Professional Documents
Culture Documents
Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol
a r t i c l e i n f o s u m m a r y
Article history: Hydrological modeling uncertainties are the results of many factors such as input error, calibration accu-
Received 25 March 2010 racy, parameter uncertainty, model structure, and so on. Wherein, input errors and parameter uncertain-
Received in revised form 2 September 2010 ties are the two of the major factors influencing the uncertainties of hydrological modeling. TOPMODEL is
Accepted 17 September 2010
a rainfall–runoff model that bases its distributed predictions on analysis of watershed topography, which
This manuscript was handled by Geoff
is widely used in hydrological modeling practices. In this study, the effects of DEM resolution and param-
Syme, Editor-in-Chief, with the assistance of eter correlation on TOPMODEL modeling uncertainties are evaluated by using GLUE technique. The
Soroosh Sorooshian, Associate Editor uncertainty evaluation is performed by modeling the rainfall–runoff processes of three tributaries in
the Hanjiang River, one of the major tributaries of the Yangtze River, China. The results show no evident
Keywords: effects of the DEM resolution on the uncertainty intervals of the TOPMODEL simulation. This can be
Modeling uncertainty attributed to the fact that the modeling uncertainty is due solely to changes of DEM resolution by fixing
GLUE the parameter values to avoid the artifacts resulted from interactions between ln(a/tan(B)) and the
DEM resolution parameters. In addition, the copula functions are used to produce more behavioral parameter sets for
Parameter correlation the same sample time intervals when the model parameters are in good correlation, and which can ben-
TOPMODEL efit thorough evaluation of effects of parameter correlation on the hydrological modeling uncertainty.
With the same number of the behavioral parameter sets, after putting the parameter correlation under
consideration, the simulated runoff series by the TOPMODEL with the behavioral parameter sets can
fit reasonably better the observed runoff series. Thus, the uncertainty due to parameter correlation of
the TOPMODEL modeling can be considerably removed. This study is of great theoretical and practical
merits in sound understanding of the modeling behaviors of the TOPMODEL under the influences of
inputs and parameter correlation.
Ó 2010 Elsevier B.V. All rights reserved.
1. Introduction generating the prediction limits for the modeled streamflow series
and a set of behavioral parameter sets (Freer et al., 1996; Blazkova
Hydrological models have been widely used to investigate and Beven, 2002; McMichael et al., 2006; Montanari, 2005, 2007;
many practical and pressing issues that arise during planning, de- Yang et al., 2007, 2008; Xiong and O’Connor, 2008; Jin et al., 2010).
sign, operation, and management of water resources systems The hydrological system is complicated, being affected by the
(Benke et al., 2007; Lin et al., 2007). The crucial step for hydrolog- climate changes such as atmospheric circulation, precipitation,
ical modeling is to identify values of model parameters and this air temperature, the underlying surface properties such as the geo-
procedure is also referred to as calibration procedure (Sorooshian logical conditions, vegetation and soil conditions, and also human
and Gupta, 1995). Estimation and specification of parameters are activities such as water reservoirs and land use changes (Zhang
the two important procedures for calibration of hydrological et al., 2009, 2010). Generally, a hydrological model consists of a
model. However, the parameter redundancy and correlations be- large number of mathematical equations describing changing
tween parameters result in universal equifinality in modeling properties of hydrological processes, e.g. streamflow series, and
behaviors of the hydrological models (Beven and Binley, 1992). estimating the streamflow variations of the future. Additional in-
Accordingly the generalized likelihood uncertainty estimation puts represent the spatial mosaic of climate, soil type, topography
(GLUE) method proposed by Beven and Binley (1992) is devoted and land use (Benke et al., 2007). Topography was taken as an
to the investigation of the hydrological modeling uncertainty by important factor in the evaluation of the hydrological responses
of the upland and forested watersheds to precipitation changes
⇑ Corresponding author. Tel./fax: +86 20 84113730. (e.g. Beven and Wood, 1983; Wolock and Price, 1994) due to the
E-mail address: zhangq68@mail.sysu.edu.cn (Q. Zhang). effects of gravity on the movement of water in a watershed. Many
0022-1694/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2010.09.012
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 371
Table 1
Ranges of parameters used in TOPMODEL model.
Parameter Unit Physical meaning Minimum value Maximum value Sampling strategy
Szm m Maximum moisture deficit 0.01 1 Uniform
T0 m2 h1 Transmissivity of the soil at saturated 2.33a 3.4a Uniformb
Td h Time parameter 1 20 Uniform
SRmax m Maximum allowable root storage deficit 0.01 0.3 Uniform
a
Ranges for T0 are shown in log in relation to the graph scales.
b
The sampling method for ln(T0).
their respective uncertainty intervals to the total number of obser- Copula functions have good performance in description of correla-
vations (Beven and Binley, 1992; Montanari, 2005; Xiong and tion structure of variables. In addition, one of the advantages of
O’Connor, 2008). The CR is formulated as: Copula functions is that the marginal distributions are beyond nor-
Pn mal and extreme value distributions. Gumbel–Hougaard Copula is
i¼1 J½Q obs ðiÞ another kind of Copula function having been taken as the right
CR ¼ ð4Þ
n technique in the description of the joint probability behaviors of
where the hydrological variables (Salvadori and De Michlele, 2004; Zhang
and Singh, 2006). This study uses Gumbel–Hougaard Copula to de-
1; Q low ðiÞ < Q obs ðiÞ < Q up ðiÞ scribe the correlative structure of the model parameters, and can
J½Q obs ðiÞ ¼ ð5Þ
0; otherwise be expressed as
Qobs(i), Qlow(i) and Qup(i) in Eqs. (4) and (5) denote the observed run- Cðu; v Þ ¼ expf½ð ln uÞh þ ð ln v Þh 1=h g; h1 ð8Þ
off, the lower and the upper uncertainty intervals at time i respec-
tively; and n is the length of the observed series. where u and v represent two random variables U and V, and C is the
Interval width (IW) is defined as the average width of the uncer- copula function, a kind of multivariate distribution function, and its
tainty intervals (Choi and Beven, 2007; Blasone and Vrugt, 2008), one-dimensional margins are uniform on the interval (0, 1), and h is
which can be calculated by the parameter of Copula. Larger h value implies stronger correlation,
Pn and vice versa. h = 1 denotes independency of variables.
i¼1 ½Q up ðiÞ Q low ðiÞ
IW ¼ ð6Þ The parameter of Copula, h, is related to the Kendall’s rank cor-
n relation coefficient s by 1/(1 s). Once an estimate s ^ of s is ob-
The Nash–Sutcliffe efficiency index (R2) (Nash and Sutcliffe, 1970) is tained, an estimate ^ h of h can be calculated. In the CGLUE
used to judge whether or not the median values MQ0.5 and the technique, correlative parameter sets can be generated directly
uncertainty intervals are effective crisp simulations of the observa- by the Copula function.
tions, which is defined as: The theory behind the link between copula and bivariate distri-
Pn butions was addressed by Sklar’s theorem (Sklar, 1959): Let H be a
2
i¼1 ½Q obs ðiÞ MQ 0:5 ðiÞ joint distribution function with margins F and G. Then there exists
R2 ¼ 1:0 Pn 2
ð7Þ
i¼1 ½Q obs ðiÞ Q obs
a copula C for x, y,
where Q obs is the mean value of the observed runoff series. Hðx; yÞ ¼ CðFðxÞ; GðyÞÞ ð9Þ
By Eqs. (8) and (9), the bivariate joint distribution H(x, y) of X and Y
2.3. Copula function
can be written as
In this study, a method based on the GLUE technique (CGLUE) is Hðx; yÞ ¼ expf½ð lnðFðxÞÞÞh þ ð lnðGðyÞÞh 1=h g ð10Þ
adopted using the Copula function to describe the correlations
amongst parameters (e.g. Lin et al., 2009). Copula technique, after The interesting point is that the properties of H(x, y) can be dis-
it was proposed in 1959, has been a kind of important method ded- cussed in terms of the structure of C. In fact, it is precisely the cop-
icated to analyze the multivariate joint distribution (Nelson, 1999). ula that captures many of the features of a joint distribution, and
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 373
ity for the water supply for the north China (Fig. 3). The basin is
GLUE dominated by the subtropical monsoon climate and hence the
Select model and determine water resource is subject to large variability. Flood occurs quite of-
the range of each parameter
ten in the Hanjiang River basin and effective water resource man-
agement is a must in this river basin. Daily rainfall and runoff data
Select likelihood function, L(θ ) , and set a of these three river basins covering 1978 and 1987 are analyzed in
this study. The geographical properties and data in these three riv-
threshold of likelihood function, choose a
er basins are introduced briefly as the follows:
desired % of observations to be contained
Mumahe River basin is 1224 km2 in drainage area. The topo-
in the uncertainty bounds
graphical properties of this river are characterized by low hills
and basins with moderate slope. Mumahe River basin is the largest
Monte Carlo Sampling with independent parameters basin in the south bank of the Hanjiang River.
Xunhe River basin is 6448 km2 in drainage area. The topography
is dominated by mountains with the altitude of the mountain peak
Repeatedly run the model, obtain the between 2000 m and 3000 m. The vegetation coverage of the Xun-
behavioral parameter sets and the he River basin is good and the soil erosion is not serious.
posterior distribution of parameters Laoguanhe River basin is 3219 km2 in drainage area. The upper
Laoguanhe River is dominated by hill landforms, and the lower
Generate the uncertainty bounds Laoguanhe River is the Danjiangkou Reservoir. The slope of the ter-
according to the likelihood weights rain of this river basin is steep. Due to the frequent rainstorms and
steep terrains, soil erosion is serious.
Fig. 3. Locations of the study region, the Hanjiang River basin and three major tributaries considered in this study, i.e. Mumahe, Xunhe and Laoguanhe Rivers. SNWTP denotes
South-to-North Water Transfer Project of China.
Table 2
Topographic characteristics of different DEM resolutions.
where ln is the Napierian logarithm, a is the upslope area per unit ln(a/tan(B)), implying that the asymmetry of the ln(a/tan(B)) distri-
contour length, and tan(B) is the slope gradient. The DEMs are used bution tends to be smaller.
to calculate a and tan(B) for each grid cell in the watershed. To do so
required the calculation of the total area draining into each cell (A),
as well as the contour length (C) and slope gradient (tan B) along 4.2.1. Effects on uncertainty of TOPMODEL
which drainage from the cell occurs (Wolock and McCabe, 1995). An implicit assumption in using topography data in TOPMODEL
The results are illustrated in Fig. 4 and Table 3. It can be ob- is that the water table configuration is related to the land surface
served from Fig. 4 and Table 3 that all statistics of the ln (a/tan B) topography. Previous study (e.g. Wolock and Price, 1994) has
distribution are heavily affected by DEM resolutions. Coarser DEM shown that it should not be concluded that coarse resolution is
resolution tends to cause greater mean ln(a/tan(B)) values within inappropriate source of topographic information for TOPMODEL.
three catchments considered in this study, being represented by In the current study, we investigate how the DEM resolution af-
increased areas covered by higher ln(a/tan(B)) values. Coarser fects the efficiency of uncertainty interval in TOPMODEL simula-
DEM resolution causes larger mean variance of the ln(a/tan B) with tion. Three indices, i.e. containing ratio (CR), interval width (IW),
smaller variation coefficients. Thus, it is evident that coarser grid and the Nash–Sutcliffe efficiency of the median MQ0.5, are selected
resolution of the DEM tends to increase absolute dispersion degree, to evaluate the efficiency of uncertainty. For the datasets consid-
but tends to decrease relative dispersion degree. The coarser the ered in this paper, the GLUE methodology is applied for the likeli-
DEM resolution is, the smaller the skewness coefficient of the hood function selected as the global Nash–Sutcliffe efficiency R2
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 375
Ac/A
Ac/A
Ac/A
0.15 200m 100m
200m
0.1 400m 400m 200m
800m 0.1 0.1 400m
0.05 800m
800m
0 0 0
0 5 10 15 20 0 5 10 15 20 25 0 5 10 15 20 25
ln (a/tanB) ln (a/tanB) ln (a/tanB)
Table 3
Statistical characteristics of topographical index corresponding to different DEM resolutions.
Table 4
Assessing indices of uncertainty for different DEM resolutions in the GLUE technique.
(Nash and Sutcliffe, 1970). The analysis uses 50,000 parameter sets different DEM resolutions in the Mumahe, Xunhe, and Laoguanhe
generated by the uniform Monte Carlo sampling schemes. The River basins respectively. The numbers 1–6 showing the DEM res-
uncertainty intervals at a given confidence level of 90% are ob- olutions in Figs. 6–8 denote 25 m, 50 m, 100 m, 200 m, 400 m, and
tained with the GLUE technique by setting the threshold value, 800 m respectively. Figs. 6–8 show that the posterior distributions
R2, as 70%. Table 4 displays the results of GLUE-based evaluation of parameters are affected by DEM resolutions, but this effect is rel-
of the uncertainties of the TOPMODEL with respect to different atively minor. The major influence of DEM resolution on the pos-
DEM resolutions. R2 (MQ0.5) in Table 4 represents the Nash– terior parameter distributions is the difference of proration of
Sutcliffe efficiency of the median MQ0.5 produced from the parameter values. The modeling uncertainty should be due solely
GLUE-based analysis by fitting the observed runoff series. The vari- to DEM resolution changes by fixing the parameter values to avoid
ations of the containing ratio (CR), and interval width (IW) with the artifacts resulted from interactions between ln(a/tan(B)) and
different the DEM resolutions are demonstrated in Fig. 5. the parameters. Besides, it should be noted here that Table 4 indi-
It can be seen from Table 4 and Fig. 5 that changes of the con- cates larger IW of the Xunhe basin compared to the other two ba-
taining ratio (CR), and interval width (IW) are not significant when sins, which should be due to larger drainage area of the Xunhe
the DEM resolutions are changing, showing no evident influences basin when compared to other two river basins considered in this
of DEM resolutions on the uncertainty intervals of TOPMODEL sim- study. Correspondingly, the discharge of the Xunhe River basin is
ulations. It should be noted here that DEM resolution can affect the larger than that of the other two basins. The same phenomenon
topographic characteristics and the ln(a/tan(B)) distribution, how- can also be identified in Tables 5–7.
ever the effects of changing DEM resolutions on the ln(a/tan(B))
distribution can be compensated by the proper adjustments of
the parameters with sampling techniques. Availability of numer- 5. Effects of parameter correlation on uncertainty of TOPMODEL
ous parameter sets can greatly reduce the impacts of DEM resolu-
tions on the modeling uncertainty, and the similar modeling 5.1. Parameter correlation in TOPMODEL
results by hydrological models with different sets of parameters
is also called as equifinality. Figs. 6–8 illustrate comparisons Complexity and correlations of the parameter space are the
amongst the posterior distributions of parameters with respect to two important factors that can trigger the parameter uncertainty.
376 K. Lin et al. / Journal of Hydrology 394 (2010) 370–383
0.66 70
0.65 60
Mumahe
IW (m3/s)
0.64 Xunhe
50
CR
0.63 Laoguanhe
Mumahe 40
0.62 Xunhe
Laoguanhe 30
0.61
0.60 20
0 200 400 600 800 0 200 400 600 800
DEM Resolution (m) DEM Resolution (m)
Parameter distribution
0.1
0.06
0.08
0.06
0.04
0.04
0.02
0.02
0 0
0 −3
0.2
−2
0.4
−1
Pa 0.6 Pa
ram ram 0
ete 0.8 ete
r ra 1 r ra 1
ng ng
e 1.2 e 2
56 56
34 ion
1.4 12
re solut 3 12
34
lution
DEM reso
DEM
Parameter distribution
0.1 0.1
0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
0 0
5 0.05
0.1
Pa
ram10 Pa
ram0.15
ete 15 ete 0.2
r ra r ra
ng ng0.25
e 20 e
56 0.3 56
25 12
34
eso lution 34
1 2 esolutio
n
EM r
0.35
D DE Mr
Fig. 6. Posterior distribution of parameters against DEM resolution in Mumahe River basin.
The following procedures are taken in this study to identify the formed on the parameters. Figs. 9–11 illustrate the correlation
parameter correlations: (1) the GLUE technique is applied to relations amongst the parameters from posterior distributions ob-
achieve the posterior distributions of the parameters by setting a tained by the GLUE technique in the three river basins respectively
threshold efficiency as 70% (Nash–Sutcliffe efficiency criterion); considered in this study. Not significant correlations can be found
(2) a total of 100,000 parameter sets are generated by the uniform from Figs. 9–11. Panels showing correlations relations of {Szm, T0}
Monte Carlo sampling schemes; and (3) correlation analysis is per- are the exceptions, which show evident correlations with correla-
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 377
Parameter distribution
Parameter distribution
0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
0 −3
0.2 −2
0.4
−1
Pa 0.6 Pa
ram ram 0
ete 0.8 ete
r ra r ra 1
ng 1 ng
e e
1.2 2
56 56
34 ion 34
1.4 12 solut 3 12 lution
M re reso
DE DEM
Parameter distribution
0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
0 0
5 0.05
0.1
Pa
ram10 Pa 0.15
ete 15 ram
r ra ete 0.2
ng r ra 0.25
e 20 ng
56
e 0.3
34 56
12 tion 34
esolu lution
25 0.35 12
r eso
DEM D EM r
Fig. 7. Posterior distribution of parameters against DEM resolution in Xunhe River basin.
tion coefficients of 0.66, 0.62 and 0.62, and the Kendall correlation same number of behavioral parameter sets, i.e. for the sake of com-
coefficients, s, of 0.69, 0.67 and 0.67, and the corresponding param- parison, the number of behavioral parameter sets of CGLUE tech-
eter of copula function h, of 3.23, 3.03, 3.03 in Mumahe, Xunhe, and nique is set to be the same as that of the GLUE technique. Firstly,
Laoguanhe River basins respectively. The evident correlation be- comparison is carried out under the same threshold criterion. Nine
tween Szm and T0 can be well explained according to the hypothe- scenarios are made by setting the sampling numbers as 10,000,
ses behind TOPMODEL. Generally, three hypotheses were 50,000, and 100,000, and threshold values as 50%, 60%, and 70%.
introduced in TOPMODEL, and the third hypothesis assumes that The number of behavioral parameter sets of these nine scenarios
transmissivity is the negative exponent function of the water depth is listed in Table 5. The variations of the assessing indices corre-
of the saturated groundwater, see Eq. (3) for more details. sponding to different DEM resolutions with parameter correlation
in the first scheme are displayed in Table 6, where 50,000 are ac-
5.2. Uncertainty of TOPMODEL simulation cepted as the sampling numbers, 90% as confidence level, and
70% as threshold value.
In this study, the GLUE and the CGLUE techniques are used to The GLUE and CGLUE techniques behave in evidently different
perform the uncertainty analysis. The objective is to investigate ways in terms of analysis of the parameter correlation: (1) evident
the effects of parameter correlation on modeling uncertainty of differences can be found for GLUE and CGLUE techniques with
TOPMODEL by comparing the results from these two methods. respect to the uncertainty estimation of the model parameters.
Two schemes are done to investigate the effects of parameter cor- Table 5 shows that the number of behavioral parameter sets
relation on uncertainty of TOPMODEL simulation in this study: (1) decreases when the threshold increases, and increases when the
comparison is carried out between the GLUE and CGLUE tech- sampling times increase, and for any given scenario, more behav-
niques with the same threshold value of R2, 70%; (2) comparison ioral parameter sets are obtained in CGLUE technique than in GLUE
is conducted between the GLUE and CGLUE techniques with the technique; (2) the containing ratios, CR, and the interval width IW
378 K. Lin et al. / Journal of Hydrology 394 (2010) 370–383
0.08
0.1
Parameter distribution
Parameter distribution
0.06 0.08
0.06
0.04
0.04
0.02
0.02
0 0
0 −3
0.2 −2
0.4
−1
Pa 0.6 Pa
ram ram 0
ete 0.8 ete
r ra r ra 1
ng 1 ng
e e
1.2 2
56 56
1.4 34
1 2 esolutio
n 3 12
34
eso lution
DE Mr DEM r
Parameter distribution
0.06 0.06
0.04 0.04
0.02 0.02
0 0
0 0
5 0.05
0.1
Pa 10
ram Pa 0.15
ete ram 0.2
r ra15 ete
ng r ra 0.25
e 20 ng
56 e 0.3
34 n 56
25 1 2 resolutio 0.35 12
34
lution
DE M eso
DEM r
Fig. 8. Posterior distribution of parameters against DEM resolution in Laoguanhe River basin.
Table 5
List of the total number of the behavioral parameter sets.
obtained by CGLUE technique are larger when compared to those one million parameters are required due to the large amount of
by GLUE technique (Table 6 and Fig. 12). Besides, more sampling parameter sets involved in the model structures. In this case,
times are necessary to evaluate the uncertainty of the parameters Monte Carlo procedure usually consumes massive computing re-
in a more comprehensive way. In practice, ten thousand or even sources. If the samples are not adequate, the sampling procedure
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 379
Table 6
Assessing indices of uncertainty in the CGLUE technique considering parameter correlation in the first scheme.
Note: In the first scheme, comparison is carried out between the GLUE and CGLUE techniques with the same threshold value of R2, 70%.
Table 7
Assessing indices of uncertainty in the CGLUE technique considering parameter correlation in the second scheme.
Note: In the second scheme, the number of behavioral parameter sets of the CGLUE technique is set to be the same as that of the GLUE technique.
3 20 0.3
2 0.25
15
0.2
1
SRmax
ln (To)
Td
10 0.15
0
0.1
5
−1
0.05
−2 0 0
0 0.5 1 0 0.5 1 0 0.5 1
Szm Szm Szm
20 0.3 0.3
0.25 0.25
15
0.2 0.2
SRmax
SRmax
Td
10 0.15 0.15
0.1 0.1
5
0.05 0.05
0 0 0
−2 0 2 −2 0 2 0 10 20
ln (To) ln (To) Td
Fig. 9. Correlations amongst the parameters from posterior distribution in Mumahe River basin.
which assumes independence of the parameters may not include parameter sets by the same sampling times, and thus the uncer-
all possible behavioral parameter combinations, and which is ex- tainty of parameters could be estimated thoroughly.
pected to underestimate the uncertainty. The results showed that In addition, comparison is conducted with the same number of
the CGLUE technique performs well in terms of estimation of the behavioral parameter sets to study the effects of parameter corre-
parameter correlation when the model parameters are in good cor- lation on uncertainty of TOPMODEL simulation. The variations of
relations, and this method can help to obtain more behavioral the assessing indices corresponding to different DEM resolutions
380 K. Lin et al. / Journal of Hydrology 394 (2010) 370–383
3 20 0.3
2 0.25
15
1 0.2
SRmax
ln (To)
Td
0 10 0.15
−1 0.1
5
0.05
−2
0 0
0 0.5 1 0 0.5 1 0 0.5 1
Szm Szm Szm
20 0.3 0.3
0.25 0.25
15
0.2 0.2
SRmax
SRmax
Td
10 0.15 0.15
0.1 0.1
5
0.05 0.05
0 0 0
−2 0 2 −2 0 2 0 10 20
ln (To) ln (To) Td
Fig. 10. Correlations amongst the parameters from posterior distribution in Xunhe River basin.
3 20 0.3
2 0.25
15
1 0.2
SRmax
ln (To)
Td
0 10 0.15
−1 0.1
5
−2 0.05
−3 0 0
0 0.5 1 0 0.5 1 0 0.5 1
Szm Szm Szm
20 0.3 0.3
0.25 0.25
15
0.2 0.2
SRmax
SRmax
Td
10 0.15 0.15
0.1 0.1
5
0.05 0.05
0 0 0
−2 0 2 −2 0 2 0 10 20
ln (To) ln (To) Td
Fig. 11. Correlations amongst the parameters from posterior distribution in Laoguanhe River basin.
with parameter correlation considered in the first scheme are dis- runoff uncertainty intervals by the CGLUE technique with the same
played in Table 7. Fig. 13 shows comparison of the assessing indi- number of behavioral parameter sets obtained by the GLUE tech-
ces between independent and correlative parameters in the second nique are illustrated in Fig. 15.
scheme. The runoff uncertainty intervals generated by the GLUE It can be observed from Table 7 and Figs. 13–15 that the Nash–
technique for the Xunhe catchment during the time period of July Sutcliffe efficiency of the median value, MQ0.5, R2 (MQ0.5) increases
3th, 1983–December 10th, 1983 when threshold of likelihood accompanying the enhancing parameter correlations. Increase of
equals to 70%, are demonstrated in Fig. 14, and the corresponding the Nash–Sutcliffe efficiency of the median value, MQ0.5, indicates
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 381
Fig. 12. Comparison of the assessing indices between the GLUE and the CGLUE techniques with the same threshold value of R2, 70% (the first scheme).
Fig. 13. Comparison of the assessing indices between the GLUE and the CGLUE techniques with the same number of behavioral parameter sets, i.e. for the sake of comparison,
the number of behavioral parameter sets of CGLUE technique is set to be the same as that of the GLUE technique (the second scheme).
4000
3500 observed data
discharge (m3/s)
Fig. 14. The runoff uncertainty intervals generated by the GLUE technique for the Xunhe catchment during the time period of July 3th, 1983–December 10th, 1983 with the
threshold of likelihood of 70%.
4000
3500 observed data
discharge (m3/s)
Fig. 15. The runoff uncertainty intervals generated by the CGLUE technique for the Xunhe catchment during the time period of July 3th, 1983–December 10th, 1983 with the
same number of behavioral parameter sets obtained by the GLUE method.
model structural improvements and uncertainty reduction No.: 1132381) and by a grant from the Research Grants Council
(Blasone and Vrugt, 2008). Beven and Freer (2001a,b) addressed of the Hong Kong Special Administrative Region, China (Project
the effects of some factors on modeling uncertainty, such as model No. CUHK405308). The last but not the least, cordial gratitude
nonlinearity, covariation of parameters and drawbacks in model should be extended to the editor, Prof. Dr. Geoff Syme, and the
structure, input error or observed variables, within the GLUE pro- two anonymous reviewers for their pertinent and professional
cedure. Considering the importance of modeling uncertainty and comments which greatly help to improve the quality of this
related underlying causes, based on the GLUE technique, the cur- manuscript.
rent study thoroughly investigated the effects of DEM resolution
and parameter correlations on uncertainties of TOPMODEL simula-
References
tions and some interesting, novel and important conclusions have
been obtained: Ajami, N.K., Duan, Q., Sorooshian, S., 2007. An integrated hydrologic bayesian multi-
model combination framework: confronting input, parameter and model
structural uncertainty. Water Resources Research 43, W01403. doi:10.1029/
(1) DEM resolution has the potential to affect the topographic
2005WR004745.
characteristics and the ln(a/tan(B)) distribution. However, Benke, K.K., Lowell, K.E., Hamilton, A.J., 2007. Parameter uncertainty, sensitivity
the results of this study indicates little influence of DEM res- analysis and prediction error in a water-balance hydrological model.
olutions on the modeling uncertainty of the TOPMODEL, Mathematical and Computer Modeling 47 (11–12), 1134–1149.
Beven, K.J. (Ed.), 1997. Distributed Modeling in Hydrology: Applications of
which is represented by no evident differences of the uncer- TOPMODEL. Wiley, Chichester.
tainty intervals corresponding to different DEM resolutions. Beven, K.J., Binley, A., 1992. The future of distributed models, model calibration and
It should be attributed to the fact that the modeling uncer- uncertainty prediction. Hydrological Processes 6, 279–298.
Beven, K.J., Freer, J., 2001a. Equifinality, data assimilation, and uncertainty
tainty should be due solely to DEM resolution changes by estimation in mechanistic modeling of complex environmental systems using
fixing the parameter values to avoid the artifacts resulted the GLUE methodology. Journal of Hydrology 249, 11–29.
from interactions between ln(a/tan(B)) and the parameters. Beven, K.J., Freer, J., 2001b. A dynamic TOPMODEL. Hydrology Processes 15, 1993–
2011.
(2) The Copula function is used to describe the correlation struc- Beven, K.J., Kirkby, M.J., 1979. A physically-based variable contributing area model
tures of the model parameters with the CGLUE technique. of basin hydrology. Hydrology Science Bulletin 24 (1), 43–69.
Comparison of the results by the GLUE and CGLUE technique Beven, K., Wood, E.F., 1983. Catchment geomorphology and the dynamics of runoff
contributing areas. Journal of Hydrology 65, 139–158.
shows that, under the same threshold, more behavioral Beven, K.J., Lamb, R., Quinn, P.F., Romanowicz, R., Freer, J., 1995. TOPMODEL. In:
parameter sets are required when the model parameters Singh, V.P. (Ed.), Computer Models of Watershed Hydrology. Water Resources
are in good correlation relations, and in so doing the uncer- Publications, Highlands Ranch, CO, pp. 627–668.
Blasone, R.S., Vrugt, J.A., 2008. Generalized likelihood uncertainty estimation (GLUE)
tainty of parameters could be evaluated in a thorough way.
using adaptive Markov Chain Mote Carlo sampling. Advances in Water
(3) The Nash–Sutcliffe efficiency of the median value, MQ0.5, and Resources 31, 630–648.
R2 (MQ0.5) increases when the parameter correlation is put Blazkova, S., Beven, K.J., 2002. Flood frequency estimation by continuous simulation
under consideration, implying that, when considering the for a catchment treated as ungauged (with uncertainty). Water Resources
Research 38 (8), 1139. doi:10.1029/2001WR000500.
parameter correlation, the simulated runoff series by the Choi, H.T., Beven, K., 2007. Multi-period and multi-criteria model conditioning to
TOPMODEL with the behavioral parameter sets can fit better reduce prediction uncertainty in an application of TOPMODEL within the GLUE
to the observed runoff series, and correspondingly, the framework. Journal of Hydrology 332, 316–336.
Duan, J., Miller, N.L., 1997. A generalized power function for the subsurface
abstracted median value, MQ0.5, can be improved for better transmissivity analysis of the TOPMODEL. Journal of Hydrology 175, 293–338.
forecasts of the runoff. Besides, the results of this current Freer, J., Beven, K.J., Ambroise, B., 1996. Bayesian estimation of uncertainty in runoff
study also confirm that the CGLUE technique considering prediction and the value of data: an application of the GLUE approach. Water
Resources Research 32 (7), 2161–2173.
the parameter correlation can greatly reduce the modeling Gupta, H.V., Sorooshian, S., Yapo, P.O., 1998. Toward improved calibration of
uncertainty of the TOPMODEL as a result of parameter corre- hydrologic models: multiple and noncommensurable measures of information.
lation to some degree. Water Resources Research 34, 751–763.
Jin, X., Xu, C., Zhang, Q., Singh, V.P., 2010. Parameter and modeling uncertainty
simulated by GLUE and a formal Bayesian method for a conceptual hydrological
The hydrological model structure becomes increasingly compli- model. Journal of Hydrology 383, 147–155.
cated to satisfy the requirements of the gradually complicated Keesman, K.J., 1990. Set theoretic parameter estimation using random scanning and
principal component analysis. Mathematics and Computers in Simulation 32,
environment problems we are facing. However, complicated
535–543.
hydrological model structure is easy to produce more uncertainties Kuczera, G., Parent, E., 1998. Monte Carlo assessment of parameter uncertainty in
within the hydrological modeling results. In this study, impacts of conceptual catchment models: the Metropolis algorithm. Journal of Hydrology
DEM resolution and parameter correlations of the TOPMODEL on 211, 69–85.
Lin, K., Guo, S., Zhang, W., Liu, P., 2007. A new baseflow separation method based on
the uncertainties of the modeling results are thoroughly investi- analytical solutions of the Horton infiltration capacity curve. Hydrological
gated by using CGLUE technique. The results of this study will be Processes 21 (13), 1719–1736.
of theoretical and practical merits in scientific understanding of Lin, K., Chen, X., Zhang, Q., Chen, Z., 2009. A modified generalized likelihood
uncertainty estimation method by using Copula function, IAHS Publication 335,
uncertainties of the modeling behaviors of the TOPMODEL, and will 51–56.
also be helpful for further improvement of the TOPMODEL in terms Madsen, H., 2003. Parameter estimation in distributed hydrological catchment
of hydrological modeling. modeling using automatic calibration with multiple objectives. Advance in
Water Resource 26, 205–216.
McMichael, C.E., Hope, A.S., Loaiciga, H.A., 2006. Distributed hydrological modeling
Acknowledgements in California semi-arid shrublands: MIKE SHE model calibration and uncertainty
estimation. Journal of Hydrology 317 (3–4), 307–324.
Montanari, A., 2005. Large sample behaviors of the generalized likelihood
This study was financially supported by the National Natural uncertainty estimation (GLUE) in assessing the uncertainty of rainfall–runoff
Science Foundation of China (Grant Nos.: 50809078; 41071020; simulations. Water Resources Research 41, W08406. doi:10.1029/
50839005), the Open Foundation of the State Key Laboratory of 2004WR003826.
Montanari, A., 2007. What do we mean by ‘uncertainty’? the need for a consistent
Water Resources, Hydropower Engineering Science (Wuhan
wording about uncertainty assessment in hydrology. Hydrological Processes 21,
University) (Grant No.: 2008B043), the Scientific Research Founda- 841–845.
tion for Yong Scholars of Sun Yat-sen University (Grant No.: Moradkhani, H., Hsu, K.-L., Gupta, H., Sorooshian, S., 2005. Uncertainty assessment
3161395), the Natural Science Foundation of Guangdong province of hydrologic model states and parameters: sequential data assimilation using
the particle filter. Water Resources Research 41 (5), 1–17.
(Grant Nos.: 4203384; 2009-37000-4203384), the Program for Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through the conceptual models,
Outstanding Young Teachers of the Sun Yat-sen University (Grant 1: a discussion of principles. Journal of Hydrology 10 (3), 282–290.
K. Lin et al. / Journal of Hydrology 394 (2010) 370–383 383
Nelson, R.B. (Ed.), 1999. An introduction to Copulas. Springer, New York. Walter, M.T., Steenhuis, T.S., Mehta, V.K., Thongs, D., Zion, M., Schneiderman, E.,
Quinn, P., Beven, K.J., Lamb, R., 1995. The ln(a/tan b) index: how to calculated it and 2002. Refined conceptualization of TOPMODEL for shallow subsurface flows.
how to use it within the TOPMODEL framework. Hydrological Processes 9, 161– Hydrology Processes 16, 2041–2046.
182. Wolock, D.M., McCabe Jr., G.J., 1995. Comparison of single and multiple flow
Salvadori, G., De Michlele, C., 2004. Frequency analysis via copulas, theoretical direction algorithms for computing topographic parameters in TOPMODEL.
aspects and applications to hydrological events. Water Resources Research 40 Water Resources Research 31 (5), 1315–1324.
(12), 1–17. Wolock, D.M., Price, C.V., 1994. Effects of digital elevation model map scale and data
Scanlon, T.M., Raffensperger, J.P., Hornberger, G.M., Clapp, R.B., 2000. Shallow resolution on a topography-based watershed model. Water Resources Research
subsurface storm flow in a forested headwater catchment: observations and 30 (11), 3041–3052.
modeling using a modified TOPMODEL. Water Resources Research 36 (9), 2575– Xiong, L., Guo, S., 2004. Effects of the catchment runoff coefficient on the
2586. performance of TOPMODEL in rainfall–runoff modeling. Hydrological
Sklar, K., 1959. Fonctions de repartition à n dimensions et leura marges. Paris, Processes 18, 1823–1836.
Publications de l’Institut de de l’Université, vol. 8, pp. 229–231. Xiong, L., O’Connor, K.M., 2008. An empirical method to improve the prediction
Sørensen, R., Seibert, J., 2007. Effects of DEM resolution on the calculation of limits of the GLUE methodology in rainfall–runoff modeling. Journal of
topographical indices: TWI and its components. Journal of Hydrology 347, 79–89. Hydrology 349, 115–124.
Sorooshian, S., Gupta, V.K., 1995. Model calibration. In: Singh, V.P. (Ed.), Computer Yang, J., Reichert, P., Abbaspour, K.C., 2007. Bayesian uncertainty analysis in
Models of Watershed Hydrology. Water Resources Publication, Highlands distributed hydrologic modeling: a case study in the Thur River basin
Ranch, CO. (Switzerland). Water Resources Research 43, W10401. doi:10.1029/
Thiemann, M., Trosset, M., Gupta, H., Sorooshian, S., 2001. Bayesian recursive 2006WR005497.
parameter estimation for hydrological models. Water Resources Research 7 Yang, J., Reichert, P., Abbaspour, K.C., Xia, J., Yang, H., 2008. Comparing uncertainty
(10), 21–35. analysis techniques for a SWAT application to the Chaohe Basin in China.
Vrugt, J.A., Robinson, B.A., 2007. Treatment of uncertainty using ensemble Journal of Hydrology 358, 1–23.
methods: comparison of sequential data assimilation and Bayesian model Zhang, W., Montgomery, D.R., 1994. Digital elevation model grid size, landscape
averaging. Water Resources Research 43, W01411. doi:10.1029/ representation, and hydrologic simulations. Water Resources Research 30,
2005WR004838. 1019–1028.
Vrugt, J.A., Gupta, H.V., Bouten, W., Sorooshian, S., 2003. A Shuffled complex Zhang, L., Singh, V.P., 2006. Bivariate flood frequency analysis using the copula
evolution metropolis algorithm for optimization and uncertainty assessment of method. Journal of Hydrologic Engineering 11, 150–164.
hydrologic model parameters. Water Resources Research 39 (8), 1201. Zhang, Q., Xu, C.-Y., Singh, V.P., Yang, T., 2009. Multi-scale variability of sediment
doi:10.1029/2002WR001642. load and streamflow of the Lower Yangtze River basin: possible causes and
Vrugt, J.A., Diks, C.G.H., Gupta, H.V., Bouten, W., Verstraten, J.M., 2005. Improved implications. Journal of Hydrology 368, 96–104.
treatment of uncertainty in hydrologic modeling: combining the strengths of Zhang, Q., Xu, C.-Y., Zhang, Z., Chen, Y.D., 2010. Changes of atmospheric water vapor
global optimization and data assimilation. Water Resources Research 41 (1), budget in the Pearl River basin and possible implications for hydrological cycle.
1–17. Theoretical and Applied Climatology. doi:10.1007/s00704-010-0257-z.