Middle-East Journal of Scientific Research 15 (9): 1255-1263, 2013

ISSN 1990-9233
© IDOSI Publications, 2013
DOI: 10.5829/idosi.mejsr.2013.15.9.11470

Application of Three Different Calibration-Uncertainty Analysis Methods

in a Semi-Distributed Rainfall-Runoff Model Application
1
Mohsen Pourreza Bilondi, 2Karim C. Abbaspour and 3Bijan Ghahraman

Department of Water Engineering, University of Birjand, Birjand, Iran

1

Swiss Federal Institute for Aquatic Science and Technology, EAWAG,

2

Ueberlandstrasse 133, P.O. Box 611, CH-8600, Dübendorf, Switzerland

3
Department of Water Engineering, College of Agriculture,
Ferdowsi Univ. of Mashhad, Mashhad, Iran

Abstract: Often parameters in hydrologic models cannot be measured directly and can only be inferred by a
calibration process. This work addresses the application and comparison of three parameter uncertainty
methods and their effects on the prediction of streamflow in Qezel Ozan watershed (43,000 km2) located in
northwestern Iran. Methods of Partcle Swarm Optimization (PSO), DiffeRential Evolution Adaptive Metropolis
(DREAM) and Sequential Uncertainty Fitting ver. 2 (SUFI2) were used in this study to calibrate a rainfall-runoff
model created using the Soil and Water Assessment Tool (SWAT). The calibration and validation results
indicated statistically insignificant differences among the three algorithms. The main difference was in their
number of runs, where DREAM converged after 36000 runs, PSO after 6000 runs and SUFI2 after 1500 runs.
SUFI2 proved to be a very efficient optimization algorithm, while PSO had the largest NS for calibration (0.59)
and validation (0.74) periods.

Key words: Uncertainty analysis SWAT Differential Evolution Adaptive Metropolis algorithm PSO
SUFI2

INTRODUCTION activities in the watershed and model inaccuracy due to

over-simplification of the processes considered in the
Rainfall-runoff models are widely used in Hydrology model. Parameter uncertainties arise due to a large number
to simulate river basin water quality and quantity and play of unknown parameters in distributed models.
an important role in management of water resources. Furthermore, errors may come from the imprecise
Hydrologic models such as HSPF (Hydrological measured data used for calibration [6].
simulation program- FORTRAN) [1], SHETRAN [2] and Different calibration-uncertainty analysis techniques
SWAT (Soil and Water Assessment Tool) [3] require have different levels of mathematical complexity and data
many parameters that cannot be measured directly and requirements. Topliceanu [7] classified them into two main
must only be estimated by calibration against a historical categories: analytical approaches and approximation
record of measured output data. Due to the problem of approaches. The selection of an appropriate technique to
non-uniqueness the uncertainty in the model prediction be used depends on the nature of the problem including
must also be estimated [4]. The major sources of availability of information, resources constraints, model
uncertainty are input data, model structure and model complexity and type and accuracy of desired results. As
parameters [5]. Input uncertainty is often related to most of the models used in hydrology are nonlinear and
imprecise or spatially interpolated measurements of highly complex, the analytical techniques do not apply
model input and lack of knowledge of initial conditions. because they are rather restrictive in practical
Sources of model structural uncertainty include processes applications. Yang et al. [6] divided calibration techniques
not accounted for in the model such as unknown into three main categories: (i) approaches without

Corresponding Author: Mohsen Pourreza Bilondia, Department of Water Engineering, University of Birjand, Birjand, Iran.

1255
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

rigorous statistical assumptions such as GLUE [8] and

SUFI2 [9, 10] (ii) approaches that account for the effect of
input errors on the output by an additive error model
which introduces temporal correlation of the residuals like
autoregressive error models [11,12] and (iii) methods that
use improved likelihood functions that explicitly represent
input errors and/or model structural error of the
underlying hydrological model [6].
Despite the large number of suggested techniques,
only a few papers on comparison of different uncertainty
analysis techniques are available [6,13,14]. Yang et al. [6]
compared GLUE, parameter Solution (ParaSol) [15], SUFI2
[10] and MCMC [4,12,16,17] methods in an application to
a watershed in China with SWAT model and found that
different methods converge to different solutions with
more or less the same calibration and validation results.
Zhang et al [18] also compared five global optimization
algorithms (genetic algorithms, shuffled complex
evolution, particle swarm optimization, differential
evolution and artificial immune system) and showed that
particle swarm optimization (PSO) can obtain better
parameter solutions than other algorithms given fewer
number of model runs (less than 2000).
In this paper we apply a relatively new MCMC Fig. 1: Location of study site as well hydrometric stations,
procedure entitled DiffeRential Evolution Adaptive rain and temperature stations
Metropolis (DREAM) [19]. We applied DREAM to a
rainfall-runoff model built using SWAT and compare the daily discharge for over a 30-years period (1975-2004) at
results with PSO [20] and SUFI2, where the latter is the basin outlet is 60 m3s 1 and varies irregularly between
reported by Yang et al. [6] to need a smaller number of 790 m3 s 1 during flood season to 17 m3 s 1 in the dry
model runs in comparison with other methods. As Yang season. Grassland covers more than 85% of the area.
et al. [6] explained there are various difficulties in Figure 1 shows the location of the watershed along with
comparing uncertainty analysis techniques. In this paper, temperature gauge, rain gauge and discharge stations.
we compare the final calibration and validation statistics
(e.g. R2, NS, MSE) and the uncertainty statistics Hydrological Model (SWAT): The Soil and Water
suggested by Abbaspour et al. (2004, 2007) [9,10] Assessment Tool (SWAT) [3] was used in the current
(e.g. R-factor and P-factor), which compare the study. We chose SWAT because of its availability and
observation signal with the 95% prediction uncertainty user-friendliness in handling input data. Swat is
band. continuous time, spatially distributed simulator of water,
sediment, nutrients and pesticides transport at a
MATERIALS AND METHODS catchment scale. In SWAT, a watershed is divided into a
number of sub-basins based on given digital elevation
Description of the Study Area: This study was conducted model (DEM) and then a sub-basin is allowed to be
in a section of Qezel Ozan watershed (Fig 1) with a divided into a number of unique hydrologic response
drainage area of 43000 km2 located in northwestern Iran. units (HRU) based on variability in soil, slope and land
The river originates from the mountains of Kurdistan use characteristics.
(from South to north) and discharges to the Caspian Sea. The hydrologic model is based on the water balance
The average annual precipitation, average daily minimum for the four storage volumes snow, soil profile (0-2 m),
and maximum temperatures respectively are 319 mm, shallow aquifer (2-20 m) and deep aquifer (>20 m). The
5.5°C and 20.5 °C. According to De Martonne formula this simulated processes include surface runoff, infiltration,
region has a climate zone of arid to semi-arid. Average evaporation, plant water uptake, lateral flow and

1256
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

Table 1: The 9 most sensitive parameters for uncertainty analysis and their initial ranges
1Initial range
------------------------------------------------
Aggregate parameters Description Min Max
r__SOL_BD(1).sol Soil bulk density (g cm 3) -0.8 0.8
r__SOL_AWC(1).sol Soil available water storage capacity (mm H 2O/mm soil) -0.9 0.9
v__ALPHA_BNK.rte Base flow alpha factor for bank storage (days) 0 1
v__CH_K2.rte Effective hydraulic conductivity in the main channel (mm h 1) 0 150
v__CH_N2.rte Manning's n value for main channel 0 0.3
v__ESCO.hru Soil evaporation compensation factor 0.01 1
r__CN2.mgt SCS runoff curve number for moisture condition II -0.7 0.7
r__HRU_SLP.hru Average slope steepness (m/m) -0.8 0.8
v__CANMX.hru Maximum canopy index 0 100

percolation to shallow and deep aquifers. Depending on successfully have applied it as a search engine to optimize
data availability, the potential evapotranspiration (PET) parameters of the models [24-26]. PSO shares many
can be computed using different methods. In this study, similarities with evolutionary computation techniques
potential evapotranspiration (PET) was simulated using such as Genetic Algorithms (GA). The basic PSO
the Hargreaves method [21] Based on the PET and algorithm consists of three steps: (i) generate the
additional soil and landuse parameters, the actual plant positions of particles (coordinate in the parameter space)
transpiration and the actual soil evaporation are estimated and their velocities (‘flying’ direction and speed); (ii)
separately. The surface runoff from daily rainfall amounts update the velocity of each particle using the information
is modelled using a modification of the SCS curve number from the best solution it has achieved so far (personal
method [22] taking into account landuse, soil type and best) and another solution with the best fitness value that
antecedent soil moisture. Based on the literature and one has been obtained so far by all the particles in the
factor-at-a-time sensitivity analysis we selected 9 population (global best); (iii) finally, the new position of
parameters for model calibration (Table 1). each particle is calculated by adding the updated velocity
to the current position [18].
SUFI2 Procedure: SUFI2 [10] is a tool for sensitivity
analysis, multi-site calibration and uncertainty analysis. Markov chain Monte Carlo (MCMC): We took advantage
It is capable of analyzing a large number of parameters of a novel Markov chain Monte Carlo (MCMC) sampler,
and measured data from many gauging stations entitled differential evolution adaptive Metropolis
simultaneously. In SUFI-2, parameter uncertainty is (DREAM). This approach is known as the latest MCMC
described by a multivariate uniform distribution in a schemes developed by Vrugt et al. [19] and is applicable
parameter hypercube, while the output uncertainty is to complex, multi-modal, search problems and on a wide
quantified by 95PPU calculated at the 2.5% and 97.5% range of model calibration and uncertainty studies to
levels of the cumulative distribution function of the estimate optimal parameter values and their underlying
output variables. Latin hypercube sampling is used to posterior probability density function [27-31]. Vrugt et al.
draw independent parameter sets [10] SUFI2 is linked to [19] showed that DREAM works really well as compared
SWAT (in the SWAT-CUP software; Abbaspour, 2011) to the other existing MCMC schemes.
[23] through an interface that also includes the programs The DREAM sampling scheme in fact is an adaptation
Generalized Likelihood Uncertainty Estimation (GLUE) [8], of the Shuffled Complex Evolution Metropolis (SCEM-
PSO, Parameter Solution (ParaSol) [15] and a Monte Carlo UA) global optimisation algorithm [17].
Markov Chain (MCMC) [17] algorithm. A full detailed In the DREAM algorithm differential evolution as a
description of SUFI2 is presented by Abbaspour et al. genetic algorithm is used for population evolution. Within
[9, 10]. this technique a preset number (N) of Markov Chains
(a chain refers to a vector containing one parameter
Particle Swarm Optimization: Particle swarm optimization realization) are simultaneously run in parallel. The chains
(PSO) is a population based stochastic optimization are initialized by latin hypercube sampling the parameter
technique developed by Eberhart and Kennedy [20], space using uniform distribution. These chains form a
inspired by social behavior of bird flocking or fish population, conveniently stored as a N × d matrix X, with
schooling [23]. Many researcher in water science d the dimension of the parameter space. For each chain, i

1257
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

i {1, 2,... N}, a candidate point zi (vector) is generated by RESULT AND DISCUSSION
taking a fixed multiple of the difference between
randomly chosen pairs of chains (without replacement) of Results of SUFI2: SUFI2 implementation is convenient
X-i (X without xi) [32] : and easy. The modeler, however, should check a set of
suggested posterior parameters to be prepared for next
  iterations. Parallelizing the runs in SUFI2 [35] has
zi = xi + (1 + e ) ( )∑
, deff  xr ( j ) −
 1 ∑xr ( n )
2
+

substantially decreased the calibration time of SUFI2 in
=  j 1 =n 1  SWAT-CUP. SUFI2 is an iterative procedure. In this study
(1) we performed three iterations of 500 simulations each.
The calibration period was 9 years (1993-2001) with three
Where signifies the number of pairs used to generate years used as warm-up period. In the third iteration, the
the proposal, is a jumprate and r1(j), r2(n) {1, 2,...,N} but 95PPU brackets 48% of the observations and r-factor
r1(j) r2(n) i. The value of e is drawn from Ud(-b, b) with equals 0.51. For the validation period (1990-1992) the
|b|< 1 and ~Nd(0,b*) is a white noise term with b* small P-factor and R-factor were, respectively, 0.31 and 0.53.
compared to the width of the target distribution. Figure 2 describes the 95PPU (light grey area) in both
The Metropolis ratio is used to decide whether to accept calibration and validation periods as well as the
the candidate point or not. If accepted, the chain moves observation points. The solid line is the best simulation
from xi to zi, otherwise the location of the chain remains based on the run with the largest NS. Same periods for
unchanged. From the guidelines of in Random Walk calibration and validation like SUFI2 are considered for
Metropolis (RWM), a good choice of = 2.38 / 2 .d eff , two other methods.
Posterior distributions in SUFI2 are aways
where deff denotes the number of dimensions that will be
independent and mostly uniformly distributed and
updated. With this approach, a Markov chain is obtained,
expressed as narrowed parameter ranges. Also, there are
the stationary distribution of which is the posterior
no correlations between different parameters (all R2s less
distribution. After a so-called burn-in period, the
than 0.25), as samples are taken randomly using Latin
convergence of a DREAM run can be monitored with the
hypercube sampling.
R̂ -statistic of Gelman and Rubin [33], which compares
the variance within and between the chains. A value
Result of PSO: Like SUFI2, PSO also is linked to SWAT
of R̂ smaller than 1.2 for each parameter ( R̂k < 1.2, k=1,
(in the SWAT-CUP software, Abbaspour, 2011) and its
2,..., d) diagnoses convergence to a limiting implementation is easy. We carried out totally 6000
distribution. The samples generated after convergence simulation runs (e.g. 6 iterations with 1000 simulations).
can be used to summarize the posterior distribution and The 95PPU brackets 66% of the observations and r-factor
communicate parameter and model predictive uncertainty. equals 0.73. For the validation period the R-factor and
The number of steps in each chain required to reach P-factor were, respectively, 0.61 and 0.83. Figure 3
stationarity (convergence) is commonly called “burn-in” describes the 95PPU (light grey area) in both calibration
and these samples are removed from the analysis [28] and validation periods as well as the observation points.
A detailed description of DREAM appears in Vrugt et al. The solid line is the best simulation based on the run with
[32]. the largest NS.
DREAM is actually available now as a Marginal posterior distribution just in connection
package in R developed by Guillaume and Andrews [34] with v__ALPHA_BNK.rte parameter is illustrated in
(from R-forge) and in this study it was linked with SWAT. Figure 4. Remaining parameters have irregular posterior
distributions and are not shown.
Criteria for the Comparison: We use the commonly used
Nash-Sutcliffe (NS) as objective function. Two indices are Result of DREAM: We used a number of chains or
used to quantify the goodness of calibration/uncertainty population size of N = 2d, d as the number of parameters,
performance [9, 10] the P-factor, which is the percentage with a maximum total of 36,000 model evaluations for
of data bracketed by the 95% prediction uncertainty band SWAT. Samples with NS more than 0.4 (like SUFI2) in
(95PPU) (maximum value 100%) and the R-factor, which is each of the 18 chains were selected as the behavioral
the average width of the band divided by the standard samples and 95PPU, P-factor and R-factor are based on
deviation of the corresponding measured variable the results that adopted from behavioral parameters.
(minimum 0).

1258
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

Fig. 4: Histogram approximating the marginal posterior

distributions of aggregate SWAT behavioral
parameter conditioning with PSO

Fig. 2: 95PPUs derived by SUFI2 (dark gray area) during

the calibration period (top) and validation period
(bottom). The dots correspond to the observed
discharge at the basin outlet, while the solid line
represents the best simulation obtained by SUFI2.

Fig. 5: 95PPUs derived by DREAM (dark gray area)

during the calibration period (top) and validation
period (bottom). The dots correspond to the
observed discharge at the basin outlet, while the
solid line represents the best simulation obtained
by DREAM.

Fig. 3: 95PPUs derived by PSO (dark gray area) during

the calibration period (top) and validation period
(bottom). The dots correspond to the observed
discharge at the basin outlet, while the solid line
represents the best simulation obtained by PSO.

Figure 5 presents the 95PPU with light grey area in both Fig. 6: Histograms approximating the marginal posterior
calibration and validation periods. The results for both distributions of aggregate SWAT behavioral
calibration and validation periods show observational parameters conditioning with Bayesian MCMC -
discharge values bracketed partly well within 95PPU with DREAM.

1259
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

Fig. 7: Box plots of parameters (on a ratio scale) uncertainty range for SUFI2, DREAM and PSO

Table 2: Measures values for three uncertainty methods for parameters on a ratio scale
Parameters
---------------------------------------------------------------------------------------------------------------------------------------------------
Method Measure v__ALPHA_BNK V__CH_K2 v__CH_N2 V__ESCO v__CANMX
SUFI2 Mean 0.28 129.25 0.24 0.84 52.15
SD 0.1 11.7 0.04 0.1 15.22
CV 0.34 0.09 0.18 0.12 0.29
PSO Mean 0.33 76.12 0.20 0.55 36.59
SD 0.27 46.53 0.10 0.38 36.47
CV 0.82 0.61 0.52 0.69 1.00
DREAM Mean 0.4 77.03 0.14 0.86 57.41
SD 0.26 41.23 0.09 0.15 27.1
CV 0.65 0.54 0.6 0.18 0.47

Table 3: Comparison of some measures of performance for both SUFI2 and DREAM in calibration and validation periods
Best simulation Number of
------------------------------------- Number of Behavioral
Period Method R2 NS MSE P-factor R-Factor simulation Simulations
Calibration DREAM 0.61 0.59 6014 0.63 0.67 36000 1599
SUFI2 0.59 0.57 6311 0.48 0.51 1500 351
PSO 0.62 0.59 5980 0.66 0.73 6000 202
Validation DREAM 0.69 0.69 4118 0.53 0.84 7200 -
SUFI2 0.71 0.71 3768 0.31 0.53 500 -
PSO 0.74 0.74 3419 0.61 0.83 202 -

P-factor 0.63 and 0.67, respectively for the calibration Posterior distributions in connection with two
period. For the validation, R-factor was 0.53 and P-factor parameters (v__ALPHA_BNK.rte and v__ESCO.hru) are
equaled 0.84. illustrated in Figure 6. Concerning posterior distribution
Because of the large number of evaluations in of v__ALPHA_BNK.rte parameter, it shows some
DREAM (36000 vs. 1500 in SUFI2), it can explores the similarity with its distribution in PSO (figure 4). Remaining
parameter space in more detail and the marginal posterior parameters have irregular posterior distributions and are
pdfs can be inferred for some sensitive parameters. not shown.
Posterior distributions of these individual parameters are The correlations between all parameters in all
well defined and occupy only a relatively small region methods are small and therefore there are no significant
interior to the uniform prior distributions. correlations.

1260
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

Comparison the Results: Table 2 lists the mean, standard In spite of the larger number of simulation in DREAM
deviation and coefficient of variation (CV) identified for rather than PSO and SUFI2, it can not provide results
five most sensitive parameters on a ratio scale which can better than other, except that two marginal posterior
only take non-negative values. CV is defined as the ratio distribution parameters can be obtained from DREAM. In
of the standard deviation to the mean of values. From that case of v__ALPHA_BNK.rte parameter, PSO also has a
table it is resulted that SUFI2 has smallest CVs between posterior distribution some like with that parameter in
two other methods for all five ratio scale parameters. DREAM. In respect to performance of their best
We get the same results by box-plots for all five ratio estimates, PSO results have largest R2 and NS and
scale parameters (also called box and whisker plots) in smallest MSE among three methods in both periods.
Figure 7. The line across the box represents the median, In according to CV values, it can be resulted that
whereas the bottom and top of the box show the location three parameters of V_CH_K2, V_CH_N2 and V_ESCO in
of the first and third quartiles (Q1 and Q3). The whiskers most cases have smallest CV values and then are
are the lines that extend from the bottom and top of the specified as most sensitive parameters.
box to the lowest and highest observations inside the
region defined by Q1 - 1.5(Q3 - Q1) and Q3 + 1.5(Q3 - Q1). ACKNOWLEDGMENTS
Individual points with values outside these limits are
plotted with asterisks. Box-plot provides a visual Authors would like to thank Dr. John Joseph who
comparison of the uncertainty methods in connection helped us to implement SWAT with DREAM package in
with different parameters. As seen for these 5 parameters, R. We also would like to have special thanks for Dr.
SUFI2 box-plots, show small spread around the median Jasper Vrugt for his useful comments.
(small range of variability and fewer parameter
REFERENCES
uncertainties) thus revealing the effectiveness and
efficiency of SUFI2 as compare to DREAM and PSO.
1. Bicknell, B.R., J. Imhoff, J. Kittle, T. Jobes and
Only the ESCO parameter has as spread around the
A.S. Donigian, 2000. Hydrological Simulation
median in DREAM method which this matter also
Program - Fortran User’s Manual. Release 12, US
understandable form table 2 that shows the closeness of
EPA.
their CVs (0.12 compared to 0.18). Since sensitive
parameters with smaller CVs show very small range of
2. Ewen, J., G. Parkin and P.E. O'Connell, 2000.
uncertainty compared to less sensitive parameters, it can
SHETRAN: Distributed River Basin Flow and
be used to specify most sensitive parameters.
Transport Modeling System. J. Hydrol. Eng.,
Finally, a summary comparison between the three
5: 250-258.
uncertainty methods is provided in Table 3. As seen,
3. Arnold, J.G., R. Srinivasan, R.S. Muttiah and
P-factor for calibration period in DREAM and PSO is
J.R. Williams, 1998. Large area hydrologic modeling
larger than SUFI2, but this is because the R-factor in and assessment part i: model development1. Jawra
SUFI2 is narrower. In validation, PSO with smaller Journal of the American Water Resources
simulation runs than DREAM slightly has better results Association, 34: 73-89.
than it. But from best simulation perspective, PSO results 4. Kuczera, G. and E. Parent, 1998. Monte Carlo
have largest R2 and NS and smallest MSE among three assessment of parameter uncertainty in conceptual
methods in both periods. catchment models: the Metropolis algorithm. Journal
of Hydrology, 211(1-4): 69-85.
CONCLUSION 5. Refsgaard, J.C., J.P. Van der Sluijs, J. Brown and
P. Van der Keur, 2006. A framework for dealing with
As Yang et al [6] concluded that SUFI2 technique uncertainty due to model structure error. Adv Water
could be run with the smallest number of model runs to Resour, 29: 1586-1597.
achieve good prediction uncertainty ranges in the sense 6. Yang, J., P. Reichert, K.C. Abbaspour, J. Xia and
of a reasonable coverage of data points by the prediction H. Yang, 2008. Comparing uncertainty analysis
uncertainty bands. This characteristic is very important techniques for a SWAT application to the Chaohe
for computationally demanding models. Basin in China. J. Hydrol., 358: 1-23.

1261
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

7. Topliceanu, L., 2007. Water resources engineering- 19. Vrugt, J.A., C.J.F.T. Braak, C.G.H. Diks,
Analysis of uncertainty. Romanian Technical B.A. Robinson, J.M. Hyman and D. Higdon, 2009b.
Sciences Academy, 1: 295-300. Accelerating Markov chain Monte Carlo simulation
8. Beven, K. and A. Binley, 1992. The Future of by differential evolution with self-adaptive
Distributed Models-Model Calibration and randomized subspace sampling. International Journal
Uncertainty Prediction. Hydrol Process, 6: 279-298. of Nonlinear Sciences and Numerical Simulation,
9. Abbaspour, K.C., C.A. Johnson and M.T. Van 10: 273-290.
Genuchten, 2004. Estimating uncertain flow and 20. Eberhart, R.C. and J. Kennedy, 1995. A new optimizer
transport parameters using a sequential uncertainty using particle swarm theory. Proceedings of the sixth
fitting procedure. Vadose Zone J., 3: 1340-1352. international symposium on micro machine and
10. Abbaspour, K.C., J. Yang, I. Maximov, R. Siber, human science pp: 39-43. IEEE service center,
K. Bogner, J. Mieleitner, J. Zobrist and R. Srinivasan, Piscataway, NJ, Nagoya, Japan.
2007. Modelling hydrology and water quality in the 21. Hargreaves, G.L., G.H. Hargreaves and J.P. Riley,
pre-alpine/alpine Thur watershed using SWAT. J 1985. Agricultural benefits for Senegal River Basin.
Hydrol, 333: 413-430. Journal of Irrigation and Drainage Engineering,
11. Campbell, E.P. and B.C. Bates, 2001. Regionalization 111: 113-124.
of rainfall-runoff model parameters using Markov 22. USDA Soil Conservation Service. 1972. National
chain Monte Carlo samples. Water Resour Res., Engineering Handbook Section 4 Hydrology,
37: 731-739. Chapter, pp: 4-10.
12. Yang, J., P. Reichert, K.C. Abbaspour and H. Yang, 23. Abbaspour, K.C., 2011. Swat-Cup4 SWAT Calibration
2007. Hydrological modelling of the chaohe basin in and Uncertainty Programs Manual Version 4,
china: Statistical model formulation and Bayesian Department of Systems Analysis, Integrated
inference. J. Hydrol., 340: 167-182. Assessment and Modelling SIAM, Eawag. Swiss
13. Makowski, D., D. Wallach and M. Tremblay, 2002. Federal Institute of Aquatic Science and Technology,
Using a Bayesian approach to parameter estimation; Duebendorf, Switzerland.
comparison of the GLUE and MCMC methods. 24. Reddy, M.J. and D. Nagesh Kumar, 2007.
Agronomie, 22: 191-203. Multi-objective particle swarm optimization for
14. Mantovan, P. and E. Todini, 2006. Hydrological generating optimal trade-offs in reservoir operation.
forecasting uncertainty assessment: Incoherence of Hydrol Process, 21: 2897- 2909.
the GLUE methodology. J. Hydrol., 330: 368-381. 25. Gaur, S., D. Mimoun and D. Graillot, 2011.
15. Van Griensven, A. and T. Meixner, 2006. Methods to Advantages of the analytic element method for the
quantify and identify the sources of uncertainty for solution of groundwater management problems.
river basin water quality models. Water Sci. Technol., Hydrol Process, 25: 3426- 3436.
53: 51-59. 26. Lu, H., Z. Yu, R. Horton, Y. Zhu, Z. Wang, Z. Hao and
16. Marshall, L., D. Nott and A. Sharma 2004. A L. Xiang, 2011. Multi-scale assimilation of root zone
comparative study of Markov chain Monte Carlo soil water predictions. Hydrol Process, 25: 3158-3172.
methods for conceptual rainfall-runoff modeling. 27. Vrugt, J.A., C.J.F. ter Braak, M.P. Clark, J.M. Hyman
Water Resources Research, 40, W02501, and B.A. Robinson, 2008. Treatment of input
doi:10.1029/2003WR002378. M.D. McKay, R.J. uncertainty in hydrologic modeling: Doing hydrology
Beckman, W. Conover, backward with Markov chain Monte Carlo simulation.
17. Vrugt, J.A., H.V. Gupta, W. Bouten and Water Resour. Res., 44: W00B09.
S. Sorooshian, 2003. A Shuffled Complex Evolution 28. Dekker, S.C., J.A. Vrugt and R.J. Elkington, 2010.
Metropolis algorithm for optimization and uncertainty Significant variation in vegetation characteristics and
assessment of hydrologic model parameters. Water dynamics from ecohydrological optimality of net
Resour. Res., 39: 1201. carbon profit. Ecohydrology, na-na.
18. Zhang, X., R. Srinivasan, K. Zhao and M. Van Liew, 29. Laloy, E., D. Fasbender and C.L. Bielders, 2010a.
2008. Evaluation of global optimization algorithms for Parameter optimization and uncertainty analysis for
parameter calibration of a computationally intensive plot-scale continuous modeling of runoff using a
hydrologic model. Hydrol. Process, 23: 430-441. formal Bayesian approach. J. Hydrol., 380: 82-93.

1262
 Middle-East J. Sci. Res., 15 (9): 1255-1263, 2013

30. Laloy, E., M. Weynants, C.L. Bielders, 33. Gelman, A. and D.B. Rubin, 1992. Inference from
M. Vanclooster and M. Javaux, 2010b. How efficient iterative simulation using multiple sequences, Stat.
are one-dimensional models to reproduce the Sci., 7: 457-472.
hydrodynamic behavior of structured soils subjected 34. Guillaume, J. and F. Andrews, 2010. https://r-forge.r-
to multi-step outflow experiments? J. Hydrol., project.org/scm/viewvc.php /pkg/DESCRIPTION ?
393: 37-52. view=markup&revision=33&root=dream&pathrev=33
31. Scharnagl, B., J.A. Vrugt, H. Vereecken and 35. Rouholahnejad, E., K.C. Abbaspour, M. Vejdani,
M. Herbst, 2010. Information content of incubation R. Srinivasan, R. Schulin and A. Lehmann, 2011.
experiments for inverse estimation of pools in the Parallelizing SWAT calibration in Windows using the
Rothamsted carbon model a Bayesian perspective. SUFI2 program. Environmental Modelling and
Biogeosciences, 7: 763-776. Software.
32. Vrugt, J., C. ter Braak, H. Gupta and B. Robinson,
2009a. Equifinality of formal DREAM and informal
GLUE Bayesian approaches in hydrologic modeling?
Stoch Env Res Risk A, 23: 1011-1026.

1263