KONKUK UNIVERSITY

The Uncertainty Analysis of SWAT Simulated

Streamflow and Water Quality Applied to
Chungju Dam Watershed of South Korea
15 June 2011
JOH, Hyung-Kyung1
1Graduate

Student

PARK, Jong-Yoon2 / JUNG, Hyun-Kyo2 / SHIN, Hyung-Jin3 / KWON, Hyung-Joong3

2 Ph.

D. Candidate / 3Ph. D.

KIM, Seong-Joon4
4Professor,

Corresponding Author

Dept. of Civil and Environmental System Eng.,

Konkuk University, Seoul, South Korea

Contents
I.

Introduction

II.

Material and Methods

III.

IV.

Study Watershed

SWAT Model Description

SWAT-CUP Model Description

Results

Sensitivity Analysis of Model Parameters

Model Calibration and Validation

Comparing Each Result

Summary and Conclusion

CHUNCHEON
1 / 18 GLOBAL WATER FORUM 2009
1/26

Konkuk University

Purpose of this study

SWAT (Soil and Water Assessment Tool, Arnold et al., 1998) model is
a very useful tool for both making decisions and research purposes
of watershed management .
To fulfill this applicability, the model should pass through a careful &
well-done calibration.

If the model is calibrated using streamflow data at the watershed

outlet, can the model be called as calibrated one? What about the
case of multisite calibration?
What if we add streamflow data from stations inside the watershed?
Will the new model give correct results from various parameter set in
the watershed? Perhaps NOT (Abbaspour et al., 2007).

Thus, we need to assess and understand the model

uncertainty occurred by parameter set. This is possible
through SWAT-CUP (Soil and Water Assessment ToolCalibration Uncertainty Program) Modeling.
CHUNCHEON
2 / 18 GLOBAL WATER FORUM 2009
2/26

Konkuk University

Prediction Uncertainty

Parameter Uncertainty vs. Simulation Uncertainty

(a)

(b)

(c)

(d)

Parameter Uncertainty
Prediction Uncertainty
Measured Value
(Karim C. Abbaspour, 2009, SWAT-CUP User manual)

CHUNCHEON
2 / 18 GLOBAL WATER FORUM 2009
3/26

Konkuk University

Prediction Uncertainty

Combination of Non-unique Parameter Uncertainty

vs. Simulation Uncertainty
GW_DELAY

3.46
0.34

CASE 1
CASE 2

CN2

0.0098
0.131

CASE 1

50
20

REVAPMN

SOL_AWC

0.8
2.4

0.11
0.23

CASE 2

Discharge (cms)

Obs.

CH_N2

Time
(Karim C. Abbaspour, 2009, SWAT-CUP User manual)

CHUNCHEON
2 / 18 GLOBAL WATER FORUM 2009
4/26

Konkuk University

Prediction Uncertainty
Swiss Cheese Effect

Goal

Goal

(Karim C. Abbaspour, 2009, SWAT-CUP User manual)

CHUNCHEON
2 / 18 GLOBAL WATER FORUM 2009
5/26

Konkuk University

Study Procedure
Meteorological Data
Precipitation,
Temperature,
Wind Speed,
Solar Radiation,
Relative Humidity

GIS Data

DEM,
Land Use,
Soil type

Observed Data

Streamflow,
Water Quality

Setup
SWAT Model
SWAT-CUP Model
Auto-Calibration
SUFI2, GLUE, ParaSol, MCMC

No

Compare
each result
Is Calibration Criteria
Satisfied?

Yes
CHUNCHEON
3 / 18 GLOBAL WATER FORUM 2009
6/26

STOP
Konkuk University

Study Area

Chungju-Dam Watershed

South Korea

Watershed Boundary
Subbasin
Stream
Calibration Points

Watershed Area : 6,661.3 km2

Forest Area : 82.2 %

Annual Average Precipitation : 1,359.5 mm

Annual Average Temperature : 9.4

CHUNCHEON
4 / 18 GLOBAL WATER FORUM 2009
7/26

Konkuk University

Input and Measured Data

Data Set for SWAT model

Data Type

Source

Scale /
Periods

Terrain

Korea National
Geography Institute

30 m

Soil

Korea Rural
Development
Administration

1/25,000

Soil classification and physical properties

viz. texture, porosity, field capacity, wilting
point, saturated conductivity, and soil
depth

Land use

2004 Landsat TM
Satellite Image

1/25,000

Landsat land use classification

Weather

Korea Institute of
Construction
Technology / WAter
Management
Information System

1971-2009

CHUNCHEON
5 / 18 GLOBAL WATER FORUM 2009
8/26

Data Description / Properties

Digital Elevation Model (DEM)

Daily precipitation, minimum and

maximum temperature, mean wind speed
and relative humidity data

Konkuk University

GIS Input Data

SWAT Input Data

Clay
Clay Loam
Loam
Loamy Sand
Sandy Loam
Silt
Silt Loam
Silty Clay Loam

(m)

1562

112

(a) DEM

CHUNCHEON
6 / 18 GLOBAL WATER FORUM 2009
9/26

(b) Soil

Water
Urban
Paddy
Pasture
Forest Mixed
Forest Evergreen
Forest Deciduous

(c) Landuse

Konkuk University

Model theory

SWAT (Soil and Water Assessment Tool)

Water Balance Equation :

SWt = Final soil water content (mm)

SW0 = Initial soil water content on day i (mm)
Rday = Amount of precipitation on day i (mm)
Qsurf = Amount of surface runoff on day i (mm)
Ea = Amount of evapotranspiration on day i (mm)
Wseep = Amount of water entering the vadose zone from the soil profile on day i (mm)

Qgw = Amount of return flow on day i (mm)

CHUNCHEON
7 / 18 GLOBAL WATER FORUM 2009
10/26

Konkuk University

Model theory

SWAT-CUP (SWAT Calibration Uncertainty Program)

Developed by Eawag* Swiss Federal Institute, to analyze the prediction
uncertainty of SWAT model calibration and validation results

Equipped Algorithms :
a. .SUFI2 (Abbaspour, et al., 2007) : Sequential Uncertainty FItting ver.2, the
parameter uncertainty accounts for all sources of uncertainties such as uncertainty
in driving variables (e.g., rainfall), conceptual model, parameters, and measured
data.
b. GLUE (Beven and Binley, 1992) : Generalized Likelihood Uncertainty Estimation is
based on the estimation of the weights or probabilities associated with different
parameter sets, based on the use of a subjective likelihood measure to derive a
posterior probability function, which is subsequently used to derive the predictive
probability of the output variables.

*Eidgenssische Anstalt fr Wasserversorgung, Abwasserreinigung und Gewsserschutz

CHUNCHEON
7 / 18 GLOBAL WATER FORUM 2009
11/26

Konkuk University

Model theory

SWAT-CUP (SWAT Calibration Uncertainty Program)

Developed by Eawag* Swiss Federal Institute, to analyze the prediction
uncertainty of SWAT model calibration and validation results

Equipped Algorithms :

c. ParaSol (van Griensven and Meixner, 2006) : Parameter Solution method

aggregates objective functions (OF) into a global optimization criterion (GOC) and
then minimizes these OFs or a GOC using the SCE-UA (Shffled Complex Evolution,
Duan, et al., 1992) algorithm.
d. MCMC : Markov Chain Monte Carlo generates samples from a random walk which
adapts to the posterior distribution (Kuczera and Parent, 1998). This simple
techniques from this class is the Metropolis Hasting algorithm (Gelman et al.
1995), which is not applied in this study.

*Eidgenssische Anstalt fr Wasserversorgung, Abwasserreinigung und Gewsserschutz

CHUNCHEON
7 / 18 GLOBAL WATER FORUM 2009
11/26

Konkuk University

Model theory

SWAT-CUP (SWAT Calibration Uncertainty Program)

Remarkable Words :

a. 95PPU : 95 Percent Prediction Uncertainty, This value is calculated at the 2.5%

and 97.5% levels of an output variable, disallowing 5% of the very bad simulations.
b. Objective Function : Coefficient of determination (R2), Nash-Sutcliffe (1970)
coefficient Etc.
c. p-factor : The percentage of observations covered by the 95PPU.
d. r-factor : Relative width of 95% probability band.
e. t-Stat : Provides a measure of sensitivity, larger absolute values are more sensitive.

f.

P-Value : Determined the significance of the sensitivity. A values close to zero has
more significance.
- A P-factor of 1 and R-factor of zero is a simulation that exactly
corresponds to measured data.
(Karim C. Abbaspour, 2009, SWAT-CUP User maunal)

CHUNCHEON
7 / 18 GLOBAL WATER FORUM 2009
12/26

Konkuk University

Sensitivity Analysis (SA)

t-Stat

CHUNCHEON GLOBAL WATER FORUM 2009

13/26

Konkuk University

Sensitivity Analysis (SA)

P-Value

CHUNCHEON GLOBAL WATER FORUM 2009

14/26

Konkuk University

Calibration and Validation

SUFI2

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.71, E= 0.59

R2= 0.79, E= 0.67

R2= 0.91, E= 0.90

CHUNCHEON GLOBAL WATER FORUM 2009

15/26

Konkuk University

Calibration and Validation

GLUE

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.80, E= 0.72

R2= 0.79, E= 0.67

R2= 0.93, E= 0.90

CHUNCHEON GLOBAL WATER FORUM 2009

16/26

Konkuk University

Calibration and Validation

ParaSol

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.81, E= 0.72

R2= 0.79, E= 0.67

R2= 0.92, E= 0.90

CHUNCHEON GLOBAL WATER FORUM 2009

17/26

Konkuk University

Calibration and Validation

Objective Function and 95PPUs for streamflow

R2

No. 5

No. 7

No. 15

No. 5

No. 7

No. 15

SUFI2

0.71

0.79

0.91

0.59

0.67

0.90

GLUE

0.80

0.79

0.93

0.72

0.67

0.90

ParaSol

0.81

0.79

0.92

0.72

0.67

0.90

Subbasin

p-factor

r-factor

No. 5

No. 7

No. 15

No. 5

No. 7

No. 15

SUFI2

0.33

0.66

0.79

0.30

0.38

0.52

GLUE

0.26

0.56

0.44

0.30

0.35

0.37

ParaSol

0.02

0.14

0.07

0.03

0.03

0.05

Subbasin

CHUNCHEON
17 / 18 GLOBAL WATER FORUM 2009
18/26

Konkuk University

Calibration and Validation

Comparing Each Result in Random Section

The 95PPU bands (model uncertainty):
green color range were caused by the
variable parameter set range.

SUFI2

Sensitive and
well distributed

GLUE

The SUFI2 was the most suitable way

to find the SWAT Uncertainty under
the condition that the parameter
range was specified. Responded to
parameter set more sensitive than any
others.
CHUNCHEON GLOBAL WATER FORUM 2009
19/26

ParaSol

Konkuk University

Sensitivity Analysis (SA)

t-Stat

<SS>

<T-N>

<T-P>

CHUNCHEON GLOBAL WATER FORUM 2009

20/26

Konkuk University

Sensitivity Analysis (SA)

P-Value

<SS>

<T-N>

<T-P>

CHUNCHEON GLOBAL WATER FORUM 2009

21/26

Konkuk University

Calibration and Validation

SUFI2

<SS>

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.87, E= 0.85

<T-N>
R2= 0.70, E= 0.53

<T-P>
R2= 0.87, E= 0.81

CHUNCHEON GLOBAL WATER FORUM 2009

22/26

Konkuk University

Calibration and Validation

GLUE

<SS>

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.87, E= 0.85

<T-N>
R2= 0.62, E= 0.53

<T-P>
R2= 0.87, E= 0.82

CHUNCHEON GLOBAL WATER FORUM 2009

23/26

Konkuk University

Calibration and Validation

ParaSol

<SS>

R2 : Determination Coefficient
E2 : Model Efficiency

R2= 0.22, E= 0.15

<T-N>
R2= 0.28, E= 0.22

<T-P>
R2= 0.06, E= -0.09

CHUNCHEON GLOBAL WATER FORUM 2009

24/26

Konkuk University

Calibration and Validation

Objective Function and 95PPUs for Water Quality

R2

SS

T-N

T-P

SS

T-N

T-P

SUFI2

0.87

0.70

0.87

0.85

0.53

0.81

GLUE

0.87

0.62

0.87

0.85

0.53

0.82

ParaSol

0.22

0.28

0.06

0.15

0.22

-0.09

p-factor

r-factor

SS

T-N

T-P

SS

T-N

T-P

SUFI2

0.67

0.54

0.23

0.53

2.73

1.30

GLUE

0.40

0.17

0.10

0.35

0.30

0.25

ParaSol

0.19

0.22

0.07

0.15

0.40

0.15

CHUNCHEON
17 / 18 GLOBAL WATER FORUM 2009
25/26

Konkuk University

Summary and Conclusion

This study tried to evaluate the prediction uncertainty of SWAT
model for streamflow and water quality.

As a result, there was no significant difference in the R2, Nash-Sutcliffe

coefficient (NSE) values for each procedure.

However, the p-factor and r-factor indicated that the numerical 95PPUs
bands showed clear distinction in this analysis. In this study, ParaSol
algorithm showed the lowest p-factor and r-factor, while SUFI-2 algorithm
was the highest in the p-factor and r-factor.

The SUFI2 algorithm showed the most sensitive and evenly distributed
95PPUs bands for streamflow and water quality.

With the SUFI2 algorithm, we could assess the most sensitive results
and reduce model uncertainty effectively.

This study gave us an obvious information that the evaluation of

model uncertainty caused by SWAT hydrologic modeling could be
predictable and quantified by SWAT-CUP.

CHUNCHEON
18 / 18 GLOBAL WATER FORUM 2009
26/26

Konkuk University

This work was supported by Mid-career Researcher Program through NRF grant funded
by the MEST (No. 2009-0080745).

Thank You
For further information, please contact:
JOH, Hyung-Kyung
Graduate student, Dept. of Civil & Environmental System Engineering, Konkuk University
[email protected]

Dept. of Civil and Environmental System Eng.,

Konkuk University, Seoul, South Korea
CHUNCHEON GLOBAL WATER FORUM 2009

Konkuk University