and Merwade (2009) with respect to the effect
of watershed subdivision on streamflow simulation accuracy, the data presented in Table
4, pg. 1187, in Kumar and Merwade (2009) were re-analyzed using the same methods in
this study (Table 24). Goodness-of-fit measures for total mass fit and hydrograph fit were
available in Kumar and Merwade (2009), but measures for flow duration fit were not
available. Kumar and Merwade subdivided the modeled watershed using various
threshold critical source areas (CSAs). The CSA represents the minimum area for a
stream channel to originate. The CSA was defined by the percentage of overall watershed
area. Only the four CSA percentages (2%, 3%, 5%, and 7%) applied consistently (four
configurations per CSA percentage) in Kumar and Merwade (2009) were analyzed in this
study. Thus, 16 of 24 total model configurations used by Kumar and Merwade were taken
into account. Lower percent CSA values result in finer watershed subdivision (i.e. more
sub-basins). Results are presented for uncalibrated, calibrated, and validated runs of the
16 model configurations. Kumar and Merwade (2009) used a single objective automatic
calibration (sum of the squared error (SSQ) only) method with the built-in SWAT auto-
calibration tool.
The results of the re-analysis confirm the findings determined in this study;
watershed subdivision had a negligible effect on streamflow simulation accuracy. Mean
absolute values of PBIAS for daily streamflow ranged from 6.47% to 8.51% for 16
uncalibrated models; ranged from 3.28% to 9.17% in the single objective (SSQ)
automatically calibrated models; and ranged from 12.42% to 14.11% in the validated
114
models. Mean values of NSE ranged from -0.01 to 0.13 for the uncalibrated models;
ranged from 0.59 to 0.69 for the calibrated models; and ranged from 0.57 to 0.60 for the
validated models. These differences in goodness-of-fit may be considered negligible.
Table 24. Goodness-of-fit model evaluation statistics comparing SWAT model configurations with varying levels of watershed subdivision (Kumar and Merwade 1999). Watershed subdivisions are defined by percent contributing source area (CSA). Error measures shown are the mean of the measures for all 24 models.
Statistic 2% CSA 3% CSA 5% CSA 7% CSA
Uncalibrated Models
Total Mass Fit Daily PBIAS (%) 6.47 7.62 8.51 7.89
Hydrograph Fit Daily NSE -0.01 0.05 -0.01 0.13
Calibrated Models
Total Mass Fit Daily PBIAS (%) 6.46 4.21 9.17 3.28
Hydrograph Fit Daily NSE 0.69 0.66 0.59 0.65
Validated Models
Total Mass Fit Daily PBIAS (%) 12.42 13.00 12.89 14.11
Hydrograph Fit Daily NSE 0.60 0.58 0.57 0.58
* Optimal goodness-of-fit values are in bold.
** PBIAS shown is the average of the absolute values of the PBIAS for all model configurations.
*** Lower percent CSA values result in finer watershed subdivision (i.e. more sub-basins).
Soil Data Resolution
Differences between goodness-of-fit values between models configured with
STATSGO soil data and with SSURGO soil data were also negligible and also varied
with the goodness-of-fit evaluation measure used (Table 13). The findings suggested that
the choice between STATSGO and SSURGO soil data had a negligible effect on
streamflow simulation. For both daily and monthly streamflow the difference in PBIAS
was 0.75%. For daily streamflow, the difference in NSE (NSE1) was 0.06 (0.02) for daily
115
flow and was 0.03 (0.02) for monthly streamflow. The value of R-NSE (R-NSE1) for
daily streamflow was 0.03 (0.02) for daily flow and for monthly flow was 0.00 (0.01).
Published research corroborates the results of this study in that the resolution of
the soil dataset had negligible influence on the modeled streamflow output (Kumar and
Merwade 2009; Ye et al. 2011). In the 15,535 km2 Xinjiang River basin in China, Ye et
al. (2011) found that varying soil data resolution did not lead to considerable differences
in SWAT streamflow simulation accuracy. For the 2800 km2 St. Joseph River Watershed
in Michigan, Indiana, and Ohio, U.S.A, Kumar and Merwade (2009) report average NSE
(PBIAS) for 12 SWAT model configurations varying in watershed discretization and soil
data resolution (STATSGO vs. SSURGO) of 0.56 (19.4%) with STATSGO soil data and
0.56 (19.0%) for models with SSURGO soil data, a negligible difference of 0.00 (0.4%).
These directly compared goodness-of-fit statistics reported in Kumar and Merwade
(2009) were only calculated for model validation following a single objective automatic
calibration (sum of the squared error (SSQ) only) using the built-in SWAT auto-
calibration tool.
The modeling results presented in Kumar and Merwade (2009) were further
analyzed in this study with respect to the effect of soil resolution on streamflow
simulation accuracy. The data presented in Table 4, pg. 1187, in Kumar and Merwade
(2009) were re-analyzed using the same methods used for direct comparison in this study
(Table 25). All 24 model configurations were taken into account, including 12 model
configurations of the aforementioned 2800 km2 St. Joseph River Watershed (SJRW) and
also 12 model configurations of the 700 km2 Cedar Creek Watershed, a sub-watershed of
the SJRW. The uncalibrated, calibration, and validation model runs were all taken into
116
account. As mentioned previously, goodness-of-fit measures for total mass fit and
hydrograph fit were available in Kumar and Merwade (2009), but measures for flow
duration fit were not available. For this re-analysis, a direct comparison was made
between all 12 models configured with STATSGO soil data and all 12 models configured
with SSURGO soil data.
The results of the soil resolution re-analysis confirm the findings determined in
this study; soil resolution had a negligible effect on streamflow simulation accuracy.
Mean absolute values of PBIAS for daily streamflow were 9.71% for STATSGO-
configured models and 7.37% for SSURGO-configured uncalibrated models, a difference
of 2.34%. For the single objective (SSQ) automatically calibrated models, PBIAS for
daily streamflow was 3.88% for STATSGO models and 5.88% for SSURGO models, a
difference of 2.00%. For the validation models, PBIAS for daily streamflow was 13.82%
for STATSGO models and 12.22% for SSURGO models, a difference of 1.60%.
Mean values of NSE for daily streamflow were 0.10 for STATSGO-configured
models and -0.16 for SSURGO-configured uncalibrated models, a difference of 0.26. For
the automatically calibrated model runs, NSE for daily streamflow was 0.66 for
STATSGO models and 0.66 for SSURGO models, a difference of 0.00. For the
validation models, NSE for daily streamflow was 0.58 for STATSGO models and 0.59
for SSURGO models, a difference of 0.01. These differences in goodness-of-fit (i.e. both
PBIAS and NSE) may also be considered negligible.
117
Table 25. Goodness-of-fit model evaluation statistics comparing SWAT models configured with low resolution soil data (STATSGO) and high resolution soil data (SSURGO) (Kumar and Merwade 2009). Error measures shown are the mean of the measures for all 24 models.
Statistic STATSGO SSURGO
Uncalibrated Models
Total Mass Fit Daily PBIAS (%) 9.71 7.37
Hydrograph Fit Daily NSE 0.10 -0.16
Calibrated Models
Total Mass Fit Daily PBIAS (%) 3.88 5.88
Hydrograph Fit Daily NSE 0.66 0.66
Validated Models
Total Mass Fit Daily PBIAS (%) 13.82 12.22
Hydrograph Fit Daily NSE 0.58 0.59
It is surmised here that soil resolution does not considerably affect SWAT
streamflow simulation accuracy due to SWAT‟s use of the empirical SCS curve number
method for surface runoff / infiltration volume partitioning (Neitsch et al. 2005). The
SCS curve number method requires input of an empirical, indirectly calculated non-
physical parameter, the curve number (CN). Default curve numbers are supplied in the
SWAT model using a built-in database. The CN is a function of soil permeability, land
use, and antecedent soil moisture conditions. The CN is dynamically adjusted in SWAT
in response to topographic slope and daily changes in soil water content (Neitsch et al.
2005). It is surmised here that a more process-based approach that uses physically
measureable soil parameters to partition surface runoff and infiltration volumes would
increase SWAT‟s sensitivity to input soil data resolution. Increased model sensitivity to
soil data resolution may provide more accurate representation of surface runoff and
infiltration patterns in the modeled watershed.
118
While a second optional method for surface runoff / infiltration portioning, the
Green-Ampt-Mein-Larsen infiltration method, is made available in SWAT (Neitsch et al.
2005), its use in the SWAT modeling literature is rare according to Gassman et al.
(2007). Summarizing the few studies that used the Green-Ampt-Mein-Larsen method,
Gassman et al. (2007) concluded that there was no discernible benefit to its use over the
SCS curve number method. Furthermore, use of the Green-Ampt-Mein-Larsen
infiltration method requires sub-daily (e.g. hourly) precipitation input data, which in
many cases may not be available to end-users.
The negligible difference in SWAT modeled streamflow accuracy between the
low and high resolution soil datasets may also be attributed to the high runoff potential of
the soils in the HCW (primarily hydrologic soil groups C and D (Figure 3)) in
combination with the well above mean precipitation during the study period (29% above
historical average when excluding Site #1). A combination of high runoff potential soils
and high soil available water content (due to above average precipitation) may be
expected to cause high levels of surface runoff. Since surface runoff flows overland
rather than through the soil, it may be expected that in this watershed (HCW) during this
study period (2009-2010), the accuracy of the representation of soil hydraulic properties
in the model was not critical to accurate model streamflow simulation. In watersheds with
soils with lower runoff potential (hydrologic soil groups A and B) and under more arid
conditions, soil data resolution may take on increased importance in model streamflow
simulation.
119
Climate Data
While study results indicated mixed (according to evaluation statistic used) and
negligible differences in model fit between watershed subdivision schemes and soil
datasets, direct comparisons made between models configured with the five different
climate datasets clearly demonstrated considerable differences in model fit (Table 14). A
direct comparison of PBIAS, the statistical measure of modeled versus observed total
mass, for models configured with each climate dataset indicated that the models
configured with the HCW-4 climate dataset performed the best (mean absolute value of
PBIAS for daily (monthly) flow of 13.53% (13.39%)). This result may be attributed to
the more accurate representation of spatial heterogeneity in climate parameters (e.g.
precipitation) by the multiple station HCW-4 dataset and the HCW-4 dataset‟s exclusion
of the problematic (i.e. precipitation undercatch) Site #1 climate data. Thus, it may be
concluded that multiple station input climate datasets that well-represent spatial climate
heterogeneity may be best suited to modeling applications requiring close fit with the
observed total mass (i.e. mass balance).
The precipitation undercatch in the Site #1 climate dataset further resulted in the
HCW-5 climate dataset obtaining values of PBIAS values for daily and monthly flow
(absolute value of PBIAS for daily (monthly) flow of 22.35% (22.52%)) indicated that
the HCW-5 climate dataset performed further from observed in terms of total mass than
the single station, homogenous Sanborn Field (absolute value of PBIAS for daily
(monthly) of 19.52% (19.32%)) and South Farms climate datasets (absolute value of
PBIAS for daily (monthly) flow of 17.16% (16.93%)). As discussed previously, the poor
120
performance (streamflow underestimation) of the HCW-5 climate dataset may be
attributed to the aforementioned precipitation undercatch at Site #1.
Based on measures of hydrograph fit, the squared error NSE and the absolute
error NSE1, however, the HCW-5 climate dataset performed best at the daily scale (NSE
(NSE1) for daily flow of 0.15 (0.29)) while the South Farms climate dataset performed
best on the monthly scale (NSE (NSE1) for monthly flow of 0.66 (0.46)). These results
suggest that multiple climate station datasets, through their spatially heterogeneous
accounting of climate patterns, may be best suited for modeling daily event timing and
magnitude, even if, as in this case, one of the climate stations (Site #1) accounts
inappropriately for the total quantity of precipitation. At a broader monthly time scale,
the results suggest that a single well-sited (i.e. negligible fetch) and representative climate
station dataset like South Farms, even if located 0.5 km outside the watershed under
study, was capable of best modeling monthly event timing and magnitude. With respect
to measures of flow duration fit, the R-NSE and R-NSE1, study results suggest that full-
suite measured climate datasets, whether using a single or multiple stations, are similarly
capable of modeling the distribution of flows through time (R-NSE (R-NSE1) ranging
from 0.77 to 0.86 (0.53 to 0.62), a negligible difference of 0.09 (0.09).
Published studies on the effect of varying quantity and quality of climate stations
in the SWAT model are not available. Studies using other models have found that
multiple climate stations for precipitation measurement reduce the error in modeled
streamflow output. Michaud and Sorooshian (1994) found that precipitation
representation errors resulting from too few climate stations accounted for approximately
half of the errors in streamflow simulation using a rainfall/runoff model. Similarly, Van
121
Werkhoven et al. (2008) concluded that representation of the spatial characteristics of
rainfall events strongly controls the value of streamflow observations in distributed
watershed models. The results of this study, however, suggest that the effects of climate
station input depend upon the topographic position of the climate station, the number of
climate stations, and the time-scale and evaluation method used (total mass fit,
hydrograph fit, and flow duration fit).
EVALUATION OF THE SWAT AUTOMATIC CALIBRATION PROCEDURE
For model users seeking to calibrate their models, this study provides support for
use of limited observed data time-series (two years) in model calibration and validation.
This finding was previously supported by Du et al. (2009), who used a discontinuous set
of 41 measures of daily streamflow for calibration and 70 measures of daily streamflow
for validation and achieved satisfactory results (NSE >0.5 for daily flow for both the
calibration and validation datasets). Similarly, based on the criteria set forth by Moriasi et
al. (2007), in this study the SSQ calibrated model performed “satisfactory” in terms of
PBIAS and “satisfactory” in terms of NSE, while the SSQ and SSQR calibrated model
performed “very good” in terms of PBIAS (PBIAS <10% for monthly flow) and “very
good” in terms of NSE (NSE >0.75 for monthly flow) at the calibrated gauging station,
Site #4. The successful values for goodness-of-fit indicators in the study by Du et al.
(2009) and in this study suggest that calibration and validation may be performed
successfully on limited time-series of observed data. The evidence supports use of time
series as limited as two years (730 days) for both calibration and validation at different
122
streamflow gauge sites as was done in this study; and 41 days for calibration and 71 days
for validation
Engel et al. (2007) (via Gan et al. 1997) stated that model calibration should
incorporate three to five years of observed data including wet, dry, and average years.
They go on to argue, however, that the amount of data necessary to calibrate a model is
project specific, indicating that shorter time-series may be appropriate. The years
included in this study were constrained by gaps in the streamflow record at the USGS
gauge (inactive between 1991 and mid-2007) and time limits on data collection at the
other four HCW gauging sites. Accordingly, this study is reflective of common
constraints in model evaluation and calibration, in which observed datasets are often
limited (Du et al. 2009; Engel et al. 2007). These constraints may further limit H/WQ
modeling efforts in the future, because, as reported by Simon et al. (2007), the number of
active USGS streamflow gauges decreased from approximately 8,500 in 1989 to
approximately 7,500 in 2006. Thus, the amount of observed streamflow data publically
available for model calibration is decreasing in the U.S.A. and future model calibration
efforts may be further limited. In regards to the HCW, however, future studies will be
able to further address H/WQ model performance during average and dry years as data
become available.
In this study, a single objective automatic calibration using the SSQ as an
objective function, and a multiple objective automatic calibration using both the SSQ and
the SSQR were compared. When van Griensven (2003) originally developed and tested
the automatic calibration method included with SWAT, the model was calibrated using
three different objective functions combined into a single global objective criterion that
123
was then optimized using the SCE-UA method (Shuffled Complex Evolution method
developed at the University of Arizona) (Duan et al. 1994): (1) the sum of the squared
error (SSQ) to minimize error in the timing and magnitude of the daily streamflow time-
series, (2) the sum of the squared error after ranking (SSQR) to minimize error in the
ranked distribution of daily streamflow, and (3) the total mass controller (TMC), a
measure of overall model bias, for the mass balance. The three objective functions were
weighted equally when combined into the single global objective criterion for
optimization.
However, only the SSQ and SSQR objective functions are made available in the
publically-distributed version of SWAT (van Griensven 2005). Furthermore, published
studies that have since evaluated the auto-calibration procedure in SWAT have limited
their optimization of streamflow parameters to a single objective function, rather than
using both available objective functions, SSQ and SSQR (Van Liew et al. 2005; Van
Liew et al. 2007; Kumar and Merwade 2009; Setegn et al. 2010).
The goodness-of-fit evaluation at Site #4 for the selected uncalibrated model, the
single objective calibrated model, and the multiple objective calibrated model indicated
that the multiple objective calibrated model performed best in five of six measures. These
measures include the daily and monthly total mass fit (PBIAS), monthly hydrograph fit
(NSE and NSE1), and daily and monthly flow duration fit (R-NSE and NSE1). The single
objective calibrated model only performed best in one of the six measures, daily
hydrograph fit (NSE and NSE1). This finding is unsurprising considering that the single
objective calibration only optimized for the sum of the squared error in daily flow, which
is represented by the daily NSE and NSE1. Furthermore, with respect to the other five of
124
six goodness-of-fit measures, the single objective calibrated model generally performed
worse than the uncalibrated model with respect to the other five of six measures. These
results provide a strong argument for use of multiple objective automatic calibration
methods for streamflow predictions including reduction of both error associated with
event timing and magnitude as well as error associated with the distribution of flows
through time. The set of validation statistics calculated on the remaining four gauging
stations show a similar pattern as the calibration statistics calculated using Site #4 only;
however, the values for each goodness-of-fit statistic are generally worse than the set of
calibration statistics, a common finding in H/WQ modeling efforts (Gassman et al. 2007).
Studies that have assessed the auto-calibration tool built into SWAT have
generally used a single objective function for streamflow calibration. Similar to the
results in this study, it was generally found that single objective function calibration
yielded poor overall model accuracy. Van Liew et al. (2005) ran the auto-calibration tool
to optimize streamflow in one scenario using only the SSQ, in another scenario using
only the SSQR, and another using only the TMC. They reported that preliminary results
from the optimization of the TMC alone were so poor that the approach was not included
in the study. They further report poor representation of the mass balance (poor PBIAS
values) when using only the SSQ or using only the SSQR for optimization. Accordingly,
published results suggest that overall model accuracy (incorporating the total mass fit,
hydrograph fit, and flow duration fit) is not well-achieved when using a single objective
function for streamflow calibration.
In this study, the selected best set of parameters found using each automatic
calibration were substantially different, particularly with respect to differences in surface
125
and ground water partitioning. The single objective calibration reduced curve numbers
(CN2) basin-wide by the maximum bound, 50%, set by the automatic calibration
procedure; the multiple objective calibration, on the other hand, increased the curve
numbers by 7%. Accordingly, the single objective calibration strongly reduced surface
runoff from storm events, while the multiple objective calibration slightly increased event
surface runoff. The GW_DELAY parameter, which represents the lag time in the vadose
zone between percolation out of the soil profile and recharge of the shallow aquifer
(Neitsch 2005), was reduced from the default value in the single objective calibration,
while it was increased in the multiple objective calibration. Thus, the single objective
calibration (SSQ objective function only) reduced the groundwater response time to
precipitation events, while the multiple objective calibration (SSQ and SSQR objective
functions) lengthened the response time of groundwater flow to precipitation events.
Similarly, Van Liew et al. (2005) found that single objective automatic calibration
resulted in sharp departures in CN2 values, ranging from -13% to -50%. The results in
this study suggest that multiple objective function automatic calibration resulted in a
more conservative change in the CN2 values of 7%.
COMPARISON OF H/WQ MODEL FIT EVALUATION METHODS
The results of the rankings of the twenty uncalibrated model configurations
according to five different goodness-of-fit statistical indicators at both a daily and
monthly time scale suggest that the selection of the optimal model configuration depends
both on the evaluation statistic being used and the temporal scale of the evaluation. The
10 different evaluation methods (daily and monthly PBIAS, daily and monthly NSE,
126
daily and monthly NSE1, daily and monthly R-NSE, and daily and monthly R-NSE1
yielded 7 different selected optimal configurations.
Four of the seven optimal model configurations were configured with 6 sub-
basins and three with 34 sub-basins. Six of the seven optimal model configurations were
configured with the low-resolution STATSGO soils data while one was configured the
high-resolution SSURGO soil data. With respect to climate input dataset, three of the
seven optimal configurations were configured with the HCW-4 climate dataset (Sites #2-
5), two configurations were configured with the South Farms dataset, and one optimal
configuration each was configured with the HCW-5 climate dataset (Sites #1-5) and the
Sanborn Field dataset, respectively. The rankings accordingly suggest that no single
statistical evaluation method can fully account for model fit.
The results of this study support the supposition that model evaluation should
incorporate measures of total mass fit, hydrograph fit, and flow duration fit, as was
conducted by van Griensven and Bauwens (2003) and suggested by Van Liew et al.
(2005). This support is provided by the different selection of best input parameter and
model configuration rankings given by each set of statistical measures across multiple
analyses in the study. Support is also provided by the better overall performance of the
multiple objective calibration (better daily and monthly PBIAS, monthly NSE (NSE1),
daily and monthly R-NSE (R-NSE1)) which incorporated a measure of daily flow
duration fit in addition to daily hydrograph fit, compared to the single objective
calibration (better daily NSE (NSE1) only), which solely accounted for hydrograph fit at
the daily scale.
127
Willmott (1984 and 1985) and others (e.g. Legates and McCabe 1999) explained
that the squared error indices inflate large-magnitude errors and minimize small-
magnitude errors, thus distorting the error between observed and modeled streamflow
through the square function as shown by Figure 21. With respect to differences in
absolute error and squared error goodness-of-fit measures, this study‟s results showed
that the absolute error measures, NSE1 and R-NSE1, produced narrower ranges of values
for the twenty uncalibrated model configurations, 0.23 and 0.50, respectively, compared
to their squared error counterparts, NSE and R-NSE, which produced ranges of 0.36 and
0.98. It is surmised here that the absolute value measures provide a more conservative
estimate of model performance; they result in lower index values for high-performing
models and higher index values for poor-performing models as was reported by Legates
and McCabe (1999). The results also suggest that model configurations may perform
equally well according to a squared error based measure while performing considerably
different using an absolute error based measure. Thus, absolute value measures may be
well-suited for determining differences in model performance that are obscured by
squared error measures.
128
Figure 21. Graphical diagram showing differences between absolute error and squared error relative to actual errors between modeled and observed data.
129
CHAPTER V
CONCLUSION
FINDINGS ON HYDROCLIMATE IN THE HINKSON CREEK WATERSHED
When excluding the Site #1 precipitation data that was affected by undercatch, the
other six climate stations used in modeling averaged 1314 mm per year (29% above
historical average) for the two-year study period 2009-2010. These unusually high levels
of precipitation during the study period resulted in perennial daily streamflow at all five
gauging stations located on Hinkson Creek during the study period, from Site #1, located
at Rogers Road, Columbia, Missouri, to Site #5, located at Scott Boulevard, Columbia,
Missouri. The lowest daily mean streamflow, 0.020 m3/s, was recorded at Site #4, the
USGS-operated gauge.
The flow duration curves and total two-year total water yields at Sites #2 and #3
suggest that further refinement is needed in the Site #2 and Site #3 rating curves. Further
refinement may be achieved by conducting regular (e.g. monthly or seasonal) cross-
sections at Sites #2 and #3 particularly during periods of low flow (less than 0.2 m3/s
discharge). Additionally, investigation of the variable backwater phenomenon at Site #5
is warranted. Investigation into backwater at Site #5 may permit a more direct
determination of the streamflow at Site #5; rather than the indirect regression modeling
with Site #4 used in this study. It is recommended that a more direct determination of
streamflow at Site #5 should be obtained through the use of a velocity-index rating curve
as described by the USGS (1982). Development of a velocity-index rating curve will
130
require installation of a continuously recording velocity meter(s) at a point in the stream
or at multiple points along a transverse line (USGS 1982). An accurate and more directly
determined estimate of streamflow at Site #5 will assist in efforts to quantify the effects
of urbanization in the City of Columbia, Missouri, on streamflow processes in the
Hinkson Creek Watershed (HCW). Such findings may have considerable impact on
future development practices and implementation of Hinkson Creek TMDL‟s (EPA
2011).
RECOMMENDATIONS FOR H/WQ MODEL CONFIGURATION
When the application for model implementation warrants prediction of monthly
values, it is recommended, based on results from this work, that the most representative
climate station (well sited; minimal fetch problem) in or near the watershed be used to
both produce better modeled predictions and also reduce time needed for model
configuration (pre-processing of climate data input files). In the HCW, the South Farms
climate station provided the best overall monthly streamflow predictions in terms of
hydrograph fit (NSE (NSE1)=0.66 (0.46)). When daily estimates are needed, the results of
this study suggest that multiple station climate input is needed for optimal model
performance. Both multiple climate station datasets performed better than all other
climate datasets with respect to hydrograph fit. The HCW-5 climate dataset performed
the best, surprisingly, with an NSE (NSE1)=0.15 (0.29), followed by the HCW-4 climate
dataset with a NSE (NSE1)=0.13 (0.18). The better performance of the HCW-5 over the
HCW-4 climate dataset at the daily scale suggest that better representation of climate
131
heterogeneity, despite precipitation undercatch at one of five climate stations, is
necessary for accurate simulation at fine time scales (i.e. daily).
With respect to watershed subdivision and soil dataset resolution, this study
showed that increased watershed subdivision and soil resolution do not considerably
affect model performance with respect to total mass fit, hydrograph fit, and flow duration
fit. These conclusions are confirmed by a re-analysis of the data in Kumar and Merwade
(2009). Accordingly, to reduce time and effort required for model configuration and
reduce simulation run-time, the lowest spatial watershed discretization and soil resolution
required by the intended modeling application, is recommended. This recommendation is
a particularly important consideration when running automatic calibration routines, which
may require 100s to 1000s of model simulations.
RECOMMENDATIONS FOR H/WQ MODEL CALIBRATION
Based on the proposed standard criteria (Moriasi et al. 2007) for judging H/WQ
model performance on a monthly time scale using Percent Bias (PBIAS) and the Nash-
Sutcliffe Efficiency (NSE), 16 of the 20 uncalibrated model configurations performed
„satisfactorily‟ (PBIAS < 25%; NSE > 0.5). All model configuration using full-suite
observed climate data (i.e. precipitation, air temperature, solar radiation, wind speed, and
relative humidity) met this criteria. Therefore, it is concluded that SWAT is capable of
performing at satisfactory levels with respect to total mass fit and hydrograph fit without
calibration in the HCW and potentially other urbanizing watersheds in the Central U.S.A.
It was shown in this study that the built-in automatic calibration included with
SWAT is capable of running successfully in less than 24 hours on a single desktop-class
132
computer, lending support for the wider use of this technology in management
applications. Care must be taken to constrain the automatic calibration routine to
optimize a limited set of parameters (six parameters in this study). In addition, the study
strongly supports the use of multiple objective automatic calibration methods (SSQ and
SSQR) over single objective methods (SSQ only) to ensure optimal model performance
across the evaluation suite of total mass fit, hydrograph fit, and flow duration fit. The
multiple objective calibrated model better predicted monthly flow (PBIAS=0.75%, NSE
(NSE1)=0.81 (0.59), R-NSE (R-NSE1)=0.89 (0.73)) than the single-objective calibrated
model (PBIAS=16.08%, NSE (NSE1)=0.54 (0.40); R-NSE (R-NSE1) = 0.82 (0.67).
Contrary to the conclusions of Van Liew et al. (2005) and Kumar and Merwade
(2009), who recommend use of single objective calibration (SSQ) combined with manual
calibration to correct for poor estimation of the total mass and flow distribution, it is
recommended here that multiple objective streamflow calibration be used to optimize
model performance. Manual calibration is not recommended due to excessive labor
requirements (e.g. Van Liew et al (2005) reported four to six weeks of expert labor
required for manual SWAT calibration).
It is further recommended that the SWAT development team incorporate the Total
Mass Controller (TMC) objective function into the next version of the SWAT model. The
TMC was originally included in Enhanced SWAT (ESWAT), the proprietary version of
SWAT developed by van Griensven and Bauwens (2003). The enhancements made to
SWAT in ESWAT were later incorporated into SWAT 2005, the version of SWAT used
in this study. However, the TMC objective function, which directly accounts for error in
the mass balance, was surprisingly not included in SWAT2005. It is expected that
133
inclusion of the TMC objective function will further improve SWAT multiple objective
automatic calibration by enabling simultaneous streamflow parameter optimization of the
sum of the squared error (SSQ), the sum of the squared error after ranking (SSQR), and
the TMC as was done by van Griensven and Bauwens (2003).
RECOMMENDATIONS FOR H/WQ MODEL EVALUATION
In the modeling literature, primary emphasis has been placed on optimizing the
Nash Sutcliffe Efficiency (Nash and Sutcliffe 1970). Results of this study suggest that the
choice of goodness-of-fit objective function or statistical indicator can have considerable
effect on the model development process. Thus, more effective assessment of model
accuracy for research and management applications would be served by accounting for
measures of error in the mass balance (total mass), the hydrograph (event timing and
magnitude), and the flow duration curve (percent of time flow exceeded) as was
presented in this study and was done by van Griensven and Bauwens (2003). As
suggested by Van Liew et al. (2005), optimization should be designed to take into
account model fit with the mass balance, the hydrograph, and the flow duration curve,
allowing development of models that best represent the needs determined by the intended
application, e.g., short-duration storm runoff analysis, low flow assessments, and water
availability evaluations. What Van Liew et al. (2005) ignored was that this capability is
already partially included in the SWAT automatic calibration tool (SSQ and SSQR
objective functions may be run optimized simultaneously), and the software code for
including a third objective function, the TMC, was previously written for use in ESWAT.
134
The results of this study also provide an argument for the consideration of
absolute-error-based measures such as the Modified Nash-Sutcliffe Efficiency introduced
by Legates and McCabe (1999). In addition to the strong theoretical argument for the use
of absolute error based goodness-of-fit statistics presented in the literature (Willmott
1984, 1985; Legates and McCabe 1999), this study demonstrated that the choice of NSE
over NSE1, for example, can change the selection of the optimal model configuration, and
consequently, the implementation of H/WQ models. Non-optimal selection of a model
configuration or a set of optimized parameters may result in less accurate H/WQ models.
Reduced model accuracy may lead to poor management decision-making. For example,
the U.S. Department of Agriculture‟s Conservation Effects Assessment Program (CEAP),
which is using the SWAT model to account for the environmental impact of U.S.
Department of Agriculture (USDA) policies, may provide improper estimates of
environmental impact and thus result in inefficient future public policy.
RECOMMENDATIONS FOR FUTURE RESEARCH IN H/WQ MODELING
Continuing data-intensive, multi-site monitoring efforts like the ongoing effort in
the nested HCW will provide opportunities to further evaluate SWAT and other H/WQ
models for simulation of water pollutant transport (e.g. suspended sediment and
nutrients), and to test in-depth multi-site (multiple stream gauges), multi-criteria
(streamflow, sediment, nutrients) (Vazquez 2008), and more-detailed multi-scalar (i.e.
yearly, monthly, and daily) model calibration approaches. Furthermore, many more
multi-configuration modeling studies, (an approach taken here, by Kumar and Merwade
(2009), by Di Luzio et al. (2005), and by Ye et al. (2011)) need to be carried out and
135
published in the scientific literature. It is also recommend that further testing of the four
new goodness-of-fit criteria, introduced in this study, be conducted to further test
appropriateness for use in H/WQ model evaluation.
Finally, the scientific community is encouraged to take into account real-world
limitations natural resource managers (e.g. labor, time, and computational power) must
face when conducting modeling research as was done in this study. As noted by van
Griensven, H/WQ models are in need of further development to reach their potential to
be simple and useful tools for water resource management decision-making.
CONTRIBUTIONS TO SCIENCE
In this thesis, a practical approach to model development (configuration,
calibration, validation, and evaluation methods) is presented. This thesis represents the
first (to the author‟s knowledge) published H/WQ model study to present a
comprehensive multi-configuration, multi-site, and multi-objective approach in SWAT
model implementation. Four introduced goodness-of-fit indicators were introduced to
measure the model fit with the distribution of flows in time (analogous to the flow
duration curve); these new goodness-of-fit indicators are inspired by the work of van
Griensven and Bauwens (2003). A novel model performance ranking method was used to
determine the effect of input resolution, including watershed discretization spatial
resolution, soil dataset spatial resolution, and climate station quantity and quality, and the
effect of temporal scale (daily vs. monthly evaluation). In short, we find that some
applications of H/WQ modeling, particularly daily simulations, will require multiple
climate station input. In addition, multiple objective automatic calibration, which is
136
shown in this study to improve SWAT model performance, requires observed streamflow
data to optimize the model. It is thus recommended that public policy decision makers
and natural resource managers increase investment in hydroclimatic data collection to
help develop and implement models (tools) that can assist in managing many pressing
water resource problems throughout the world.
137
LITERATURE CITED
Ahl, R.S., S.W. Woods, and H.R. Zuuring. 2008. Hydrologic calibration and validation of
SWAT in a snow-dominated Rocky Mountain watershed, Montana, U.S.A.
Journal of the American Water Resources Association 44:1-21.
Arabi, M., R.S. Govindaraju, M.M. Hantush, and B.E. Engel. 2006. Role of watershed
subdivision on modeling the effectiveness of best management practices with
SWAT. Journal of the American Water Resources Association 2: 513-528
ASCE. 1993. Criteria for evaluation of watershed models. Journal of Irrigation and
Drainage Engineering 119:429-442
Bae, D-H, I-W. Jung, and D.P. Lettenmaier. In Press, accepted 11 Feb 2011. Hydrologic
uncertainties in climate change from IPCC AR4 GCM simulations of the Chungju
Basin, Korea. Journal of Hydrology DOI: 10.1016/j.jhydrol.2011.02.012
Benham, B.L., C. Baffaut, R.W. Zeckowski, K.R. Mankin, Y.A. Pachepsky, A.M.
Sadeghi, K.M. Brannan, M.L. Soupir, and M.J. Habersack. 2006. Modeling
bacteria fate and transport in watershed models to support TMDLs. Transactions
of the American Society of Agricultural and Biological Engineers 49:987-1002
Borah, D.K. and M. Bera. 2003. Watershed-scale hydrologic and nonpoint-source
pollution models: Review of mathematical bases. Transactions of the American
Society of Agricultural Engineers 46:1553-1566
Borah, D.K. and M. Bera. 2004. Watershed-scale hydrologic and nonpoint-source
pollution models: Review of applications. Transactions of the American Society
of Agricultural Engineers 47:789-803.
Bougeard, M., J. Le Saux, N. Perenne, C. Baffaut, M. Robin, M. Pommepuy. 2011.
Modeling of Escherichia coli fluxes on a catchment and the impact on coastal
water and shellfish quality. Journal of the American Water Resources Association
DOI:10.1111/j.1752-1688.2011.00520.x
Chapman, S.S., J.M. Omernik, G.E. Griffith, W.A. Schroeder, T.A. Nigh, and T.F.
Wilton. 2002. Ecoregions of Iowa and Missouri (color poster with map,
descriptive text, summary tables, and photographs): Reston, Virginia, U.S.
Geological Survey (map scale 1:1,800,000).
138
Chaubey, I., C.T. Haan, S. Grunwald, and J.M. Salisbury. Uncertainty in the model
parameters due to spatial variability of rainfall. 1999. Journal of Hydrology 220:
48-61
Di Luzio, M., J.G. Arnold, and R. Srinivasan. 2005. Effect of GIS data quality on small
watershed stream flow and sediment simulations. Hydrological Processes 19: 629-
650
Di Luzio, M., R. Srinivasan, and J.G. Arnold. 2004. A GIS-coupled hydrological model
system for the watershed assessment of agricultural nonpoint and point sources of
pollution. Transactions in GIS 8: 113-136
Di Luzio, M., R. Srinivasan, J.G. Arnold and S.L. Neitsch. 2002. ArcView Interface for
SWAT2000 User‟s Guide. USDA-ARS Grassland, Soil, and Water Research
Laboratory, Temple, Texas.
Dingman, S.L. 2002. Physical Hydrology, 2nd
Edition. Waveland Press, Inc. Long Grove,
Illinois. 646 pp.
Du, B., X. Ji, R.D. Harmel, and L.M. Hauck. 2009. Evaluation of a watershed model for
estimating daily flow using limited flow measurements. Journal of the American
Water Resources Association 45: 475-484
Duan, Q., S. Sorooshian, and V.K. Gupta. 1994. Optimal use of the SCE-UA global
optimization method for calibrating watershed models. Journal of Hydrology 158:
265-284
Eckhardt, K. and J.G. Arnold. 2001. Automatic calibration of a distributed catchment
model. Journal of Hydrology 251: 103-109
Engel, B., D. Storm, M. White, J. Arnold and M. Arabi. 2007. A hydrologic/water quality
model application protocol. Journal of the American Water Resources Association
43:1223-1236
Gassman, P.W., M.R. Reyes, C.H. Green, and J.G. Arnold. 2007. The Soil and Water
Assessment Tool: Historical development, applications, and future research
directions. Transactions of the American Society of Agricultural and Biological
Engineers 50:1211-1250
Ghaffari, G., S. Keesstra, J. Ghodousi, and H. Ahmadi. 2010. SWAT-simulated
hydrological impact of land-use change in the Zanjanrood Basin, Northwest Iran.
Hydrological Processes 24: 892-903
Harmel, R.D., R.J. Cooper, R.M. Slade, R.L. Haney, and J.G. Arnold. 2006. Cumulative
uncertainty in measured streamflow and water quality data for small watersheds.
139
Transactions of the American Society of Agricultural and Biological Engineers
49:689-701.
Harmel, R.D. and P.K. Smith. 2007. Consideration of measurement uncertainty in the
evaluation of goodness-of-fit in hydrologic and water quality modeling. Journal of
Hydrology 337:326-336
Im, S., K. Brannan, S. Mostaghimi, and J. Cho. 2003. A comparison of SWAT and HSPF
models for simulating hydrologic and water quality responses from an urbanizing
watershed. An American Society of Agricultural Engineers Meeting Presentation
- Paper Number 032175.
Kumar, S. and V. Merwade. 2009. Impact of watershed subdivision and soil data
resolution on SWAT model calibration and parameter uncertainty. Journal of the
American Water Resources Association 45: 1179-1196
Legates, D.R., and G.J. McCabe. 1999. Evaluating the use of “goodness-of-fit” measures
in hydrologic and hydroclimatic model validation. Water Resources Research
35:233-241
Li, Z., Z. Xu, and Z. Li. 2010. Performance of WASMOD and SWAT on hydrological
simulation in Yingluoxia watershed in northwest of China. Hydrological
Processes DOI:10.1002/hyp.7823
MDNR. 2006. Stream Survey Sampling Report. Phase III Hinkson Creek Stream Study,
Columbia, Missouri, Boone County. Prepared by the Missouri Department of
Natural Resources, Field Services Division, Environmental Services Program,
Water Quality Monitoring Section.
Mehta, V.M., N.J. Rosenberg, and K. Mendoza. 2011. Simulated impacts of three decadal
climate variability phenomena on water yields in the Missouri River Basin.
Journal of the American Water Resources Association 47: 126-135
Merritt, W.S., R.A. Letcher, and A.J. Jakeman. 2003. A review of erosion and sediment
transport models. Environmental Modeling and Software 18:761-799
Michaud, J. and S. Sorooshian. 1994. Comparison of simple versus complex distributed
runoff models on a midsized semiarid watershed. Water Resources Research 30:
593-605
Migliaccio, K.W. and P. Srivastava. 2007. Hydrologic components of watershed-scale
models. Transactions of the American Society of Agricultural and Biological
Engineers 50:1695-1703
Moriasi, D.N., J.G. Arnold, M.W. Van Liew, R.L. Binger, R.D. Harmel, and T.L. Veith.
2007. Model evaluation guidelines for systematic quantification of accuracy in
140
watershed simulations. Transactions of the American Society of Agricultural and
Biological Engineers 50:885-900
Muleta, M.K. and J.W. Nicklow. 2005. Sensitivity and uncertainty analysis coupled with
automatic calibration for a distributed watershed model. Journal of Hydrology
306:127-145
Nash, J.E. and J.V. Sutcliffe. 1970. River flow forecasting through conceptual models.
Part 1-A: Discussion of principles. Journal of Hydrology 10:282-290
Neitsch, S.L., J.G. Arnold, J.R Kiniry, R. Srinivasan, and J.R. Williams. 2004. Soil and
Water Assessment Tool Input/Output File Documentation, Version 2005. USDA-
ARS Grassland, Soil, and Water Research Laboratory, Temple, Texas.
Neitsch, S.L., J.G. Arnold, J.R Kiniry, and J.R. Williams, 2005. Soil and Water
Assessment Tool Theoretical Documentation, Version 2005. USDA-ARS
Grassland, Soil, and Water Research Laboratory, Temple, Texas.
Nelder, J.A., and R. Mead. 1965. A simplex method for function minimization. Computer
Journal 7: 308-313
Nigh, T.A. and W.A. Schroeder. 2002. Atlas of Missouri Ecological Sections. Missouri
Department of Conservation.
O‟Donnell, T.K., C. Baffaut, and D.L. Galat. 2008. Predicting effects of best
management practices on sediment loads to improve management in the Midwest,
USA. International Journal of River Basin Management 6: 243-256
O‟Neill, M.P., R. Butts, D.A. Bucks, M.A. Weltz, and K.R. Hinga. 2006. Addressing
agricultural water resources issues in North America: The role of research and
education. Tri-National Initiative on Environmentally Sustainable Agriculture and
Water Quality, International Institute for Sustainable Development.
Parajuli, P.B. 2010. Assessing sensitivity of hydrologic responses to climate change from
forested watershed in Mississippi. Hydrological Processes 24: 3785-3797
Perkins, B. 1995. Temporal variability in a Midwestern stream during spring. Master‟s
Thesis, Graduate School, University of Missouri – Columbia.
Pflieger, W.L. 1989. Aquatic Community Classification System for Missouri. Missouri
Department of Conservation. Jefferson City, Missouri. Aquatic Series No. 19, 70
pp.
Rachman, A., S.H. Anderson, E.E. Alberts, A.L. Thompson, and C.J. Gantzer. 2008.
Predicting runoff and sediment yield from a stiff-stemmed grass hedge system for
141
a small watershed. Transactions of the American Society of Agricultural and
Biological Engineers 51: 425-432
Radcliffe, D.E., Z. Lin, L.M. Risse, J.J. Romeis, and C.R. Jackson. 2009. Modeling
phosphorus in the Lake Allatoona Watershed using SWAT: I. Developing
phosphorus parameter values. Journal of Environmental Quality 38: 111-120
Rickett, H.W. 1931. Notes on the Vegetation of Columbia, Mo. American Midland
Naturalist, 12(10): 411-419
Santhi, C., J.G. Arnold, J.R. Williams, W.A. Dugas, R. Srinivasan, and L.M. Hauck.
2001. Validation of the SWAT model on a large river basin with point and
nonpoint sources. Journal of the American Water Resources Association 37:1169-
1188
Santhi, C., R. Srinivasan, J.G. Arnold, and J.R. Williams. 2006. A modeling approach to
evaluate the impacts of water quality management plans implemented in a
watershed in Texas. Environmental Modelling and Software 21:1141-1157
Setegn, S.G., R. Srinivasan, A.M. Melesse, and B. Dargahi. 2010. SWAT model
application and prediction uncertainty analysis in the Lake Tana Basin, Ethiopia.
Hydrological Processes 24: 357-367
Sivapragasam, C. and N. Muttil. 2005. Discharge Rating Curve Extension – A New
Approach. Water Resources Management 19: 505-520
Srinivasan, R., J.G. Arnold, and C.A. Jones. 1998. Hydrologic modeling of the United
States with the Soil and Water Assessment Tool. Water Resources Development
14: 315-325
Srinivasan, R., T.S. Ramanarayanan, J.G. Arnold, and S.T. Bednarz. 1998. Large Area
Hydrologic Modeling, Part II: Model Application. Journal of the American Water
Resources Association 37: 91-101
Srivastava, P., K.W. Migliaccio and J. Simunek. 2007. Landscape models for simulating
water quality at point, field, and watershed scales. Transactions of the American
Society of Agricultural and Biological Engineers 50: 1683-1693.
Stone, M.C., R.H. Hotchkiss, C.M. Hubbard, T.A. Fontaine, L.O. Mearns, and J.G.
Arnold. 2001. Impacts of climate change on Missouri river basin water yield.
Journal of the American Water Resources Association 37: 1119-1129
Thom, R.H. and J.H. Wilson. 1980. The Natural Division of Missouri. Transactions,
Missouri Academy of Sciences. Vol. 14, 23 pp.
142
USACE. 1999. Review of watershed water quality models. USACE Technical Report W-
99-1.
USEPA. 2006. Wadeable streams assessment: A collaborative survey of the nation‟s
streams. EPA Report 841-B-06-002.
USEPA. 2008. Overview of Impaired Waters and Total Maximum Daily Loads Program.
Accessed May 28, 2009. http://www.epa.gov/owow/tmdl/intro.html
USEPA. 2011. Total Maximum Daily Load (TMDL) for Hinkson Creek. EPA Report
number not assigned.
Van Griensven, A. 2002. Untitled Ph.D. Dissertation. Free University of Brussels,
Belgium. 236 pp.
Van Griensven, A. and W. Bauwens. 2003. Multiobjective autocalibration for
semidistributed water quality models. Water Resources Research 39: 1348
Van Griensven, A, A. Francos, and W. Bauwens. 2002. Sensitivity analysis and auto-
calibration of an integral dynamic model for river water quality. Water Science
and Technology 45: 325-332
Van Griensven, A. 2005. Sensitivity, auto-calibration, uncertainty and model evaluation
in SWAT2005. Draft Report. USDA-ARS Grassland, Soil, and Water Research
Laboratory, Temple, Texas. 48 pp.
Van Liew, M.W., J.G. Arnold, and D.D. Bosch. 2005. Problems and potential of
autocalibrating a hydrologic model. Transactions of the American Society of
Agricultural Engineers 48: 1025-1040
Van Werkhoven, K., T. Wagener, P. Reed, and Y. Tang. 2008. Rainfall characteristics
define the value of streamflow observations for distributed watershed model
identification. Geophysical Research Letters 35
White, K.L. and I. Chaubey. 2005. Sensitivity analysis, calibration, and validations for a
multisite and multivariable SWAT model. Journal of the American Water
Resources Association 41: 1077-1089
White, M.J., D.E. Storm, M.D. Smolen and H. Zhang. 2009. Development of a
quantitative pasture phosphorus management tool using the SWAT model.
Journal of the American Water Resources Association 45:397-406
Willmott, C.J. 1981. On the validation of models. Physical Geography 2: 184-194
143
Willmott, C.J. 1984. On the evaluation of model performance in physical geography. In
Spatial Statistics and Models, pp. 443-460. G.L. Gaile and C.J. Willmott, eds.
Boston, Mass.: D. Reidel Publishing.
Willmott, C.J., S.G. Ackleson, R.E. Davis, J.J. Feddema, K.M. Klink, D.R. Legates, J.
O‟Donnell, and C.M. Rowe. 1985. Statistics for the Evaluation and Comparison
of Models. Journal of Geophysical Research 90: 8995-9005
Winchell, M., R. Srinivasan, M. Di Luzio, and J. Arnold. 2009. ArcSWAT 2.3 Interface
for SWAT2005 User‟s Guide. USDA-ARS Grassland, Soil, and Water Research
Laboratory, Temple, Texas. 465 pp.
Ye, X., Q. Zhang, and N.R. Viney. 2011. The effect of soil data resolution on
hydrological process modeling in a large humid watershed. Hydrological
Processes 25: 130-140
Zhang, Z., R. Srinivasan, and M. Van Liew. 2008. Multi-site calibration of the SWAT
model for hydrologic modeling. Transactions of the American Society of
Agricultural and Biological Engineers 51: 2039-2049
Zhang, X., R. Srinivasan, K. Zhao, and M. Van Liew. 2009. Evaluation of global
optimization algorithms for parameter calibration of a computationally intensive
hydrologic model. Hydrological Processes 23: 430-441
Zhang, X., R. Srinivasan, J. Arnold, R.C. Izaurralde, and D. Bosch. 2011. Simultaneous
calibration of surface flow and baseflow simulations: a revisit of the SWAT
automatic calibration framework. Hydrological Processes DOI:10.1002/hyp.8058
APPENDIX A: STAGE-DISCHARGE RATING CURVES
Table 26. Equations for rating curves applied in the study in Hinkson Creek Watershed, Missouri, U.S.A. The „y‟ variable is equivalent to discharge in m
3/s; the „x‟ variable is equivalent to stage in m.
HCW
Gauge Site
Linear Rating Polynomial Rating Power Rating
Stage (m) Equation Stage (m) Equation Equation Stage (m)
HCW #1 0 - 0.5523 y = 0.0472x 0.5523 - max y = 12.794x3-17.502x
2+10.57x-2.6548 N/A N/A
HCW #2 0 - 0.464 y = 0.4163x 0.464 - max y = 4.3007x3+21.944x
2-20.288x+4.4526 N/A N/A
HCW #3 0 - 0.11976 y = 1.3590180x 0.11976 - 0.363 y = 17.571x3-7.0615x
2+3.94x-0.238 0.363 - max y = 15.061x
2.5797
HCW #5 N/A N/A N/A N/A entire range y = 5.1365x1.8118
144
145
Figure 22. Detailed plot of Site #1 rating curve and flow measurements.
146
Figure 23. Detailed plot of Site #2 rating curve and flow measurements.
147
Figure 24. Detailed plot of Site #3 rating curve and flow measurements.
148
Figure 25. Detailed plot of Site #5 rating curve and flow measurements.
149
APPENDIX B: COMPLETE MODELING RESULTS
Table 27. Full daily streamflow descriptive and goodness-of-fit statistics for all uncalibrated SWAT model configurations. Average Percent Bias is the average of the absolute values of each Percent Bias measure.
Statistic Observed Daily Streamflow Data
Site #1 Site #2 Site #3 Site #4 Site #5
MIN (m3/s) 0.024 0.115 0.046 0.020 0.099
MEAN (m3/s) 1.184 2.105 2.066 3.516 4.540
MEDIAN (m3/s) 0.196 0.178 0.362 0.510 0.708
MAX (m3/s) 77.518 185.347 115.218 177.830 141.523
Statistic 6 Subbasin - STATSGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.013 0.015 0.011 0.026 0.015
MEAN (m3/s) 0.854 1.195 1.412 2.347 2.749
MEDIAN (m3/s) 0.357 0.547 0.647 1.196 1.378
MAX (m3/s) 29.590 30.830 31.770 39.700 50.630
PBIAS (%) 27.894 43.245 31.636 33.257 39.447 35.096
RMSE (m3/s) 5.583 10.688 8.250 13.294 15.195
MAE (m3/s) 1.602 2.723 2.644 4.458 5.595
NSE -0.285 -0.088 -0.169 -0.145 -0.178 -0.173
NSE1 0.009 0.137 0.060 0.070 0.110 0.078
R-RMSE (m3/s) 2.670 7.958 5.021 8.700 9.186
R-MAE (m3/s) 0.471 1.293 0.928 1.789 2.402
R-NSE (AVG) 0.741 0.420 0.603 0.536 0.601 0.580
R-NSE (MED) 0.751 0.440 0.622 0.562 0.629 0.601
R-NSE1 (AVG) 0.709 0.590 0.670 0.627 0.618 0.643
150
Statistic 6 Subbasin - STATSGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.079 0.099 0.107 0.173 0.205
MEAN (m3/s) 1.734 2.262 2.594 4.060 4.666
MEDIAN (m3/s) 0.815 1.037 1.182 1.901 2.141
MAX (m3/s) 73.680 98.950 115.200 181.800 211.400
PBIAS (%) -46.444 -7.466 -25.588 -15.453 -2.766 19.543
RMSE (m3/s) 4.639 8.947 7.239 11.001 15.459
MAE (m3/s) 1.505 2.473 2.402 3.624 5.397
NSE 0.113 0.237 0.100 0.216 -0.219 0.090
NSE1 0.070 0.217 0.146 0.244 0.142 0.164
R-RMSE (m3/s) 1.828 5.942 2.448 3.933 3.876
R-MAE (m3/s) 0.651 0.986 0.717 1.155 1.744
R-NSE (AVG) 0.931 0.799 0.960 0.963 0.909 0.912
R-NSE (MED) 0.934 0.806 0.962 0.965 0.915 0.916
R-NSE1 (AVG) 0.597 0.688 0.745 0.759 0.723 0.702
Statistic 6 Subbasin - STATSGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.076 0.093 0.102 0.174 0.225
MEAN (m3/s) 1.680 2.196 2.531 3.976 4.621
MEDIAN (m3/s) 0.786 1.004 1.150 1.867 2.141
MAX (m3/s) 64.110 86.150 100.600 158.700 185.500
PBIAS (%) -41.841 -4.328 -22.525 -13.062 -1.778 16.707
RMSE (m3/s) 4.771 9.134 7.432 11.319 15.673
MAE (m3/s) 1.439 2.424 2.325 3.564 5.343
NSE 0.062 0.205 0.051 0.170 -0.253 0.047
NSE1 0.110 0.232 0.173 0.257 0.150 0.185
R-RMSE (m3/s) 1.371 5.317 1.845 3.240 4.522
R-MAE (m3/s) 0.573 0.991 0.687 1.304 1.971
R-NSE (AVG) 0.945 0.794 0.961 0.952 0.884 0.907
R-NSE (MED) 0.947 0.801 0.963 0.955 0.892 0.912
R-NSE1 (AVG) 0.646 0.686 0.756 0.728 0.686 0.700
151
Statistic 6 Subbasin - STATSGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.008 0.004 0.031 0.175 0.219
MEAN (m3/s) 0.949 1.239 1.553 3.061 3.732
MEDIAN (m3/s) 0.315 0.394 0.520 1.476 1.817
MAX (m3/s) 67.170 90.160 106.300 167.300 196.800
PBIAS (%) 19.913 41.127 24.808 12.935 17.804 23.317
RMSE (m3/s) 4.346 8.902 6.797 10.461 14.440
MAE (m3/s) 1.186 2.052 1.964 3.152 4.772
NSE 0.222 0.245 0.206 0.291 -0.064 0.180
NSE1 0.267 0.350 0.302 0.343 0.241 0.300
R-RMSE (m3/s) 2.356 6.602 3.667 5.399 6.150
R-MAE (m3/s) 0.454 1.245 0.827 1.427 2.117
R-NSE (AVG) 0.902 0.725 0.886 0.905 0.831 0.850
R-NSE (MED) 0.906 0.735 0.891 0.911 0.842 0.857
R-NSE1 (AVG) 0.719 0.606 0.706 0.702 0.663 0.679
Statistic 6 Subbasin - STATSGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.077 0.100 0.112 0.181 0.216
MEAN (m3/s) 1.607 2.101 2.415 3.923 4.593
MEDIAN (m3/s) 0.801 1.013 1.150 1.985 2.279
MAX (m3/s) 75.940 101.900 118.000 179.000 208.500
PBIAS (%) -35.723 0.196 -16.892 -11.561 -1.160 13.106
RMSE (m3/s) 4.425 8.747 6.894 10.598 14.992
MAE (m3/s) 1.436 2.345 2.288 3.506 5.257
NSE 0.193 0.271 0.184 0.272 -0.147 0.155
NSE1 0.112 0.257 0.187 0.269 0.164 0.198
R-RMSE (m3/s) 1.820 6.049 2.530 4.104 4.237
R-MAE (m3/s) 0.533 1.013 0.685 1.205 1.932
R-NSE (AVG) 0.947 0.798 0.968 0.963 0.897 0.915
R-NSE (MED) 0.949 0.805 0.969 0.965 0.904 0.919
R-NSE1 (AVG) 0.670 0.679 0.756 0.749 0.693 0.709
152
Statistic 6 Subbasin - SSURGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.036 0.042 0.046 0.074 0.062
MEAN (m3/s) 0.868 1.200 1.415 2.356 2.821
MEDIAN (m3/s) 0.444 0.723 0.868 1.437 1.648
MAX (m3/s) 29.110 30.180 30.980 33.560 45.160
PBIAS (%) 26.706 42.992 31.483 32.985 37.872 34.408
RMSE (m3/s) 5.416 10.573 8.074 13.068 14.984
MAE (m3/s) 1.573 2.681 2.584 4.365 5.526
NSE -0.209 -0.065 -0.120 -0.106 -0.145 -0.129
NSE1 0.028 0.151 0.081 0.090 0.121 0.094
R-RMSE (m3/s) 3.098 8.373 5.516 9.106 9.536
R-MAE (m3/s) 0.613 1.500 1.181 2.111 2.702
R-NSE (AVG) 0.671 0.378 0.532 0.482 0.562 0.525
R-NSE (MED) 0.684 0.399 0.554 0.510 0.593 0.548
R-NSE1 (AVG) 0.621 0.525 0.580 0.560 0.570 0.571
Statistic 6 Subbasin - STATSGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.053 0.063 0.067 0.093 0.090
MEAN (m3/s) 1.732 2.261 2.596 4.087 4.743
MEDIAN (m3/s) 1.052 1.342 1.520 2.353 2.552
MAX (m3/s) 73.210 97.510 113.500 180.100 212.500
PBIAS (%) -46.227 -7.404 -25.663 -16.222 -4.456 19.995
RMSE (m3/s) 4.379 8.739 6.806 10.543 15.035
MAE (m3/s) 1.529 2.496 2.418 3.652 5.384
NSE 0.209 0.273 0.204 0.280 -0.153 0.163
NSE1 0.055 0.209 0.140 0.238 0.144 0.157
R-RMSE (m3/s) 1.927 6.209 2.782 4.567 4.681
R-MAE (m3/s) 0.703 1.337 0.978 1.616 2.274
R-NSE (AVG) 0.938 0.775 0.956 0.952 0.877 0.900
R-NSE (MED) 0.941 0.783 0.958 0.955 0.886 0.905
R-NSE1 (AVG) 0.565 0.576 0.652 0.663 0.638 0.619
153
Statistic 6 Subbasin - SSURGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.047 0.056 0.061 0.087 0.091
MEAN (m3/s) 1.692 2.212 2.551 4.025 4.721
MEDIAN (m3/s) 0.992 1.278 1.467 2.280 2.552
MAX (m3/s) 64.340 85.560 99.770 158.300 187.800
PBIAS (%) -42.825 -5.116 -23.475 -14.476 -3.978 17.974
RMSE (m3/s) 4.575 8.981 7.086 10.999 15.394
MAE (m3/s) 1.480 2.468 2.365 3.621 5.369
NSE 0.137 0.232 0.138 0.216 -0.209 0.103
NSE1 0.085 0.218 0.159 0.245 0.146 0.171
R-RMSE (m3/s) 1.398 5.509 2.089 3.607 4.907
R-MAE (m3/s) 0.620 1.298 0.987 1.716 2.362
R-NSE (AVG) 0.963 0.781 0.965 0.949 0.869 0.906
R-NSE (MED) 0.965 0.789 0.967 0.952 0.878 0.910
R-NSE1 (AVG) 0.617 0.589 0.649 0.642 0.624 0.624
Statistic 6 Subbasin - SSURGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.004 0.000 0.030 0.106 0.113
MEAN (m3/s) 0.964 1.259 1.576 3.116 3.825
MEDIAN (m3/s) 0.439 0.559 0.766 1.775 2.065
MAX (m3/s) 70.970 94.390 110.300 171.300 203.400
PBIAS (%) 18.571 40.185 23.700 11.384 15.759 21.920
RMSE (m3/s) 4.422 8.959 6.858 10.468 14.441
MAE (m3/s) 1.207 2.070 1.993 3.195 4.771
NSE 0.194 0.235 0.192 0.290 -0.064 0.170
NSE1 0.254 0.344 0.292 0.334 0.241 0.293
R-RMSE (m3/s) 2.546 6.811 4.034 5.980 6.713
R-MAE (m3/s) 0.547 1.357 0.999 1.761 2.424
R-NSE (AVG) 0.876 0.715 0.848 0.873 0.791 0.821
R-NSE (MED) 0.881 0.725 0.855 0.880 0.806 0.829
R-NSE1 (AVG) 0.662 0.570 0.645 0.633 0.614 0.625
154
Statistic 6 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.060 0.071 0.078 0.105 0.111
MEAN (m3/s) 1.628 2.128 2.445 3.985 4.693
MEDIAN (m3/s) 1.009 1.298 1.489 2.377 2.667
MAX (m3/s) 79.430 105.500 121.400 182.400 214.500
PBIAS (%) -37.491 -1.093 -18.354 -13.319 -3.371 14.726
RMSE (m3/s) 4.327 8.653 6.705 10.357 14.785
MAE (m3/s) 1.479 2.397 2.324 3.555 5.268
NSE 0.228 0.287 0.228 0.305 -0.115 0.187
NSE1 0.086 0.241 0.174 0.259 0.162 0.184
R-RMSE (m3/s) 1.909 6.289 2.856 4.782 5.083
R-MAE (m3/s) 0.638 1.355 0.997 1.720 2.465
R-NSE (AVG) 0.945 0.783 0.951 0.943 0.855 0.896
R-NSE (MED) 0.947 0.791 0.954 0.946 0.865 0.901
R-NSE1 (AVG) 0.606 0.571 0.646 0.641 0.608 0.614
Statistic 34 Subbasin - STATSGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.002 0.001 0.000 0.004 0.000
MEAN (m3/s) 0.892 1.232 1.450 2.383 2.792
MEDIAN (m3/s) 0.388 0.532 0.659 1.193 1.328
MAX (m3/s) 26.340 27.580 28.520 46.710 58.370
PBIAS (%) 24.672 41.462 29.819 32.227 38.499 33.336
RMSE (m3/s) 5.445 10.662 8.256 13.416 15.382
MAE (m3/s) 1.613 2.750 2.678 4.501 5.655
NSE -0.222 -0.083 -0.171 -0.166 -0.207 -0.170
NSE1 0.003 0.129 0.048 0.061 0.101 0.068
R-RMSE (m3/s) 2.943 8.144 5.111 8.511 8.765
R-MAE (m3/s) 0.470 1.262 0.894 1.733 2.294
R-NSE (AVG) 0.682 0.393 0.579 0.571 0.644 0.574
R-NSE (MED) 0.694 0.413 0.599 0.594 0.669 0.594
R-NSE1 (AVG) 0.710 0.600 0.682 0.639 0.635 0.653
155
Statistic 34 Subbasin - STATSGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.067 0.086 0.093 0.159 0.175
MEAN (m3/s) 1.719 2.246 2.578 4.047 4.645
MEDIAN (m3/s) 0.759 0.973 1.122 1.841 2.030
MAX (m3/s) 81.010 106.600 122.800 191.600 221.700
PBIAS (%) -45.170 -6.702 -24.809 -15.104 -2.311 18.819
RMSE (m3/s) 5.026 9.227 7.662 11.517 16.133
MAE (m3/s) 1.588 2.537 2.496 3.735 5.537
NSE -0.041 0.189 -0.008 0.141 -0.328 -0.010
NSE1 0.019 0.196 0.113 0.221 0.119 0.134
R-RMSE (m3/s) 1.717 5.710 2.297 3.674 3.489
R-MAE (m3/s) 0.607 0.876 0.702 1.129 1.643
R-NSE (AVG) 0.925 0.827 0.957 0.964 0.907 0.916
R-NSE (MED) 0.928 0.833 0.959 0.966 0.913 0.920
R-NSE1 (AVG) 0.625 0.723 0.750 0.765 0.739 0.720
Statistic 34 Subbasin - STATSGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.063 0.080 0.089 0.165 0.200
MEAN (m3/s) 1.667 2.182 2.517 3.976 4.623
MEDIAN (m3/s) 0.718 0.932 1.078 1.800 2.025
MAX (m3/s) 70.410 92.710 107.100 167.400 195.000
PBIAS (%) -40.724 -3.673 -21.863 -13.068 -1.816 16.229
RMSE (m3/s) 5.188 9.457 7.886 11.883 16.402
MAE (m3/s) 1.531 2.493 2.431 3.692 5.508
NSE -0.110 0.148 -0.068 0.085 -0.372 -0.063
NSE1 0.054 0.210 0.136 0.230 0.124 0.151
R-RMSE (m3/s) 1.259 5.019 1.716 3.019 4.392
R-MAE (m3/s) 0.520 0.910 0.656 1.254 1.934
R-NSE (AVG) 0.936 0.822 0.955 0.952 0.877 0.908
R-NSE (MED) 0.938 0.828 0.957 0.955 0.886 0.913
R-NSE1 (AVG) 0.678 0.712 0.767 0.738 0.692 0.718
156
Statistic 34 Subbasin - STATSGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.000 0.007 0.034 0.164 0.197
MEAN (m3/s) 0.944 1.279 1.593 3.048 3.686
MEDIAN (m3/s) 0.286 0.395 0.503 1.374 1.639
MAX (m3/s) 73.770 97.740 113.800 181.300 210.800
PBIAS (%) 20.273 39.213 22.862 13.314 18.820 22.896
RMSE (m3/s) 4.632 9.071 7.076 10.891 14.976
MAE (m3/s) 1.253 2.095 2.027 3.260 4.864
NSE 0.116 0.216 0.140 0.232 -0.144 0.112
NSE1 0.226 0.336 0.279 0.320 0.226 0.278
R-RMSE (m3/s) 2.153 6.337 3.423 5.185 5.953
R-MAE (m3/s) 0.411 1.191 0.757 1.378 2.027
R-NSE (AVG) 0.925 0.762 0.905 0.914 0.828 0.867
R-NSE (MED) 0.928 0.770 0.910 0.919 0.840 0.873
R-NSE1 (AVG) 0.746 0.623 0.731 0.713 0.678 0.698
Statistic 34 Subbasin - STATSGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.067 0.090 0.100 0.172 0.194
MEAN (m3/s) 1.597 2.090 2.404 3.858 4.496
MEDIAN (m3/s) 0.733 0.941 1.083 1.890 2.149
MAX (m3/s) 83.380 109.700 125.800 193.200 222.700
PBIAS (%) -34.879 0.714 -16.365 -9.727 0.982 12.533
RMSE (m3/s) 4.793 9.006 7.299 11.119 15.624
MAE (m3/s) 1.516 2.408 2.382 3.649 5.375
NSE 0.053 0.227 0.085 0.199 -0.245 0.064
NSE1 0.063 0.237 0.153 0.239 0.145 0.167
R-RMSE (m3/s) 1.653 5.809 2.333 3.882 3.994
R-MAE (m3/s) 0.519 0.937 0.681 1.177 1.847
R-NSE (AVG) 0.948 0.827 0.968 0.966 0.888 0.920
R-NSE (MED) 0.950 0.833 0.970 0.968 0.896 0.923
R-NSE1 (AVG) 0.679 0.703 0.758 0.754 0.706 0.720
157
Statistic 34 Subbasin - SSURGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.023 0.029 0.024 0.032 0.010
MEAN (m3/s) 0.882 1.214 1.429 2.385 2.849
MEDIAN (m3/s) 0.503 0.724 0.839 1.406 1.620
MAX (m3/s) 24.560 25.590 26.390 39.070 51.550
PBIAS (%) 25.542 42.338 30.817 32.186 37.247 33.626
RMSE (m3/s) 5.285 10.535 8.053 13.160 15.135
MAE (m3/s) 1.559 2.682 2.591 4.397 5.567
NSE -0.151 -0.057 -0.114 -0.122 -0.169 -0.123
NSE1 0.036 0.150 0.079 0.083 0.115 0.093
R-RMSE (m3/s) 3.356 8.585 5.611 8.917 9.189
R-MAE (m3/s) 0.625 1.484 1.156 2.039 2.598
R-NSE (AVG) 0.599 0.337 0.501 0.513 0.600 0.510
R-NSE (MED) 0.615 0.359 0.524 0.540 0.628 0.533
R-NSE1 (AVG) 0.614 0.530 0.589 0.575 0.587 0.579
Statistic 34 Subbasin - SSURGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.047 0.056 0.058 0.078 0.077
MEAN (m3/s) 1.722 2.252 2.587 4.094 4.743
MEDIAN
(m3/s) 0.989 1.281 1.459 2.250 2.437
MAX (m3/s) 77.060 101.400 117.400 186.100 218.600
PBIAS (%) -45.387 -6.990 -25.238 -16.430 -4.472 19.703
RMSE (m3/s) 4.596 8.890 7.041 10.860 15.461
MAE (m3/s) 1.563 2.522 2.459 3.704 5.465
NSE 0.130 0.247 0.149 0.236 -0.220 0.108
NSE1 0.034 0.201 0.126 0.228 0.131 0.144
R-RMSE (m3/s) 1.788 6.042 2.610 4.258 4.259
R-MAE (m3/s) 0.658 1.250 0.925 1.486 2.113
R-NSE (AVG) 0.947 0.795 0.963 0.961 0.886 0.910
R-NSE (MED) 0.949 0.802 0.965 0.963 0.894 0.915
R-NSE1 (AVG) 0.593 0.604 0.671 0.690 0.664 0.645
158
Statistic 34 Subbasin - SSURGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.043 0.050 0.053 0.075 0.082
MEAN (m3/s) 1.681 2.203 2.541 4.039 4.727
MEDIAN (m3/s) 0.945 1.231 1.423 2.220 2.448
MAX (m3/s) 67.670 88.980 103.200 163.900 193.500
PBIAS (%) -41.914 -4.660 -23.011 -14.867 -4.114 17.713
RMSE (m3/s) 4.809 9.157 7.338 11.341 15.842
MAE (m3/s) 1.517 2.495 2.408 3.682 5.456
NSE 0.047 0.201 0.075 0.167 -0.280 0.042
NSE1 0.062 0.210 0.144 0.232 0.132 0.156
R-RMSE (m3/s) 1.260 5.319 1.939 3.373 4.680
R-MAE (m3/s) 0.583 1.213 0.934 1.637 2.265
R-NSE (AVG) 0.967 0.800 0.968 0.955 0.872 0.912
R-NSE (MED) 0.968 0.807 0.970 0.957 0.881 0.917
R-NSE1 (AVG) 0.640 0.616 0.668 0.659 0.640 0.644
Statistic 34 Subbasin - SSURGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.000 0.011 0.041 0.105 0.110
MEAN (m3/s) 0.962 1.303 1.620 3.113 3.793
MEDIAN (m3/s) 0.407 0.563 0.777 1.664 1.917
MAX (m3/s) 75.070 99.110 115.000 183.900 215.300
PBIAS (%) 18.754 38.083 21.559 11.461 16.449 21.261
RMSE (m3/s) 4.589 9.035 6.993 10.761 14.814
MAE (m3/s) 1.241 2.082 2.019 3.250 4.825
NSE 0.132 0.222 0.160 0.250 -0.119 0.129
NSE1 0.233 0.340 0.282 0.322 0.233 0.282
R-RMSE (m3/s) 2.437 6.663 3.870 5.760 6.506
R-MAE (m3/s) 0.509 1.332 0.961 1.681 2.313
R-NSE (AVG) 0.891 0.738 0.864 0.884 0.791 0.833
R-NSE (MED) 0.895 0.747 0.870 0.890 0.806 0.842
R-NSE1 (AVG) 0.685 0.578 0.658 0.650 0.632 0.641
159
Statistic 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.055 0.069 0.073 0.102 0.107
MEAN (m3/s) 1.620 2.120 2.437 3.930 4.610
MEDIAN (m3/s) 0.966 1.252 1.441 2.277 2.479
MAX (m3/s) 84.060 110.200 126.100 195.000 226.400
PBIAS (%) -36.772 -0.711 -17.964 -11.753 -1.526 13.745
RMSE (m3/s) 4.541 8.795 6.935 10.739 15.236
MAE (m3/s) 1.514 2.417 2.366 3.636 5.326
NSE 0.150 0.263 0.174 0.253 -0.184 0.131
NSE1 0.064 0.234 0.159 0.242 0.153 0.170
R-RMSE (m3/s) 1.764 6.129 2.696 4.553 4.855
R-MAE (m3/s) 0.610 1.270 0.932 1.621 2.330
R-NSE (AVG) 0.952 0.804 0.957 0.947 0.851 0.902
R-NSE (MED) 0.954 0.811 0.959 0.950 0.861 0.907
R-NSE1 (AVG) 0.623 0.598 0.669 0.662 0.629 0.636
160
Table 28. Full monthly streamflow descriptive and goodness-of-fit statistics for all
uncalibrated SWAT model configurations. Average Percent Bias is the average of the absolute values of each Percent Bias measure.
Statistic Observed Monthly Streamflow Data
Site #1 Site #2 Site #3 Site #4 Site #5
MIN (m3/s) 0.026 0.130 0.119 0.238 0.300
MEAN (m3/s) 1.186 2.105 2.069 3.520 4.538
MEDIAN
(m3/s) 0.836 1.373 1.699 2.430 3.349
MAX (m3/s) 3.833 9.556 6.497 10.309 12.799
Statistic 6 Subbasin - STATSGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.066 0.083 0.092 0.130 0.186
MEAN (m3/s) 0.851 1.191 1.409 2.343 2.745
MEDIAN (m3/s) 0.629 0.961 1.119 2.024 2.474
MAX (m3/s) 2.630 3.008 3.304 5.246 6.182
PBIAS (%) 28.291 43.416 31.900 33.442 39.503 35.310
RMSE (m3/s) 1.490 2.786 2.174 3.433 4.430
MAE (m3/s) 1.139 1.790 1.581 2.442 3.130
NSE -0.635 -0.284 -0.297 -0.181 -0.191 -0.317
NSE1 -0.206 0.005 -0.061 0.032 0.089 -0.028
R-RMSE (m3/s) 0.630 1.970 1.281 2.269 3.184
R-MAE (m3/s) 0.381 1.117 0.864 1.446 2.158
R-NSE (AVG) 0.708 0.358 0.550 0.484 0.385 0.497
R-NSE (MED) 0.732 0.410 0.566 0.539 0.433 0.536
R-NSE1 (AVG) 0.596 0.379 0.421 0.427 0.372 0.439
161
Statistic 6 Subbasin - STATSGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.172 0.220 0.243 0.367 0.409
MEAN (m3/s) 1.733 2.260 2.592 4.057 4.662
MEDIAN
(m3/s) 1.649 2.145 2.465 3.868 4.417
MAX (m3/s) 4.309 5.650 6.464 10.081 11.605
PBIAS (%) -46.093 -7.375 -25.321 -15.243 -2.744 19.355
RMSE (m3/s) 0.899 1.743 1.292 1.894 2.567
MAE (m3/s) 0.672 1.138 0.931 1.417 1.796
NSE 0.405 0.497 0.542 0.641 0.600 0.537
NSE1 0.288 0.367 0.375 0.439 0.477 0.389
R-RMSE (m3/s) 0.616 1.175 0.677 1.019 1.225
R-MAE (m3/s) 0.551 0.837 0.619 0.916 1.008
R-NSE (AVG) 0.721 0.771 0.874 0.896 0.909 0.834
R-NSE (MED) 0.744 0.790 0.879 0.907 0.916 0.847
R-NSE1 (AVG) 0.417 0.535 0.585 0.637 0.707 0.576
Statistic 6 Subbasin - STATSGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.154 0.193 0.216 0.345 0.404
MEAN (m3/s) 1.678 2.194 2.529 3.972 4.617
MEDIAN (m3/s) 1.475 1.930 2.240 3.515 4.090
MAX (m3/s) 3.541 4.639 5.351 8.362 9.778
PBIAS (%) -41.473 -4.219 -22.239 -12.836 -1.742 16.502
RMSE (m3/s) 0.736 1.554 1.085 1.592 2.356
MAE (m3/s) 0.540 1.004 0.799 1.242 1.753
NSE 0.601 0.600 0.677 0.746 0.663 0.658
NSE1 0.429 0.442 0.464 0.508 0.490 0.466
R-RMSE (m3/s) 0.573 1.371 0.716 1.066 1.510
R-MAE (m3/s) 0.531 0.903 0.666 0.976 1.250
R-NSE (AVG) 0.758 0.689 0.859 0.886 0.862 0.811
R-NSE (MED) 0.778 0.714 0.864 0.898 0.872 0.826
R-NSE1 (AVG) 0.438 0.498 0.553 0.613 0.636 0.548
162
Statistic 6 Subbasin - STATSGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.031 0.039 0.122 0.411 0.461
MEAN (m3/s) 0.947 1.238 1.551 3.058 3.727
MEDIAN (m3/s) 0.687 0.905 1.164 2.770 3.353
MAX (m3/s) 3.485 4.577 5.388 8.434 9.881
PBIAS (%) 20.143 41.208 25.024 13.130 17.858 23.473
RMSE (m3/s) 0.724 1.782 1.134 1.729 2.652
MAE (m3/s) 0.514 1.096 0.850 1.358 1.948
NSE 0.614 0.474 0.647 0.701 0.573 0.602
NSE1 0.455 0.390 0.430 0.462 0.433 0.434
R-RMSE (m3/s) 0.444 1.600 0.859 1.351 1.967
R-MAE (m3/s) 0.282 0.908 0.573 0.900 1.480
R-NSE (AVG) 0.855 0.576 0.798 0.817 0.765 0.762
R-NSE (MED) 0.867 0.611 0.805 0.837 0.784 0.781
R-NSE1 (AVG) 0.701 0.495 0.615 0.643 0.569 0.605
Statistic 6 Subbasin - STATSGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.181 0.233 0.260 0.408 0.458
MEAN (m3/s) 1.604 2.096 2.409 3.916 4.585
MEDIAN (m3/s) 1.464 1.894 2.160 3.589 4.139
MAX (m3/s) 4.464 5.870 6.743 10.280 11.878
PBIAS (%) -35.205 0.430 -16.468 -11.253 -1.050 12.881
RMSE (m3/s) 0.779 1.656 1.103 1.724 2.484
MAE (m3/s) 0.612 1.089 0.817 1.335 1.793
NSE 0.553 0.546 0.667 0.702 0.626 0.619
NSE1 0.352 0.395 0.452 0.471 0.478 0.429
R-RMSE (m3/s) 0.489 1.079 0.518 1.030 1.356
R-MAE (m3/s) 0.437 0.718 0.468 0.826 1.069
R-NSE (AVG) 0.824 0.807 0.926 0.894 0.888 0.868
R-NSE (MED) 0.839 0.823 0.929 0.905 0.897 0.879
R-NSE1 (AVG) 0.537 0.601 0.686 0.673 0.689 0.637
163
Statistic 6 Subbasin - SSURGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.062 0.079 0.094 0.165 0.231
MEAN (m3/s) 0.865 1.197 1.412 2.353 2.817
MEDIAN (m3/s) 0.736 0.991 1.187 2.186 2.556
MAX (m3/s) 2.762 3.192 3.492 4.880 6.072
PBIAS (%) 27.070 43.149 31.737 33.166 37.926 34.610
RMSE (m3/s) 1.466 2.766 2.146 3.393 4.376
MAE (m3/s) 1.107 1.777 1.551 2.446 3.144
NSE -0.583 -0.266 -0.263 -0.153 -0.162 -0.286
NSE1 -0.172 0.012 -0.041 0.031 0.085 -0.017
R-RMSE (m3/s) 0.634 1.964 1.289 2.276 3.109
R-MAE (m3/s) 0.369 1.114 0.839 1.395 2.100
R-NSE (AVG) 0.704 0.361 0.544 0.481 0.414 0.501
R-NSE (MED) 0.728 0.413 0.561 0.536 0.460 0.540
R-NSE1 (AVG) 0.609 0.380 0.437 0.447 0.389 0.453
Statistic 6 Subbasin - SSURGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.127 0.158 0.172 0.239 0.258
MEAN (m3/s) 1.730 2.258 2.593 4.083 4.738
MEDIAN
(m3/s) 1.721 2.237 2.566 4.032 4.639
MAX (m3/s) 3.977 5.223 5.998 9.490 11.107
PBIAS (%) -45.837 -7.286 -25.359 -15.981 -4.414 19.775
RMSE (m3/s) 0.893 1.726 1.280 1.901 2.516
MAE (m3/s) 0.703 1.148 0.994 1.491 1.755
NSE 0.413 0.507 0.551 0.638 0.616 0.545
NSE1 0.255 0.361 0.333 0.409 0.489 0.370
R-RMSE (m3/s) 0.643 1.286 0.740 1.127 1.370
R-MAE (m3/s) 0.571 0.904 0.692 1.010 1.116
R-NSE (AVG) 0.696 0.726 0.850 0.873 0.886 0.806
R-NSE (MED) 0.721 0.749 0.855 0.886 0.895 0.821
R-NSE1 (AVG) 0.395 0.497 0.535 0.600 0.675 0.541
164
Statistic 6 Subbasin - SSURGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.108 0.135 0.150 0.221 0.250
MEAN (m3/s) 1.689 2.210 2.547 4.020 4.715
MEDIAN (m3/s) 1.632 2.138 2.441 3.845 4.382
MAX (m3/s) 3.581 4.708 5.494 8.742 10.346
PBIAS (%) -42.404 -4.970 -23.140 -14.208 -3.908 17.726
RMSE (m3/s) 0.754 1.549 1.103 1.636 2.297
MAE (m3/s) 0.576 1.007 0.871 1.359 1.748
NSE 0.581 0.603 0.667 0.732 0.680 0.652
NSE1 0.390 0.440 0.415 0.462 0.491 0.440
R-RMSE (m3/s) 0.625 1.429 0.794 1.212 1.621
R-MAE (m3/s) 0.571 0.953 0.736 1.090 1.384
R-NSE (AVG) 0.713 0.662 0.827 0.853 0.841 0.779
R-NSE (MED) 0.736 0.690 0.834 0.868 0.853 0.796
R-NSE1 (AVG) 0.395 0.470 0.506 0.568 0.597 0.507
Statistic 6 Subbasin - SSURGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.020 0.022 0.116 0.303 0.341
MEAN (m3/s) 0.963 1.257 1.573 3.113 3.821
MEDIAN (m3/s) 0.696 0.920 1.235 2.730 3.602
MAX (m3/s) 3.661 4.815 5.658 8.930 10.552
PBIAS (%) 18.838 40.291 23.947 11.573 15.794 22.089
RMSE (m3/s) 0.783 1.807 1.207 1.802 2.613
MAE (m3/s) 0.556 1.123 0.899 1.383 1.909
NSE 0.549 0.459 0.601 0.675 0.586 0.574
NSE1 0.412 0.376 0.397 0.452 0.445 0.416
R-RMSE (m3/s) 0.443 1.596 0.878 1.423 1.977
R-MAE (m3/s) 0.254 0.888 0.551 0.881 1.466
R-NSE (AVG) 0.856 0.578 0.789 0.797 0.763 0.757
R-NSE (MED) 0.868 0.613 0.796 0.819 0.782 0.775
R-NSE1 (AVG) 0.731 0.506 0.630 0.651 0.573 0.618
165
Statistic 6 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.147 0.183 0.202 0.302 0.341
MEAN (m3/s) 1.625 2.123 2.439 3.978 4.687
MEDIAN (m3/s) 1.575 2.048 2.329 3.885 4.591
MAX (m3/s) 4.141 5.441 6.281 9.705 11.372
PBIAS (%) -36.939 -0.843 -17.907 -13.021 -3.283 14.398
RMSE (m3/s) 0.814 1.670 1.143 1.775 2.461
MAE (m3/s) 0.670 1.125 0.889 1.410 1.755
NSE 0.512 0.538 0.641 0.684 0.633 0.602
NSE1 0.290 0.374 0.404 0.441 0.489 0.400
R-RMSE (m3/s) 0.529 1.178 0.599 1.132 1.460
R-MAE (m3/s) 0.477 0.785 0.533 0.939 1.202
R-NSE (AVG) 0.794 0.770 0.902 0.872 0.871 0.842
R-NSE (MED) 0.811 0.789 0.905 0.885 0.881 0.854
R-NSE1 (AVG) 0.495 0.564 0.642 0.628 0.650 0.596
Statistic 34 Subbasin - STATSGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.056 0.071 0.081 0.118 0.185
MEAN (m3/s) 0.890 1.229 1.447 2.380 2.789
MEDIAN (m3/s) 0.709 0.988 1.175 2.189 2.640
MAX (m3/s) 2.328 2.730 3.218 5.410 6.354
PBIAS (%) 25.020 41.603 30.056 32.393 38.540 33.522
RMSE (m3/s) 1.372 2.681 2.083 3.356 4.353
MAE (m3/s) 1.019 1.692 1.472 2.357 3.033
NSE -0.387 -0.190 -0.191 -0.129 -0.150 -0.209
NSE1 -0.079 0.059 0.013 0.066 0.118 0.035
R-RMSE (m3/s) 0.669 2.001 1.282 2.220 3.115
R-MAE (m3/s) 0.420 1.133 0.837 1.404 2.106
R-NSE (AVG) 0.670 0.337 0.549 0.506 0.411 0.495
R-NSE (MED) 0.698 0.391 0.566 0.559 0.458 0.534
R-NSE1 (AVG) 0.556 0.370 0.438 0.443 0.387 0.439
166
Statistic 34 Subbasin - STATSGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.157 0.202 0.225 0.349 0.382
MEAN (m3/s) 1.718 2.244 2.577 4.045 4.642
MEDIAN (m3/s) 1.629 2.116 2.432 3.838 4.365
MAX (m3/s) 4.288 5.628 6.441 10.056 11.603
PBIAS (%) -44.844 -6.624 -24.557 -14.902 -2.296 18.645
RMSE (m3/s) 0.883 1.732 1.277 1.880 2.556
MAE (m3/s) 0.655 1.128 0.920 1.407 1.782
NSE 0.426 0.504 0.553 0.646 0.604 0.546
NSE1 0.307 0.373 0.383 0.442 0.482 0.397
R-RMSE (m3/s) 0.600 1.167 0.661 1.008 1.209
R-MAE (m3/s) 0.535 0.821 0.603 0.902 0.985
R-NSE (AVG) 0.735 0.774 0.880 0.898 0.911 0.840
R-NSE (MED) 0.757 0.793 0.885 0.909 0.918 0.852
R-NSE1 (AVG) 0.434 0.543 0.595 0.643 0.713 0.586
Statistic 34 Subbasin - STATSGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.138 0.177 0.200 0.333 0.384
MEAN (m3/s) 1.665 2.180 2.515 3.972 4.619
MEDIAN (m3/s) 1.462 1.915 2.226 3.513 4.074
MAX (m3/s) 3.569 4.666 5.362 8.422 9.859
PBIAS (%) -40.379 -3.577 -21.589 -12.848 -1.784 16.036
RMSE (m3/s) 0.720 1.544 1.070 1.583 2.348
MAE (m3/s) 0.525 1.000 0.788 1.235 1.746
NSE 0.618 0.606 0.686 0.749 0.666 0.665
NSE1 0.444 0.444 0.472 0.511 0.492 0.472
R-RMSE (m3/s) 0.557 1.362 0.702 1.057 1.495
R-MAE (m3/s) 0.516 0.890 0.651 0.969 1.246
R-NSE (AVG) 0.771 0.693 0.865 0.888 0.864 0.816
R-NSE (MED) 0.790 0.718 0.870 0.900 0.875 0.830
R-NSE1 (AVG) 0.453 0.505 0.563 0.616 0.637 0.555
167
Statistic 34 Subbasin - STATSGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.024 0.056 0.139 0.388 0.428
MEAN (m3/s) 0.943 1.278 1.591 3.043 3.680
MEDIAN (m3/s) 0.689 0.926 1.193 2.769 3.245
MAX (m3/s) 3.510 4.639 5.448 8.740 10.154
PBIAS (%) 20.505 39.305 23.091 13.549 18.901 23.070
RMSE (m3/s) 0.725 1.745 1.091 1.739 2.665
MAE (m3/s) 0.514 1.084 0.831 1.377 1.941
NSE 0.612 0.496 0.673 0.697 0.569 0.610
NSE1 0.456 0.397 0.442 0.454 0.435 0.437
R-RMSE (m3/s) 0.439 1.574 0.819 1.300 1.932
R-MAE (m3/s) 0.275 0.888 0.550 0.874 1.460
R-NSE (AVG) 0.858 0.590 0.816 0.831 0.773 0.774
R-NSE (MED) 0.870 0.623 0.823 0.849 0.791 0.791
R-NSE1 (AVG) 0.708 0.506 0.631 0.653 0.575 0.615
Statistic 34 Subbasin - STATSGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.169 0.219 0.246 0.385 0.426
MEAN (m3/s) 1.594 2.085 2.398 3.850 4.487
MEDIAN (m3/s) 1.433 1.862 2.127 3.485 4.056
MAX (m3/s) 4.466 5.870 6.743 10.435 12.078
PBIAS (%) -34.365 0.946 -15.944 -9.387 1.115 12.351
RMSE (m3/s) 0.767 1.646 1.090 1.722 2.490
MAE (m3/s) 0.603 1.081 0.807 1.327 1.793
NSE 0.566 0.552 0.674 0.703 0.624 0.624
NSE1 0.361 0.399 0.459 0.474 0.478 0.434
R-RMSE (m3/s) 0.474 1.070 0.504 0.955 1.374
R-MAE (m3/s) 0.428 0.705 0.453 0.719 1.059
R-NSE (AVG) 0.834 0.811 0.930 0.909 0.885 0.874
R-NSE (MED) 0.848 0.826 0.933 0.918 0.894 0.884
R-NSE1 (AVG) 0.547 0.608 0.696 0.715 0.692 0.651
168
Statistic 34 Subbasin - SSURGO - Weather Generator Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.056 0.071 0.087 0.158 0.233
MEAN (m3/s) 0.880 1.211 1.426 2.381 2.846
MEDIAN (m3/s) 0.729 1.003 1.258 2.167 2.577
MAX (m3/s) 2.391 2.803 3.103 5.169 6.358
PBIAS (%) 25.857 42.465 31.042 32.348 37.286 33.800
RMSE (m3/s) 1.357 2.675 2.062 3.319 4.301
MAE (m3/s) 0.998 1.678 1.444 2.357 3.041
NSE -0.356 -0.184 -0.167 -0.104 -0.123 -0.187
NSE1 -0.057 0.067 0.031 0.066 0.115 0.044
R-RMSE (m3/s) 0.688 2.031 1.313 2.244 3.064
R-MAE (m3/s) 0.423 1.141 0.836 1.355 2.057
R-NSE (AVG) 0.651 0.317 0.527 0.496 0.430 0.484
R-NSE (MED) 0.680 0.373 0.545 0.549 0.475 0.524
R-NSE1 (AVG) 0.552 0.365 0.439 0.463 0.401 0.444
Statistic 34 Subbasin - SSURGO - Sanborn Field Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.121 0.150 0.164 0.231 0.248
MEAN (m3/s) 1.720 2.250 2.585 4.090 4.739
MEDIAN (m3/s) 1.698 2.216 2.544 4.020 4.609
MAX (m3/s) 3.950 5.202 5.977 9.530 11.152
PBIAS (%) -45.018 -6.881 -24.943 -16.194 -4.435 19.494
RMSE (m3/s) 0.879 1.721 1.269 1.888 2.505
MAE (m3/s) 0.688 1.141 0.980 1.478 1.744
NSE 0.431 0.510 0.558 0.643 0.619 0.552
NSE1 0.271 0.365 0.342 0.414 0.493 0.377
R-RMSE (m3/s) 0.629 1.282 0.729 1.113 1.344
R-MAE (m3/s) 0.558 0.894 0.682 0.999 1.093
R-NSE (AVG) 0.708 0.728 0.854 0.876 0.890 0.811
R-NSE (MED) 0.732 0.750 0.860 0.889 0.899 0.826
R-NSE1 (AVG) 0.409 0.503 0.542 0.604 0.682 0.548
169
Statistic 34 Subbasin - SSURGO - South Farms Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.101 0.128 0.142 0.214 0.240
MEAN (m3/s) 1.679 2.200 2.538 4.034 4.721
MEDIAN (m3/s) 1.612 2.113 2.416 3.813 4.311
MAX (m3/s) 3.572 4.694 5.481 8.805 10.411
PBIAS (%) -41.516 -4.525 -22.688 -14.603 -4.048 17.476
RMSE (m3/s) 0.740 1.544 1.091 1.622 2.288
MAE (m3/s) 0.562 1.000 0.858 1.342 1.732
NSE 0.597 0.605 0.674 0.736 0.682 0.659
NSE1 0.405 0.444 0.424 0.468 0.496 0.448
R-RMSE (m3/s) 0.611 1.422 0.783 1.198 1.597
R-MAE (m3/s) 0.557 0.942 0.725 1.083 1.362
R-NSE (AVG) 0.725 0.665 0.832 0.856 0.845 0.785
R-NSE (MED) 0.748 0.693 0.838 0.872 0.857 0.801
R-NSE1 (AVG) 0.410 0.476 0.513 0.571 0.604 0.515
Statistic 34 Subbasin - SSURGO - HCW-5 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.015 0.051 0.145 0.292 0.324
MEAN (m3/s) 0.961 1.301 1.617 3.108 3.788
MEDIAN (m3/s) 0.700 1.026 1.299 2.749 3.416
MAX (m3/s) 3.677 4.857 5.700 9.270 10.854
PBIAS (%) 19.021 38.196 21.816 11.694 16.515 21.449
RMSE (m3/s) 0.777 1.768 1.159 1.799 2.623
MAE (m3/s) 0.551 1.096 0.877 1.401 1.903
NSE 0.555 0.483 0.631 0.676 0.583 0.586
NSE1 0.417 0.390 0.411 0.445 0.446 0.422
R-RMSE (m3/s) 0.437 1.570 0.847 1.360 1.930
R-MAE (m3/s) 0.256 0.873 0.531 0.832 1.434
R-NSE (AVG) 0.860 0.592 0.803 0.815 0.774 0.769
R-NSE (MED) 0.871 0.625 0.810 0.834 0.792 0.786
R-NSE1 (AVG) 0.728 0.515 0.644 0.670 0.583 0.628
170
Statistic 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.141 0.177 0.197 0.290 0.322
MEAN (m3/s) 1.616 2.115 2.431 3.922 4.602
MEDIAN (m3/s) 1.544 2.017 2.298 3.696 4.400
MAX (m3/s) 4.152 5.443 6.286 9.885 11.569
PBIAS (%) -36.233 -0.468 -17.524 -11.421 -1.413 13.412
RMSE (m3/s) 0.799 1.659 1.130 1.774 2.469
MAE (m3/s) 0.658 1.115 0.877 1.384 1.740
NSE 0.529 0.544 0.650 0.685 0.630 0.608
NSE1 0.303 0.380 0.412 0.451 0.494 0.408
R-RMSE (m3/s) 0.517 1.171 0.588 1.051 1.456
R-MAE (m3/s) 0.467 0.774 0.522 0.828 1.173
R-NSE (AVG) 0.803 0.773 0.905 0.889 0.871 0.848
R-NSE (MED) 0.819 0.791 0.909 0.901 0.881 0.860
R-NSE1 (AVG) 0.505 0.569 0.650 0.672 0.659 0.611
171
Table 29. Full daily and monthly streamflow descriptive and goodness-of-fit statistics for
the selected uncalibrated SWAT model configuration, the sum of square error (SSQ) optimized model, and the sum of squared error (SSQ) and sum of squared error after ranking (SSQR) optimized model. Average Percent Bias shown is the average of the absolute values of each Percent Bias measure.
Daily Statistic Uncalibrated - 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.055 0.069 0.073 0.102 0.107
MEAN (m3/s) 1.620 2.120 2.437 3.930 4.610
MEDIAN (m3/s) 0.966 1.252 1.441 2.277 2.479
MAX (m3/s) 84.060 110.200 126.100 195.000 226.400
PBIAS (%) -36.772 -0.711 -17.964 -11.753 -1.526 13.745
RMSE (m3/s) 4.541 8.795 6.935 10.739 15.236
MAE (m3/s) 1.514 2.417 2.366 3.636 5.326
NSE 0.150 0.263 0.174 0.253 -0.184 0.131
NSE1 0.064 0.234 0.159 0.242 0.153 0.170
R-RMSE (m3/s) 1.764 6.129 2.696 4.553 4.855
R-MAE (m3/s) 0.610 1.270 0.932 1.621 2.330
R-NSE (AVG) 0.952 0.804 0.957 0.947 0.851 0.902
R-NSE (MED) 0.954 0.811 0.959 0.950 0.861 0.907
R-NSE1 (AVG) 0.623 0.598 0.669 0.662 0.629 0.636
Monthly Statistic Uncalibrated - 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.141 0.177 0.197 0.290 0.322
MEAN (m3/s) 1.616 2.115 2.431 3.922 4.602
MEDIAN (m3/s) 1.544 2.017 2.298 3.696 4.400
MAX (m3/s) 4.152 5.443 6.286 9.885 11.569
PBIAS (%) -36.233 -0.468 -17.524 -11.421 -1.413 13.412
RMSE (m3/s) 0.799 1.659 1.130 1.774 2.469
MAE (m3/s) 0.658 1.115 0.877 1.384 1.740
NSE 0.529 0.544 0.650 0.685 0.630 0.608
NSE1 0.303 0.380 0.412 0.451 0.494 0.408
R-RMSE (m3/s) 0.517 1.171 0.588 1.051 1.456
R-MAE (m3/s) 0.467 0.774 0.522 0.828 1.173
R-NSE (AVG) 0.803 0.773 0.905 0.889 0.871 0.848
R-NSE (MED) 0.819 0.791 0.909 0.901 0.881 0.860
R-NSE1 (AVG) 0.505 0.569 0.650 0.672 0.659 0.611
172
Daily Statistic SSQ-Optimized 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.000 0.000 0.000 0.000 0.001
MEAN (m3/s) 1.695 2.218 2.548 4.102 4.792
MEDIAN
(m3/s) 0.232 0.311 0.368 0.741 1.012
MAX (m3/s) 45.990 59.700 68.300 104.100 123.900
PBIAS (%) -43.151 -5.402 -23.362 -16.657 -5.546 18.824
RMSE (m3/s) 3.715 7.995 5.433 8.586 9.887
MAE (m3/s) 1.287 2.052 1.963 3.139 3.819
NSE 0.431 0.391 0.493 0.522 0.501 0.468
NSE1 0.204 0.350 0.302 0.345 0.393 0.319
R-RMSE (m3/s) 2.886 7.349 4.168 6.620 6.612
R-MAE (m3/s) 0.843 1.340 1.102 1.779 2.411
R-NSE (AVG) 0.757 0.587 0.798 0.801 0.830 0.755
R-NSE (MED) 0.767 0.601 0.808 0.812 0.842 0.766
R-NSE1 (AVG) 0.479 0.576 0.608 0.629 0.617 0.582
Monthly
Statistic
SSQ-Optimized 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.106 0.144 0.173 0.293 0.347
MEAN (m3/s) 1.688 2.208 2.537 4.086 4.776
MEDIAN (m3/s) 1.084 1.420 1.636 2.712 3.300
MAX (m3/s) 6.649 8.648 9.849 15.240 17.292
PBIAS (%) -42.264 -4.907 -22.627 -16.083 -5.249 18.226
RMSE (m3/s) 1.115 1.574 1.478 2.132 2.337
MAE (m3/s) 0.807 0.935 1.066 1.526 1.717
NSE 0.084 0.590 0.401 0.545 0.669 0.458
NSE1 0.146 0.480 0.285 0.395 0.500 0.361
R-RMSE (m3/s) 0.777 0.647 0.910 1.349 1.253
R-MAE (m3/s) 0.501 0.447 0.506 0.783 0.855
R-NSE (AVG) 0.556 0.931 0.773 0.818 0.905 0.796
R-NSE (MED) 0.592 0.936 0.781 0.837 0.912 0.812
R-NSE1 (AVG) 0.469 0.751 0.660 0.690 0.751 0.664
173
Daily Statistic SSQ & SSQR-Optimized 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.219 0.271 0.294 0.478 0.506
MEAN (m3/s) 1.427 1.873 2.156 3.499 4.150
MEDIAN (m3/s) 0.414 0.532 0.614 0.990 1.083
MAX (m3/s) 90.170 118.200 135.300 209.400 242.500
PBIAS (%) -20.459 11.028 -4.361 0.501 8.600 8.990
RMSE (m3/s) 4.896 9.011 7.445 11.423 16.274
MAE (m3/s) 1.412 2.243 2.212 3.368 5.175
NSE 0.012 0.227 0.048 0.155 -0.351 0.018
NSE1 0.127 0.290 0.214 0.298 0.177 0.221
R-RMSE (m3/s) 1.126 3.917 1.347 2.085 4.764
R-MAE (m3/s) 0.416 0.756 0.548 0.878 1.344
R-NSE (AVG) 0.948 0.854 0.969 0.972 0.884 0.925
R-NSE (MED) 0.950 0.859 0.970 0.973 0.892 0.929
R-NSE1 (AVG) 0.743 0.761 0.805 0.817 0.786 0.782
Monthly
Statistic
SSQ & SSQR-Optimized 34 Subbasin - SSURGO - HCW-4 Climate Run
Site #1 Site #2 Site #3 Site #4 Site #5 Mean - All Sites
MIN (m3/s) 0.343 0.435 0.480 0.730 0.781
MEAN (m3/s) 1.424 1.869 2.152 3.494 4.145
MEDIAN (m3/s) 1.098 1.433 1.635 2.727 3.223
MAX (m3/s) 4.390 5.761 6.634 10.420 12.110
PBIAS (%) -20.035 11.208 -4.014 0.749 8.662 8.934
RMSE (m3/s) 0.606 1.513 0.917 1.378 2.302
MAE (m3/s) 0.494 1.033 0.720 1.024 1.635
NSE 0.729 0.621 0.769 0.810 0.678 0.722
NSE1 0.476 0.426 0.517 0.594 0.524 0.507
R-RMSE (m3/s) 0.384 1.118 0.516 1.065 1.582
R-MAE (m3/s) 0.348 0.729 0.377 0.683 1.191
R-NSE (AVG) 0.892 0.793 0.927 0.886 0.848 0.869
R-NSE (MED)
R-NSE1 (AVG) 0.631 0.594 0.747 0.729 0.654 0.671