Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Water Resour Manage (2010) 24:2869–2880

DOI 10.1007/s11269-010-9584-1

Predictive Analysis and Simulation Uncertainty


of a Distributed Hydrological Model

Abdolreza Bahremand · Florimond De Smedt

Received: 7 December 2008 / Accepted: 8 January 2010 /


Published online: 21 January 2010
© Springer Science+Business Media B.V. 2010

Abstract Distributed hydrological models should pass through a careful calibration


procedure backed by sensitivity, uncertainty and predictive analysis before they are
utilized as a decision making aid in watershed management and scenario studies. This
paper examines whether the uncertainty of the parameters of the spatially distributed
hydrologic model WetSpa causes significant uncertainty in the model predictions.
The WetSpa model is applied to the Torysa river basin, a rather large catchment
located in Slovakia. Parameter estimation, sensitivity and predictive analysis of the
model parameters are performed using a model-independent parameter estimator,
PEST. It is found that the correction factor for measured evaporation data has the
highest relative sensitivity. Parameter uncertainty and predictive analysis give an
insight of a proper parameter set and parameter uncertainty intervals and prove
that the parameter uncertainty of the model does not result in a significant level
of predictive uncertainty.

Keywords WetSpa model · PEST · Sensitivity/uncertainty analysis ·


Predictive analysis · Flow simulation

1 Introduction

Distributed hydrological models are usually parameterized by deriving estimates of


parameters from the topography and physical properties of the soils, aquifers and
land use of the basin. Distributed models might also include conceptual parameters,

A. Bahremand (B)
Department of Watershed Management, Gorgan University of Agricultural Sciences
and Natural Resources, Beheshti 386, Gorgan, Iran
e-mail: Abdolreza.Bahremand@yahoo.com, abdolreza.bahremand@gmail.com

F. De Smedt
Department of Hydrology and Hydraulic Engineering, Vrije Universiteit Brussel,
Pleinlaan 2, 1050 Brussels, Belgium
2870 A. Bahremand, F. De Smedt

with properties that cannot be assessed by physical measurements or characteristics


of the system. The reliability of model predictions depends on how well the model
structure is defined and how well the model is parameterized. However, estimation of
model parameters is difficult due to the large uncertainties involved in determining
the parameter values, which can not be directly measured in the field. Therefore
model calibration is necessary to improve the model performance (Liu 2004). Un-
fortunately, model calibration does not guarantee reliability of model predictions.
The parameter values obtained during calibration and the subsequent predictions
made using the calibrated model are only as realistic as the validity of the model
assumptions for the study watershed and the quality and quantity of actual watershed
data used for calibration and simulation. Therefore, even after calibration, there is
potentially a great deal of uncertainty in results that arises simply because it is too
unlikely to find error-free observational data (e.g. precipitation, streamflow, topog-
raphy) and because no simulation model is an entirely true reflection of the physical
process being modeled (Muleta and Nicklow 2004). There are some automated meth-
ods designed for autocalibration, optimization, uncertainty and sensitivity analysis.
The methods are traditionally grouped into two general classes: global sampling
based methods (global methods) and local point-estimation method (local methods).
Global methods perform a global analysis by running simulations using parameter
values sampled over the entire parameter space. Point-estimation methods search
locally to find an optimum point till no improvements are found in the neighborhood
(van Griensven 2002). To perform calibration and uncertainty analysis, in recent
years many procedures and program tools such as Parameter ESTimator (PEST),
Generalized Likelihood Uncertainty Estimation (GLUE) (Beven and Binley 1992),
Parameter Solution (ParaSol) (van Griensven and Meixner 2006), and Sequential
Uncertainty FItting (SUFI-2) (Abbaspour et al. 2007) have become available. PEST
a non-linear Parameter ESTimator (Doherty 2005) is an optimization tools based
on a local search method. PEST can be used to estimate parameters and carry out
various predictive analysis tasks and sensitivity analysis, for a wide range of model
types.
This paper mainly examines whether the uncertainty of the parameters of the
spatially distributed hydrologic model WetSpa causes significant uncertainty in the
model predictions. PEST is used for the model optimization and the predictive
analysis. In this study, the WetSpa model is applied to a rather large catchment
located in Slovakia.

2 WetSpa Model

The WetSpa model used in this study is a grid-based distributed runoff and water
balance simulation model that runs on an hourly or daily time step. The model is
suitable for use of flood prediction and watershed management on catchment scale.
It predicts hourly or daily overland flow occurring at any point in a watershed,
and provides spatially distributed hydrologic characteristics in the basin. Inputs to
the model include digital elevation data, soil type, land use data, precipitation and
evaporation time series. Stream discharge data is optional for model calibration.
The model is physically based and simulates hydrological processes of precipitation,
snowmelt, interception, depression, surface runoff, infiltration, evapotranspiration,
Analysis and Uncertainty of a Distributed Hydrological Model 2871

percolation, interflow, groundwater flow, etc. continuously both in time and space,
for which the water and energy balance are maintained on each raster cell. Surface
runoff is produced using a modified coefficient method based on the cell character-
istics of slope, land use, and soil type, and allowed to vary with soil moisture, rainfall
intensity and storm duration. Interflow is computed based on the Darcy’s law and
the kinematic approximation as a function of the effective hydraulic conductivity and
the hydraulic gradient, while groundwater flow is estimated with a linear or nonlinear
reservoir method on a small subcatchment scale as a function of groundwater storage
and a recession coefficient. Special emphasis is given to the overland flow and
channel flow routing using the method of linear diffusive wave approximation, which
is capable to predict flow discharge at any converging point downstream by a unit
response function. The model accounts for spatially distributed hydrological and
geophysical characteristics of the catchment and therefore is suitable for studying the
impact of land use change on the hydrological behaviours of a river basin. The model
structure and formulation has been documented and published in several Ph.D.
theses and papers (e.g. Wang et al. 1997; De Smedt et al. 2000; Liu 2004; Liu and De
Smedt 2004; Bahremand 2006; Bahremand et al. 2007).TheWetSpa distributed model
involves a large number of model parameters to be specified during the model setup.
Most of these parameters can be assessed from field data, e.g. hydrometeorological
observations, maps of topography, soil types, and land use, etc. However, compre-
hensive field data are seldom available to fully support specification of all model
parameters. Distributed model parameters are derived during the preprocessing
phase of the modeling process. They are derived from three basic input maps: i.e.
digital elevation model, landuse and soil type. The distributed parameters are about
20 parameters. In addition, some model parameters are of a more conceptual nature
and cannot be directly assessed. In particular, correction factors were introduced in
the model to compensate for the lack of precise field data and particular condition
that might be present in the catchment. This approach enables to simplify the
application of automated calibration procedures to improve the model performance.
In total, 11 global or lumped parameters are used in Wetspa for simplify the
process of parameter calibration. The major model parameters that can be calibrated
are as follow: (1) Ki , interflow scaling factor [–]; (2) Ks , initial soil moisture [–];
(3) Ke , correction factor for evaporation data [–]; (4) Kg , groundwater reces-
sion coefficient [day−1 ]; (5) Kgi , initial active groundwater storage [mm]; (6) Kgm ,
Maximum active groundwater storage [mm]; (7) Km , surface runoff exponent [–];
and (8) Kp , maximum rainfall intensity [mm day−1 ]. All other model parameters
(the distributed parameters) are automatically derived using GIS tools and are kept
constant. The choice of parameters to calibrate is based on earlier studies of the
WetSpa model (Liu 2004). The detailed description of the model parameters has
been already published by the authors (Bahremand and De Smedt 2008).

3 Application

TheWetSpa model is applied to the Torysa river basin (1,297 km2 ) located in
Slovakia. Daily hydrometeorological data from 1991 to 2000, including precipitation
data from 14 stations, temperature data from two stations and evaporation data
measured at one station are used as input to the model. The spatial characteristics
2872 A. Bahremand, F. De Smedt

120 0

80 20
Q (m3/s)

P (mm)
Precipitation
Q measured
40 Q simulated 40

0 60
1 31 61 91 121 151 181 211 241 271 301 331 361
Time (d)

Fig. 1 Graphical comparison between observed and calculated daily flow at Kosicke Olsany for the
year 1997

of the basin are calculated using three base maps, i.e. DEM, landuse and soil type,
which are in GIS form with 100 m cell size. Results of the simulations show a good
agreement between calculated and measured hydrographs at the outlet of the basin.
The model predicts the daily discharge values with a good accuracy, i.e. about 73%
according to the Nash–Sutcliffe criterion. Figure 1 gives a graphical comparison
between observed and calculated daily flow at the outlet (Kosicke Olsany station)
for the year 1997, showing that both the spring and summer flood hydrographs are
well reproduced by the model. Details of this simulation including full description of
the study area, input data and maps, model preparation, auto calibration, validation
and a discussion on the results of the simulation has been published in the PhD thesis
and papers of the main author (Bahremand 2006; Bahremand and De Smedt 2008).

4 Parameter Optimization

Automatic calibration, and sensitivity and uncertainty analysis of the model pa-
rameters are performed using a model-independent parameter estimator, PEST.
PEST is a nonlinear parameter estimation and optimization package, which offers
model independent optimization routines (Doherty 2005). It applies a robust Gauss–
Marquardt–Levenberg algorithm, which combines the advantages of the inverse
Hessian method and the steepest descent method and therefore provides faster and
more efficient convergence towards the objective function minimum. PEST uses a
local calibration method and searches for the best values of the model parameters by
minimizing the sum of squares of the differences between measured and calculated
model results. When undertaking this task, PEST is run in parameter estimation
mode or calibration mode. This paper presents a strategy by incorporating PEST
Analysis and Uncertainty of a Distributed Hydrological Model 2873

1.8
High sensitivity Low sensitivity

Upper and lower limits of parameter


1.6

estimates / Best estimate 1.4

1.2

0.8

0.6

0.4

0.2

0
Ke Km Ks Ki Kg Kp Kgi Kgm
Parameters

Fig. 2 Margin of uncertainty of each parameter

for automatic calibration, and sensitivity and uncertainty analysis. The results of
this study demonstrate that the use of combining a GIS-based hydrological model
with PEST can generate calibrated parameters that produce fit (or realistic) simula-
tions. The ability to calculate the derivatives of all observations with respect to all
adjustable parameters is fundamental to the Gauss–Marquardt–Levenberg method
of parameter estimation; these derivatives are stored as the elements of the Jacobian
matrix used for sensitivity analysis. The relative sensitivity of the eight major model
parameters is given using the PEST automated calibration scheme. In Fig. 2 para-
meters are sorted based on their sensitivity rank. The parameter sensitivity value is
expressed by the relative composite sensitivity obtained by multiplying its composite
sensitivity by the magnitude of the value of the parameter. The composite sensitivity
of each parameter is the normalized (with respect to the number of observations)
magnitude of the column of the Jacobian matrix pertaining to that parameter. The
Jacobian matrix comprised of m rows (one for each observation), and the n elements
of each row are the derivatives of one particular observation with respect to each of
the n parameters. The margin of uncertainty of each parameter is illustrated using the
95% confidence intervals of the calibrated parameters. It is found that the correction
factor for measured evaporation data Ke has the highest relative sensitivity, because
the margin of uncertainty of this parameter is the smallest, while parameter Kgm
has the highest uncertainty. The PEST results of sensitivity and uncertainty of the
model parameters have been already published and discussed in detail by the authors
(Bahremand and De Smedt 2008). However, Fig. 2 summarizes and depicts these
results in another way, i.e. to be able to compare the margin of uncertainty of each
parameter, the lower limit and upper limit of each estimate is divided to its best
estimates.
2874 A. Bahremand, F. De Smedt

5 Predictive Analysis

5.1 Methodology

High levels of parameter uncertainty can result from a poor fit between model
outcomes and field observations, from a high level of parameter correlation, from
insensitivity on the range of certain parameters, or from all of these. The first
possibility indicates that either field data is poor or that the model is inappropri-
ate. Parameter uncertainty resulting from high levels of parameter correlation or
caused by parameter insensitivity may or may not result in high levels of predictive
uncertainty (Doherty 2005). The extent of predictive uncertainty can be estimated
using PEST’s predictive analyzer. Predictive analysis is another use of PEST, which
is done by its “predictive mode”. When used to calibrate a model (the traditional
use of PEST), PEST is asked to minimize an objective function comprised of the
sum of weighted squared deviations between certain model outcomes and their
corresponding field-measured counterparts. When undertaking this task, PEST is run
in “parameter estimation mode”. PEST’s operation in predictive analysis mode has
much in common with its operation in parameter estimation mode. The concepts of
prediction analysis done by PEST is given in Appendix A which is summarized from
the manual of PEST (Doherty 2005).
To do predictive analysis by changing parameters (the way PEST does it), they
must be varied in such a way that the objective function hardly changes (i.e. keeping
the model under calibration conditions). To run PEST in predictive analysis mode
the user informs PEST of a value of the objective function below which the model can
be considered to be calibrated; this value is normally just slightly above the minimum
objective function value as determined in a previous PEST calibration run. The sum
of squared residuals of the discharge, , is the objective function in this optimization.
When run in predictive analysis mode, PEST must be supplied with the value of
min + δ (where δ is relatively small), henceforth referred to as 0 . This, 0 , is a
value for the objective function which, under calibration conditions, is considered
sufficient to “just calibrate” the model. A PEST predictive analysis run should be
preceded by a parameter estimation run in which min is determined. This minimum
objective function, min , obtained during the calibration is 78,631 m6 /s2 . The user
then decides on a suitable value for δ and hence 0 for a PEST predictive analysis
run. Obviously 0 should be greater than min ; however in most circumstances it
should only be a little bit greater. The value of the δ was calculated as 75 m6 /s2 based
on the formula given in the manual of PEST (Appendix A). This is about 0.1% of
min and therefore 0 was set to 78,706 m6 /s2 . PEST then maximizes or minimizes the
model predictions (the PEST user decides which), while ensuring that the objective
function is as close as possible to 0 .
As model setup for predictive analysis, a composite model is constructed that is
comprised of the model run under calibration conditions followed by the model run
under predictive conditions. There can be as many field observations (discharges)
corresponding to the former model component as desired, these being the “cali-
bration observations”. However, there should be only one output for the predictive
model component, which is used by PEST as a target observation. This observation,
to be recognized by PEST, should be the sole member of an observation group
named “predict”. It is important to note that PEST takes no notice of either the
“observed value” of this observation or of the weight assigned to this observation.
Analysis and Uncertainty of a Distributed Hydrological Model 2875

Table 1 Key target predictions, objective function, analysis of residuals and model efficiency for the
calibration and prediction mode of the model
Criteria Units Calibration Minimum Maximum
prediction prediction
Key target prediction m3 /s 59.24 53.24 64.12
Objective function (sum of squared residuals) m6 /s2 78,631 78,764 79,501
Mean residual m3 /s 0.400 0.400 0.350
Maximum residual m3 /s 75.75 75.50 73.62
Minimum residual m3 /s −35.63 −34.56 −37.57
Standard deviation of residuals m3 /s 4.64 4.64 4.65
Model efficiency (Nash–Sutcliffe criterion) – 0.73 0.73 0.73

PEST’s job is simply to raise or lower the model output corresponding to this
observation, while maintaining the objective function at or below min + δ.

5.2 Analysis

In this research the sole number of the model outputs (named as “key model
prediction”), which was considered as the predict group, was the highest peak
discharge of the year 1997, i.e. the summer flood which occurred on the 11th of July
1997 (Fig. 1). The measured discharge of this flood is 90.72 m3 /s, but the calibrated
model estimates this peak as 59.24 m3 /s. After calibration, the model is run in the
prediction mode twice, i.e. once PEST has minimized the objective function to get

Fig. 3 Calibration and prediction discharges against observations for the 10-year simulation period
2876 A. Bahremand, F. De Smedt

100 11-July

80 Min Prediction
Max Prediction
Calibration
Measured
60
Q (m3/s)

40

20

0
0 10 20 30 40 50 60
Time (d)

Fig. 4 Calibrated model results bounded by minimum and maximum prediction for the target flood
which occurred on the 11th of July 1997

the best prediction, PEST is asked to maximize or minimize the target prediction
maintaining the model calibration (see the Appendix A). The predictive mode yields
a maximum prediction of 64.12 m3 /s and a minimum value of 53.24 m3 /s for this key
target. The comparison between calibration and predictive mode for the objective
function, analysis of residuals, and the model efficiency, are presented in Table 1. The
obtained model efficiency for the two performance modes, i.e. those of calibration
mode and of prediction mode, show that the uncertainty of the parameters does not
result in high levels of predictive uncertainty. The model efficiency for the calibration
mode and for the predictive mode is the same, i.e. 73%. This value gives an idea of
the error and disagreement between model results and the observation. The model
evaluation for this simulation, based on different criteria, has been discussed in
Bahremand (2006); Bahremand and De Smedt (2008).

Table 2 The calibrated parameters and the estimated parameters after running PEST in minimum
predictive mode
Parameter Calibration mode Minimum prediction mode t test result
Estimated value Standard deviation Estimated value Standard deviation
(mx ) (sx ) (my ) (sy )
Ki 5.90 0.087 5.88 0.090 mx = my
Ks 0.95 0.015 0.96 0.015 mx = my
Ke 1.06 0.005 1.07 0.006 mx = my
Kg 7.5 × 10−5 4.2 × 10−6 7.7 × 10−5 4.9 × 10−6 mx = my
Kgi 34.5 5.26 32.2 5.37 mx = my
Kgm 1,300 414 1,649 148 mx = my
Km 2.25 0.023 2.30 0.022 mx = my
Kp 660 58.6 444 87.1 mx = my
Analysis and Uncertainty of a Distributed Hydrological Model 2877

Table 3 The calibrated parameters and the estimated parameters after running PEST in maximum
predictive mode
Parameter Calibration mode Maximum prediction mode t test result
Estimated value Standard deviation Estimated value Standard deviation
(mx ) (sx ) (my ) (sy )
Ki 5.90 0.087 6.04 0.089 mx = my
Ks 0.95 0.015 0.94 0.014 mx = my
Ke 1.06 0.005 1.05 0.005 mx = my
Kg 7.5 × 10−5 4.2 × 10−6 7.9 × 10−5 4.5 × 10−6 mx = my
Kgi 34.5 5.26 35.1 5.08 mx = my
Kgm 1,300 414 1,187 568 mx = my
Km 2.25 0.023 2.19 0.020 mx = my
Kp 660 58.6 1,033 304 mx = my

Figure 3 plots the calibration and prediction discharges against the observations
for the whole 10-year simulation period together with their linear regression equa-
tions. The figure shows that prediction results are not significantly different from
the calibration results. This implies that the parameter uncertainty does not imply
significant uncertainty to the model outputs and its predictions.
Figure 4 shows how the prediction bound around the optimized model results.
While PEST is run under predictive mode, it is either asked to minimize the
prediction which gives the best prediction results or to maximize it, and it can be
obviously seen in the figure that the minimum prediction is closer to the calibrated
results. The maximum prediction not only gives a higher, thus more realistic, value
of the peak discharge, but it has also a better performance as its line is closer to the
median. In fact it illustrates that a parameter set can give better model performance
according to one criterion (i.e. peak discharge for the maximum prediction set), while
another one can show a better overall performance on the general hydrograph (i.e.
the calibrated parameter set).
Tables 2 and 3 show the two parameter sets for the calibration results and the
minimum and maximum prediction results. The two sets are statistically the same.
The difference between the two sets (whether the two groups differ from each other)
is verified using the t test, i.e. the two sample t test with unequal variances (Helsel
and Hirsch 1992). The t test shows the new values for all parameters do not differ
from the previous calibrated values, as their 95% confidence intervals overlap each
other.
It is a fact of model usage that there are often many different sets of parameter
values for which the objective function is at its minimum or almost at its minimum.
Thus there are many different sets of parameters which could be considered to
calibrate a model. The results of this study shows that there is only one parameter set
which gives the best results and optimizes the model (considering the uncertainty of
the set).

6 Discussions and Conclusions

In general, the more similarity of condition between model calibration and model
prediction, the greater is the likelihood that parameter uncertainty arising as an
2878 A. Bahremand, F. De Smedt

outcome of the calibration process will not result in a large degree of predictive un-
certainty. If parameter uncertainty does not considerably diminish the model’s ability
to replicate historical system conditions, then it will quite possibly not undermine its
ability to predict future conditions, provided things are not too different in the future.
However where conditions are different, and where predictions are of a different
type from those used in the calibration process, it is possible that a large degree of
predictive uncertainty may be associated with model parameter uncertainty. Hence
parameter uncertainty is, in itself, not a big issue. Parameter uncertainty is important
only in so far as it affects predictive uncertainty, and this depends on the prediction
(Doherty 2005).
The results of this study demonstrate that the use of combining a GIS-based
hydrological model with PEST can generate calibrated parameters that produce fit
(or realistic) simulations. The relative sensitivity of eight lumped parameters is given
using the PEST automated calibration scheme. It is found that the correction factor
for calculating the actual evapotranspiration from the potential evapotranspiration
has the highest relative sensitivity. The margin of uncertainty of each parameter
is determined as it is found that the estimation of Parameter Kgm has the highest
uncertainty. In this research, the PEST’s algorithms and techniques were imple-
mented to verify whether these parameters’ uncertainties can actually matter in
terms of the model’s ability for predicting future, and to see if any usage of the
model for predictive scenarios should be doubted or inversely shall be permitted. The
results show the uncertainty of outputs caused by the parameter uncertainty is not
significant. Parameter uncertainty and predictive analysis give an insight of a proper
parameter set and parameter interval and prove that the parameter uncertainty of the
model does not result in a significant level of predictive uncertainty. About the PEST
results it should be noted that PEST is a local search method and does sensitivity and
parameter uncertainty analysis around the optimal parameter set. Therefore, these
analyses are limited to the vicinity of the optimal parameters and also affected by the
initial parameter set which is given to the PEST to start the parameter estimation.
To have a better idea of the model parameter sensitivity and uncertainty the use of a
global search method is recommended.

Acknowledgements The authors thank the anonymous reviewers whose comments and sugges-
tions very much improved the paper.

Appendix A: How PEST does the Predictive Analysis?

The reference of this appendix is the manual of PEST written by Doherty (2005).

A.1. The Formulation

Let X represent the action of a linear model under calibration conditions. Let b
represent the parameter vector for this model, while c is a vector of field or laboratory
observations for which there are model-generated counterparts. When calibrating a
model, it is PEST’s task to minimize an objective function defined as:

 = (c − Xb)t Q (c − Xb)
Analysis and Uncertainty of a Distributed Hydrological Model 2879

where Q is the “cofactor matrix”, a diagonal matrix whose elements are the squares
of observation weights. Let Z represent the same linear model under predictive
conditions. Thus the action of the model when used in predictive mode can be
represented by the equation: d = d = Zb, where d is a 1 × 1 vector (i.e. a scalar,
d) representing a single model outcome (i.e. prediction). Naturally, when run in
predictive mode, the model operates on the same parameter vector as that on which
it operates in calibration mode, i.e. b. The aim of predictive analysis is to maximize
(minimize) d while “keeping the model calibrated”. d will be maximized (minimized)
when the objective function associated with b lies on the min + δ contour. min
is the lowest achievable value of the objective function, while δ is an acceptable
increment to the objective function minimum such that the model can be considered
calibrated as long as the objective function is less than min + δ (more discussion and
further details in A.2). Thus the predictive analysis problem can be formulated as
follows: Find b such as to maximize (minimize) Zb subject to (c − Xb)t Q (c − Xb) =
0 where 0 = min + δ. It can be shown that the solution to this problem is:
 
 −1 Z
b = XT QX Xt Qc −

where λ is defined by the equation:
 2  −1 t
1 0 − ct Qc + ct QX Xt QX X Qc
=
2λ Zt (Xt QX)−1 Z
Where predictive analysis is carried out for a nonlinear model the same equations
are used. However in this case X is replaced by the model Jacobian matrix J, and
a parameter upgrade vector is calculated instead of a solution vector. The solution
process is then an iterative one in which the true solution is approached by repeated
calculation of an upgrade vector based on repeated linearization of the problem
through determination of the Jacobian matrix. Use of the Marquardt lambda in
solving the nonlinear parameter estimation problem is discussed in the manual of
PEST. Its use in solving the nonlinear predictive analysis problem is very similar.
As in the parameter estimation problem, PEST continually adjusts the Marquardt
lambda through the solution process such that its value is optimal at all stages of this
process. As a further numerical measure to solve the predictive analysis problem for
a nonlinear model, PEST undertakes a line search in the direction of the parameter
upgrade vector to determine the point at which this vector crosses the 0 contour.

A.2. What Should δ be?

Theory is available for calculating δ, the objective function increment for which the
model can be considered to be “uncalibrated”. At a probability level of 1 − α, δ is
given by:

δ = nσ2 Fα (n, m − n)

where σ2 is given by equation σ2 = min /(m − n) where n is the number of parame-


ters requiring estimation, m is the number of observations comprising the calibration
dataset, and F(.,.) denotes the F distribution (more details in the PEST manual,
Doherty 2005).
2880 A. Bahremand, F. De Smedt

References

Abbaspour KC, Yang J, Maximov I, Siber R, Bogner K, Mieleitner J, Zobrist J, Srinivasan R (2007)
Spatially-distributed modelling of hydrology and water quality in the pre-alpine/alpine Thur
watershed using SWAT. J Hydrol 333:413–430
Bahremand A (2006) Simulating the effects of reforestation on floods using spatially distributedhy-
drologic modeling and GIS. PhD thesis, Vrije Universiteit Brussel, Belgium
Bahremand A, De Smedt F (2008) Distributed hydrological modeling and sensitivity analysis in
Torysa watershed, Slovakia. Water Resour Manag 22:393–408
Bahremand A, De Smedt F, Corluy J, Liu YB, Poórová J, Velcická L, Kuniková E (2007) WetSpa
model application for assessing reforestation impacts on floods by in Margecany–Hornad water-
shed, Slovakia. Water Resour Manag 21:1373–1391
Beven K, Binley A (1992) The future of distributed models—model calibration and uncertainty
prediction. Hydrol Process 6(3):279–298
De Smedt F, Liu YB, Gebremeskel S (2000) Hydrological modeling on a catchment scale using GIS
and remote sensed land use information. In: Brebbia CA (ed) Risk analysis II. WTI, Boston,
pp 295–304
Doherty J (2005) PEST: model independent parameter estimation, user manual, 5th edn. Watermark
Numerical Computing, Brisbane
Helsel DR, Hirsch RM (1992) Statistical methods in water resources. Elsevier, New York, p 126
Liu YB (2004) Development and application of a GIS-based hydrological model for flood prediction
and watershed management. PhD Thesis, Vrije Universiteit Brussel, Belgium
Liu YB, De Smedt F (2004) WetSpa extension, documentation and user manual, department of
hydrology and hydraulic engineering. Vrije Universiteit Brussel, Belgium
Muleta MK, Nicklow JW (2004) Sensitivity and uncertainty analysis coupled with automatic calibra-
tion for a distributed watershed model. J Hydrol 306:127–145
van Griensven A (2002) Developments towards integrated water quality modeling for river basins.
PhD thesis, Vrije Universiteit Brussel, Belgium
van Griensven A, Meixner T (2006) Methods to quantify and identify the sources of uncertainty for
river basin water quality models. Water Sci Technol 53(1):51–59
Wang Z, Batelaan O, De Smedt F (1997) A distributed model for water and energy transfer between
soil, plants and atmosphere (WetSpa). Phys Chem Earth 21:189–193

You might also like