You are on page 1of 14

Journal of Hydrology 575 (2019) 810–823

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

A simple innovative method for calibration of Manning’s roughness T


coefficient in rivers using a similarity concept
Mohammad Attari, Seyed Mahmood Hosseini

Civil Engineering Department, Faculty of Engineering, Ferdowsi University of Mashhad, Mashhad, Iran

ARTICLE INFO ABSTRACT

This manuscript was handled by A. Bardossy, Geometric characteristics and Manning’s roughness coefficient (n) are needed in steady and unsteady flow
Editor-in-Chief analysis in open channels and rivers. Various methods are used to estimate n of a river reach while there is
Keywords: always some uncertainty in its estimation. One of the methods for estimating n is its calibration using observed
Manning’s roughness coefficient inflow and outflow hydrographs in a river reach. This calibration procedure is conducted by adjusting n such
Calibration that the outflow hydrograph calculated by an adopted flow routing model matches the observed outflow hy-
Random search optimization drograph. The available calibration methods usually consider a uniform roughness along the reach of a river. In
Similarity concept this study, a new method for calibration of n is presented that can recognize different regions with different n in a
Clustering
river reach even when only one recorded inflow and corresponding outflow hydrographs are available. In the
proposed method, at first different cases are generated where, in each case, the river reach is divided into several
sub-reaches in a systematic procedure. Then for each case, the n of all sub-reaches are optimized by using the
Monte Carlo method and random search optimization. In the next step, the best case among all cases is identified
by applying similarity criteria defined in this study. Finally, a clustering method is employed to identify different
regions with different n values for the selected best case. The method was applied to some synthetic data and also
Tiber river data and showed promising results.

1. Introduction and background field n. Coon (1995) found the differences between n values computed
from a study-site and those estimated from published n-value equa-
Information about the peak depth and discharge, especially at the tions and concluded that no one equation was capable of accurately
time of critical flood events, is needed to design structures and to estimating n values for all stages on all channels. Hosseini (2000)
protect property and life (Rashid and Chaudhry, 1995). However, discussed the key role of n and Manning’s equation in evaluating the
uncertainty is an inherent ingredient in the estimation of flood depth accuracy of energy and momentum principles for the analysis of one-
or discharge hydrographs because of the uncertainty involved in dimensional open channel flow. In general, it can be concluded from
model structure and parameters used in simulating flood events. different studies that the estimation of n is a hard task because of its
Among other parameters, the roughness coefficient is known to be a dependency on different flow resistance factors (Cowan, 1956) that
key parameter for a realistic simulation of open channel flows but can vary along a river and are hard to quantify; hence, uncertainty
remains especially difficult to determine (Ballesteros et al., 2011). For would be an inherent ingredient in estimating n. Johnson (1996)
a difference of 50% in the roughness coefficient, Kidson et al. (2002) studied the uncertainty of various hydraulic parameters, including n
calculated a discharge error of 40% for a palaeoflood in a bedrock for the Paint Branch stream in Maryland. She concluded that the
river. They concluded that initial uncertainty in the selection of probability distribution of n was uniform. Zoppou and Li (1993)
Manning’s n propagates and amplifies uncertainty in discharge esti- studied the uncertainty due to the natural variability of n in open
mates, and then in flood frequency estimates. Boyer (1954) states that channels and its influence on water surface profile calculations using
the Manning’s roughness coefficient in open channels plays an im- a developed point estimation method to quantify uncertainty propa-
portant role in the determination of discharge using the usual slope- gation.
area method. Li and Zhang (2001) refer to the Manning’s roughness Considering the uncertainties involved in estimating n, calibration
coefficient (n) as one of the most important parameters for analyzing techniques can provide a way to estimate or adjust n in a reach of a
water flow over the ground and provide a technique for calculation of river if flow data is available and a reliable simulation or flow routing


Corresponding author.
E-mail address: shossein@um.ac.ir (S.M. Hosseini).

https://doi.org/10.1016/j.jhydrol.2019.05.083
Received 6 January 2019; Received in revised form 25 May 2019; Accepted 27 May 2019
Available online 28 May 2019
0022-1694/ © 2019 Elsevier B.V. All rights reserved.
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

model is adopted. previous ones is related to item 3. In this section, a brief description of
One of the commonly used calibration methods is based on using the flood simulation or routing model used in this study is provided,
observed inflow and outflow hydrographs in the reach of a channel. In while the methodology used to calibrate n is described in detail.
this case, the reach-scale estimation of n is made by optimizing the
coefficient to minimize the difference between simulated and observed 2.1. Flood routing model, governing equations and numerical solution
outflow hydrograph(s) for the corresponding inflow hydrograph(s).
Many studies have been performed to calibrate n by assuming a uniform In the absence of lateral flow, the one-dimensional governing
or constant n value for any reach of a channel having a known or equations for analysis of gradually-varied unsteady flow in open
specified length. channels are Eqs. (1) and (2), which are commonly referred to as the
Aronica et al. (1998) applied the generalized likelihood uncertainty Saint-Venant equations.
estimation (GLUE) procedure to a complex distributed model of a study
A Q
area which is located in the southern part of Sicily and is crossed by the + =0
t x (1)
Imera river based on little information. They used two different tech-
niques to calibrate n, based around a statistical model error and fuzzy- Q ( Q 2 /A ) y
rule-based system. They concluded that the likelihood distributions + + gA + gA (S0 Sf ) = 0
t x x (2)
predicted by the two techniques are similar. Pappenberger et al. (2005)
presented the uncertainty analysis of the unsteady flow component of where Q is the discharge, A is the cross-sectional area normal to the
the one-dimensional model HEC-RAS in two different rivers within the flow, is the momentum correction factor, g is the acceleration due to
framework of GLUE, taking surface roughness as an uncertain input gravity, y is the flow depth, S0 is the bed slope, Sf is the friction slope, t
parameter. The results of varying the reach-scale roughness showed is the time and x refers to the distance from the reference point. There is
that many parameter sets could perform equally well even with extreme no analytical solution for these equations except for idealized situations
values. The necessity to distinguish between effective parameters and and numerical methods are employed to solve these equations. Dif-
real physical parameters was also emphasized. Kidson et al. (2006) ferent numerical methods are used to solve the Saint-Venant equations.
examined the uncertainty in discharge estimation of palaeoflood The numerical method used in this research is the weighted four-
events, particularly associated with Manning’s n and optimal water point implicit finite difference scheme, commonly referred to as the
surface. Their case study was the Mae Chaem river in northern Thai- Preissman scheme (Fread, 1993; Singh, 1996). In a discretized domain
land. They also assessed the influence of discharge uncertainty on flood of time and distance, the scheme approximates any function f and its
frequency analysis to estimate return periods. Boulomytis et al. (2017) derivatives with respect to time and distance as:
estimated n of the main channel and floodplain of Juqueriquere river fi +j +11 + fi j + 1 fi +j 1 + fi j
basin using Cowan method based on field observations. Then these data f= ( ) + (1 )( )
2 2 (3)
and measured flow data were inserted into the HEC-RAS model for the
calibration of the roughness coefficients in a trial and error process. f f j+1 fi j + fi +j +11 fi +j 1
They stated that both estimation and calibration methods were effi- = i
t 2 t (4)
cient, having no significant difference.
Ramesh et al. (2000) employed the sequential quadratic program- f (fi +j +11 + fi j + 1 ) (1 )(fi +j 1 + fi j )
ming optimization algorithm to calibrate n values in a hypothetical = +
x xi xi (5)
rectangular multiple reach open channel system under unsteady flow
condition. Their results showed that the model did not perform sa- where i and j refer to distance and time, respectively, and is a
tisfactorily when the number of observation stations was less than the weighting factor that varies between 0 and 1 (Hosseini and Joy, 2007).
number of parameters to be estimated. Timbadiya et al. (2011) showed In this study, the weighting factor was considered 0.55 based on a re-
that different n were required for upper and lower reaches of the Tapi commendation made by Fread for flood waves (Fread, 1974). Initial
river in India for finding a satisfactory agreement between observed and boundary conditions are required to solve Eqs. (1) and (2). The
hydrographs and those calculated by HEC-RAS. initial conditions are the specified depth and discharge along the river
In the present study, an innovative method for calibrating n is at the start. The boundary conditions used in this study were the im-
proposed that is capable of identifying the length and n value of any posed depth hydrograph upstream and a stage-discharge relationship
sub-reach along a river reach between two gauging stations. Briefly, downstream.
the proposed methodology applies a pattern recognition procedure
coupled with an optimization procedure and clustering that enhance 2.2. Methodology for calibrating n
the capabilities and accuracy of the method to find the length and n
value of all sub-reaches in a river reach between two gauging stations. Fig. 1 shows a river reach with an overall length of L = Li where
In other words, the method identifies possible distinct regions in a Li refers to the length of regions i having Manning’s roughness coeffi-
river reach. cient ni. The existence of regions with different n values can be the
Section 2 presents the methodology presented in this study. In characteristic of long natural rivers. Fig. 1 also presents an inflow depth
Section 3, the results of applying the methodology to three synthetic or discharge hydrograph entering the reach and the corresponding
data sets and also Tiber river field data are reported and discussed. outflow hydrograph. While previous research focused on calibrating n,
Section 4 highlights the general conclusions made in this study. have ignored the variation of n along the river reach between the two
gauging stations 1 and 2, the main objective of the current study was to
2. Methodology introduce a method to recognize the pattern shown in Fig. 1, identify
the distinct regions with different length and n along the river reach and
In this study, the framework used to calibrate n is similar to the determine the length and n of each region. The data required in this
framework used by many other researchers. This framework has three method is at least one observed inflow and outflow depth or discharge
main elements: 1- a flood simulation or routing model, 2- a set of ob- hydrograph, as shown in Fig. 1.
served inflow and corresponding outflow hydrographs and 3- a meth- The following steps were introduced and followed to achieve the
odology for calibrating or optimizing n to make a good agreement be- objective of this study. Although in this study the steps were followed to
tween the simulated and observed outflow hydrographs. However, as analyze some illustrative examples described in Section 3, they are
described in Section 1, the main difference of the current study with the general and can be applied to any problem under study.

811
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 1. A river reach with N distinct regions of different length and n.

Step 1: n value of each sub-reach is first estimated based on the results of the
Main point- The river is divided into two equal sub-reaches, and the previous two steps and then optimized.
best n of each sub-reach is determined. Fig. 2 shows a schematic of view of the different runs used in this
In the proposed method, the river is first divided into two equal sub- study. In Fig. 2, k refers to the run number while c is the number of sub-
reaches, and for each sub-reach, a set of random data of n is generated reaches. As Fig. 2 shows, Step 3 is the key step. In Step 3, the river is
using Monte Carlo simulation assuming a uniform distribution. It is divided into four sub-reaches, and the mean n value of each sub-reach is
better to consider a wide interval for n distribution based on experience calculated using Eqs. (7)–(10). In these equations, the n of each sub-
or field observation. For all n values of the generated data, the outflow reach is related to the corresponding n values obtained at the previous
hydrograph is developed using the imposed inflow hydrograph as the two runs. For example, as a general equation, Eq. (8) is the equation
input to the river reach. This can be done using any flood routing that is used to estimate n2(3) , the mean n value of sub-reach number 2 at
model, such as the one described in Section 2.1. Then, the best si- run 3. Eq. (8) indicates that a weighted mean value of n2(3) is de-
multaneous n values of all sub-reaches are those that minimize the error termined, where the weighting parameter depends on the length and
between calculated and observed outflow hydrographs based on Eq. SSE values of the previous two runs that are related to n2(3) and l3 .
(6).
n1(1) . l3 n1(2) . l3 n1(1) n1(2)
w (1 + SSE1)
+ (1 + SSE2) (1 + SSE1)
+ (1 + SSE2)
SSE = (hoi hci )2 n1(3) = l3 l3
= 1 1
+ + (7)
i=1 (6) (1 + SSE1) (1 + SSE2) (1 + SSE1) (1 + SSE2)

where hoi and hci are the ordinates of observed and computed depth or
n1(1) . l3 n1(2) . x1 n2(2) . x2
discharge outflow hydrograph, respectively, w is the total number of (1 + SSE1)
+ (1 + SSE2)
+ (1 + SSE2)
data points and SSE is the sum of the square of errors. n2(3) = l3 x1 x2
+ + (8)
Step 2: (1 + SSE1) (1 + SSE2) (1 + SSE2)

Main point- The river is divided into three equal sub-reaches, and
the best n of each sub-reach is determined. n2(1) . l3 n3(2) . x 1 n2(2) . x2
The same procedure described in Step 1 is followed in Step 2, while (1 + SSE1)
+ (1 + SSE2)
+ (1 + SSE2)
n3(3) =
the number of sub-reaches is three in this step. l3
+
x1
+
x2
(9)
Step 3:
(1 + SSE1) (1 + SSE2) (1 + SSE2)

Main point- The river is divided into four sub-reaches, and the mean

Fig. 2. The procedure of increasing the number of divisions (sub-reaches) in a river reach (k: run number, c: number of sub-reaches).

812
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

n2(1) . l3 n3(2) . l3 n2(1) n3(2) Step 5:


+ +
n4(3) =
(1 + SSE1) (1 + SSE2)
=
(1 + SSE1) (1 + SSE2) Main point- A calculation stopping condition is proposed.
l3 l3 1 1
(1 + SSE1)
+ (1 + SSE2) (1 + SSE1)
+ (1 + SSE2) (10) The calculation termination condition was defined according to Eq.
(16).
In the above equations, is the n of sub-reach i at run 3, while
ni(3)
SSE1 and SSE2 refer to the least sum of the square of errors at runs 1 and lk lk 1 (16)
2, respectively. The lengths x1 and x2 having different n values that have where is the difference between the length of two sub-reaches at runs
an effect on n2(3) and n3(3) can be observed in Fig. 2. After determining k and k + 1. A of less than one percent of the length of the river reach
the mean n value of each sub-reach, different probable n values in each is suggested for the pattern recognition. As a result, as will be discussed
sub-reach are generated using Monte Carlo simulation, assuming a in Section 3, all calculations were performed in eleven runs in all ex-
uniform distribution where the interval of the distribution is calculated amples provided in this study.
by Eq. (11). Step 6:Main point- The best run from all runs is chosen.
The best run must be determined after stopping the calculations. If
(11)
(3) (3) (3) (3)
(1 i ) ni n (1 + i ) ni
at run k, the river is divided into c sub-reaches, the number of river sub-
where denotes the interval definition parameter of the sub-reach i
i
(3)
reaches is equal to 2c at run 2 k + 1. In other words, each sub-reach at
at run 3. The choice of i(3) is very important. It must be bigger for the run k is divided into two sub-reaches with equal length at run 2 k + 1. If
sub-reaches that are influenced by more different n values such as sub- the n values of all sub-reaches at run 2 k + 1 are not significantly dif-
reaches 2 and 3 in Fig. 2. Eq. (12) is used to find i(3) , where i(3) is ferent from the corresponding n values at run k, it can be accepted that
determined based on the maximum and minimum values of n that n has been identified well. On the other hand, another factor that plays
control ni(3) . a role in the selection of the best run is the SSE between the computed
and observed outflow hydrographs. Eq. (17) includes the combined
max[n(1) , n(2)] min[n(1), n(2) ]
i
(3)
= effects of the similarity between the n values and SSE.
2ni(3) (12)
c c
where i is the interval definition parameter of the sub-reach i and n(1)
(3)
E= |n 2(2i k + 1) ni(k)| + |n2(2i k +1 1) ni(k ) | l2k + 1
(2)
and n refer to the n values in runs 1 and 2 that have an effect on the i=1 i=1

determination of ni(3) based on Eq. (7)–(10). (SSEk + SSE2k + 1) (17)


It is possible that the value of i(k ) in some sub-reaches tend to zero
at the initial runs and consequently n is not accurately identified in where c is the number of river sub-reaches at run k. The E value can be
these sub-reaches. Therefore, in each sub-reach, a minimum value of calculated for all conjugate runs k and 2 k + 1 starting from k = 1. The
(k )
was considered to increase the accuracy of the estimated n value. best conjugate runs were suggested to be those with the minimum E
i
This minimum value adds more degrees of freedom in approaching the value. Between the two best conjugate runs k and 2 k + 1, the one with
real n in each sub-reach. Theoretically, this minimum value must de- less SSE was considered to be the final answer.
crease as the number of sub-reaches, c or equivalently k + 1, increases. Step 7:
After testing different case studies, general Eq. (13) was proposed to Main point- Different regions having different roughness coeffi-
find i(min , which results in a value of 0.0625 for i(3) cients are identified by applying a clustering algorithm.
min .
k)

It must be considered that the n values of sub-reaches resulting from


0.25 0.25
(k )
= = Step 6 can be close to each other in some sub-reaches. Practically, the
i min
c k+1 (13)
proximity of n values highlights the need for the identification of dis-
where c represents the number of river sub-reaches and k is run tinct regions having the same n. To recognize the pattern of n, it can be
number. accepted that adjacent sub-reaches with nearly similar n values can be
In conclusion, the mean n values of all sub-reaches can be found classified as one region by the use of a clustering algorithm such as the
using Eqs. (7)–(10), and the corresponding uniform distribution inter- K-means method used in this study. Hartigan and Wong (1979) provide
vals can be specified by Eqs. (11)–(13). Therefore, in each sub-reach, a details of the K-means method. The following procedure was suggested
set of random n values can be generated using Monte Carlo simulation. and employed to apply the K-means method in this study:
By a random combination of the n values in sub-reaches, as previously
discussed in Step 1, the outflow hydrographs are calculated by em- a- After the termination of calculations and choosing the best answer
ploying the flood routing model and using the imposed inflow hydro- as described in Step 6, n values at some locations are selected as
graph as the input to the river reach and a specified downstream seeds of the clusters. The number of clusters is identified based on
boundary condition in subcritical flow. Finally, Eq. (6) can be used to the variation of n observed in the best answer. Each seed contains
find the best scenario and the corresponding n values of all sub-reaches. information about the distance from the beginning of the river reach
Step 4: (x) and the corresponding n value.
Main point- The number of sub-reaches is increased and analysis b- Each sub-reach is assigned to the nearest cluster based on the
similar to that of Step 3 repeated. Euclidian distance between the cluster seed and the center of that
The same procedure described in Step 3 was continued in each run sub-reach. The Euclidian distance is calculated using normalized n
by increasing the number of sub-reaches. After calculating the mean n values and log(x). To normalize n, the n value in each sub-reach is
values of all sub-reaches, its interval is determined according to the divided by the maximum n found along the river reach. The nor-
analyses of the two previous runs for each sub-reach using Eqs. malized log(x) is calculated similarly.
(13)–(15). Then, a set of random data for each sub-reach is generated c- The new centroid of each cluster is determined which basically in-
using Monte Carlo simulation. Finally, the best set of the simultaneous n cludes the arithmetic mean of the normalized n and log(x) of the
values of all sub-reaches is identified by multiple runs of the flood sub-reaches positioned in one cluster. These centroids are the new
routing model and criterion defined by Eq. (6). seeds.
d- The items b and c, described above, are repeated until the con-
(14)
(k ) (k ) (k ) (k )
(1 i ) ni n (1 + i ) ni vergence is achieved and the centroids of all clusters remain un-
changed.
max[n(k 2) , n(k 1)] min[n(k 2) , n(k 1) ]
i
(k )
= e- Finally, the arithmetic mean of n values of all sub-reaches positioned
2ni(k ) (15) in one cluster is considered as the n value of that cluster or region.

813
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 3. Schematic view of a hypothetical river used in Example 1.

example was to use the inflow and outflow hydrographs as available


data, and as an inverse problem, recognize or identify the pattern in
Manning’s roughness distribution shown in Fig. 3 by using the method
and steps introduced in Section 2. A uniformly distributed random error
in the interval [−1; 1] cm was added to the ordinates of the outflow
hydrograph.
To apply the steps introduced in Section 2.2, the parameters of the
method were selected as follows.

1. In Steps 1 and 2, the n values were randomly sampled from the


uniform distribution [0.015; 0.05].
2. A thousand random n values were generated using Monte Carlo si-
mulation in all runs.
3. A value of 25 m was considered for the termination condition, as
discussed in Step 5 of Section 2.2.

The summary of the results obtained in all 11 runs by applying the


proposed methodology can be observed in Fig. 5.
The sum of the square of the errors for all 11 runs is shown in Fig. 6.
Fig. 7 shows the E values calculated by Eq. (17).
According to Fig. 7, the minimum E is related to the runs when the
river reach is divided into 6 and 12 sub-reaches, followed by the cases
when the river is divided into 5 and 10 sub-reaches. Therefore, the
cases when the river reach was divided into 6 and 12 sub-reaches were
selected as the best pattern candidates. The selection between these two
sub-reach numbers was made based on the SSE between the ordinates of
the outflow hydrographs. Therefore, as Fig. 6 shows, the run 5 results
corresponding to 6 sub-reaches had the best rank. Similarly, Figs. 6 and
7 show that the run 4 results corresponding to 5 sub-reaches ranked
Fig. 4. Depth initial condition and inflow and outflow depth hydrographs in second.
Example 1, (a) Flow initial depth, (b) Inflow and outflow depth hydrographs. A review of the n values shown in Fig. 5, especially those related to
the two best cases, indicated three distinct regions, regions with higher
n value in the first and last parts of the river and a region with lower n
3. Application of the method to some case Studies, results and
value in the middle part of the river. Therefore, three seeds were se-
discussion
lected for clustering. Initial seeds for clustering were selected arbitrarily
in sub-reaches 1, 3 and 6 of the river based on the results of run 5. The
To show the capabilities of the method, three examples based on
clustering results are shown in Fig. 8. Fig. 8 also shows similar results
synthetic data and one real-world example are presented in this section.
for run 4 where the initial seeds were selected in sub-reaches 1, 3, and
5.
3.1. Example 1 using synthetic data This example shows that the proposed methodology was able to
recognize three distinct regions having different lengths and n values.
Fig. 3 shows the schematic view of a hypothetical river used in this
example. It is a rectangular river with 10 m width and 3000 m length,
and a bed slope of 0.0001. The river has three distinct regions having 3.2. Example 2 using synthetic data
different n values. The first and last regions are 500 m long with an n of
0.03, while the middle region is 2 km in length and has an n of 0.02. Fig. 9 shows the schematic view of a hypothetical river used in this
The initial condition before the arrival of the flood wave is a subcritical example. It is a rectangular river with 10 m width and 3000 m length,
flow with a constant discharge of 3.308 m3/s. The water surface profile and a bed slope of 0.0001. The river has two distinct regions having
and depths corresponding to the initial discharge, shown in Fig. 4(a), different n values. The first region is 1000 m long with n of 0.02, while
were calculated using the conventional standard step method. Fig. 4(b) the second region is 2 km in length and has n of 0.03. The initial con-
shows the hypothetical inflow hydrograph and the corresponding out- dition before the arrival of the flood wave is a subcritical flow with a
flow hydrograph. The outflow hydrograph was calculated using the constant discharge of 3.21 m3/s. The water surface profile and depths
Priessman scheme discussed in Section 1 assuming a downstream corresponding to the initial discharge, shown in Fig. 10(a), were cal-
boundary condition of S0 Sf = 0 where S0 and Sf are the longitudinal culated using the conventional standard step method. Fig. 10(b) shows
slope of the river and friction slope, respectively. This boundary con- the hypothetical inflow hydrograph and the corresponding outflow
dition basically means that a Manning equation type stage-discharge hydrograph. The outflow hydrograph was calculated using the
relationship controls the flow downstream. The objective of this Priessman scheme discussed in Section 2.1 assuming a downstream

814
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 5. The summary of the results of eleven runs in Example 1.

Fig. 6. The sum of the square of errors in all 11 runs – Example 1.


Fig. 7. The E values related to different conjugate runs – Example 1.

boundary condition of S0 Sf = 0 where S0 and Sf are the longitudinal


slope of the river and friction slope, respectively. Similar to the Ex- the application of the proposed method and the results were presented.
ample 1, the main objective of this example was to use the inflow and However, in Example 2, a sensitivity analysis on some arbitrarily se-
outflow hydrographs as available data, and as an inverse problem, re- lected numerical parameters was conducted. The selected parameters
cognize or identify the pattern in Manning’s roughness distribution are the support of the uniform distribution considered for n in Steps 1
shown in Fig. 9 by using the method and steps introduced in Section and 2 of the proposed methodology and the initial seeds considered for
2.2. A uniformly distributed random error in the interval [−1; 1] cm clustering in Step 7.
was added to the ordinates of the outflow hydrograph. In Example 1, As mentioned, an interval for n must be considered for performing

815
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 8. Recognition of 3 distinct regions based on runs 4 and 5 after clustering – Example 1.

Fig. 9. Schematic view of a hypothetical river used in Example 2.

simulation, and in Eq. (14), a of 25 m was considered as the calcu-


lation termination condition. The summary of the results obtained in
this example for all 11 runs of analysis is shown in Fig. 11.
The sum of the square of the errors for all 11 runs of analysis is
shown in Fig. 12, while Fig. 13 shows the E values calculated by Eq.
(17).
The minimum E and SSE are related to the cases when the river is
divided into 6 and 12 sub-reaches, which correspond to k = 5 and
k = 11, respectively. However, the 12 sub-reach results are somewhat
better.
Fig. 11 indicates the presence of two distinct regions at k = 11. By
selecting the n in the first and last sub-reaches as the two initial arbi-
trary seeds in the clustering procedure, the results shown in Fig. 14
were obtained. The pattern in Fig. 14 reasonably matches with the real
pattern shown in Fig. 9.
The reason for such a reasonable pattern recognition regardless of
initially out of range selected interval of the n, is imposing a minimum
value for based on Eq. (13). According to Fig. 11, the results of the
analysis performed in the first and second runs indicate a value of 0.025
for the n in the first sub-reach of the river which results in value of
zero based on Eq. (12) in the third run for this sub-reach. However, the
imposed minimum based on Eq. (13) contributes to the better cali-
bration of n and its convergence to the real n value in the first sub-
reach.
In another study, the sensitivity of the clustering results to the se-
lection of initial seeds was investigated. Table 1 presents the clustering
results when the two initial seeds were randomly selected in different
sub-reaches. As Table 1 shows the clustering results did not show any
sensitivity to the initial seeds. This is because the clustering procedure
introduced in Section 2.2 has the capability of approaching the real
physical system.
Fig. 10. Depth initial condition and inflow and outflow depth hydrographs in
Example 2, (a) Flow initial depth, (b) Inflow and outflow depth hydrographs.
3.3. Example 3 – application of the method to the Tiber river data

the analysis in Steps 1 and 2. This interval can be estimated based on To evaluate the application of the method proposed in this study to
experience or field observations. It is possible that the real value of n in the real-world problems, the field data of the Tiber river in central Italy
one sub-reach or more is out of the selected interval if this interval is for the river reach between the gauged section of Pierantonio (up-
estimated with some errors. In this example, in the first place, the effect stream) and Ponte Felcino (downstream) was used. The overall length
of such errors on the pattern recognition results was studied. of the reach is 15 km with a bed slope of 0.0016. A Manning roughness
The interval considered for the uniformly distributed n was [0.025; coefficient of 0.049 was proposed for this river reach (Perumal et al.,
0.04] in Steps 1 and 2. As Fig. 9 shows, the n value in region 1 is out of 2007). The cross sections of the Tiber river at Pierantonio and Ponte
the selected interval. Additionally, for all runs, including Steps 1 and 2, Felcino gauging stations are shown in Fig. 15.
five hundred random data were generated using Monte Carlo The data related to the floods in December 1996 and November

816
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 11. The summary of the results of eleven runs in Example 2.

Fig. 12. The sum of the square of errors in all 11 runs – Example 2.
Fig. 13. The E values related to different conjugate runs – Example 2.

1997 were used in this study. The data of flood in 1996 was used to
simulation. The method presented in this research was applied to the
recognize any possible pattern in the distribution of n if it exists. The
data of the 1996 flood. Fig. 18 shows a summary of the results after 11
data of flood in 1997 were used to verify the results. The observed
runs.
inflow and corresponding outflow depth hydrographs related to the two
Fig. 19 shows the SSE values of all 11 runs while Fig. 20 shows the E
floods are shown in Fig. 16. The downstream boundary condition was
values for all conjugate runs calculated by Eq. (17).
the stage-discharge relationship (Q vs. y) depicted in Fig. 17 (Perumal
Fig. 20 shows the E values calculated by Eq. (17).
et al., 2007).
Although the SSE values are not significantly different in different
The interval considered for the uniformly distributed n was [0.04;
runs, the E values change significantly as Fig. 20 shows. By considering
0.06] in Steps 1 and 2. Additionally, for all runs, including Steps 1 and
E criterion, as the main criterion, and SSE criterion, it was concluded
2, two hundred random data were generated using Monte Carlo
that the best result was achieved when the river reach was divided into

817
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 14. Recognition of 2 distinct regions


based on run 11 after clustering – Example
2.

Table 1
The sensitivity of the results to the selection of initial seeds locations – Example
2.
Initial seeds location Region 1 Region 2

L (m) n L (m) n

Sub-reaches 1 and 12
Sub-reaches 4 and 12
Sub-reaches 4 and 5 1000 0.0207 2000 0.0298
Sub-reaches 5 and 8
Sub-reaches 6 and 11

Fig. 15. Cross section of the Tiber river at Pierantonio and Ponte Felcino
gauging stations (Perumal et al., 2007).
Fig. 16. Observed inflow and outflow stage hydrographs related to the floods in
December 1996 and in November 1997 (Perumal et al., 2007).
4 sub-reaches, which corresponds to k = 3 in Fig. 18. By looking at the
relatively close n values in Fig. 18 at k = 3 and arbitrarily selecting two
initial seeds, the clustering analysis indicated the following two regions:
1- region 1 with a length of 3750 m and n of 0.0485 and 2- region 2
with a length of 11250 m and n of 0.0524. In other words, although
from a practical point of view, a uniform roughness coefficient for the
river reach can be accepted, the current analysis reveals that the up-
stream part of the river reach has somewhat lower roughness coefficient
than the rest of the river reach. Fig. 21 shows the computed depth (y)
and discharge (Q) hydrographs based on the pattern selected in this
study, and also independent uniform n = 0.049 suggested by Perumal
et al. (2007) versus the corresponding observed hydrographs at the
Ponte Felcino station. As Fig. 21 shows the model with calibrated n
values performs reasonably well both in calibration and verification
phases. This indicates that the river system has not changed sig-
nificantly in a time scale of about one year.
Although the Tiber river n value does not show a significant change
Fig. 17. The depth-discharge relationship at Ponte Felcino gauging station
over time, this cannot be true for all natural river systems. The n values (Perumal et al., 2007).
in natural river systems can change between different seasons, flood
events, and with channel changes over time, given the dependence of n
on discharge, vegetation, and local conditions. This issue highlights the calibration and uncertainty analysis can shed more light on the extent
fact that the effective n values calibrated for one flood using the hy- of the problem for any river under study.
draulic approach and method presented in this study, cannot be applied
to another flood with certainty. Although this issue of non-unique n 3.4. Example 4 using synthetic data
values can be considered as a potential limitation of the hydraulic ap-
proach used in this study, collection of data under different conditions, The river reach used in this example is similar to the river used in
Examples 1 and 2 with a difference in the distribution of n. The river

818
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 18. The summary of the results of eleven runs in Example 3 – Tiber river.

Fig. 19. The sum of the square of errors in all 11 runs – Tiber river.
Fig. 20. The E values related to different conjugate runs – Tiber river.

has three distinct regions having different n values. The first and last
hydraulic routing model assuming the same boundary conditions as
parts are 500 m long with n of 0.02 and 0.03, respectively. The middle
those considered in Examples 1 and 2. Although the method presented
region is 2 km in length, and n linearly varies from 0.02 to 0.03 in this
in this research was for the patterns with sharp boundaries, the objec-
region. The initial condition before the arrival of the flood wave is a
tive of this example was to see how the performance of the proposed
subcritical flow with a constant discharge of 1.69 m3/s. The boundary
method would be in analyzing such a complex river system.
conditions are the same as those considered in Examples 1 and 2 re-
To apply the steps introduced in Section 2.2, the parameters of the
ported in Sections 3.1 and 3.2. The water surface profile and depths
method were selected as follows.
corresponding to the initial discharge, shown in Fig. 22 (a), were cal-
culated using the conventional standard step method. Fig. 22(b) shows
1. In Steps 1 and 2, the n values were randomly sampled from the
the hypothetical inflow hydrograph and the corresponding outflow
uniform distribution [0.015; 0.04].
hydrograph. The outflow hydrograph was calculated using the

819
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 21. Calculated and observed depth (y) and discharge (Q) hydrographs for different floods in Tiber river – 1996 y data for calibration, 1997 y and Q data for
verification.

2. One thousand random n values were generated using Monte Carlo


simulation in all runs. In comparison with the previous examples, it
was required to increase the number of random n values in Example
4 to achieve better accuracy. Better accuracy facilitates the selection
of the best case as the following analysis shows.
3. A value of 25 m was considered for the termination condition, as
discussed in Step 5 of Section 2.2.

A uniformly distributed random error in the interval [-1;1] cm was


added to the ordinates of the outflow hydrograph and the same pro-
cedure used in Examples 1 and 2 was followed in this example to find
the results. The summary of the results obtained in all 11 runs by ap-
plying the proposed methodology can be observed in Fig. 23.
The E and SSE diagrams are not shown here for the sake of brevity.
The E diagram showed that the best pattern was related to the con-
jugate run (5,11) with an E value of 0.129. However, the SSE values for
runs 5 and 11 were 0.0247, and 0.0242 m2, respectively. Therefore, the
12 sub-reach case, corresponding to k = 11, results were slightly better
and selected as the best case for further analysis. Fig. 24 demonstrates
the distribution of n captured by the proposed method against the n of
the real system. As Fig. 24 indicates, there is a good match between the
discretized distribution of n and the linear trend in variation of the real
system n. The distribution of n can be used by any practicing engineer
involved in modeling flow system in this river. Although Fig. 24 in-
dicates that the clustering analysis was not required in this case, three
initial arbitrary seeds of 0.0199, 0.0236, and 0.03, based on Fig. 24,
were selected, and clustering analysis was also performed to further
investigate the performance of the method. The clustering analysis in-
dicated an equivalent system consisting of three distinct regions, the
Fig. 22. Depth initial condition and inflow and outflow depth hydrographs in
first 750 m of the river with n of 0.0196, the second or middle region
Example 4, (a) Flow initial depth, (b) Inflow and outflow depth hydrographs. with a length of 1000 m and n of 0.0241 and the remaining 1250 m with
n of 0.0292. The results were not surprising because it was expected

820
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Fig. 23. The summary of the results of eleven runs in Example 4.

Fig. 24. The distribution of n resulting from run 11 vs. Real system n values –
Example 4. Fig. 25. Outflow depth hydrographs calculated by n values from random search
optimization and clustering method and calibrated single n value vs. corre-
sponding observed hydrographs, under different inflow conditions – Example 4.
that the method would result in equivalent or effective roughness
coefficients that approximate the roughness coefficient in the region
flow system. The sensitivity analysis indicated that in this case, the
with linearly varying n.
outflow depth hydrograph was more controlled by the roughness
To evaluate the performance of the hydraulic model under equiva-
coefficient in the upstream region of the river than those in the middle
lent roughness coefficients, the roughness coefficient was also cali-
and last regions.
brated using conventional methods, assuming a uniform roughness
Fig. 25 shows the performance of the hydraulic routing model under
along the river. A single n value of 0.0203 was found in this case.
different inflow hydrographs. As discussed, only the inflow and corre-
Although a higher value for the single n was expected, the interaction of
sponding outflow hydrographs shown in the middle were used in the
many factors including roughness distribution, initial and boundary
calibration phase, while the other two inflow hydrographs were used to
conditions may justify the constant n calibrated for such a non-linear
evaluate the performance of the calibrated hydraulic model under

821
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

Table 2 along a river reach between two gauged stations. The application of the
Errors in predicting outflow depth hydrographs using equivalent n values re- conventional calibration methods would be computationally hard and
sulting from random search optimization and clustering method, and calibrated time-consuming if different regions of a river reach have different n
single n value of 0.0203, under different inflow conditions – Example 4. values separated by sharp boundaries. In these situations, the conven-
n = 0.0203 Random search optimization and tional methods may require a predefined length for each region to be
clustering able to calibrate the n values in different regions. Such information may
not be available or be subject to high uncertainty in most situations. In
Outflow 1 SSE (m2)* 4.87 0.14
AREPD (%)** 3.8 0.7
this study, a new method was proposed that can identify different re-
AREPDT 1.2 0.6 gions with different roughness coefficients along the long river reaches
(%)*** just based on at least one inflow and corresponding outflow hydro-
Outflow 2 SSE (m2) 3.33 0.09 graph. The method has the following three main components:
AREPD (%) 4.6 1.0
AREPDT (%) 1.1 0.5 1. Dividing the river reach into a different number of sub-reaches in a
Outflow 3 SSE (m2) 3.40 0.09 systematic manner and in any sub-reach arrangement or case, cali-
AREPD (%) 5.5 1.1 brating n in each sub-reach by applying a random search optimi-
AREPDT (%) 1.6 0.5
zation using the Monte Carlo simulation,
2. Finding the best case among different cases considering the simi-
* Sum of the square of errors.
** Absolute value of relative error in peak depth. larity between different cases and applying the similarity criteria
*** Absolute value of relative error in peak depth time. defined in this study, and
3. Using clustering methods to recognize different regions with dif-
significantly different flow conditions. Table 2 quantifies the difference ferent n.
between the performance of the two equivalent hydraulic systems using
three standard error measures, SSE, the absolute value of the relative Results from applying the proposed method to two test examples
error in peak depth (AREPD), and the absolute value of the relative developed by synthetic data and Tiber river data indicated that the
error in peak depth time (AREPDT). In each case, the absolute value of method could recognize different regions with different n along the
the relative error was calculated by dividing the absolute value of the river reach when only one recorded inflow and corresponding outflow
difference between the predicted and corresponding real value by the hydrograph were available. The method is particularly useful in long
real value, reported in percentage. As Fig. 25 and Table 2 show the river reaches when the number of gauging stations are limited. The
hydraulic model calibrated with equivalent roughness coefficients main advantage of using the proposed method in comparison with the
performs better than the model calibrated with the single n value. conventional methods is that information about the variation of n in a
The results of the four examples presented above indicate that if there river system is generated with the same amount of data. This in-
is no information regarding the number of distinct regions, the number of formation not only improves the prediction accuracy of the flood
initial seeds can be selected based on the model outputs. For instance, in routing models but also can be useful in river restoration, dredging, and
Example 1, the results indicated that n first decreased and then increased management projects by spotting regions with high roughness. Also,
along the river reach. As a result, three initial seeds were selected. information about the variation of n along a river may reduce the cost
Another advantage of the proposed method is that it can appropriately of river engineering projects by reducing the amount of field work.
recognize distinct regions, even if the interval of n for starting the ana-
lysis is not known properly. Meanwhile, depending on the field condi- Declaration of Competing Interest
tion, as long as a proper hydraulic routing model that can model any flow
condition (e.g., subcritical and supercritical flow) and handle any None.
channel geometry is employed, the developed method can be used to
estimate the n values within the framework proposed in this study. Acknowledgment
Although the method proposed in this study was developed to recognize
patterns in n in river reaches having different regions with constant n The authors are thankful to the Department of Environment,
values separated by sharp boundaries, as discussed in Example 4, the Planning and Infrastructure, Umbria Region and IRPI, Perugia, Italy for
method can result in reasonable equivalent roughness coefficients along providing the Tiber River data used in this study. The authors would
a river reach when there is a linear trend in variation of n. also like to thank the anonymous reviewers for their detailed and va-
It can be stated that based on the examples discussed in this article, a luable comments and suggestions.
few limitations can be considered for the hydraulic approach and
methodology presented in this research for calibration of n in rivers. References
Uncertainty in estimated n values in natural rivers discussed in the Tiber
river case and piecewise estimation of n values in the presence of a trend Aronica, G., Hankin, B., Beven, K., 1998. Uncertainty and equifinality in calibrating
distributed roughness coefficients in a flood propagation model with limited data.
in variation of n, discussed in Example 4 can be given as examples. Also, Adv. Water Resour. 22 (4), 349–365. https://doi.org/10.1016/S0309-1708(98)
like any other numerical method, the numerical parameters of the pro- 00017-7.
posed method must be selected properly to preserve both accuracy and Ballesteros, J., Bodoque, J., Díez-Herrero, A., Sanchez-Silva, M., Stoffel, M., 2011.
Calibration of floodplain roughness and estimation of flood discharge based on tree-
computational efficiency. Another limitation of the proposed metho- ring evidence and hydraulic modelling. J. Hydrol. 403 (1–2), 103–115. https://doi.
dology is that a uniform or constant n value in a river cross-section is org/10.1016/j.jhydrol.2011.03.045.
assumed. In other words, the proposed method cannot handle compound Boulomytis, V.T.G., Zuffo, A.C., Dalfré Filho, J.G., Imteaz, M.A., 2017. Estimation and
calibration of Manning’s roughness coefficients for ungauged watersheds on coastal
channels, where the roughness of the floodplains is different from that of
floodplains. Int. J. River Basin Manage. 15 (2), 199–206. https://doi.org/10.1080/
the main channel. This issue is currently under study. 15715124.2017.1298605.
Boyer, M., 1954. Estimating the Manning coefficient from an average bed roughness in
open channels. Eos. Trans. Am. Geophys. Union 35 (6), 957–961. https://doi.org/10.
1029/TR035i006p00957.
4. Conclusions
Coon, W.F., 1995. Estimates of roughness coefficients for selected natural stream channels
with vegetated banks in New York. USGS Water Supply Paper 2441, open file report
Optimization based conventional methods used for calibration of n 93-161. DOI: https://doi.org/10.3133/wsp2441.
in a river usually assume a uniform Manning’s roughness coefficient Cowan, W.L., 1956. Estimating hydraulic roughness coefficients. Agric. Eng. 37 (7),

822
M. Attari and S.M. Hosseini Journal of Hydrology 575 (2019) 810–823

473–475. Manage. 49 (2), 153–161. https://doi.org/10.1016/S0378-3774(00)00139-6.


Fread, D.L., 1993. Flow Routing in Handbook of Hydrology. McGraw-Hill Inc., New York. Pappenberger, F., Beven, K., Horritt, M., Blazkova, S., 2005. Uncertainty in the calibration
Fread, D.L., 1974. Numerical properties of implicit four-point finite difference equations of effective roughness parameters in HEC-RAS using inundation and downstream
of unsteady flow. Washington, DC: NOAA Technical Memorandum NWS HYDRO-18. level observations. J. Hydrol. 302 (1–4), 46–69. https://doi.org/10.1016/j.jhydrol.
Hartigan, J.A., Wong, M.A., 1979. Algorithm AS 136: a k-means clustering algorithm. J. 2004.06.036.
Roy. Stat. Soc.: Ser. C (Appl. Stat.) 28 (1), 100–108. https://doi.org/10.2307/ Perumal, M., Moramarco, T., Sahoo, B., Barbetta, S., 2007. A methodology for discharge
2346830. estimation and rating curve development at ungauged river sites. Water Resour. Res.
Hosseini, S.M., 2000. Discussion of: energy and momentum in one dimensional open 43 (2), 1–22. https://doi.org/10.1029/2005WR004609.
channel flow by W.G. Field, M.F. Lambert and B.J. Williams. J. Hydraul. Res. 38 (3), Ramesh, R., Datta, B., Bhallamudi, S.M., Narayana, A., 2000. Optimal estimation of
233–234. https://doi.org/10.1080/00221680009498340. roughness in open-channel flows. J. Hydraul. Eng. 126 (4), 299–303. https://doi.org/
Hosseini, S.M., Joy, D., 2007. Development of an unsteady model for flow through coarse 10.1061/(ASCE)0733-9429(2000) 126:4(299).
heterogeneous porous media applicable to valley fills. Int. J. River Basin Manage. 5 Rashid, R.M., Chaudhry, M.H., 1995. Flood routing in channels with floodplains. J.
(4), 253–265. https://doi.org/10.1080/15715124.2007.9635325. Hydrol. 171 (1–2), 75–91. https://doi.org/10.1016/0022-1694(95)02693-J.
Johnson, P.A., 1996. Uncertainty of hydraulic parameters. J. Hydraul. Eng. 122 (2), Singh, V.P., 1996. Kinematic Wave Modeling in Water Resources, Surface-Water
112–114. https://doi.org/10.1061/(ASCE)0733-9429(1996) 122:2(112). Hydrology. John Wiley & Sons.
Kidson, R., Richards, K., Carling, P., 2002. Hydraulic model calibration using a modern Timbadiya, P.V., Patel, P.L., Porey, P.D., 2011. Calibration of HEC-RAS model on pre-
flood event: the Mae Chaem River, Thailand, Workshop, Barcelona. Citeseer. diction of flood for lower Tapi River, India. J. Water Resour. Prot. 3 (11), 805–811.
Kidson, R., Richards, K., Carling, P., 2006. Hydraulic model calibration for extreme floods https://doi.org/10.4236/jwarp.2011.311090.
in bedrock-confined channels: case study from northern Thailand. Hydrol. Process. Zoppou, C., Li, K., 1993. New point estimate method for water resources modeling. J.
20 (2), 329–344. https://doi.org/10.1002/hyp.6086. Hydraul. Eng. 119 (11), 1300–1307. https://doi.org/10.1061/(ASCE)0733-
Li, Z., Zhang, J., 2001. Calculation of field Manning’s roughness coefficient. Agric. Water 9429(1993) 119:11(1300).

823

You might also like