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

Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: https://www.tandfonline.com/loi/thsj20

Stochastic and analytical approaches for sediment


accumulation in river reservoirs

Tanju Akar & Hafzullah Aksoy

To cite this article: Tanju Akar & Hafzullah Aksoy (2020) Stochastic and analytical approaches
for sediment accumulation in river reservoirs, Hydrological Sciences Journal, 65:6, 984-994, DOI:
10.1080/02626667.2020.1728474

To link to this article: https://doi.org/10.1080/02626667.2020.1728474

Published online: 19 Feb 2020.

Submit your article to this journal

Article views: 176

View related articles

View Crossmark data

Citing articles: 4 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
HYDROLOGICAL SCIENCES JOURNAL
2020, VOL. 65, NO. 6, 984–994
https://doi.org/10.1080/02626667.2020.1728474

Stochastic and analytical approaches for sediment accumulation in river reservoirs


Tanju Akara and Hafzullah Aksoy b

a
Department of Civil Engineering, Akdeniz University, Antalya, Turkey; bDepartment of Civil Engineering, Istanbul Technical University, Istanbul,
Turkey

ABSTRACT ARTICLE HISTORY


Sediment accumulation in a river reservoir is studied by stochastic time series models and analytical Received 3 April 2019
approach. The first-order moving average process is found the best for the suspended sediment Accepted 10 December 2019
discharge time series of the Juniata River at Newport, Pennsylvania, USA. Synthetic suspended sediment EDITOR
discharges are first generated with the chosen model after which analytical expressions are derived for A. Castellarin
the expected value and variance of sediment accumulation in the reservoir. The expected value and
variance of the volume of sediment accumulation in the reservoir are calculated from a thousand ASSOCIATE EDITOR
synthetic time series each 38 years long and compared to the analytical approach. Stochastic and J. Rodrigo-Comino
analytical approaches perfectly trace the observation in terms of the expected value and variability. KEYWORDS
Therefore, it is concluded that the expected value and variance of sediment accumulation in a reservoir Juniata River; moving
could be estimated by analytical expressions without the cost of synthetic data generation mechanisms. average model; river
reservoir; storage volume;
suspended sediment
discharge

1 Introduction
(Garde and Ranga Raju 1977) or soft computational techni-
Suspended sediment load corresponds to that fraction of sedi- ques such as artificial neural networks (Tfwala and Wang
ment moving continuously within flow at a velocity which is 2016); monitoring, sampling, surveying; or remote sensing
closely the same as that of water, while the remaining fraction and using geographical information systems (Jain and Das
is denoted as bedload transported intermittently near the river 2010). Also, statistical analysis and stochastic time series mod-
bed by sliding and rolling (Bogardi 1974). In other words, elling techniques (Phien 1981, Aksoy et al. 2004, Hao et al.
rivers carry some sediment in suspension and transport larger 2017) have always been attractive alternatives due to the fact
solids as bedload which forms a delta at the head of reservoir as that monitoring and sampling are not easy tasks to use and, as,
the velocity and turbulence are greatly reduced in the reservoir in most cases, the simultaneous streamflow and sediment dis-
(Yanmaz 2006). However, the majority of the sediment load is charge records are not available for comparatively long periods
in the form of suspended sediment in plain rivers with mild to arrive at conclusive deterministic relationships between
slopes particularly (Bayazit 1997). Therefore, it could be a good sediment concentration and streamflow discharge. It was also
approximation to take the suspended sediment load into stated that analytical approaches and simulation techniques
account for the determination of the dead volume of river yield commensurate results to estimate the accumulated sedi-
reservoirs that is expected to be filled with sediment trapped ment yield stemming from erosion (Bogardi et al. 1977).
during the project life. The accumulation can be simply calcu- Rather than the time series simulation models, the so-called
lated by multiplying the annual load, reduced by the trapping soft computational techniques or computational intelligence
coefficient, to the project life of the reservoir although estima- approaches (Olyaie et al. 2015) have recently been more pop-
tion of reservoir sedimentation can be best performed by ular and preferable in the modelling complex hydrological
continuous hydrologic simulation models (Yanmaz 2006). processes including the suspended sediment discharge. For
Neither underestimation of accumulation nor its overestima- example; Melesse et al. (2011) estimated suspended sediment
tion is desired as the underestimation shortens the economic loads for three major rivers in the US using a multilayer
life of the reservoir while the overestimation results in unne- perceptron artificial neural network (ANN) modelling
cessary costs (Ranzi et al. 2012). approach to accommodate the historical daily and weekly
Numerous methods are available in the literature to quan- hydro-climatological data, current discharge, antecedent dis-
tify sediment transported within streamflow. It can be deter- charge, and antecedent sediment load. Chen and Chau (2016)
mined by analysing the streamflow and sediment discharge proposed a hybrid double feed forward neural network model
time series statistically (Phien and Arbhabhirama 1979, Gao for daily suspended sediment load estimation by combining
et al. 2010); correlating the collection of available streamflow fuzzy pattern-recognition and continuity equation into
and sediment discharge data through rating curve established a structure of double neural networks. Tabatabaei et al.
between the two datasets (Rosen and Xu 2014); employing (2019) proposed a multi-objective optimization approach
empirical approaches (Bogardi 1974), traditional equations using the genetic algorithm to increase the commonly used

CONTACT Hafzullah Aksoy haksoy@itu.edu.tr


© 2020 IAHS
HYDROLOGICAL SCIENCES JOURNAL 985

conventional model of the sediment rating curve. Azamathulla any time series model can be replaced for the model itself to
et al. (2013) applied gene expression programming, which is calculate the average and variance of the sediment accumulation
an extension of genetic programming, to the modelling of the in river reservoirs. Application on the 38 year-long suspended
suspended sediment load for the three Malaysian rivers. Wang sediment discharge data of the hydrological watershed of Juniata
et al. (2014) used different fuzzy logic models for the assess- River at Newport, Pennsylvania, USA is presented below to demon-
ment of river water quality data in China. Aksoy et al. (2004) strate the applicability of the methodology proposed.
applied a wavelet-based simulation method to generate statis-
tically similar annual suspended sediment discharge data. In
a similar approach, Nourani and Andalib (2015) examined the 2 Stochastic models for suspended sediment
efficiency of wavelet-based least square support vector discharge
machine model for the prediction of daily and monthly sus- Stochastic models are used in hydrology quite frequently to
pended sediment load of Mississippi River in the United States. generate synthetic time series of random sequences with sta-
Besides the above data-based time series and soft computa- tistical characteristics similar to the observation (Salas et al.
tional techniques, watershed-scale models solely calculating 1980). Among the synthetic data generation techniques, auto-
sediment transported within flow have also been developed regressive moving average models are very common in hydrol-
due to the importance of the issue in the engineering practice ogy for data generation (Salas et al. 1980). An autoregressive
(Merritt et al. 2003, Aksoy and Kavvas 2005). Also, such model of order p, AR(p) and a moving average model of order
hydrological models as SHESED (Wicks 1988), WEPP q, MA(q) are combined in an autoregressive moving average
(Flanagan et al. 2007), SWAT (Ayele et al. 2017) and WEHY model, ARMA(p,q) which is defined by:
(Kavvas et al. 2004, 2006) among many others accommodate
erosion and sediment transport modules to predict sediment X
p
  Xq
xt ¼ ϕi xt  i  μx  θj εt  j þ εt (1)
eroded by rainfall or flow, and transported to the reservoir i¼1 j¼1
through the existing river channel.
In terms of methodology, this study is close to the work of The model has p + q + 2 parameters; μx, the mean value of x; σε2,
Phien (1981) who calculated sedimentation from the generated the variance of the random variable ε; ϕ1, . . ., ϕp, autoregressive
sediment discharge time series. The generation mechanism for coefficients, and θ1, . . ., θq, moving average coefficients, all deter-
sediment discharge in Phien (1981) is the same as that of the mined from the existing data. The AR model is a particular case of
streamflow but it takes into account the basic relationship the ARMA model in which all θ coefficients are zero, while the
between the sediment and streamflow. This study however MA model is another particular case where all ϕ coefficients are
differs in the sense that it uses the autoregressive moving zero. For the sake of simplicity, AR(1), AR(2), MA(1), MA(2) and
average time series modelling approach directly for the sedi- ARMA(1,1) were considered as potential models in this study. No
ment discharge together with analytical expressions of the further detail was supplied here on the models as they are exten-
expected value and variance of sediment accumulation. By sively available in many textbooks (Salas et al. 1980).
doing so, the behaviour of sediment discharge is incorporated
into the model directly instead of transferring it from the
3 Data and case study
behaviour of the streamflow discharge as it was performed by
Phien (1981). A case on the suspended sediment discharge data was per-
The major goal of this study is to determine the expected value formed in this study. A brief information on the study area and
and variance of the sediment accumulation under the assumption statistical characteristics of the available suspended sediment
that the sediment is trapped with a full efficiency in the river discharge time series are first given below. The most appro-
reservoir. This is done in two ways; the first is by taking the expected priate model was then decided for which analytical expressions
value and variance of the suspended sediment time series simulated were derived to calculate the expected value and variance of
based on the autoregressive-moving average type model, and sediment accumulation along the project life of the reservoir.
the second is by using analytical expressions derived for the selected Annual suspended sediment discharge time series from
model. For the former case, the first- and second-order autoregres- 1952 to 1989, 38 years in length, taken from the US
sive models – AR(1) and AR(2) – the first- and second-order Geological Survey station (ID: 01567000, latitude: 40º28′44″,
moving average models – MA(1) and MA(2) – and the first-order longitude: 77º07′42″) installed on the Juniata River at Newport,
autoregressive-moving average model – ARMA(1,1) – were tested Pennsylvania, USA, was used in this study (Fig. 1). The Juniata
in choosing the most proper time series model for the simulation of River is the second-largest tributary of the Susquehanna River
suspended sediment discharge in the river. Analytical expressions of and is approximately 145 km long. It has a drainage basin that
the selected time series model are derived for the latter case to encompasses an area of 8687 km2 at the upstream of the
calculate the expected value and variance of the sediment accumu- station, about one-eighth of the drainage area of the
lation as a second goal of this study. The expected value and Susquehanna. The topography in the drainage basin is diverse.
variance of the sediment accumulation calculated from the simu- The drainage basin is underlain by sedimentary rocks com-
lated time series and those determined by using the analytical posed mainly of sandstone, siltstone, shale, and carbonate
expressions are then compared to understand if the analytical (Sevon 1989). Land use and land cover in the Juniata River
expressions could be replaced for the time series simulation models drainage basin are, respectively, forest (66%), cultivated land
which are time-consuming and costly in this sense. In other words, (14%), grassland (17%) and urban areas (3%) (Williams and
the goal of the study is to demonstrate that analytical expressions of Reed 1972); and forest (68%) agriculture (23%) urban (5%)
986 T. AKAR AND H. AKSOY

Figure 1. Hydrological watershed of the USGS station (ID: 01567000) on Juniata River at Newport, Pennsylvania, USA.

and other (4%) (Taylor et al. 1982). Although the trend has
been toward more urbanization, the mixed land use in the
Juniata basin still primarily includes forested areas (about two-
thirds of the basin) concentrated on the ridges, with agricul-
tural and urban areas in the valleys (SRBS 2005).
Time series of annual suspended sediment discharge at the
gauging station is given in Fig. 2. Inspection of the suspended
sediment discharge time series showed no evidences of a trend
or a jump. No periodicity exists in the time series as it is at the
annual timescale. The statistical characteristics of the annual
suspended sediment discharge are presented in Table 1. The
dataset is positively skewed; i.e. it is non-normal. Therefore,
a transformation would be needed to make the dataset nor- Figure 2. Annual suspended sediment discharge time series in Juniata River at
Newport.
mally distributed. As shown in Fig. 3, the dataset turns into
normal probability distribution after the logarithmic transfor-
mation has been applied: The correlation structure of the suspended sediment discharge
time series was checked to decide the best-fit model among poten-
y ¼ ln x (2) tial autoregressive-type stochastic models, which were taken in this
study as AR(1), AR(2), MA(1), MA(2) and ARMA(1,1) for the
where x corresponds to the original data while y is its trans- sake of simplicity. The correlogram and cross-correlogram of the
formed version. annual suspended sediment discharge time series of Juniata River,

Table 1. Characteristics of annual suspended sediment discharge of Juniata River at Newport, Pennsylvania (x is measured suspended sediment discharge, and y is its
log-transformed value). SD: standard deviation; CV: coefficient of variation, CS: coefficient of skewness.
Variable Average SD CV CS Autocorr. coeff. Max Min
(m3) (m3) r1 r2 r3 (m3) (m3)
x 88,476.1 60,224.6 0.681 2.939 0.055 0.052 −0.101 367,906 21,249.4
y 11.231 0.553 0.049 0.290 0.095 −0.004 −0.111 12.816 9.964
HYDROLOGICAL SCIENCES JOURNAL 987

annual suspended sediment discharge time series of Juniata River,


the AR(1), AR(2), MA(1), MA(2) and ARMA(1,1) models were
determined by Akar (2000), respectively, as:
zt ¼ 0:0949zt1 þ εt (3)

zt ¼ 0:0961zt1  0:0130zt2 þ εt (4)

zt ¼ 0:0958εt1 þ εt (5)

zt ¼ 0:0028εt1  0:0083εt2 þ εt (6)

zt ¼ 0:0404zt1  0:1366εt1 þ εt (7)


where z is the difference between the log-transformed variable
and its mean value is calculated as
zt ¼ yt  y (8)
Among the models above, the least value of the Akaike informa-
tion criterion (AIC) was obtained for the MA(1) model, which
Figure 3. Histogram of annual suspended sediment discharge in Juniata River at
Newport fitted to the normal probability distribution after logarithmic
makes it the best-fit model for the data in hand (Table 2). To
transformation. check the independence and the normality of the MA(1) model,
the Porte Manteau and the probability plot correlation coeffi-
cient (PPCC) tests were applied, respectively. The test statistics
(χ 2 for the Porte Manteau and r for the PPCC tests) are given in
Table 2 and compared to their critical values at 95% of con-
fidence level for L – p – q degrees of freedom, where L is the
maximum lag and is taken as 10, which usually is 10–30% of the
time series length (Salas et al. 1980). The model passed both tests
(χ 2 < χ2cr and r > rcr) indicating that it has independent residuals
with normal distribution.
Among the tested models, MA(1) was found the best for the
case studied here using the annual suspended sediment dis-
charge time series of Juniata River at Newport. Having decided
that the type and order of the model is MA(1), the question
remains why an MA(1) model should be established, while
a simple random number generator might do the same or
similar work. The reason is that the MA(1) model is compar-
able with a random number generation mechanism due to the
fact that it has a structure with zero-correlation for lags higher
than zero. The question can be responded by making
a connection to the two-fold goal of the study in which
a time series model is first determined for the simulation of
the suspended sediment discharge data, and analytical expres-
sions for the average and variance of the sediment accumula-
tion are then derived for the selected model. The question is
understandable when the MA(1) model is obtained for the data
in hand; however, when the goal of the study is considered, it
Figure 4. Correlogram (upper panel) and partial-correlogram (lower panel) of
annual suspended sediment discharge of Juniata River at Newport with 95% becomes void as analytical expressions for the average and
confidence interval. variance of the suspended sediment accumulation can only

together with the 95% confidence intervals, are given in Fig. 4. The Table 2. Statistics (AIC, and χ 2 and r) of the Porte Manteau and PPCC tests
calculated for the models tested and the critical values.
correlation and partial correlation coefficients up to lag 20 except-
ing lag-zero are all within the 95% confidence interval. Low corre- Model AIC χ2 χ 2cr r rcr
lation coefficients suggest the use of lower-order autoregressive- AR(1) 2.081 5.92 16.92 0.9878 0.971
AR(2) 1.694 9.37 15.51 0.9820 0.971
type models, such as AR(1), MA(1) and ARMA(1,1), that were MA(1) 1.645 5.13 16.92 0.9884 0.971
tested to alternate with each other. Using the steps applied on the MA(2) 3.984 5.32 15.51 0.9855 0.971
derivation of the autoregressive models (Salas et al. 1980) for the ARMA(1,1) 3.432 5.64 15.51 0.9869 0.971
988 T. AKAR AND H. AKSOY

 
be derived for a given model. Therefore, due to the specific cov xi xj
goals of this study, rather than a random number generator, ρij ¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi (15)
σ xi 2 σ xj 2
a model is needed to determine even if it is the MA(1) with
zero-correlation at lags one and higher.
Although σ x 2 , the variance of the variable x in the denominator
in Equation (15), has been indexed differently, both are the
same as the variable x comes from the same population. With
4 Analytical expressions for the stochastic model
the same idea, the expected values in the second term of the
In the hydraulic engineering practice, it is a common assump- right-hand side in Equation (14) are the mean value of the
tion that sediment transported within flow in the river is fully same variable x which are equal to the population value μx .
trapped in the reservoir behind the hydraulic structure such as Equations (9–15) are all valid when the normal probability
a dam. Because, with the reduced velocity of flow, sediment distribution can be applied on the random variable x.
motion becomes such slow that sediment particles settle down- The expected value and variance of sediment accumulation
ward in the reservoir under gravity. Therefore, sediment accu- in the reservoir can be determined with analytical expressions
mulation in the river reservoir can be calculated by taking the that can be derived for a set of normally distributed random
sum of the suspended sediment discharge. Sediment accumu- numbers as well as correlated processes such as autoregressive
lation in the reservoir through an N-year period is given by: moving average. In this study, an approach similar to that used
by Phien (1981) was followed for deriving analytical expres-
X
N
sions for the expected value and variance of sediment accu-
SN ¼ xi (9)
mulation when a first-order moving average model, MA(1), is
i¼1
applied on the normal and log-normal suspended sediment
where xi is the annual suspended sediment volume flowing into discharge sequences.
the reservoir during the year i. Analytical expressions can also be The MA(1) model has the general form of:
derived for the variance of the sediment accumulation by using
simple statistical manipulations. The annual suspended sedi- xi ¼ μx  θ1 εi1 þ εi (16)
ment volume is a random process which is the decomposition in which θ1 is the model parameter, ε is the residual with zero-
of sediment accumulation, another random variable. The var- mean and variance σ ε 2 . Analytical expressions were derived for
iance of the random variable SN can be calculated by: the MA(1) process that was found the best-fit model to the 38-
  year-long annual suspended sediment discharge time series of
varðSN Þ ¼ E SN 2  E2 ðSN Þ (10)
Juniata River.
The first term on the right-hand side of Equation (10) is: In order to derive the expected value and variance of the
sediment accumulation in river reservoirs by using analytical
  X N   X N X
N   expression of the generation mechanism, the MA(1) model in
E SN 2 ¼ E xi 2 þ E xi xj i Þ j (11) this study, Equation (16) was inserted into Equation (9) and its
i¼1 i ¼ 1j ¼ 1
expected value is taken:
which is the expected value of the product of the sediment X
N
accumulation. By rearranging the lower and upper limits in the EðSN Þ ¼ Eðμx  θ1 εi1 þ εi Þ (17)
summation, Equation (11) can be written as: i¼1

  X N   X N 1 X
N   X N X
i1   Considering that the expected value of the residual is zero:
E SN 2 ¼ E xi 2 þ E xi xj þ E xi xj
i¼1 i ¼ 1j ¼ i þ 1 i ¼ 2j ¼ 1
EðSN Þ ¼ Nμx (18)
(12) is obtained. For the variance of sediment accumulation,
Equation (10) is valid. When Equation (12) is inserted into
which is a matrix that is composed of the elements on the Equation (10) and the MA(1) process is considered:
diagonal as in the first term of the right-hand side plus the
upper and lower triangles as in the second and the third terms, varðSN Þ
respectively. h
X
N 2 i
Equation (10) is used for the diagonal term in Equation ¼ E μx  θ1 εi  1 þ εi
(12) as: i¼1
  N 1 X
X N   
varðxi Þ ¼ E xi 2  E2 ðxi Þ (13) þ E μx  θ1 εi  1 þ εi μx  θ1 εj  1 þ εj
i ¼ 1j ¼ i þ 1
while
X
N iX
 1   
      þ E μx  θ1 εi  1 þ εi μx  θ1 εj  1 þ εj
cov xi xj ¼ E xi xj  Eðxi ÞE xj (14)
i ¼ 2j ¼ 1
 2
is valid for the non-diagonal terms. The left-hand side in  Nμx
Equation (14) can be obtained from the definition of the
(19)
correlation coefficient given by:
HYDROLOGICAL SCIENCES JOURNAL 989

is obtained. It should be noticed also that Equation (18) has


already been replaced in Equation (19) as the last term in the
right-hand side.
The independent residuals are correlated at lag zeroonly. In
Equation (19), this condition is satisfied N times for E ðεi1 Þ2
   
and E ðεi Þ2 in the first term, N  1 times for E εi εj1 in
 
the second term, and N  1 times for E εi1 εj in the third
term. Other terms with the residuals are either not correlated
or have zero-mean, and they are therefore omitted. The MA(1)
process follows:
σx2
σε2 ¼ (20)
1 þ θ1 2

θ1
ρxð1Þ ¼  (21)
1 þ θ1 2
where ρxð1Þ denotes the lag-one autocorrelation of variable x.
The variance of the sedimentation accumulation is finally
obtained as:
h i
varðSN Þ ¼ N þ 2ðN  1Þρxð1Þ σ x 2 (22)

When the non-normal variable x is transformed to the log-


normal variable y, it can be proven that:
σy 2
μx ¼ e μy þ 2 (23)
 2 
σ x 2 ¼ e2μy þσ y eσy  1
2
(24)

eσ y ρyðkÞ  1
2

ρxðkÞ ¼ σ2 (25)
e y 1
are valid. In Equations (23–25), μy and σ y are the mean and
standard deviation of the log-normal variable y, respectively,
and ρyðkÞ is its autocorrelation coefficient at lag k. Equations
(23–25) are inserted into Equations (18) and (22) to derive the
analytical expressions:
σy 2
EðSN Þ ¼ Ne μy þ 2 (26)
h  2   i
varðSN Þ ¼ e2μy þσ y N eσ y  1 þ 2ðN  1Þ eρyð1Þ σ y  1
2 2

(27)
for the expected value and variance of sediment accumulation,
respectively, when the distribution of the suspended sediment
discharge is log-normal.

5 Results
Figure 5. Step-by-step application of the time series model.
The application of the model on the dataset described above
was performed by following steps in Fig. 5. The suspended
sediment discharge data are first transformed into the nor- finally checked against the independence and normality
mal distribution by the logarithmic transformation as in conditions. With the chosen model in hand, random num-
Equation (2). The time series model, MA(1), was selected bers were generated for the synthetic suspended sediment
among the alternatives considered and its parameters were discharge time series at the length required. At the last step,
determined as in Equation (5) after which its residuals were back-transformation was applied on the generated
990 T. AKAR AND H. AKSOY

suspended sediment discharge sequences to obtain the time


series in the original domain.

5.1 Synthetic suspended sediment discharge


A set of 1000 suspended sediment discharge time series,
each 38 years long, was generated by the MA(1) model in
Equation (5). The sum of the generated suspended sediment
discharge over the period of time was calculated as the
volume of sediment accumulation in the river reservoir.
The generated suspended sediment discharge time series
was considered to make a comparison to the 38-year
observed time series. Comparison of the average of 1000
generated suspended sediment discharge time series with
the observation in Table 3 shows that the statistical char-
acteristics are mostly well preserved. The average was per-
fectly simulated with a relative error much less than 1%. The
standard deviation of the generated time series is about 12%
higher than its observed counterpart. The maximum and
minimum of the generated sequences are higher and lower
than the observed maximum and minimum values as Figure 6. Histogram of the observed 38-year-long annual suspended sediment
expected; because the generated time series are 1000-fold discharge and the average of 1000 generated time series each 38 years long.
longer than the observed time series. Although a lower
coefficient of skewness was obtained in the generated time
series, the distribution of the suspended sediment discharge
is positively skewed as in the observation. Correlation coef-
ficient was found at the same order of magnitude and close
to zero again as in the observed time series. This statistical
comparison proves that the MA(1) model has a good per-
formance in simulating the observed suspended sediment
discharge process.
To further show the similarity between the generated and
observed suspended sediment discharge time series, the his-
togram, the cumulative frequency curve and the duration
curve of the suspended sediment discharge were respectively
given in Figs. 6–8 all depicting the quite good performance of
the generation mechanism. The histogram in Fig. 6 shows the
similarity between the distribution of the observed and gen-
erated annual suspended sediment discharge sequences. The
Figure 7. Cumulative frequency of the observed 38-year-long annual suspended
distribution of the generated sequence is based on the average sediment discharge and maximum, minimum and average of 1000 generated
of 1000 time series. It is seen that the positively skewed time series each 38 years long.
distribution of the suspended sediment discharge is preserved
in the time series generated by the MA(1) model. The most
frequently observed interval of the suspended sediment dis- plotted in Fig. 7 from which it is seen that the average curve
charge is the same both in the observed and generated time of the generated MA(1) process fits the observation perfectly.
series; however, the higher frequency in the generated time The maxima and minima curves are, respectively, larger and
series should be noticed. The cumulative frequency diagrams lower than the average curve as they should be. In Fig. 8, three
of the maximum, average and minimum values of annual suspended sediment discharge duration curves are compared;
suspended sediment discharges of 1000 time series are the observed time series (38 years long), the average time
series (average of 1000 single generated time series each
38 years long), and the whole time series (1000 generated
Table 3. Statistical characteristics of the observed and generated annual sus-
pended sediment discharge. time series taken at once, i.e. 38,000 years in length). It is seen
Characteristic Observed Generated that the average curve matches the observed duration curve
Average (m3) 88,476.1 88,204.9 perfectly. The duration curve that takes the 1000 suspended
Standard deviation (m3) 60,224.6 52,738.3 sediment discharge time series at once fits the observed dura-
Coefficient of variation 0.681 0.598 tion curve as well as the average curve. Higher maxima and
Maximum (m3) 367,906.0 771,792.2
Minimum (m3) 21,249.4 7850.3 lower minima are found in this particular curve for low and
Coefficient of skewness 2.939 2.000 high exceedance probabilities as expected due to its much
Correlation coefficient 0.055 0.083 larger size.
HYDROLOGICAL SCIENCES JOURNAL 991

Table 4. Observed sediment accumulation, (SN)o, expected value, E(SN)s, and


variance var(SN)s, of accumulation calculated from the synthetic time series
generated by the MA(1) model (Equation (5)), and expected value, E(SN)a, and
variance, var(SN)a, of accumulation calculated by analytical expressions
(Equations (26–27)) of the log-transformed MA(1) process.
Sediment accumulation
S25 S38 S50 S75 S100
(SN)o (106 m3) 3.362
E(SN)s (106 m3) 2.208 3.352 4.393 6.588 8.781
E(SN)a (106 m3) 2.197 3.340 4.395 6.592 8.790
var(SN)s (1012 m6) 0.084 0.128 0.167 0.247 0.334
var(SN)a (1012 m6) 0.080 0.122 0.161 0.241 0.321

total sediment accumulation of the observed time series (shown as


a dot). The 38-year accumulation at the end of observation (from
1952 to the end of 1989) is 3.362 × 106 m3, while the median value
of the 38-year-long simulations is 3.327 × 106 m3. The relative
error between the observed sediment accumulation over 38 years
and the median value of simulation is only 1.04%. This shows how
the expected sediment accumulation was perfectly reproduced by
the MA(1) process of the annual suspended sediment discharge at
Figure 8. Duration curve of the observed and generated suspended sediment Juniata River at Newport.
discharge time series. Sediment accumulation calculated from the 1000 generated
time series of 25, 38, 50, 75 and 100 years in length were
considered to make a comparison with the analytical expres-
5.2 Sediment accumulation
sions with which the expected value and variance of the accu-
In addition to the 38-year-long generated suspended sediment mulation are determined. The average and variance of the
discharge, the MA(1) model (Equation (5)) was re-run to generate sediment accumulation in the reservoir were calculated from
synthetic time series at lengths of 25, 50, 75 and 100 years. the generated time series and compared to the expected value
A thousand time series were generated at each of the lengths as and variance of the accumulation calculated analytically by
in the 38-year-long generated sequence. Sediment accumulation using Equations (26–27). Results are compared in Table 4. It
at the end of each generated time series was calculated simply by is seen that the average and the variance of the sediment
taking the summation of the total volume of the simulated annual accumulation calculated from the stochastic suspended sedi-
sediment discharges. It was assumed that sediment within flow is ment discharge time series and those calculated with the ana-
trapped fully at the reservoir as mentioned earlier. Calculated lytical expressions match perfectly. The volume of sediment
based on this concept, the cumulative frequency of sediment accumulation (3.362 × 106 m3) over the 38-year observation
accumulation at the end of each simulation length considered period in the Juniata River was again perfectly replaced by both
(i.e. 25, 38, 50, 75 and 100 years) are plotted in Fig. 9. The median the stochastic and analytical approaches. The stochastic
value of each curve in Fig. 9 might be taken as the expected value approach calculated the 38-year accumulation volume as
of the sediment accumulation at the end of each simulation length. 3.352 × 106 m3, while the analytical calculation gave
Clearly, median values of the curves increase with increase in the 3.340 × 106 m3. The relative error between the simulated
simulation length. It is seen in Fig. 9 that the median value of the time series and the observation is 0.30% only, while it is
38-year-long curve of the generated time series is the same as the 0.65% for the analytical approach. Both approaches indicate
a perfect match in terms of the expected value of the sediment
accumulation in the reservoir. It is important to emphasize
that the close similarity of the expected values to the observed
accumulation was equally successful in both the stochastic and
analytical approaches. In terms of the variance of the sediment
accumulation, both approaches are identical. In other words,
no difference exists between the stochastic and analytical
approaches. There is only one value observed for the sediment
accumulation; therefore, the variance of the observed accumu-
lation could not be determined to make a comparison with.

6 Discussion
The autoregressive moving average (ARMA)-type stochastic
models have been useful tools in many hydrological applica-
Figure 9. Cumulative frequency of sediment accumulation based on 1000 gen-
erated time series of 25, 38, 50, 75 and 100 years in length. The circle dot is the tions including the generation of suspended sediment dis-
observed 38-year volume of sediment accumulation. charge time series as in this study. This is an important gain
992 T. AKAR AND H. AKSOY

against short records of suspended sediment discharge in the watershed in the United States is possible to mention as an
rivers. Short records of the suspended sediment discharge example (Mukundan et al. 2013). It is important to close this
prevent their use in the determination of the volume of sedi- brief discussion by concluding that any effort to develop mod-
ment accumulation in the reservoir design. Therefore, stochas- els using either data-based techniques or physically based
tic modelling techniques, particularly the ARMA processes, models are good opportunities for the data-available sites,
have been important tools in hydrology since the pioneering and to transfer them to data scarce-regions of the world or to
studies going back to the 1960s (Thomas and Fiering 1962, regions with no data at all.
Yevjevich 1972, Salas et al. 1980), and they are still in use for Another important issue in the stochastic modelling is that
hydrological problems (Wang et al. 2018). the time series is assumed stationary. In other words, gener-
Analytical expressions for the expected value and variance ated sequences of the hydrological process under investigation,
of sediment accumulation in river reservoirs can be derived for the suspended sediment discharge in this study, are considered
ARMA processes. This study is a particular case which is based stationary; i.e. they have the same statistical characteristics as
on MA(1) process applied on annual suspended sediment the observed sequence of the suspended sediment discharge
discharge time series of a single gauging station. The study even though they are much longer compared to the observed
can be extended to higher order stochastic models by also counterpart. However, on one hand, the stationarity is con-
taking the periodicity into account, considering monthly data sidered dead (Milly et al. 2008) with the change in hydrological
for instance. Another possibility to extend this study could be processes (Montanari et al. 2013, McMillan et al. 2016, Ceola
the use of stochastic and analytical approaches in the stream- et al. 2016, Bu et al. 2018), while, on the other hand, it is still
flow and suspended sediment discharge processes jointly and alive and inevitably useful in modelling the hydrological pro-
to derive the rating curve between the two processes. cess under investigation (Koutsoyiannis 2011, Lins and Cohn
Analytical expressions were derived for the best-fit stochas- 2011, Matalas 2012, Koutsoyiannis and Montanari 2014);
tic model to calculate the expected value and variance of because it is convenient to use in making reliable predictions
sediment accumulation in the reservoir for any given project for engineering design. As used in this study, the modelling
life. They are simple to use without needing the costly syn- concept based on the observed data is a useful practice due to
thetic data generation mechanisms of the stochastic approach. the fact that the past is representative of the future (Montanari
This is one of the advantages of the analytical expressions and Koutsoyiannis 2014). Yet, any gradual and sudden
which are available to use as soon as the stochastic model of changes due either to natural variation of the hydrological
the observed annual suspended sediment discharge has been process under investigation or to anthropogenic interventions
decided after statistical tests. should not be ignored as they are likely to cause increased
Many case studies have been reported in the literature using frequency and intensity of extreme hydrological events such as
the suspended sediment discharge data from the United States floods and droughts (Kundzewicz et al. 2014, Cavus and Aksoy
(Jones et al. 2001, Alp and Cigizoglu 2005, Rajaee 2011, 2019, 2020). Such changes that could arise in the form of either
Melesse et al. 2011, Leisenring and Moradkhani 2012, a trend or a jump could be considered through the existing
Mukundan et al. 2013, Olyaie et al. 2015, Taormina and trend detection mechanisms and segmentation tools (Gedikli
Chau 2015, Chen and Chau 2016) together with similar et al. 2007, 2010, Aksoy et al. 2008, Grimaldi et al. 2011) to
attempts from other countries such as Taiwan (Kao et al. eliminate the non-stationarity of the process. Any jump or
2005), China (Wang et al. 2014), Turkey (Buyukyildiz and trend, when exist, are determined and added into the modelled
Kumcu 2017) and Iran (Tabatabaei et al. 2019). As it is fol- hydrological variable such that the nonstationary behaviour of
lowed from the mentioned recent literature, modelling tech- the process is taken into account.
nology uses mainly the data-based soft computational methods
which are all black-box approaches that process the input data
7 Conclusion
to get the output. This has so far been a good practice in the
sediment transport research to develop models for the estima- Based on the case study on the annual suspended sediment
tion of suspended sediment sequences to calculate sediment discharge time series measured in streamflow gaging station on
accumulation. Juniata River at Newport, Pennsylvania, USA, the first-order
For processing the input, the black-box approach requires moving average model, MA(1), is found the best among the
either no physics at all for the hydrological problem in hand or candidate models to simulate the suspended sediment discharge.
it is needed at minimum. On the other hand, physically based It is assumed that the suspended sediment is trapped fully in the
models use such basin characteristics as topography, geology, river reservoir. Statistical moments of the generated suspended
land use, land cover, etc., at different space and time resolu- sediment discharge time series are used to calculate the expected
tions. Large volume of geomorphological datasets is required value and variance of sediment accumulation. Also analytical
about the river basin which can only be made available after expressions are derived for the expected value and variance of
great efforts with high costs particularly in non-developed the accumulation for the MA(1) model. The expected value and
countries. Even when the data are available, it might be ques- variance of the sediment accumulation are found to be similar, no
tionable to use because of data quality concerns. Therefore, matter whether they are calculated by taking the statistical
application of the physically based models has not been as moments of the artificial sediment discharge time series generated
simple as the soft computational techniques. However, there by the chosen stochastic model or by using its analytical expres-
are studies that use physically based models for which the case sions. Large number of synthetic data sequences are generated
study of SWAT model on the New York City water supply (1000 time series in this study) in the stochastic approach while
HYDROLOGICAL SCIENCES JOURNAL 993

synthetic data generation is not needed if the analytical expres- Bu, J., et al., 2018. Attribution of runoff reduction in the Juma River basin
sions are used. Based on the results of the case study, it is seen that to climate variation, direct human intervention, and land use change.
the statistical moments of the stochastic model and its analytical Water, 10 (12), 1775. doi:10.3390/w10121775
Buyukyildiz, M. and Kumcu, S.Y., 2017. An estimation of the suspended
expressions could replicate the observed expected value of sedi- sediment load using adaptive network based fuzzy inference system, sup-
ment accumulation perfectly, and both methods demonstrate the port vector machine and artificial neural network models. Water Resources
same variability. The conclusion is that, by using the analytical Management, 31 (4), 1343–1359. doi:10.1007/s11269-017-1581-1
expressions of the best-fit model of the suspended sediment Cavus, Y. and Aksoy, H., 2019. Spatial drought characterization for
discharge, the expected value and variance of the sediment accu- Seyhan River basin in the Mediterranean region of Turkey. Water,
11, 1331. doi:10.3390/w11071331
mulation over a given project life of a river reservoir can be
Cavus, Y. and Aksoy, H., 2020. Critical drought severity/intensity-dura-
estimated with a simple hand-calculator without the costly syn- tion-frequency curves based on precipitation deficit. Journal of
thetic data generation mechanism which requires extra efforts in Hydrology, 124312. doi:10.1016/j.jhydrol.2019.124312
the stochastic approach. Ceola, S., et al., 2016. Adaptation of water resources systems to changing
society and environment: a statement by the international association
of hydrological sciences. Hydrological Sciences Journal, 61 (16),
Acknowledgements 2803–2817. doi:10.1080/02626667.2016.1230674
Chen, X.Y. and Chau, K.W., 2016. A hybrid double feedforward neural
The authors are grateful to the anonymous reviewers and Associate network for suspended sediment load estimation. Water Resources
Editor, Dr Jesús Rodrigo-Comino, whose constructive comments Management, 30 (7), 2179–2194. doi:10.1007/s11269-016-1281-2
improved the paper significantly. The authors thank also Asst. Prof. Flanagan, D.C., Gilley, J.E., and Franti, T.G., 2007. Water erosion predic-
Halil Ibrahim Burgan, Istanbul Kultur University, and Ms Yonca Cavus, tion project (WEPP): development history, model capabilities, and
PhD candidate at Istanbul Technical University, for the figures. future enhancement. Transactions of the American Society of
Agricultural and Biological Engineers, 50 (5), 1603–1612.
doi:10.13031/2013.23968
Gao, P., et al., 2010. Trend and change-point analyses of streamflow and
Disclosure statement sediment discharge in the Yellow River during 1950–2005.
No potential conflict of interest was reported by the authors. Hydrological Sciences Journal, 55 (2), 275–285. doi:10.1080/
02626660903546191
Garde, R.J. and Ranga Raju, K.G., 1977. Mechanics of sediment transpor-
tation and alluvial stream problems. New Delhi: Wiley Eastern.
ORCID Gedikli, A., Aksoy, H., and Unal, N.E., 2007. Segmentation algorithm for
long time series analysis. Stochastic Environmental Research and Risk
Hafzullah Aksoy http://orcid.org/0000-0001-5807-5660 Assessment, 22 (3), 291–302. doi:10.1007/s00477-007-0115-4
Gedikli, A., et al., 2010. Modified dynamic programming approach for
offline segmentation of long hydrometeorological time series.
References Stochastic Environmental Research and Risk Assessment, 24 (5),
547–557. doi:10.1007/s00477-009-0335-x
Akar, T., 2000. Time series modelling of suspended sediment discharges and
Grimaldi, S., et al., 2011. Statistical hydrology, treatise on water science.
dead storage volume estimation. Thesis (PhD). Istanbul Technical
P. Wilderer, ed. Oxford: Academic Press, Vol. 2, 479–517.
University, Institute of Science and Technology, Istanbul, Turkey (in
Hao, C.F., Qiu, J., and Li, F.F., 2017. Methodology for analysing and
Turkish with English summary).
predicting the runoff and sediment into a reservoir. Water, 9, 440.
Aksoy, H., Akar, T., and Unal, N.E., 2004. Wavelet analysis for modeling
doi:10.3390/w9060440
suspended sediment discharge. Nordic Hydrology, 35 (2), 165–174.
Aksoy, H., et al., 2008. Fast segmentation algorithms for long hydrome- Jain, M.K. and Das, D., 2010. Estimation of sediment yield and areas of
teorological time series. Hydrological Processes, 22 (23), 4600–4608. soişl erosion and deposition for watershed prioritization using GIS and
doi:10.1002/hyp.v22:23 remote sensing. Water Resources Management, 24 (10), 2091–2112.
Aksoy, H. and Kavvas, M.L., 2005. A review of hillslope and watershed doi:10.1007/s11269-009-9540-0
scale erosion and sediment transport models. Catena, 64, 247–271. Jones, K.B., et al., 2001. Predicting nutrient and sediment loadings to
doi:10.1016/j.catena.2005.08.008 streams from landscape metrics: a multiple watershed study from the
Alp, M. and Cigizoglu, H.K., 2005. Suspended sediment load simulation United States Mid-Atlantic region. Landscape Ecology, 16, 301–312.
by two artificial neural network methods using hydrometeorological doi:10.1023/A:1011175013278
data. Environmental Modelling and Software, 22, 2–13. doi:10.1016/j. Kao, S.J., Lee, T.Y., and Milliman, J.D., 2005. Calculating highly fluctuated
envsoft.2005.09.009 suspended sediment fluxes from mountainous rivers in Taiwan.
Ayele, G.T., et al., 2017. Streamflow and sediment yield prediction for Terrestrial, Atmospheric, and Oceanic Sciences, 16 (3), 653–675.
watershed prioritization in the Upper Blue Nile river basin, Ethiopia. doi:10.3319/TAO.2005.16.3.653(T)
Water, 9, 782. doi:10.3390/w9100782 Kavvas, M.L., et al., 2004. Watershed environmental hydrology (WEHY)
Azamathulla, H., et al., 2013. Suspended sediment load prediction of river model based on upscaled conservation equations: hydrologic module.
systems: GEP approach. Arabian Journal of Geoscience, 6, 3469–3480. ASCE Journal of Hydrologic Engineering, 9 (6), 450–464. doi:10.1061/
doi:10.1007/s12517-012-0608-4 (ASCE)1084-0699(2004)9:6(450)
Bayazit, M., 1997. Design and operation of river reservoirs. 1st ed. Istanbul, Kavvas, M.L., et al., 2006. Watershed environmental hydrology model:
Turkey (in Turkish): Istanbul Technical University Press Office. environmental module and its application to a California watershed.
Bogardi, J., 1974. Sediment transport in alluvial streams. Budapest: ASCE Journal of Hydrologic Engineering, 11 (3), 261–272. doi:10.1061/
Akademiai Kiado. (ASCE)1084-0699(2006)11:3(261)
Bogardi, J., et al., 1977. Reservoir sedimentation under uncertainty: ana- Koutsoyiannis, D., 2011. Hurst-Kolmogorov dynamics and uncertainty.
lytic approach versus simulation. Hydrological Sciences Journal, 22(4), Journal of the American Water Resources Association, 47, 481–495.
545–553, doi: 10.1080/02626667709491759 doi:10.1111/j.1752-1688.2011.00543.x
994 T. AKAR AND H. AKSOY

Koutsoyiannis, D. and Montanari, A., 2014. Negligent killing of scien- Ranzi, R., Le, T.H., and Rulli, M.C., 2012. A RUSLE approach to model
tific concepts: the stationarity case. Hydrological Sciences Journal. suspended sediment load in the Lo River (Vietnam): effects of reser-
doi:10.1080/02626667.2014.959959 voirs and land use changes. Journal of Hydrology, 422–423, 17–29.
Kundzewicz, W., et al., 2014. Flood risk and climate change: global and doi:10.1016/j.jhydrol.2011.12.009
regional perspectives. Hydrological Sciences Journal, 59 (1), 1–28. Rosen, T. and Xu, Y.J., 2014. A hydrograph-based sediment availability
doi:10.1080/02626667.2013.857411 assessment: implications for Mississippi River sediment diversion.
Leisenring, M. and Moradkhani, H., 2012. Analysing the uncertainty of Water, 6, 564–583. doi:10.3390/w6030564
suspended sediment load prediction using sequential data Salas, J.D., et al., 1980. Applied modeling of hydrologic time series.
assimilation. Journal of Hydrology, 468–469, 268–282. doi:10.1016/j. Littleton, CO: Water Resources Publications.
jhydrol.2012.08.049 Sevon, W.D., 1989. Erosion in the Juniata River drainage basin, Pennsylvania.
Lins, H.F. and Cohn, T.A., 2011. Stationarity: wanted dead or alive? Geomorphology, 2, 303–318. doi:10.1016/0169-555X(89)90017-2
Journal of the American Water Resources Association, 47, 475–480. SRBS (Susquehanna River Basin Commission), 2005. Juniata River
doi:10.1111/j.1752-1688.2011.00542.x Subbasin survey, a water quality and biological assessment.
Matalas, N.C., 2012. Comment on the announced death of stationarity. Harrisburg, Pennsylvania, Publication 240, July
Journal of Water Resources Planning and Management, 138, 311–312. Tabatabaei, M., Jam, A.S., and Hosseini, S.A., 2019. Suspended sediment
doi:10.1061/(ASCE)WR.1943-5452.0000215 load prediction using non-dominated sorting genetic algorithm II.
McMillan, H., et al., 2016. PantaRhei 2013–2015: global perspectives on International Soil and Water Conservation Research, 7, 119–129.
hydrology, society and change. Hydrological Sciences Journal, 61 (7), doi:10.1016/j.iswcr.2019.01.004
1174–1191. doi:10.1080/02626667.2016.1159308 Taormina, R. and Chau, K.W., 2015. Data-driven input variable selection
Melesse, A.M., et al., 2011. Suspended sediment load prediction of river for rainfall–runoff modeling using binary-coded particle swarm opti-
systems: an artificial neural network approach. Agricultural Water mization and extreme learning machines. Journal of Hydrology, 529,
Management, 98 (5), 855–866. doi:10.1016/j.agwat.2010.12.012 1617–1632. doi:10.1016/j.jhydrol.2015.08.022
Merritt, W.S., Letcher, R.A., and Jakeman, A.J., 2003. A review of erosion Taylor, L.E., et al., 1982. Groundwater resources of the Juniata River basin,
and sediment transport models. Environmental Modelling and Pennsylvania. Pennsylvania Geological Survey, 4th Series, Water
Software, 18 (8–9), 761–799. doi:10.1016/S1364-8152(03)00078-1 Resources Report 54.
Milly, P.C.D., et al., 2008. Stationarity is dead: whither water
management? Science, 319, 573–574. doi:10.1126/science.1151915 Tfwala, S.S. and Wang, Y.M., 2016. Estimating sediment discharge using
Montanari, A. and Koutsoyiannis, D., 2014. Modeling and mitigating sediment rating curves and artificial neural networks in the Shiwen
natural hazards. Stationarity Is Immortal! Water Resources Research, River, Taiwan. Water, 8, 53. doi:10.3390/w8020053
50, 9748–9756. doi:10.1002/2014WR016092 Thomas, H.A. and Fiering, M.B., 1962. Mathematical synthesis of stream-
Montanari, A., et al., 2013. Panta Rhei-everything flows: change in hydrology flow sequences for the analysis of river basins by simulation. In:
and society-The IAHS scientific decade 2013–2022. Hydrological Sciences A. Maass, et al., ed. Design of water resource systems. Cambridge,
Journal, 58 (6), 1256–1275. doi:10.1080/02626667.2013.809088 MA: Harvard University Press, 459–493.
Mukundan, R., et al., 2013. Suspended sediment source areas and future Wang, W., et al., 2018. A stochastic simulation model for monthly river
climate impact on soil erosion and sediment yield in a New York City flow in dry season. Water, 10 (11), 1654. doi:10.3390/w10111654
water supply watershed, USA. Geomorphology, 183, 110–119. doi:10.1016/j. Wang, W.C., et al., 2014. Assessment of river water quality based on
geomorph.2012.06.021 theory of variable fuzzy sets and fuzzy binary comparison method.
Nourani, V. and Andalib, G., 2015. Daily and monthly suspended sedi- Water Resources Management, 28 (12), 4183–4200. doi:10.1007/
ment load predictions using wavelet based artificial intelligence s11269-014-0738-4
approaches. Journal of Mountain Science, 12 (1), 85–100. doi:10.1007/ Wicks, J.M., 1988. Physically-based mathematical modelling of catchment
s11629-014-3121-2 sediment yield. Thesis submitted for the degree of doctor of philoso-
Olyaie, E., et al., 2015. A comparison of various artificial intelligence phy. Department of Civil Engineering, University of Newcastle upon
approaches performance for estimating suspended sediment load of Tyne, UK. doi:10.3168/jds.S0022-0302(88)79586-7
river systems: a case study in United States. Environmental Monitoring Williams, K.F. and Reed, L.A., 1972. Appraisal of stream sedimentation
and Assessment, 187 (4), 189. doi:10.1007/s10661-015-4381-1 in the Susquehanna RiverBasin US Geological Survey Water Supply
Phien, H.N., 1981. Reservoir sedimentation with correlated inflows. Paper 1532-FPrepared in cooperation with the Pennsylvania
Journal of Hydrology, 53, 327–341. doi:10.1016/0022-1694(81)90009-3 Department of Forests and Waters, Pennsylvania Department of
Phien, H.N. and Arbhabhirama, A., 1979. A statistical analysis of the Agriculture, State Soil and Water Conservation Commission, and
sediment accumulation in reservoirs. Journal of Hydrology, 44, the US Army Corps of Engineers, Baltimore District.
231–240. doi:10.1016/0022-1694(79)90133-1 Yanmaz, A.M., 2006. Applied water resources engineering. 3rd ed. Ankara,
Rajaee, T., 2011. Wavelet and ANN combination model for prediction of Turkey: Middle East Technical University Press.
daily suspended sediment load in rivers. Science of the Total Yevjevich, V., 1972. Stochastic processes in hydrology. Littleton, CO: Water
Environment, 409 (15), 2917–2928. doi:10.1016/j.scitotenv.2010.11.028 Resources Publications.

You might also like