Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

J. Environ. Treat. Tech. ISSN:

2309-1185

Journal weblink: http://www.jett.dormaj.com

Comparing the Estimation of Suspended Load using Two

Methods of Sediments Rating Curve and Artificial Neural

Network (A Case Study: Cham Anjir Station, Lorestan Province)

Bahador Abbaspour 1 , Amir Hamzeh Haghiabi 2

1- M.S. Student of Water Structures, Water Engineering Department, Lorestan University

2- Associate Professor ,Water Engineering Department, Lorestan University

Received: 15/01/2015

Accepted: 17/05/2015

Published: 30/12/2015

Abstract

It is significantly important to predict and estimate the sediment load of the rivers to manage rivers and dam reservoirs in

water projects. In this study, the suspended load of the river is predicted using artificial neural network. In this paper, it is

attempted to evaluate the performance of artificial neural networks in predicting the suspended sediments. Using ANN

(Multilayer Layer Perceptron Model), the suspended sediment in hydrometric station of Cham Anjir river of Khorramabad

has been predicted and the results have been compared with sediment rating curve. Based on the obtained results, ANN

presents acceptable results in simulating the suspended load in Cham Anjir station, in such a way that it is of higher accuracy

compared to sediment rating curve. The results showed that ANN could be employed to estimate the sediment suspended

load with appropriate accuracy and more confidence compared to the rating curve. Here, it should be noted that neural

network could not predict the peaks accurately, and this is regarded as a weak point of this model.

Keywords: Suspended Sediment, ANN, Sediment Rating Curve, Khorramabad River.

1 Introduction 1

concentration

and

he

indicated

that

multi-layer

As

an

intensifying

process,

erosion

and

perceptron models performed better than the generalized

sedimentation result in the loss of agricultural fertile soil

regression neural networks and radial basis function

and cause irreparable damages to constructive water

networks [9].

projects, such as sediment accumulation behind dams

Kisi (2005) employed ANN to model the suspended

and reducing their useful volume, destructing the

sediment load of flow. He also used sediment

structures, damaging the coasts and harbors, reducing the

rating

curves

and a

multi-regression

model to

capacity and increasing the maintenance cost of irritation

predict the sediment load and showed that ANN

canals to name a few. On the other hand, sediment

produced the best results [10].

transfer affects the water quality indices with regard to

Using the measurement data of sediment suspended

potable and agricultural water. Therefore, estimating the

load and with the help of sediment rating models, it is

amount of sediment is required in soil protection

possible to predict the transfer sediment. However,

projects, designing and implementing water structures,

estimating the sediment using this model always

watershed and utilizing water resources [7].

associated some errors [2]. In addition, other researches

Jain (2001) used ANN to introduce the relation of

have shown that estimating the suspended sediment

sediment density in Mississippi river and proved that it

using rating curves always associates with errors and it

gives more reliable results in comparison with other

appears that the ANN model leads to fewer errors using

methods [8].

rating curves [12]. Therefore, it is required to utilize

Cigizoglu (2002) compared the results from ANN

smart methods such as ANN in estimating the amount of

with sediment rating curves to predict the density

suspended load. To simulate the suspended load using

of suspended sediments [3].

ANN, two algorithms of back-propagation training

The history of scientific investigations in the field of

algorithm and radial functions, they used the values

suspended sediment transfer in rivers is over a century, in

obtained in these two methods to draw the sediment

such a way that the first sampling of suspended load

diagram and they realized that back-propagation training

from the rivers was conducted in 1845 in Mississippi

algorithms estimate the discharge of river’s sediment

River [16].

load more accurately [13]. Dehghani et al. 2009

Kisi (2004) used different ANN techniques for

compared the estimation of suspended load using two

predicting and estimating daily suspended sediment

methods of sediment rating curve and ANN. The results

showed that ANN could be utilized more accurately and

confidently to estimate the sediment suspended load

Corresponding author: Bahador Abbaspour, M.S.

compared to rating curve with and without data

Student of Water Structures, Water Engineering

classification [6]. Using the data of water discharge and

Department,

Lorestan

University

E-mail:

sediment discharge simultaneously, in Galinak station in

b.a1366@yahoo.com.

Talghan River, along with some geomorphologic

215

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

parameters of the catchment area of this river, ANN was

values of total sediment load. The reason might be the

used to model the estimation of daily suspended

effect of suspended load on total load value, and the

sediment. The results of this study showed the higher

suggested equations are not able to calculate it [7].

accuracy of NN relative to regression models [15]. The

Melesse et al. (2011) estimated Suspended sediment

common sampling method in most hydrometric stations

loads for three major rivers (Mississippi, Missouri and

is measuring the sediment suspended load. Then, the bed

Rio Grande) in USA using MLP modeling approach.

load is estimated as a percent of suspended load with

Results from ANN model were compared with results

regard to geomorphologic conditions of the river.

from multiple linear regressions (MLR), multiple non-

However, the results of this procedure are not reliable

linear

regressions

(MNLR)

and

Autoregressive

[12]. It is required to have accurate information about the

integrated moving average (ARIMA) using correlation

amount of sediment load of the whole river in many

coefficient (R), mean absolute percent error (MAPE) and

water resources projects such as estimating the

model efficiency (E). The results show that the ANN

sedimentation volume in dam reservoirs. In order to

predictions for most simulations were superior compared

determine the total load of the floodway, numerous

to predictions using MLR, MNLR and ARIMA

procedures have been proposed, each one of which are

[14].Therefore, there is a need to present new

applicable in special conditions. Some of these relations

procedures, which could estimate total sediment load

are presented based on extending the equations of bed

more accurately, without requiring special conditions to

load. However, in addition to the need for special

be applied. The aim of this study has been improving the

conditions, the calculated value of total sediment load

estimation of total sediment load of the rivers using ANN

using these relations is different from the measured

(multilayer perceptron and radial base function).

Nomenclature

discharge of suspended sediment (tons/day)

୫ ୬

minimum value of observed data

discharge of its corresponding flow (m 3 /s)

୫ୟ୶

maximum value of observed data

a , b

coefficients of the equation

ୱ୲

mean of estimated data

observational discharge

୭ୠୱ

mean of observational data

ୱ୲

estimated discharge of neural network

ୱ୲

estimated data

Target discharge of normalized suspended sediment

୭ୠୱ

observational data

Output estimated discharge of suspended sediment from neural

correlation between data

network

Root mean square error

୬

expressive of normalized data

Nash coefficient

expressive of observational data

2 Material and Methods

responses from the network [5]. The artificial neural

2.1 Features of Catchment under Study

network consists of the following sections:

This

study

was

conducted

in

Cham

Anjir

1-

Input layer: In this layer, inputs are connected

hydrometric station in the catchment of Khorramabad

to the outside world and the next layer of the network.

River, regarding the detailed statistics of the available

No processing is performed in this layer.

concentration. The station is located in 33 o and 22 ’ north

2-

Hidden layer: a layer in which processing

latitude with 1122 meter height and upstream catchment

occurs. The network could consist of one or more middle

area of 1280 km 2 . The statistical details of the catchment

layers. The designer, often through trial and error

under study are shown in Table 1 and the catchment

process, obtains the number of layers and the number of

location is shown in Fig. 1.

nodes in each layer.

3-

Output layer: in this layer, the outputs are

2.2 Artificial Neural Network

connected to the outside world, in which the outside

Artificial neural network is a variant of artificial

vectors are mapped and established. Often, back-

intelligence, which performs similar to human brain,

propagation learning rule is employed to train the

albeit general and incomplete. In fact, ANN is an idea for

Multilayer Layer Perceptron (MLP) Neural Networks. In

data processing, which is inspired by bio-neural-system

other words, the topology of the network is completed

and processes the data similar to the brain. This system

through back-propagation learning rule. Neural networks

consists of numerous processing elements called

generally lack good interpolation. Therefore, this should

neurons, which act coordinately to solve a problem.

be considered when selecting the training pattern. To do

These neurons learn the problem’s process through an

so, the patterns are classified into training and test

example (observational data). In other words, they

patterns, before the neural network starts working. The

transfer data to the network structure through processing

training patterns should cover the data space as far as

experimental data, knowledge or hidden rules. In fact,

possible to produce better learning. It is obvious that the

ANN is a math model, which is capable of modeling and

more the training patterns, the higher the generalization

creating nonlinear math relations for interpolation. In

capability of promoting the network. Although training is

fact, ANN is trained through a limited series of real data

a process that occurs during a relatively long time, but

and if the parameters, which affect the phenomenon

after generalization, it could present an output in lieu of

under study, are selected accurately and given to the

each input rapidly.

network, it could be expected to receive rational

216

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

Fig. 1: The location of catchment area under study .

Table 1: Statistical details of flow discharge and sediment data in Cham Anjir station.

variable

mean

minimum

maximum

Standard deviation

flow discharge (m 3 /s)

8.84

0.33

34.89

6.30

Sediment discharge (tons/day)

123.87

0.469

1597.844

217.1286

Fig. 2: The structure of Artificial Neural Network.

2.3 Sediment Rating Curve

ୠ

୵

(1)

Among common methods in estimating discharge of

suspended sediment load in hydrometric stations is

In addition, various methods are proposed to increase

creating a connection between discharge data of the

the accuracy of estimating the sediment through rating

sediment and their corresponding discharge data, which

curve, among which is classifying the data in different

is obtained through fitting of the curve between flow

shapes. Classifying the data could be used in the form of

discharge values and their corresponding suspended

annual, seasonal, monthly, intermediate categories,

sediment Eq. (1).

217

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

classifieds discharge, similar hydrological cycle, and

ሺ

୭ୠୱ

ୱ୲ ሻ

periods of low and high precipitation.

(5)

ሺ

ୱ୲

୭ୠୱ ሻ

2.4 Data and Methods Used

R 2 shows the correlation between data and the more

In this study, the data and information available in

and close to one is R , the more correlation exists among

2

Cham Anjir hydrometric station, Khorramabad, have

observational and estimated data. RMSE shows how

been used. To do so, 473 corresponding data of flow

much the estimated values deviate from observational

discharge

and sediment

discharge

are

measured

data and the less is RMSE, the less is the deviation

simultaneously. In artificial network methods, first, some

among the data and the more accurate are the results.

data are selected to train the network, which are

Nash coefficient (r 2 ) is a statistical index, which shows

expressive of the conditions of the problem and other

the conformation ratio in hydrologic models. The value

data are used to test the performance of trained network.

of this index is between 1 and

. When it approaches

The main point in selecting the training data is that it

one, the model has a better performance; if it is zero, it

should cover a wide range of assorted data. In this study,

means that the model corresponds to the mean of the

70% of the data were used as training data, data 15% for

data; and when it is negative, the mean of the data is

testing and 15% for testing the accuracy of the neural

more efficient than the model [5].

network model. An important fact in training neural

networks is normalizing the data before being used in the

model. This fact leads to better and faster training of the

3 Results and Discussion

model, especially when the changing range of the inputs

3.1 Results from Artificial Neural Network

is high, because entering data in raw form reduces the

The results from the final model of neural network,

speed and accuracy of the network. To normalize the data

used in this study, are shown in Fig. 3, 4 and 5 for test

in this study Eq. (2), is employed.

stage data along with related observational data.

As can be seen in Fig. 3, the neural network has

lower or more estimation compared to observational

ሺ

୫ ୬

୬

ሻ

(2)

୫ୟ୶

୫ ୬

values, which could be one of the weak points of neural

network in estimating the suspended sediment. In other

Neural network was designed after selecting training

words, the neural network has simulated the suspended

data, testing and validation. To model the neural

sediment related to the base and normal discharge, well.

network, the software of MATLAB (Matrix Laboratory)

However, in simulating the sediment in flood events, it

was used. Among 473 data available, 331, 71 and 71 data

has not seen much successful. Results from Zhu et.al

were

used

for

training,

testing and

validation,

show this fact well.

respectively. To achieve a suitable structure of artificial

neural network, different models are designed and tested

3.2 Results from Sediment Rating Curve

with different hidden layers and nodes and the related

As the study continued, the sediments were estimated

results were compared. After creating and testing

through sediment rating curve to compare the results of

different structures of the network and evaluating the

neural network with the results of the common sediment

obtained results, the final structure of the employed

rating curve method. In sediment rating curve, a

neural network for this study was selected, which was

regression relation is usually extracted between the data

multilayer perceptron and back-propagation training

corresponding to the flow and sediment discharge. Then,

method was used, which was simulated using MATLAB.

based on this relation, the daily discharge value of

The input and output layer of the network, which results

suspended sediment is estimated for those days that the

in water discharge and output layer of sediment

concentration of sediment is not measured.

discharge, was used to evaluate the results from different

In this study, the discharge and sediment data related

models of artificial neural network and sediment rating

to those days, which were used for modeling neural

curve method in addition to comparing the final results

network for training, were drawn on a coordinates system

with observed values, along with 3 statistical parameters

to determine the sediment rating curve and obtain the line

as follows:

and equation related to the relation between discharge

and suspended sediment. Fig. 6 shows the line and

ሺ(

equation related to the rating curve. After obtaining the

ୱ୲

ୱ୲ ) (

୭ୠୱ

୭ୠୱ )ሻ

(3)

line equation in sediment rating curve method, data of the

ට ሺ

flow discharge related to the days, used in test stage of

ୱ୲

ୱ୲ ሻ ሻ ሺ

୭ୠୱ

୭ୠୱ ሻ

neural network, are placed in this equations, in such a

ሺ

୭ୠୱ

ୱ୲ ሻ

way that all of the discharge data of test stage are placed

ඨ

(4)

separately in the equation resulted from the sediment

rating curve.

218

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

Fig. 3: Results from the simulation of suspended sediment using neural network against observational values

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

y = 0.7992x + 0.0168

0.2

R² = 0.86326

0.1

0

0

0.2

0.4

0.6

0.8

1

1.2

observed suspended sediment (tons/day)

Fig. 4: The diagram of the results from the simulation of suspended sediment using neural network against observational values.

Then, the related

corresponding sediment is

The results of different statistical parameters such as

calculated. In Fig. 7, the values of estimated sediment

R 2 , RMSE and r 2 , based on rating curve and artificial

(based on sediment rating curve) are placed against the

neural network, are shown in table 2. According to the

measured values. Comparing Fig. 4 and 7, which show

Table, it could be observed that the neural network

the results of neural network and the common method of

method presents the results with less error and higher

sediment rating curve, respectively, against observational

correlation.

data, shows the relative superiority of neural network

method compared to sediment rating curve method.

219

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

Fig. 5: Results from neural network in each stage of training, testing and validation.

1800

1600

1400

1200

1000

800

600

y = 1.7745x 1.7111

400

R² = 0.6677

200

0

0

5

10

15

20

25

30

35

40

flow discharge (m3/s)

Fig. 6: Sediment rating curve and the line and equation obtained to estimate the sediment using this method.

220

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

900

800

700

600

500

400

y = 3.7293x0.6677

300

R² = 0.6677

200

100

0

0

200

400

600

800

1000

1200

1400

1600

1800

observed sediment (tons/day)

Fig. 7: The values of estimated suspended sediment based on sediment rating curve against observational values.

Table 2: Calculated statistical parameters related to neural network and rating curve.

Statistical parameters

Rating curve

Neural network

R 2

0.6677

0.87247

RMSE

0.152

0.0187

r 2

0.654

0.945

4 Conclusions

Ahvaz Seventh National Conference of River

In this study, the application of neural network

Engineering

method and rating curve in estimating the amount of

2- Achite, M., and S. Ouillon, 2007. Suspended

suspended sediment in Cham Anjir stations was

Sediment Transport in a Semiarid Watershed, J.

investigated. To do so, two parameters were used as

Hydro, 84: 3, 187-202.

inputs and the model was simulated. The neural network

3- Cigizoglu H.K., 2002. Suspended sediment estimation

presents acceptable results for simulation in Cham Anjir

for rivers using artificial neural networks and

station in such a way that one of the statistical parameters

sediment

rating

curves.

Turkish

Journal

of

to compare the acc uracy of the model’s estimations

Engineering Environmental Sciences 26(1):27-36.

relative to observational data is R 2 . R 2 , obtained from the

4- Dastorani, M.T. and H. Afkhami, 2011. Evaluation of

neural network, nearly equals 87% and R 2 resulted from

the application of artificial neural networks on

sediment rating curve is nearly 67%, which shows the

drought prediction in Yazd (Iran), Journal of Desert,

relative superiority of neural network compared to the

16 (1): 39- 49.

rating curve. In fact, predicting and estimating this

5- Dastorani, M. T., Kh. Azimi Fashi, A. Talebi, and M.

phenomenon is difficult and sometimes inaccurate.

R. Ekhtesasi, 2012. Estimating Suspended Sediment

Since, in predicting the amount of sediments in the rivers

Using Artificial Neural Network (A Case Study:

different factors involve and in addition, it is a of a

Jamishan Catchment Area, Kermanshah Province),

complicated nature: many problems related to suspended

Journal of watershed management, N. 6.

sediments in rivers could be solved through this method,

6- Dehghani, A., M. A. Zanganeh, and N. Koohestani,

since artificial intelligent methods are used to analyze

2009, Comparing the Estimation of Suspended Load

those problems for which there is not enough knowledge

Using Two Methods of Sediment Rating Curve and

or clear description. Certainly, it should be considered

Artificial Neural Network, (A Case Study: Doogh

that neural network could not predict the peak points and

River, Golestan Province), Journal of Agricultural

this is considered as one of its weak points.

Sciences and Natural Resources, 16 (1-A): 1-12

7- Emami, S. A., 2000. Sediment Transfer, Jahad

Acknowledgement

Daneshgahi Publications, Amir Kabir Technical

The authors gratefully acknowledge the Lorestan

University, Tehran, First Edition, 716 Pages.

Regional Water Company for data contribution.

8- Jain S.K., (2001). Development of integrated sediment

rating curves using ANNs. Journal of Hydraulic

Engineering ASCE 127(1):30- 37.

References

9- Kisi

O.,

2004.

Multi-layer

perceptrons

with

Levenberg-Marquardt optimization algorithm for

1- Abbasi Shooshtari, S., and M. Kashefi Poor, 2006.

suspended sediment concentration prediction and

Estimating the Suspended Sediment Using Artificial

estimation.

Hydrological

Science

Journal

Neural Network (A Case Study: Ahvaz Station),

49(6):1025 – 1040.

221

Journal of Environmental Treatment Techniques

2015, Volume 3, Issue 4, Pages: 215-222

10- Kisi O., 2005. Suspended sediment estimation using

neuro-fuzzy

and

neural

network

approaches.

Hydrological Science Journal 50(4):683 – 696.

11- Mirabolghasemi,

H.,

and

S.

Morid,

1995.

Investigation

of Hydrological

Approaches

for

Estimation of Suspended Load in Rivers, Journal of

Water and Development, 3: 3. 54-67.

12- Mosa’edi, A., S. F. Hashemi Najafi, M. Heydarnejad,

M. Nabizadeh, and M. A. Meshkati, 2009.

Estimating the Sedimentation Volume in Karaj and

Dez Dams’ Reservoir and Comparing it With

Hydrographic Results, Journal of Agricultural

Sciences and Natural Resources, 16 (2): 1-13

13- Murat, A., and H. K. Cigizoglu, 2007. Suspended

Sediment Load Simulation by Two Artificial Neural

Network Methods Using Hydro-meteorological Data,

Environ Model Soft, 22: 2-13.

14- Melesse, A.M., S. Ahmand, M.E. McClain, X. Wang

and Y.H. Lim, 2011. Suspended sediment load

prediction of river systems: An artificial neural

network approach, Agricultural Water Management,

98: 855-866.

15- Tahmoureth, M., Ahmadi, H., Taghavi, N., and

Askari, H. M., 2009. Comparing the Accuracy of

Geomorphological

Artificial

Neural

Network

(GANNs) Models With Regression Models (RM) In

Estimating Sediments Of Talghan River, watershed

studies, 84: 19-27

16- Walling, D.E., and B. W. Webb, 1981. The

Reliability of Suspended Sediment Load Data in

Erosion And Sediment Transport, Proceeding of

Florence Congress, IAHS Pub. 133: 177-194.

17- Zhu, Y.M., X.X. Lu, and Y. Zhou, 2007. Suspended

Sediment Flux Modeling With Artificial Neural

Network: an Example of the Longchuanjiang River

in

the

Upper

Yangtze

Catchment,

China.

Geomorphology, 84: 111-125.

222