Professional Documents
Culture Documents
Journal of Hydrology: Nibret A. Abebe, Fred L. Ogden, Nawa R. Pradhan
Journal of Hydrology: Nibret A. Abebe, Fred L. Ogden, Nawa R. Pradhan
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: HBV is a conceptual hydrological model extensively used in operational hydrological forecasting and
Received 18 September 2008 water balance studies. In this paper, we apply the HBV model on the 1924 km2 semi-humid Leaf River
Received in revised form 6 May 2010 catchment near Collins, Mississippi. We analyze individual sensitivity of the parameters by calibrating
Accepted 6 June 2010
the model using the Multi-Objective Shuffled Complex Evolution (MOSCEM) algorithm, perform
Monte-Carlo based identifiability analysis and investigate the dynamic behavior of the parameters using
This manuscript was handled by
Konstantine P. Georgakakos, Editor-in-Chief, the Dynamic Identifiability Analysis (DYNIA) approach in reference to the hydrological process in the
with the assistance of Michael Brian Butts, catchment. The sensitivity analysis using two objective measures showed that there are distinct groups
Associate Editor of parameters that control total runoff volume errors and errors from the high-flow series. The DYNIA
analysis revealed that parameters have specific periods where they show higher identifiability and play
Keywords: a crucial role in representing the predicted stream flow. Temporal changes of parameter optima were
HBV observed due either to inadequacies in the model structure or possible time-varying catchment response
Sensitivity analysis subject to unsteady hydrodynamic and hydroclimatic conditions.
Identifiability analysis Ó 2010 Elsevier B.V. All rights reserved.
DYNIA
Model structure
0022-1694/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2010.06.007
302 N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310
to a higher degree, as measured by objective functions. However, plied to a catchment. They also argued that the fact that a param-
analysis of behavioral parameter sets also shows that as they give eter is identifiable does not imply the assumptions about the
equally good results, there is no unique solution to the calibration parameter in the model are correct for they found a very identifi-
of hydrological models. They argued that if there are ample Monte- able parameter was found to be time-variant. They attributed the
Carlo simulations, the likelihood-weighted predictions could be poor identifiability of some of the parameters and errors in the
evaluated to get the prediction quantiles at any time step. model structure as the main contributors to the model uncertainty
The identifiability of a parameter expresses how well defined for unobserved conditions.
the parameter is within a model structure. It is a measure of how Lee and Moon (2007) used two conceptual models (a Penman
closely grouped the values of a parameter are among the behav- model and a probability-based model) with no state variables
ioral parameter sets. In other words, if a parameter is well-defined, and analyzed model parameter identifiability and the models’
then the behavioral parameter values lie in a narrow region of the structures using the DYNIA approach. First, they found that the
parameter range and the distribution of the likelihood values of the optimal values of some of the parameters are time-invariant while
parameter is non-uniform. The optimum values of a parameter the rest exhibit time dependency. However, contrary to the obser-
may be influenced by the length and information content of the vation by Tripp and Niemann (2008), the performance of the mod-
calibration data (Wriedt and Rode, 2006). In addition to standard els in terms of representing the outflow hydrograph remained
assumptions, the optimum parameter sets can also be a function fairly the same when they improved the models’ structures by
of the values of the model state variables. Senarath et al. (2000) including soil moisture index state parameter. They concluded that
have shown that initial conditions affect the optimized parameters data quality has much to do with model performance in addition to
for single event calibration. Thus, relations between parameter op- model structure errors.
tima and state variables could result in time varying optimal Results of application of the DYNIA algorithm in assessing mod-
parameters (Wriedt and Rode, 2006). el structural problems and the corresponding parameter identifi-
If we assume that a given model structure is a combination of ability are dependent on the model used and the corresponding
different hypotheses on the workings of the natural system, then underlying assumptions. In this study, we analyzed the sensitivity
they should be individually testable and should be related to indi- of the HBV model parameters in reference to the hydrological pro-
vidual model components (Wagener et al., 2003). However, the cesses in a catchment, performed identifiability analysis of the
general assumption is that model parameters are constant in time model parameters in relation to parameter interaction and as-
given that catchment characteristics do not change with time, and sessed the dynamic variability of the optimal parameter values in
if parameter optima changes, then the inference is that there is a reference to the model structure. We used data from the
missing aspect in the model formulation and thus a model struc- 1924 km2 semi-humid Leaf River catchment, which is located in
tural error. southern Mississippi. In light of the problems mentioned above,
DYNIA is an approach to parameter identification assuming that the objectives of this work were to:
parameter optima changes in time. The algorithm uses an identifi-
ability measure based on likelihood values of parameter sets in Identify the most sensitive parameters of the model with
representing the model variable and reveals temporal changes of respect to multiple objective measures so that process based
the parameter distribution and parameter uncertainty (Wriedt calibration would be possible.
and Rode, 2006). The main motivation behind this approach is that Analyze the identifiability of parameters of the model in refer-
by aggregating model errors in time, we lose information as to the ence to parameter interaction.
performance of the individual components of a model, which be- Analyze the temporal change of the optimum parameter values
comes activated during different periods in time. Hence the ap- and the implications on the model structure.
proach helps in identification of parameters and model structures Assess the performance of the model in a semi-humid climate
by improving the amount of information that can be retrieved from dominated by convective rainfall and qualitatively analyze the
observed catchment variables through using a moving window. uncertainty thereof.
Cullmann and Wriedt (2008) tried to identify factors affecting
event-based optimum parameter sets and temporal variation of 2. HBV model
parameter values with the objective to link best parameter values
with a priori knowledge (pre-event conditions) so as to find alter- HBV model (Bergstrom, 1976) belongs to the conceptual class of
native ways of model updating. They found that the optimal value hydrological models, which are characterized by attempts to cover
of the drainage parameter of the WaSim-ETH model (Shulla, 1997) the most important runoff generating processes using a simple and
depends on discharge, which was taken as an integral measure of robust structure, and a small number of parameters. The model
catchment wetness. They concluded that the range of the optimum consists of three main modules: snow accumulation and melt, soil
parameter values varies with time and with discharge revealing a moisture accounting and river routing and response modules. The
model structure deficiency to simulate the catchment process soil moisture accounting routine, which is controlled by three
dynamics before any flood event. They also argued that state- parameters, computes an index of the wetness and soil moisture
dependent parameter changes should be considered in the formu- storage in a catchment. Parameter FC is the maximum soil storage
lation of conceptual models intended for continuous simulations. capacity in the basin and parameter b determines the relative con-
In a similar way, McCuen (1973) suggested that the dynamic nat- tribution to runoff from a millimeter of rain or snowmelt at a given
ure of sensitivity of parameters should be taken into account dur- soil moisture deficit while parameter LP controls the shape of the
ing hydrologic model formulation. reduction curve for potential evapotranspiration. The response
Tripp and Niemann (2008) analyzed two models that simulate routine transforms excess water from the soil moisture routine
the local and spatially averaged soil moisture in a watershed to to discharge to each sub-basin. It consists of two reservoirs con-
see if parameter uncertainty and model structural errors contrib- nected in series by constant percolation rate. The routine has three
ute to the model uncertainty for forecasting. From their analysis, recession coefficients, namely K0, K1 and K2, a reservoir threshold
they found that the more physically-based soil moisture model UZL, and a constant percolation rate PERC. Computed outflow from
did not produce fewer structural errors than the probabilistic mod- a catchment is transformed using a triangular weighting function,
el they used as the underlying physical assumptions (e.g., uniform the base width of which is the calibration parameter MAXBAS. The
hydraulic conductivity) of the model could be violated when ap- transformed catchment response is routed with a Muskingum
N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310 303
Table 1
Rain and Snow
evapotranspiration Daily statistics of hydrometeorological data at the gauging station.
snow
snow Measure Rainfall Potential evapotranspiration Runoff
TT, CFMAX (mm) mm) (mm)
elevation
FC
FC, LP, BETA
"Soil routine" tal number of calibration parameters was reduced from 12 to 9. A
schematic representation of the model structure is shown in Fig. 1
recharge Q0=K0(SUZ−UZL)
UZL 3. The Leaf River catchment
SUZ
The Leaf River catchment, located in the Covington County, Mis-
lumped
Q1=K1*SUZ sissippi, and bounded by the coordinates 31°420 2300 N, 32°210 3500 N
PERC "Routing routine" latitude and 89°120 3000 W, 89°410 5000 W longitude, is shown in
Fig. 2. It has a total area above the stream flow gauge of
SLZ
1924 km2 and has been extensively used as an experimental catch-
ment for various studies (e.g., Sorooshian et al., 1983, 1993; Duan
"Response function" Q2=K2*SLZ MAXBAS
et al., 1992; Yapo et al., 1998). The catchment is located in the
coastal plain geographic province and is characterized by a gentle
Computed Runoff rolling topography. The soil textures range from a clay to a coarse
sands with 15.13 cm average available water storage within the
Fig. 1. HBV model structure (reproduced from Seibert (2000)). top 100 cm depth while saturated hydraulic conductivity values
range from 0.0126 to 50.81 cm/h with a mean value of 4.331 cm/
h (STATSGO online source, August 2008). We used the same hydro-
routing model. However, in this application, given that the catch- logical and hydrometeorological data that were used in Yapo et al.
ment is snow-free and that the daily potential evapotranspiration (1996), where continuous data from 1948 through 1988 are avail-
data were available, for the version used in this application, the to- able at time steps of 6 h. Table 1 shows the daily statistics of the
Elevation (m)
63−87
N 88−102
103−114
115−126
127−136
0 10 20 137−150
Scale (km) 151−193
hydrological data derived from 40 years of recorded data between For the Dynamic Identifiability Analysis, first we ran Monte-
1948 and 1988 at the outlet of the catchment. Carlo simulations with the RMSE used as an objective measure.
In evaluating the identifiability of the parameters at each time step,
4. Methodology the algorithm first selects n time steps of observed and the Monte-
Carlo simulated data before and after the current time step (i.e., a
For the individual sensitivity analysis, first we calibrated the window size of 2n + 1 steps) and then calculates the average model
HBV model using the Multi-Objective Shuffled Complex Evolution error for this window from the best performing top 10% parameter
Metropolis algorithm (Vrugt et al., 2003) on data from October 1, sets. This error is assumed to be the average model error associated
1972 to September 30, 1983. The ranges of the model parameters to the current time step. Next a cumulative probability distribution
used in this automatic calibration are shown in Table 2 and were of the top 10% best performing simulations is computed using the
obtained from Mr. Bo Holst, International Projects Manager, Swed- method discussed in the text above. A 90% confidence interval of
ish Meteorological and Hydrological Institute, Environment and the parameter values for each time step and for each parameter
Safety Services Division (Holst, pers. comm., August, 2006). We is then calculated. In those time steps where we observe a nar-
used the absolute BIAS and the RMSE objective measures of model rower 90% confidence interval and a non-uniform probability dis-
performance and aggregated them into a single objective measure tribution, a parameter has higher identifiability and vice versa.
using the Euclidian distance approach outlined by Madsen (2000). Generally, the size of the moving window depends on the temporal
These measures are given as: length of influence of a given parameter and the quality of input
!! and observed flow data. And in view of this and based on previous
1 Xn
experiences in other applications, for all parameters, we used a
BIAS ¼ abs ½Q Q prd;i ð1Þ
n i¼1 obs;i time step of 6 h and a window size of 61 time steps.
Two plots are produced for each parameter as an output of the
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi analysis. The first plot (upper plot) shows the 90% confidence inter-
Pn 2
i¼1 ½Q obs;i Q prd;i val, the normalized observed hydrograph with respect to the max-
RMSE ¼ ð2Þ
n imum flow and the identifiability of the parameters. The second
plot (lower plot) shows the information content in time calculated
where Q obs;i and Q prd;i are the observed and predicted flow for each as one minus the normalized width of the confidence limit with re-
time step and n is the number of steps in the simulation period con- spect to the original parameter range. Hence, those time steps with
sidered. After obtaining a representative parameter set for the narrower widths in the confidence interval of the upper plot result
catchment, we divided the range of each parameter into 10 equal in higher information content values in the lower plot and vice
increments and the model was run by varying one parameter at a versa.
time and over these divisions while keeping the rest of the param-
eters at their optimized values. This helped us to see the net effect
5. Results and discussion
of each parameter without having to involve the others and as mea-
sured by BIAS and RMSE separately.
5.1. Individual sensitivity analysis
For the identifiability analysis, we used 60,000 Monte-Carlo
simulations and screened these simulations into behavioral and
A representative plot of the simulated and the observed flows
non-behavioral parameter sets using a threshold Nash–Sutcliffe
for the first two years of the calibration period is shown in Fig. 3,
efficiency value of 0.88. The Nash–Sutcliffe efficiency measure
the corresponding optimized parameter values are given in Table 2,
(R2), given by Eq. (3), was used as an objective measure so that
and the individual sensitivity analyses of each parameter are
all model parameters could be compared on a standard scale.
shown in Figs. 4 and 5 respectively. Fig. 4 shows that the recession
Pn
½Q obs;i Q prd;i 2 coefficients K0 and K1, the reservoir threshold UZL, the percolation
R2 ¼ 1 Pi¼1
n 2
ð3Þ parameter PERC and the base of the unit hydrograph function
i¼1 ½Q obs;i Q obs;i
MAXBAS are insensitive to the BIAS measure. The parameter sensi-
where Q obs;i , Q prd;i and n are as defined above and Q obs is the average tivity result based on a RMSE measure shown in Fig. 5 reveals that
observed flow. For each parameter, a cumulative probability distri- the response and transformation routine parameters K1, MAXBAS
bution of the posterior parameter value distribution is computed and UZL appear to be the most influential parameters controlling
from the screened behavioral parameter sets and the gradient of the high-flow series and thus their clear sensitivity to the RMSE
this distribution is taken as a measure the identifiability of the measure, which to a large extent depends on the high-flow series.
parameters (Beven and Freer, 2001; Wagener, 2004). The least iden- Sensitivity of K1 resembles that of UZL in the sense that both
tifiable parameter with a uniform posterior density distribution will showed a clear optima and a change in RMSE from 27 m3 s1 to
have a nearly 1:1 slope for this cumulative probability distribution. 40 m3 s1 when varied over half of their respective ranges. On
Table 2
HBV model parameters, parameter ranges and optimized values.
Rainfall (mm/6h)
20
40
60
80
1500 100
1000
Runoff (m s )
3 −1
Observed
Simulated
500
0
2
7 2 3 3 3 3 3 3 4 4 4 4 4
19 97 197 197 197 197 197 197 197 197 197 197 197
t− c −1 n − r − n − g − t − c − n − r − n − g − t −
Oc De Ja A p Ju Au O c
D e Ja A p Ju Au O c
Fig. 3. Observed flow, simulated flow and precipitation of the first two years of the ten years data used in calibrating the HBV model at Leaf River near Collins. The
corresponding statistics of this calibration are: BIAS = 0.24 m3 s1; RMSE = 27.10 m3 s1 and R2 = 0.94.
15 15 15 40 40 40
RMSE m3s−1
10 10 10 35 35 35
BIAS
5 30 30 30
5 5
25 25 25
0 0 0 0 0.5 1 0 0.5 1 0 0.5 1
0 0.5 1 0 0.5 1 0 0.5 1
UZL K0 K1 UZL K0 K1
15 40 40 40
RMSE m3s−1
15 15
10 10 35 35 35
BIAS
10
30 30 30
5 5 5
25 25 25
0 0 0 0 0.5 1 0 0.5 1 0 0.5 1
0 0.5 1 0 0.5 1 0 0.5 1
K2 PERC MAXBAS
K2 PERC MAXBAS
40 40 40
RMSE m3s−1
15 15 15
35 35 35
10 10 10
BIAS
30 30 30
5 5 5
25 25 25
0 0.5 1 0 0.5 1 0 0.5 1
0 0 0
0 0.5 1 0 0.5 1 0 0.5 1 LP FC BETA
LP FC BETA
Fig. 5. Sensitivity plots of parameters with reference to RMSE objective function
Fig. 4. Sensitivity plots of parameters with reference to BIAS objective function (m3 s1).
(m3 s1).
value of K2, the higher the volume that remains in deep aquifer
the contrary, parameters K2 and LP, which showed some degree of storage and contributes to the regional groundwater flow system.
sensitivity to the BIAS measure in Fig. 4, are less sensitive to the This in turn decreases the outflow at the catchment outlet and thus
RMSE measure. K0 and PERC, though not very significant, show the parameter’s sensitivity around its lower edge in Fig. 4. How-
some sensitivity in Fig. 5. ever, increasing the value of K2 above an optimal percolation rate
From the structure of the model, it is clear that parameters K0, has no effect on the outflow, meaning that it is insensitive after
K1, UZL, PERC and MAXBAS play roles in the response and transfor- nearly 15% of its range.
mation routines. Thus, they influence the high-flow series more In the HBV model formulation, FC plays a key role in partition-
than the volume balance. Specifically UZL, K0 and K1 are parame- ing the effective precipitation into runoff and soil moisture. A low-
ters that control the overland runoff and the quick interflow. Their er FC value means that the soil water holding capacity is very low
influence on the catchment response is through changing the (thin soil layer) and thus a smaller amount of water is available for
shape of outflow hydrograph at the outlet of the catchment and evapotranspiration and vice versa. This makes FC the main param-
hence they have negligible effect when it comes to the overall vol- eter to control volume error and thus the sensitivity of this param-
ume error. Parameter K2, which controls the base flow from deep eter to the BIAS measure. As b is an exponent on the ratio between
groundwater, affects the volume error. The lower the parameter soil moisture and field capacity, it affects the division of the net
306 N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310
CDF
pends on this exponent and hence the strong effect of the param- 0.5 0.5 0.5
eter on the volume error. A very steep gradient over its range
reveals the effect this parameter has on the volume error. Parame- 0 0 0
0 25 50 75 100 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3
ter LP, which determines the actual evapotranspiration in refer-
UZL (mm) −1 −1
ence to the available soil moisture, has moderate effect on the K0 (h ) K1 (h )
volume balance. A mild slope over its range reveals that though 1 1 1
the volume error is influenced by the change in the parameter va-
CDF
lue, the net effect is not as such very significant. Overall, in this 0.5 0.5 0.5
semi-humid catchment we found that b and FC are the most dom-
inant parameters of the model controlling the volume error. 0 0 0
0 0.05 0.1 0 0.1 0.2 13 16 19 22 24
PERC and K2 work in combination. A higher PERC value allows
−1
high flow of water from the upper to the lower reservoir and thus K2 (h ) PERC (−) MAXBAS (h)
1 1 1
more storage in the lower one. This combined with high or low K2
value determines the error in the low flow series. However, given
CDF
0.5 0.5 0.5
that K2 determines only the base flow, it has less effect on peak
flows and thus a mild variation of the RMSE measure over the
range of the parameter. On the other hand, PERC plays a role by 0 0 0
20 40 60 80 100 100 150 200 250 300 1 2.25 3.5 4.75 6
reducing the volume of water in the first tank. However, given that
LP (mm) FC (mm) BETA (−)
it has no direct role on the peak flow, it has a relatively minor effect
on the RMSE objective measure when varied over its range. Fig. 6. Identifiability plot of parameters with respect to the Nash–Sutcliffe
The parameter K0 controls the overland runoff that forms when efficiency objective measure.
cumulative precipitation is in excess of the reservoir threshold UZL.
Given that the parameter is functional only when the cumulative
Parameters UZL, MAXBAS, FC and K1 tended to be more identi-
precipitation is higher than the optimized and fixed UZL value,
fiable than the remaining parameters. A closer look at the internal
we see the effect of the parameter only when the system is already
definitions of these parameters showed that these are the main
wet and there is ample input of precipitation. However, given that
parameters in their respective model components. The plot for
the number of times where the threshold could be exceeded is
K0 showed more identifiability as compared to its sensitivity in
much smaller as compared to the total period considered, the net
the individual parameter analysis. Given that we used the effi-
effect of the parameter on the overall aggregated error measure
ciency measure, LP might not be properly identifiable in this anal-
is not so strong. The mild gradient that is shown in Fig. 5 with a to-
ysis. However, K2 and PERC consistently showed that they are the
tal change in RMSE from 27 m3 s1 to 32 m3 s1 over the range of
least identifiable when sampled simultaneously with other param-
the parameter supports the argument. The formulation of the re-
eters and the least sensitive when analyzed independently on both
sponse routine is such that K1 and UZL influence each other. Given
RMSE and BIAS measures signaling that there is a model formula-
K1 and K0 fixed at their optimum values, varying the threshold UZL
tion problem in the way the two reservoirs are connected and in
affects the amount of water that would be available in the reservoir
the way the second reservoir is parameterized. There could also
and that forms direct interflow. This indirectly affects the high-
be a component in the catchment response that is missing in the
flow series at the catchment and its observed high sensitivity to
model response routine formulation and which resulted in the
the objective measure and a clear optimum.
least identifiability of these parameters. This identification of the
In general, while FC, b, and LP determine the balance in volume,
most interacting and the least identifiable parameters helps reduce
UZL, K0, K1 and MAXBAS control the high-flow series in the catch-
the risk of having uncertainty in the model output during model
ment. In addition, FC was observed to have a strong role in both the
application.
high-flow series and the volume error, whereas PERC seems to
have a simple role of redistributing the volume between the two
reservoirs with no clear effect on either of the two objective mea- 5.3. DYNIA analysis
sures. From this result, we can see that calibration of the model
without considering this division of parameters along the objective Dynamic Identifiability Analysis results for five of the total nine
measures and using only single objectives would incur uncertainty parameters on the two-water-year data set extending from Octo-
in the model output. In cases of manual calibration of the model, ber 1, 1972, to September 30, 1974, are shown in Figs. 7–11. These
usage of the multi-step calibration approach described in Yapo parameters were selected based on whether or not each of them
et al. (2000), where each distinct group of parameters are cali- are very identifiable and/or represents one of the components of
brated sequentially considering two or more objectives, would re- the model. In the upper plots of these figures, the dotted lines rep-
duce the uncertainty in the model prediction. resent the limits of the 90% confidence interval likelihood distribu-
tion in time while the solid dark lines show the normalized
5.2. Identifiability analysis observed flow series with respect to the maximum flow. The grey
shading is the identifiability of the parameters in time. The inten-
Fig. 6 shows the identifiability of the model parameters when sity of the shading in the vertical direction at each time step shows
Nash–Sutcliffe efficiency is used as an objective measure and when the magnitude of the probability of the corresponding parameter
the model parameters were sampled simultaneously. Each of the value in representing the observed hydrograph. Hence, for each
plots in the figure shows graphs of the cumulative distributions step, those areas with darker shading represent the parameter val-
of the posterior parameter distribution and a theoretical uniform ues with higher probability and vice versa. The lower plots show
distribution of a total of 1322 behavioral parameter sets, which the information content for each time step as discussed previously.
met the threshold efficiency of 0.88. This result gave us insight In general, parameters were found to have specific periods
as to how the parameter interaction affects the identifiability of where they are more identifiable, more sensitive and where they
parameters when sampled simultaneously. play a clearer role than other periods. From the lower plot of
N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310 307
250
FC
200
150
information content
1
0.8
0.6
0.4
0.2
0
Fig. 7. Parameter identifiability, likelihood distribution and information content for FC for the period from October 1, 1972 to September 30, 1974.
5
BETA
4
3
2
1
0.8
0.6
0.4
0.2
0
Fig. 8. Parameter identifiability, likelihood distribution and information content for BETA for the period from October 1, 1972 to September 30, 1974.
22
MAXBAS
20
18
16
1
0.8
0.6
0.4
0.2
0
Fig. 9. Parameter identifiability, likelihood distribution and information content for MAXBAS for the period from October 1, 1972 to September 30, 1974.
80
LP
60
40
1
0.8
0.6
0.4
0.2
0
Fig. 10. Parameter identifiability, likelihood distribution and information content for LP for the period from October 1, 1972 to September 30, 1974.
308 N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310
80
UZL
60
40
20
information content
1
0.8
0.6
0.4
0.2
0
Fig. 11. Parameter identifiability, likelihood distribution and information content for UZL for the period from October 1, 1972 to September 30, 1974.
Fig. 7, FC is more identifiable during dry periods where the soil moisture recharge. We observe that the parameter takes low val-
moisture is less than the maximum field capacity and where there ues at periods where the soil moisture is below the field capacity
is precipitation input. The optimum value of the parameter shows and where runoff has to form from a given precipitation. On the
a variation over the period considered taking higher values from other hand, in a similar way as to the parameter FC, the informa-
October 1972 till January 1973 and lower values later on for the tion content as well as identifiability of b decreases considerably
periods around February 1973, August 1973 and December 1974. during the wet season.
During the run-up to the winter and early-spring rainy season, The DYNIA plot of MAXBAS is shown in Fig. 9. The lower infor-
FC plays a strong role in determining the rate of water accumula- mation content plot shows that the parameter is more identifiable
tion in the storage reservoirs. The higher information content from at high flow periods and during the wet season. The plots for the
October 1972 till January 1973 could be attributed to this process. periods from January 1973 till June 1973 and from December
After the soil attains a higher level of wetness, the role of the FC is 1973 till May 1974 support this argument. On the other hand, dur-
diminished. The corresponding information content of the param- ing the wetting up period and the dry summer, the parameter has
eter during the wet season (between January and May), therefore little effect on the low flow series resulting in a low information
becomes very low. After the rainy season is over, we see increasing content and weak identifiability.
information content for FC at various periods between May and Assuming a constant volume of runoff per given amount of pre-
September of the first water year. Overall, the information content cipitation, the higher the value of MAXBAS, the lower the high-flow
of FC increases at periods where the soil moisture is below maxi- series from the catchment and vice versa. However, in the event of
mum capacity and when there is precipitation input. pronounced peaks, the linear transformation function assumption
With regard to the optimal parameter values, it could be said may not be appropriate for simulating peak flows due to flows in
that parameter FC took different values over the period considered. the flood plain and the resulting non-linear effect on channel flow
This could be attributed to the formulation of the soil moisture routing, which was not activated in this application owing to a
routine in the model. In reality, the partitioning of the rainfall into lumped approach followed. Due to this, the high-flow series from
runoff and soil moisture depends on many other factors including the model could be underestimated under normal MAXBAS opti-
the type, intensity and duration of precipitation which are not mal values. To balance this effect and produce higher peaks, the va-
explicitly considered in the model. For those steps where runoff lue of MAXBAS should decrease. Thus, for the pronounced peak
formation is suppressed by incompleteness in soil moisture routine observed in April 1974, the parameter assumes a very low value.
formulation, the FC parameter must assume a very low value to The posterior distribution of the parameter also showed that the
counter balance the effect of the missing factors and thus the fluc- 90% confidence limit narrowed with a corresponding proportional
tuation in parameter optima over time. The observation at times increase in the information content signifying that a unique value
steps of around February 1973, August 1973 and December 1974 of the parameter became important during that period.
where the parameter takes lower optimal values supports this The information content plot in Fig. 10, which is the DYNIA re-
argument. Especially in high-intensity convective rainfall, runoff sult for parameter LP, shows that the parameter becomes slightly
could easily form before the soil moisture increases appreciably. more identifiable when there is a high flow during summer season.
On the contrary, if there is a catchment process that was not taken For the month of August 1973, the information content is a little
into account in the routine and the absence of which overestimates higher and the corresponding probability distribution shows that
the amount of rainfall that goes to runoff formation, then the the parameter takes a unique value in its upper part of the range.
parameter assumes higher values to balance that effect. Higher However, for the rest of the series, the parameter assumes values
optimal values that were observed at steps of around July 1973, that vary widely over the range and the information content is
September 1973 and January 1974 could be due to such factors. nearly uniform. However, given that the parameter is not sensitive
Fig. 8 is the DYNIA plot of parameter b. We observe that the to the RMSE objective measure use in the analysis, and due to the
information content is higher during wetting up periods and dry absence of a period where the parameter is uniquely identified, no
summer periods where the soil moisture is lower than the maxi- conclusion could be drawn as to the variation of the parameter LP
mum capacity. On average, the periods from October 1972 till in time.
December 1972, from June 1973 till January 1973 and from May The DYNIA analysis of parameter UZL is shown in Fig. 11. The
1974 to October 1974 are informative regions. In general, in all information content plot exhibits somewhat higher identifiability
the steps where there is a rainfall input to the system and the soil during periods when there is precipitation input and the catch-
moisture has not reached the FC level, b has a significant effect on ment is relatively wet. This pattern is seen for the periods between
the partitioning of the precipitation into soil moisture and runoff January 1973 and May 1973 and between January. 1974 and April
and on the corresponding information content. For a given precip- 1974. The corresponding posterior distribution also shows that
itation input, low b values favor runoff formation more than soil during these steps lower values of this parameter are more likely.
N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310 309
However, the contribution of UZL to high-flow series either by NIA approach to real hydrological model applications in general
favoring more overland runoff (low UZL) or more quick interflow and the first attempt to apply it to the HBV model in particular.
from the first reservoir (high UZL) is obscured by the effect of the We believe that our analysis of the model and the results of our
two recession parameters K0 and K1. Thus, the net effect of this study will be of interest to hydrologists who use HBV or similar
parameter is difficult to discern unless the recession parameters models, either for operational forecast purposes or for water bal-
are kept constant. ance studies.
To summarize, the DYNIA approach showed that depending on Individual sensitivity analysis of the parameters around cali-
their role in the model, parameters are better identified for a cer- brated optimum points showed that the model has two groups of
tain segment of the total series. This is because different modes parameters that effectively control the volume error and the error
of catchment response, which are represented by different compo- in high-flow series. The first group contains FC, b, and to a certain
nents of the model, occur at different times implying that the cor- extent LP, which are parameters of the soil moisture and evapo-
responding parameters become effective during that period. The transpiration components. We observed that they have a strong ef-
consideration of specific segment of the data series maximizes fect in the total volume error of the model. Especially, FC was found
the gain of information that comes from aggregating residuals to be the most dominant parameter of the model, which affects
from periods where a parameter is more identifiable, thus giving both the high-flow series and the volume errors showing sensitiv-
a reasonably accurate estimation of that particular parameter. ity to both RMSE and BIAS objective measures. The second group of
However, the assumption almost universally employed in model parameters constitute of UZL, K1, K0, and MAXBAS mainly from the
development is that catchment response mechanism is constant response and transformation routines. These parameters showed
in time given the corresponding catchment properties are con- higher sensitivity by the RMSE objective measure and thus their
stant. This implies that if there is a change in parameter optimal significant influence on the peak flows. The third group contains
value then there is a model structure problem (Wagener et al., parameters K2 and PERC. These parameters were found to have
2003). no dominant effect on either the high-flow series or the volume er-
We observed that the optimal values of the parameters showed ror because of either inadequate model structure of the respective
a variation in time not only because there is a model structure routines or due to the existence of parameter interaction. Given
inadequacy but also the model structure is fixed while important that our analysis has shown that there are two groups of parame-
catchment processes are unsteady. Hence, the fixed structure of ters sensitive to the objective measures used, the traditional ap-
the model makes it impossible for the parameter optima to assume proach of calibrating the HBV model using the Nash–Sutcliffe
the same values when the catchment is subjected to an unusual efficiency measure, which is much the same as the RMSE measure
hydrometeorological forcing or seasonal changes in the landscape. and which emphasizes the high-flow series, should be replaced by
For instance, the dynamics of a flood wave, which is one of the fac- a combined multi-objective approach that measures different
tors that affect peak flows, could change with the type, intensity, modes of a catchment response.
uniformity, and duration of rainfall. Thus, in very intense events The Monte-Carlo based identifiability analysis revealed that the
the linear transformation assumption in the model may not be va- main parameters of the respective components of the model show
lid and this could force the transformation parameter to assume higher identifiability as compared to the remaining parameters. On
different optima. Measured soil moisture values at 5, 20, 51 and the contrary, some of the parameters of the response routine
102 cm soil depth in an agricultural area near the catchment fur- showed the least identifiability when sampled simultaneously
ther showed that there are quasi-stable soil moisture profiles with revealing a strong parameter interaction, a model formulation
depth that vary from season to season. Seasonally-dependent vari- problem or a missing component in the catchment response
ations of soil moisture with depth call into question the assump- routine.
tion of the persistence of different flow generation mechanisms DYNIA analysis on five of the nine model parameters showed
(direct runoff, quick interflow, delayed interflow and base flow) that MAXBAS, FC and b have clearly defined periods where the
and their representations in the model formulation. information content is very high. We also found FC and b to have
There are two possible explanations for the existence of dy- the highest information content during wetting-up and dry sum-
namic variability of the optimal parameter values. One is model mer periods. On the contrary, UZL was found to have on average
formulation inadequacies, and the second is unsteadiness in dom- uniform information content in time indicating that the model
inant watershed processes. Bronstert et al. (2002) provide a com- component that the parameter belongs to does not have specific
prehensive list of catchment-scale land-use changes that periods where it plays a key role in the catchment process. A sys-
frequently occur and affect hydrologic response. These vary from tematic dependence between model parameters and the model
seasonal changes in leaf litter to land-surface changes due to agri- state variables was also observed. Analysis of parameters FC and
cultural processes. Very few hydrologic models have the ability to b showed that their identifiability is related to the state of soil
explicitly simulate unsteady catchment processes. From our analy- moisture in the catchment indicating a calibration scheme consid-
sis, we observe that the assumption that model parameters are ering state variables, where possible, might help find the true and
constant in time leads to high level of uncertainty in the model representative parameter values of a given catchment.
output. From the DYNIA analysis, we also found that parameter optima
could change in time not only because of model structure inade-
quacies, but also because of possible unsteadiness in catchment-
6. Conclusions scale processes. The current hydrological model development par-
adigm assumes that the dominant runoff generation processes is
This study performed a detailed analysis of the sensitivity and constant. Analysis of the 1974 flood event and the change in time
function of the HBV model parameters on different flow periods of the optimal MAXBAS parameter showed that the fixed linear
and different objective functions. We used data from the semi-hu- transformation function in the model structure, which does not
mid Leaf River watershed located in the sub-tropical coastal plain have the flexibility to adjust to a change in the catchment’s re-
of the southeastern United States. Our study is the first attempt sponse caused by a change in hydrodynamic and hydroclimatic
to calibrate the HBV model using multiple objectives and using conditions, results in changes in optimal parameter in time. None-
the Multi-Objective Shuffled Complex Evolution Metropolis algo- theless, the HBV model was found to perform well in a large, snow-
rithm. This is also one of the few attempts made to apply the DY- free, semi-humid catchment with frequent convective rainfall with
310 N.A. Abebe et al. / Journal of Hydrology 389 (2010) 301–310
efficiency of above 0.9 and good fit to both the low and high-flow Madsen, H., 2000. Automatic calibration of a conceptual rainfall–runoff model using
multiple objectives. J. Hydrol. 235, 276–288.
series.
McCuen, R.H., 1973. The role of sensitivity analysis in hydrologic modeling. J.
Hydrol. 18, 37–53.
Acknowledgments Seibert, J., 2000. Multi-criteria calibration of a conceptual runoff model using a
genetic algorithm. Hydrol. Earth Syst. Sci. 4 (2), 215–224.
Senarath, S.U.S., Ogden, F.L., Downer, C.W., Sharif, H.O., 2000. On the calibration and
We are thankful to Thorsten Wagener of Pennsylvania State verification of two-dimensional, distributed, hortonian, continuous watershed
University for providing the DYNIA software. This research was models. Water Resour. Res. 36, 1495–1510.
supported by the US Army Research Office through Grant Shulla, J., 1997. Hydrologische Modellierung von Flussgebieten zur Abschätzung der
Folgen von Klimaänderungen, Diss., ETH Zürick, CH, 161 pp.
W911NF-07-1-0389 to the University of Wyoming. We acknowl- Sieber, A., Uhlenbrook, S., 2005. Sensitivity analyses of a distributed catchment
edge contributions from three anonymous reviewers. model to verify the model structure. J. Hydrol. 310, 216–235.
Sorooshian, S., Gupta, V.K., Fulton, J.L., 1983. Evaluation of maximum likelihood
parameter estimation techniques for conceptual rainfall–runoff models:
References influence of calibration data variability and length on model credibility.
Water Resour. Res. 19, 251–259.
Bergstrom, S., 1976. Development and Application of a Conceptual Runoff Model for Sorooshian, S., Duan, Q., Gupta, V.J., 1993. Calibration of rainfall–runoff models:
Scandinavian Catchments. Report RHO 7, Swedish Meteorological and application of global optimization to the sacramento soil moisture accounting
Hydrological Institute, Norrkoping, Sweden, 134 pp. model. Water Resour. Res. 29 (4), 1185–1194.
Bertalanffy, L. Von, 1950. The theory of open systems in physics and biology. Science Spear, R.C., Hornberger, G.M., 1980. Eutrophication in peel inlet, II, identification of
New Series 111 (2872), 23–29. critical uncertainties via generalized sensitivity analysis. Water Res. 14, 43–49.
Beven, K.J., Binley, A., 1992. The future of distributed models: model calibration and Tripp, D.R., Niemann, J.D., 2008. Evaluating the parameter identifiability and
uncertainty prediction. Hydrol. Process. 6, 279–298. structural validity of a probability-distributed model for soil moisture. J. Hydrol.
Beven, K., Freer, J., 2001. Equifinality, data assimilation, and uncertainty estimation 353, 93–108.
in mechanistic modelling of complex environmental systems using the GLUE Uhlenbrook, S., Seibert, J., Leibundgut, C., Rodhe, A., 1999. Prediction uncertainty of
methodology. J. Hydrol. 249, 11–29. conceptual rainfall–runoff models caused by problems in identifying model
Bronstert, A., Niehoff, D., Bürger, G., 2002. Effects of climate and land-use change on parameters and structure. Hydrol. Sci. – J. Sci. Hydrol. 44 (5).
storm runoff generation: present knowledge nd modelling capabilities. Hydrol. Vrugt, J.A., Gupta, V.H., Bastidas, L.A., Bouten, W., Sorooshian, S., 2003. Effective and
Process. 16, 509–529. doi:10.1002/hyp. 326. efficient algorithm for multi-objective optimization of hydrologic models.
Butts, B.M., Payne, T.J., Kristensen, M., Madsen, H., 2004. An evaluation of the impact Water Resour. Res. 39 (8), 1214. doi:10.1029/2002WR001746.
of model structure on hydrological modeling uncertainty for stream flow Wagener, T., 2004. Monte-Carlo Analysis Toolbox User Manual. Penn State
simulation. J. Hydrol. 298, 242–266. University.
Cullmann, J., Wriedt, G., 2008. Joint application of event-based calibration and Wagener, T., McIntyre, N., Lees, M.J., Wheater, H.S., Gupta, H.V., 2003. Towards
dynamic identifiability analysis in rainfall–runoff modeling: implications for reduced uncertainty in conceptual rainfall–runoff modeling: dynamic
model parameterization. J. Hydroinform. 10 (4), 301–316. doi:10.2166/ identifiability analysis. Hydrol. Process. 17, 455–476.
hydro.2008.055. Wriedt, G., Rode, M., 2006. Investigation of parameter uncertainty and
Duan, Q., Gupta, V.K., Sorooshian, S., 1992. Effective and efficient global identifiability of the hydrological model WaSiM-ETH. Adv. Geosci. 9, 145–150.
optimization for conceptual rainfall–runoff models. Water Resour. Res. 28 (4), Yapo, P.O., Gupta, H.V., Sorooshian, S., 1996. Automatic calibration of conceptual
1015–1031. rainfall–runoff models: sensitivity to calibration data. J. Hydrol. 181, 23–48.
Harlin, J., Kung, C.S., 1992. Parameter uncertainty and simulation of design floods in Yapo, P.O., Gupta, V.H., Sorooshian, S., 1998. Multi-objective global optimization for
Sweden. J. Hydrol. 137, 209–230. hydrologic models. J. Hydrol. 204, 83–97.
Huang, M., Liang, X., 2006. On the assessment of the impact of reducing parameters Yapo, P.O., Gupta, H.V., Sorooshian, S., 2000. Towards improved calibration of
and identification of parameter uncertainties for a hydrologic model with hydrologic models: combining the strengths of manual and automatic
applications to ungauged basins. J. Hydrol. 320, 37–61. calibration methods. Water Resour. Res. 36 (12), 3663–3674.
Lee, B., Moon, Y., 2007. Analysis and development of conceptual rainfall–runoff
model structures for regionalization purposes. KSCE J. Civil Eng. 11 (1), 57–64.