You are on page 1of 45

Accepted Manuscript

Past, present and prospect of an Artificial Intelligence (AI) based model for
sediment transport prediction

Haitham Abdulmohsin Afan, Ahmed El-shafie, Wan Hanna Melini Wan


Mohtar, Zaher Mundher Yaseen

PII: S0022-1694(16)30478-4
DOI: http://dx.doi.org/10.1016/j.jhydrol.2016.07.048
Reference: HYDROL 21434

To appear in: Journal of Hydrology

Received Date: 24 April 2016


Revised Date: 3 July 2016
Accepted Date: 30 July 2016

Please cite this article as: Abdulmohsin Afan, H., El-shafie, A., Hanna Melini Wan Mohtar, W., Mundher Yaseen,
Z., Past, present and prospect of an Artificial Intelligence (AI) based model for sediment transport prediction, Journal
of Hydrology (2016), doi: http://dx.doi.org/10.1016/j.jhydrol.2016.07.048

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Past, present and prospect of an Artificial Intelligence (AI) based model for sediment

transport prediction

Haitham Abdulmohsin Afan* 1, Ahmed El-shafie1,2 , Wan Hanna Melini Wan Mohtar1, Zaher

Mundher Yaseen1
1
Civil and Structural Engineering Department, Faculty of Engineering and Built Environment,

Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia

2
Civil Engineering Department, Faculty of Engineering, University of Malaya, 50603 Kuala
Lumpur, Malaysia

*Corresponding author

Abstract

An accurate model for sediment prediction is a priority for all hydrological researchers. Many

conventional methods have shown an inability to achieve an accurate prediction of suspended

sediment. These methods are unable to understand the behaviour of sediment transport in rivers

due to the complexity, noise, non-stationarity, and dynamism of the sediment pattern. In the past

two decades, Artificial Intelligence (AI) and computational approaches have become a

remarkable tool for developing an accurate model. These approaches are considered a powerful

tool for solving any non-linear model, as they can deal easily with a large number of data and

sophisticated models. This paper is a review of all AI approaches that have been applied in

sediment modeling. The current research focuses on the development of AI application in

sediment transport. In addition, the review identifies major challenges and opportunities for

prospective research. Throughout the literature, complementary models superior to classical

modeling.

Keywords: Artificial Intelligence, sediment transport prediction, complementary model, review.


2

1 Introduction

Suspended sediment in water streams can act as a physical pollutant by increasing turbidity or as

a chemical pollutant, being the primary carrier of adsorbed chemicals, particularly for finer size

particles (Doğan et al., 2007). The estimation of erosion and transported sediment is accelerated

due to mostly anthropogenic activities such as deforestation, poor agricultural practices and

massive development. The transportation of sediment to and through the river system consists of

a number of complex phenomena including fluid-sediment interaction and the characteristics of

both flow and sediment should be taken into account. Sediment is transported from its original

position as incipient sediment motion and is kept in suspension due to turbulence. At all stages of

the transportation process, from the initial detachment of sediment particles to the arrival of

sediment at a point of deposition, a combination of highly non-linear and interacting processes

contributes to this complexity. There are several laborious field works in the estimation and

behaviour of individual components, either incipient motion or suspended sediment behaviour.

Thus, the accurate prediction of the total suspended sediment flux, taking into account the

combination of all components, is highly complicated. Despite the empirical formula established

by several researchers, for example, Lane-Kalinske (1941) and Bagnold (1966), it is extremely

difficult for an engineer to select an appropriate equation for specific case study. Furthermore,

none of these equations has gained universal acceptance for suspended sediment prediction.

Thus, an accurate forecasting of the river suspended sediment load, based on local data, is a

striking alternative and possibly plays a significant role in enhancing the prediction of suspended

sediment for specific cases. Accurate prediction of sediment load would be beneficial and useful

for a wide range of hydraulics and engineering designs such as (i) sediment transport in the river,

(ii) design of the dead storage of a dam, (iii) design of stable channels, (iv) estimation of
3

aggradation and degradation at bridge piers, (v) prediction of sand and gravel mining effects on

river bed equilibrium, (vi) determination of environmental impact assessment, and (vii) dredging

needs, to name a few.

Most recently, the AI and computational methods have shown the ability to overcome the

complicity of estimating the high complexity, dynamism, and non-stationarity of suspended

sediment loads. Artificial Intelligence approaches are characterized by having the ability to solve

non-linear models and easily handling the large number of data and sophisticated models. As is

well known, sediment transport computing is one of the most difficult problems in water

resources engineering. The estimation of sediment load does not only rely on water discharge in

rivers, but also on many characteristics of the river itself and these characteristics experience

temporal changes. In recent years, there has been noticed a significant rise in the number of data

driven approaches applied to hydrologic modeling, in particular computing the suspended

sediment. The data-driven modelling by artificial intelligence has demonstrated an aptitude to

solve and handle complexity and noise data problems (Nourani et al., 2014). Specifically, in the

field of sediment prediction and forecasting, there are many AI techniques and methods that have

been used such as artificial neural network (ANN), support vector machine (SVM), fuzzy logic

(FL), and numerous search optimization and statistical learning methods.

In summary, due to the continuous development of Artificial Intelligence and its

application to water resources engineering, specifically in the estimation of suspended sediment,

a thorough review of the research into all data-driven applications is needed. The main focus was

on relevant research studies undertaken within the past two decades. There has in recent years

been a remarkable increase in sediment-related publications, as shown in Figure 1. This shows

that AI has a high potential for estimating localized suspended sediment characteristics based on
4

individual water stream behaviour. As such, this review aimed to introduce how effectively the

applied AI methods have accomplished proper predictive suspended sediment models,

particularly in relation to the external or internal structure of the AI models, input variables and

the efficiency of AI integrated modelling. This paper is structured as follows: section 2 describes

the comprehensive ANN application in suspended sediment modeling and the application of the

FL method is presented in section 3. The application of SVM and the evolutionary computing

method is thoroughly described in section 4. Complementary modeling, which is the Wavelet-AI

model, is reviewed in section 5. The evaluation and assessment of previous studies is presented

in section 6. Recommendations for future research are discussed in section 7 and the last section

concludes the article.

(Figure 1)

2 Artificial Neural Network modeling

Artificial neural networks are data processing, modeling techniques which are generally used for

estimation, forecasting, pattern recognition, optimization, and to establish relationships between

complex featured variables. Haykin and Lippmann (1994) defined ANN as a massively parallel

distributed information processing system that has certain performance characteristics

resembling biological neural networks of the human brain. ANNs have been developed as a

generalization of mathematical models of human cognition or neural biology based on the

following rules described by Govindaraju (2000): (i) Information processing occurs at many

single elements called neurons; (ii) Signals are passed between nodes through connecting links;

(iii) Each connection link has an associated weight that represents its connection strength; (iv)
5

Each node typically applies a nonlinear transformation called an activation function to its net

input to determine its output signal.

The artificial neural network (ANN) has the ability to learn the exact behaviour between

the inputs and outputs from the examples without any kind of physical involvement. Also, it has

a superior characteristic of being able to extract the exact pattern between the input and output

variables without any additional explanation (Mustafa et al., 2012). ANN models have been

successfully applied to many tasks in hydrology, hydraulics and water resources management

such as flood forecasting, groundwater level forecasting and rainfall-runoff prediction (Abrahart

et al., 2004; Campolo et al., 2003; Chang et al., 2014a; Daliakopoulos et al., 2005; El-Shafie et

al., 2011b; El-Shafie et al., 2012; Hsu et al., 1995; Noori and Kalin, 2016; Pan et al., 2013;

Yaseen et al.) . The use of ANN is extended to the suspended sediment estimation, thus there are

a lot of studies that have been applied in sediment transport prediction (Afan et al., 2014;

Ardıçlıoğlu et al., 2007; Bayram et al., 2012; Bayram et al., 2013; Heng and Suetsugi, 2013; Isa,

2014; Jain, 2001; Kisi, 2004; Kisi, 2008; Melesse et al., 2011; Mustafa et al., 2012; Nourani et

al., 2012; Raghuwanshi et al., 2006; Rai and Mathur, 2007; Singh et al., 2013; Sivakumar and

Wallender, 2005; Talebizadeh et al., 2009; Tayfur, 2002; van Maanen et al., 2010). The

performance of these AI models generally are satisfactory and they have been able to correctly

predict the suspended sediment discharge concentration by up to 90%. However, we found that

the obvious variations are in terms of input variables and time scaling used in the models.

Therefore, this article puts an emphasis on the two main aspects mostly considered by previous

studies.
6

2.1 Types of Input Data Driven Models

Any data driven model can be defined on the basis of the connection between the system state

variables (i.e. input, internal, output variables) (Solomatine, 2005). The input layer, the most

important part of the model system, is often manipulated to enhance the performance of the

modeling process. In the field of sediment transport prediction in general there are many

parameters posed as having direct effects on the outcome of sediment discharge or concentration

estimation. Therefore, the researchers employed and tested the effects of input variable variation

in the performance of ANN predictive models. Figure 2 illustrates the structure of neural

network and the common input variables used to predict the suspended sediment load and

concentrations.

(Figure 2)

The first use of AI in the estimation of suspended sediment in rivers was done by Jain

(2001). He developed a conventional sediment rating curve model by using the feed forward

neural network. The study established an integrated stage-discharge-sediment concentration

relationship for two sites located on the Mississippi River. The back propagation based on the

generalized delta was used as the training algorithm. Water stage, discharge, and sediment

concentration were used as input data while the output data were the discharge and sediment

concentrations. The number of hidden layers was determined by trial and error. The performance

of FFNN was compared with the conventional curve fitting, where the AI technique was found to

have superior results in comparison to the conventional method. Despite limited data and not

considering the important parameters in sediment transport behaviour such as tractive shear
7

stress and velocity ration, to name a few, the ANN model outperformed the conventional method

and produced good results. As such, due to the limitation in data availability, developing the

sediment transport prediction model based on minimal input is essential and necessary.

Nagy et al. (2002) employed a multilayer feed forward neural network with a back

propagation training algorithm to predict the sediment load in three rivers of the Niobrara River,

the Middle Loup River and the Hii River. Eight input variables of tractive shear stress, velocity

ratio, suspension parameter, longitudinal slope, water depth ratio, Froude number, Reynolds

number and stream width ratio were used where the sediment concentration was the output.

Comparing the results from the artificial neural network with seven popular equations in the

sediment load concentration estimation of Engelund and Hansen (1967), Ackers and White

(1973), Yang (1996), Brownlie (1981), Shen and Hung (1972) and Laursen and Toffaleti (1969))

found that the neural network produced satisfactory results. Also, they concluded that the

conventional equations cannot accomplish an accurate result due to the uncertainty and the

stochastic nature of the sediment transport. Once again, this study relied on several common

parameters as the input pattern for the proposed model. The estimation of sediment transport

behaviour, particularly the suspended sediment, is localized to each specific river. Using similar

successful modelling profiles in another river does not guarantee high certainty of prediction.

The type of input variable is considered the main component necessary to develop an

accurate sediment model, another part of which is important for the researchers to manipulate is

the data partitions. Soft splitting of the data set is an effective technique to improve the

predictive model where the model will be based on sub-groups rather than one group. Kerem

Cigizoglu and Kisi (2006) proposed a method to improve the performance of the neural network

to estimate suspended sediment in rivers. Daily sediment and flow data for the Schuylkill River
8

in the United States were used for a case study. The K-fold partitioning method was used to

divide the data into three sub-groups. Auto-correlation as the initial statistical analysis was

employed to determine the adequate numbers of input data. The data were divided into three

groups and each group was trained individually. The range-dependent neural network (RDNN)

using FFBP was utilized to train and test the data. The study deduced that the range-dependent

neural network (RDNN) produced satisfactory results compared with the traditional ANN

applications where a single network is trained, considering the entire training data set. Actually,

utilizing the K-fold method as one of the cross-validation techniques is a vital step in developing

the ANN model. Cross-validation techniques are usually utilized to assure a similar level of

accuracy when switching from training to the testing session, commonly used for long time

series data. In some cases the cross-validation is eliminated from modelling because the data are

few, such as 3 years or less.

Alp and Cigizoglu (2007) investigated the estimation and forecasting of daily suspended

sediment load with feed-forward back-propagation and radial basis function using the previous

daily rainfall, sediment load and river flow data. They prepared the input layer using three

different combinations of hydro-meteorological data. It was deduced that the learning rate

directly affects the performance of the algorithm in the training stage. Alp and Cigizoglu (2007)

concluded that only rainfall data as input produced unsatisfactory results up to difference than

what was obtained when the flow data was added to the input variables. Also, they found that

FFNN and RBF produced close results when compared with each other.

Rai and Mathur (2007) developed a model for the computation of event-based temporal

variation of sediment yield from the watersheds. The feed forward back propagation neural

network was employed in their study. They used gradient descent algorithm with automated
9

Bayesian regularization in the training. Rainfall, runoff and sediment discharge data were

employed as the input pattern variables. They analyzed the data by using the autocorrelation and

cross-correlation analysis to select the optimal input variables. The results of prediction from the

ANN model were compared with the linear transfer function model. Using multiple performance

criteria such as Nash efficiency (CE), error in time to peak (ETP), error in peak sediment flow

rate (EPS) and error in total sediment yield (ESY), the performance of the ANN model was

found to be better than the linear transfer function model.

Multi-Layer Perceptron with back propagation algorithm was used by Jothiprakash and

Garg (2009) to estimate the volume of sediment retained in a reservoir. Annual rainfall, annual

inflow, and capacity of the reservoir were chosen as input variables. They compared the results

from ANN with conventional regression analysis and concluded that ANN produced better

results to estimate the volume of sediment retained in the reservoir.

In light of the above, it could be depicted that most of the proposed sediment prediction

models rely on the advantages of utilizing the neural network methods. The major advantages of

the neural network model are its ability to synthesize algorithms through a learning process and

no prior knowledge of the mathematical details between the input the output. However, there are

several challenges while utilizing ANN methods especially for sediment related phenomena.

The ANN model is localized and must be individually trained for each case. Moreover,

several trial and error procedures must be conducted to define the adequate architecture. In

addition, the training requirement involves large volumes of data. Special attention should be

given to the ANN model in case there is a need to make a certain modification to it, such as

adding more input parameters.


10

2.2 Time increment scale

It is known that the prediction of sediment transport phenomena depends on the time series input

variables, where the time series may be presented in several scales such as daily, weekly, ten-

day, monthly, and yearly. The multi-time scale prediction model is considered one of the

important comparative studies in sediment modeling. Agarwal et al. (2005) developed daily,

weekly, ten-day, and monthly rainfall-runoff-sediment (RQS) yield models based on the

multilayer back prorogation neural network (MLBPNN) and linear transfer function. They

applied two different learning processes (Batch-learning and Pattern-learning) to compare the

results for each model. It was found that not only do the results obtained from the pattern-learned

models have better results than the batch-learned models but the time-scale variation proved to

be insignificant and consistently produced better efficiency than the output from the linear

transfer function.

The capability of multi-layer perceptron neural network to estimate and forecast the

suspended sediments in rivers has been investigated by Cigizoglu (2004). Daily mean flow and

suspended sediment data of 29 years (from 1952 to 1981) were used for two stations on the

upstream and downstream of Schuylkill River. Two different sets of input data were applied,

where in the first set, the downstream and upstream sediment data were used as input to forecast

the one day ahead of suspended sediment discharge. The second set was the estimation of

downstream suspended sediment data by using the flow data as input. The correlation between

the flow and sediment data was analysed using the statistical analysis of mean, standard

deviation, coefficient of variation, skewness coefficient, first order autocorrelation coefficient,

overall minimum, and overall maximum. The statistical analysis can give a better expectation for

the performance of the model. Cigizoglu (2004) concluded that the data with higher cross-
11

correlation and lower skewness can give more satisfactory results (R2 value increased nearly

100% and MSE decreased 35%). Cigizoglu (2004) used conventional methods to compare the

performance of MLP. This study is rather more unique than other studies as the researcher

utilized data taken at both upstream and downstream of the same river. Analysis found that the

use of upstream sediment data gives more accurate results than using the input from downstream.

The MLP model was capable of capturing the non-linearity between the flow and suspended

sediment load, which was evidently displayed compared with the conventional sediment rating

curve.

Melesse et al. (2011) studied the suspended sediment prediction in the rivers by using

artificial neural network with back propagation algorithm. The performance of several AI

techniques such as multiple linear regressions (MLR), multiple non-linear regressions (MNLR)

and autoregressive integrated moving average (ARIMA) were tested. Daily and weekly data over

five years for three big rivers, i.e. Mississippi (1971-1975), Missouri and Rio Grande (1977-

1981) in the USA, were utilized to predict the suspended sediment load. Rainfall, discharge,

antecedent discharge and antecedent sediment load were used as input parameters. The

performance of ANN was assessed in different input combinations of data with a different time

period (i.e. 4, 3 and 2 years of training and 1, 2 and 3 years of testing, respectively). The

comparison of results showed that the performance of three years of training and two years of

testing produced better results than others. The study concluded that the prediction of daily data

was more accurate than weekly, and the performance of ANN for all input combinations was

consistently better than the results from MLR, MNLR and ARIMA.

Cigizoglu and Alp (2006) employed two neural network algorithms to estimate the

suspended sediment by using flow and sediment data. Generalized Regression Neural network
12

(GRNN) and FFNN were utilized and compared with the SRC and Multi Linear Regression

(MLR) methods. The generalized regression neural network was capable of handling the

problems of the frequently encountered local minima. Furthermore, the generalized regression

does not require an iterative training procedure which reduces the time consumption for training

the model. Based on the performance criteria used (i.e. RMSE, MAE and E%), the performance

of GRNN and FFNN was found to produce significant results compared to MLR and SRC.

Antecedent value of input variables is a common technique known to give better

prediction. Kisi (2008) designed the artificial neural network to estimate the suspended sediment

load using three different algorithms (i.e. Levenberg–Marquardt (LM), Conjugate Gradient (CG)

and Gradient Descent (GD) inside the Multi-Layer back Prorogation (MLP). The antecedent

value of water discharge and suspended sediment were made as input variables. Not only was the

normalization data applied as a preprocessing step, statistical analyses were also conducted to

identify the effective input variables. The preprocessing data step increased the performance of

the algorithm and reduced the time of computing. Comparison between the results of the three

algorithms, LM and CG algorithms produced better results than the GD algorithm.

Singh et al. (2013) simulated the suspended sediment yield by utilizing two types of

artificial neural network, i.e. standard-back propagation (SBP) and the radial basis neural

network (RBNN). Data of monthly rainfall and discharge data during the rainy season for a

watershed were used. Four different combinations of rainfall and discharge, taking into account

the antecedent value of discharge, were conducted as input parameters. By incorporating the

rainfall data to support the commonly used discharge data obviously improved the results. Their

RBNN modelling was also found to have higher accuracy than the SBP.
13

Afan et al. (2014) estimated the daily sediment load by using two types of artificial neural

network algorithms of feed forward neural network (FFNN) and radial basis function (RBF). As

suggested by Singh et al. (2013), different input combinations of varying antecedent value for

both suspended sediment load and discharge were examined to get more accurate results. Eleven

years of daily suspended sediment load and discharge were employed to estimate the sediment

load whereas ten years of data was used for training and the rest was for testing. Two different

scenarios were used for prediction, where the first scenario predicted sediment load from the

antecedent value only of sediment load and the second scenario obtained the results by utilizing

the antecedent value of both sediment load and discharge. Combining flow data with the

sediment load data gave an accurate model to predict sediment load, echoing the results from

Singh et al. (2013). As for the ANN technique capability, the FFNN produced better

performance than the RB model for this particular case.

Based on the exhaustive study particularly on the ANN sediment prediction model, the

majority of studies applied antecedent discharge data whereas some of the studies used rainfall

data. Furthermore, the efficiency of the modelling was not only restricted to one ANN technique

only, instead the prediction capability was evaluated based on different methods of ANN. The

performance of each technique was also examined based on a different time scale of daily,

weekly, and yearly. The largest number of these studies deduced that the prediction at a daily

time scale is more precise than other time scales. Where the variation data value in a daily time

scale can give a better view of sediment transport behaviour in rivers, in addition the peak value

can also be seen in a daily scale. To give a better illustration of the variability of input variables

and time scales used, Table 1 summarizes some details of each study mentioned above. It is

noticeable that all studies (except one) have utilized a daily time scale in their work. For the
14

suspended sediment modelling using ANN, the most popular performance criteria were Mean

Square Error (MSE), Root Mean Square Error (RMSE), and coefficient of determination (R2).

Table 1 also shows that for the input variables, suspended sediment load (SSL) and discharge are

the most commonly used.

(Table 1 )

3 Adaptive Neuro Fuzzy Inference System sediment model

The architecture and the learning procedure for the fuzzy inference systems (FIS) was first

introduced by Jang (1993), where the neural network learning algorithm was applied to construct

a set of fuzzy IF-THEN rules with appropriate membership functions (MFs) from the specified

input-output pairs. The architecture of the fuzzy system consists of three main components, i.e.

Fuzzifier, Fuzzy data base and Defuzzifier. The main purpose of Fuzzifier is to transform the

data into vectors before using it inside the fuzzy data base, while the Defuzzifier converts the

vector to actual data. The fuzzy data base contains two partitions, which are the Fuzzy rule base

and the Inference system. Fuzzy rule base is simply defined by the conditional statement IF-

THEN. The fuzzy interface system is divided into three main types depending on the type of

interface operation of IF-THEN rules: Sugeno’s system, Tsukamoto’s system, and Mamdani’s

system. Mamdani’s system is more commonly used, while Sugeno’s system is more compact and

computationally efficient; the output is crisp without the time consuming setback and

mathematically intractable defuzzification operation (Takagi and Sugeno, 1985).

Adaptive Neuro Fuzzy Inference System (ANFIS) is commonly employed to solve

nonlinear problems in environmental and hydrology engineering applications such as rainfall,

inflow, and dam water level, (Chang et al., 2014b; El-Shafie et al., 2011a; El-Shafie et al., 2006;
15

Firat and Güngör, 2009; Firat et al., 2009; Hipni et al., 2013; Lohani et al., 2014; Noori et al.,

2013; Si et al., 2015; Yurdusev and Firat, 2009).

The earliest research of ANFIS in the estimation of sediment in rivers was by Kisi

(2005). He estimated the suspended sediment by using Neuro-fuzzy and artificial neural network

models. The input parameters used in this study were the antecedent values of sediment

concentrations and stream-flow. Daily suspended sediment and flow data were employed for two

stations to test the performance of neuro-fuzzy and neural network. The results were compared

not only between those two models, but the analysis was extended to the multi-linear regression

and sediment rating curve. This study established that the neuro-fuzzy model outperformed the

other approaches.

Lohani et al. (2007) investigated the potential of fuzzy modeling in predicting the

sediment concentration in rivers. A fuzzy logic algorithm was developed to estimate the

sediment concentration using sediment concentrations, water level and discharge data as the

input variables. The results were compared with the output obtained from other models of ANN

and SRC. In conclusion, the fuzzy model was able to capture the inherent non-linearity in the

river gauge, discharge and sediment relationship with remarkable accuracy. Also, the

comparative analysis of predictive ability in different ranges of flow proposed superiority of the

fuzzy modelling compared to the ANN.

Kisi et al. (2008) predicted the sediment load using the ANFIS method by utilizing the

daily suspended sediment load and water discharge. Five different methods of RBNN, FFNN,

GRNN, MLR, and SRC were employed and each result was compared with findings from

ANFIS. The results indicated that the neuro-fuzzy model, in general, produced better estimates
16

than the other five techniques. Also, among the neural network techniques, the RBNN was found

to perform better than the FFNN and GRNN. Similarly, Firat and Güngör (2009) estimated the

suspended sediment by using ANFIS with the same input variables of discharge and sediment

but based on the monthly time scale data. Both studies stated the accuracy of ANFIS in

forecasting total sediment was significantly better than the result obtained from ANN and MLR.

In general, all AI models have shown better performance than SRC.

Rajaee et al. (2009) estimated the sediment concentration by using ANFIS and compared

it with the output from MLR and SRC models. The antecedent value of daily suspended

sediment concentration (SSC) and discharge were used as input parameters. Two other studies

have also utilized ANFIS to predict the SSC but with different input combinations. Cobaner et

al. (2009) utilized rainfall data with the discharge and SSC while Demirci and Baltaci (2012)

used mean water temperature instead of rainfall to predict the SSC. Based on these three studies,

we can conclude that ANFIS has better performance compared with other methods of AI and

SRC. The comparison of input variables, time scale and performance criteria employed using the

ANFIS model is summarized in Table 2. It was noted that the daily data is the most often used

time scale, whereas the input variables consistently consist of flow discharge. RMSE is one of

the most popular performance criteria that has been used by most researchers where RMSE

measures the difference between the actual and predicted values.

(Table 2 )

By assessing the performance of ANFIS compared with other predictive models, it is

ascertained that ANFIS is a better option due to its better learning ability for a similar network

complexity and a much smaller convergence error is achieved. Although ANFIS is capable of

highly nonlinear mapping, and is far superior to MPL and other common linear methods of
17

similar complexity, ANFIS often presents a sudden convergence preceded by a region of

considerable instability. ANFIS still requires adjustable parameters that should be assumed and

achieved by using trial and error methods. It is very complicated to integrate ANFIS with

optimization methods to identify the optimal values of these parameters.

4 Other AI and evolutionary methods

In this section, a review of the individual methods that have been applied in the sediment

transport prediction will be presented. There are many optimization algorithms and methods that

have been examined by researchers, where the common goal is to find the optimal solution. A

few of these algorithms were employed in the suspended sediment prediction such as linear

genetic programming (LGP), genetic programming (GP), gene expression programming (GEP),

support vector machine (SVM), least square support vector machine (LSSVM), neural different

evolution (NDE), boosted regression tree (BRT), principle component analysis (PCA), and M5

Model Tree Algorithms. Table 3 below illustrates the details of the studies that have been applied

in sediment estimation by using the other evolutionary methods.

Genetic programming (GP) is an evolutionary algorithm inspired by biological evolution

to find computer programs that perform a user-defined task. It is a method for evolving equations

by taking various mathematical building blocks such as functions, constants and arithmetic

operations and combining them into a single expression and this method was developed by Koza

(1992). LGP and GEP are the extensions of the conventional tree-based GP. The first study using

the evolutionary method in sediment estimation was conducted by Guven and Kişi (2010). They

investigated the use of a machine-coded linear genetic programming (LGP) in the daily

suspended sediment load estimation. LGP, GP, ANN, and GEP were employed to evaluate the
18

performance of each method and a comparison between each technique was made. They

concluded that LGP showed better performance than other methods, especially in the estimation

of the output (i.e. the peak values of suspended sediment load). The GEP was also found to be

better than the ANN.

Table 3 presents many types of individual studies that have been applied in sediment

transport prediction. Due to the low number of applications using these AI techniques in

sediment modeling, they are grouped together in this section. Two types of input variables are

shown in Table 3: time scale and non-time scale. It is noticeable that a daily time scale is more

applicable while only one study (Goyal, 2014) applied a monthly time scale. The most popular

input that have been utilized are SSL, SSC and discharge which are also commonly used in most

ANN and Fuzzy sediment models.

(Table 3)

Ab. Ghani and Azamathulla (2012) adopted GEP to predict the sediment load transport in

the Kurau, Langat and Muda rivers in Malaysia with ten different input parameters of flow

velocity, flow depth, discharge, suspended load, bed load, water surface slope, width, total bed

material load, mean sediment size, and Manning value. The results obtained from the GEP model

were compared with the conventional rating curve, where it was established that the GEP has

better performance. Kisi and Shiri (2012) estimated the daily suspended sediment concentration

using hydro-metrological data (rainfall, streamflow) using GEP, ANN and ANFIS. The results

showed that the GEP method had better performance than other AI techniques. Similar results

were also found by Kisi et al. (2012), who applied the GP technique to estimate the daily

suspended sediment load based on the daily flow and sediment data from 1972 to 1989. Their
19

work also deduced that utilizing GP produced more accurate daily SSL estimation than ANFIS,

ANN and SVM models. Typical recent studies did not only evaluate the performance of one

model, but instead sought the best option by comparing it with several AI techniques.

Genetic Algorithm is a popular evolutionary method and in spite of good performance it

still has some drawbacks. GAs are particularly notorious for their long training times.

Furthermore, since some of the solutions can be extremely complicated, another drawback is that

it can be too slow for real-time applications. One of the major drawbacks of utilizing GA for

even non-real-time application is random convergence of solutions, regardless of the fitness

function used in it. Besides, the time consumed by the optimization algorithm is also high since it

involves lots of parameters.

The identical rudimentary disadvantage of GA is its unguided metamorphosis and

mutation. The mutation operator in GA functions usually engenders a randomly generated

number to a parameter of an individual of the population. As a result, a very slow convergence of

genetic algorithm is experienced. In fact, this issue could be solved by combining it with other

algorithms which perform guided search such as differential evolution. However, the efficiency

of this method is yet to be asserted with confidence in any sedimentation prediction model.

Support vector machines (SVM) is one of the AI methods that is based on the extension

of the idea of identifying a line, a plane or some surface that separates two classes in the

classification. It is based on the statistical learning theory initiated by Vapnik in the 1970s

(Vapnik and Vapnik, 1998). There are many studies employing SVM in the sediment transport

estimation, but most of these studies have shown that the performance of SVM is worse than

other AI methods such as ANN or GP (Haji et al., 2013; Kakaei Lafdani et al., 2013; Nourani
20

and Andalib, 2015a). However, the work of Kisi (2012) which estimated the suspended sediment

concentration found out LSSVM output is more accurate compared to ANN and SRC.

Tayfur et al. (2013) studied the sediment load prediction and generalization from

laboratory scale to field scale using the principle component analysis (PCA). Another two

methods (ANN and GP) were also applied for comparison purposes. Five parameters, i.e.

channel width, relative roughness, sediment size, geometric standard, and sediment concentration

were made as input variables. In their work, PCA essentially produced better performance than

ANN and GA for estimation of both total load and SSL. They deduced that ANN and GA

produced a more accurate prediction of total load than SSL.

From this exhaustive literature review, we found that all prediction model techniques

have advantages and disadvantages, which are more or less vital according to the data under

study and analysed. Generally, SVM could be considered a useful tool for insolvency analysis

and cases with non-regularity data. The attractiveness of SVM as a prediction method is due to

its four main advantages: firstly, it has a regularisation parameter to overcome the problem of

overfitting which is common in ANN and ANSIF modelling. Secondly, SVM is defined by a

convex optimisation problem which overcomes the problem of having local minima by efficient

methods. With the selection of an appropriate kernel, such as the Gaussian kernel, SVMs provide

a good out-of-sample generalization. This means that, by picking a suitable generalization

evaluation values, SVMs can be robust, even when in the training phase the model’s sampling

data experienced a certain level of bias.

Despite the substantial superiority of SVM, a common disadvantage of this technique is

the lack of transparency of results due to an excessively high dimension. The first limitation of
21

the SVM methodology is its wide range of choice of the kernel values. An important practical

question that is not entirely solved is the selection of the kernel function parameters for Gaussian

kernels and the value of insensitive loss function. A second limitation is both training and testing

sessions are time consuming which is not convinient for real time application. Although SVMs

have good generalization, they can be abysmally slow during the test phase. In addition, from a

practical point of view perhaps the most serious problem with SVMs is the high algorithmic

complexity and extensive memory requirements of the required quadratic programming in large-

scale applications.

5 Wavelet–AI Integrated model

Recently, considerable attention has been given to the data preprocessing for the model input-

output patterns. This is due to the fact that data preprocessing might help for special feature

extraction from the data before further processing. Data preprocessing designates most of the

processing performed on raw data to reformulate and/or re-shape it in a more suitable fashion for

further processing procedures. Actually, data preprocessing transforms the data into a different

shape and style that will be more straightforwardly and successfully processed for the purpose of

the user, especially for an artificial intelligence model. There are several methods and techniques

used for preprocessing such as sampling, transformation, de-noising and normalization.

Sampling is basically concerned with focusing on a representative subset from a large population

of data. Transformation which handles the raw data to create the input variable is based on

frequent manner. In addition, a de-noising method is used to eliminate noise to attain much

clearer data in order to achieve a higher accuracy in the model output. Normalization organizes

data for more efficient access and feature extraction, which is necessary and essential for
22

developing the prediction and forecasting models in order to detect mapping between the input-

output of the data.

In reality, the sediment transport time series can be viewed as a combination of two

components, one is stochastic (or random) and the other is the deterministic components. Once

the stochastic (noise) component is appropriately eliminated, the deterministic component can

then be easily modeled. For the purpose of cleaning up the sediment time series, many data

preprocessing techniques, including Principal Component Analysis (PCA), singular spectrum

analysis (SSA) and wavelet analysis (WA) have been employed. In fact, most of the recent

studies on developing a sediment prediction model were focused on utilizing a wavelet method

as one of the advanced data preprocessing techniques.

Wavelet transformation is one of the most powerful tools in the time-frequency

transformations in recent years. Although the wavelet method was based on the Fourier

transform, it was found to do better than Fourier transform in studying the measured non-

stationary time series, transient phenomenon like hydrological processes (Polikar, 1999). The

wavelet transform method breaks up the signals into wavelets, which are the waveforms of

efficiently limited duration and zero mean. There are two types of wavelet transforming:

continuous wavelet transform (CWT) and discrete wavelet transform (DWT). In the hydrological

related phenomena, the discrete wavelet transform is most used by researchers because they need

to access a discrete time signal, rather than a continuous time signal. Wavelet transform deals

with time localization and frequency localization, which allows running different window sizes

depending on the scale of frequency. Wavelet has the ability to reduce the noise by decomposing

the time series to sub series. Many researchers who applied the DWT in the sediment transport

prediction in rivers produced improved results compared with the single AI models (Haji et al.,
23

2013; Kakaei Lafdani et al., 2013; Kişi, 2009; Liu et al., 2013; Mirbagheri et al., 2010; Nourani

and Andalib, 2015a; Nourani and Andalib, 2015b; Partal and Cigizoglu, 2008; Rajaee, 2010;

Rajaee et al., 2010a; Rajaee et al., 2010b; Ramezani et al., 2014; Sadeghpour Haji et al., 2014;

Shiri and Kişi, 2012).

The first application of wavelet in the estimation of suspended sediment was conducted

by Partal and Cigizoglu (2008). They estimated and predicted the suspended sediment load in

rivers by a combined wavelet–ANN method. In the first part of their study, they predicted the

suspended sediment by using sediment load data and in the second part they utilized the

sediment and flow data to estimate the suspended sediment. The discrete wavelet transform was

employed to decompose the data sum up and served as an input for the ANN model. Results

showed more accurate predictions and the wavelet-ANN model performed better than the

conventional ANN model and the SRC model. The result illustrates that Wavelet-ANN could

capture better approximation for the peak values.

By combining two methods of ANN and discrete wavelet transform, Kişi (2009)

estimated the daily suspended sediment load. Daily stream flow and suspended sediment data

were used in different combinations as input variables. The work concluded that not only was the

developed neuro-wavelet model capable of estimating the suspended sediment to a high degree

of accuracy, but it also has better performance than the conventional ANN.

Rajaee (2010) employed neuro-fuzzy (NF), a combination of wavelet transforms and

neuro-fuzzy (WNF), multi-linear regression (MLR), and the conventional sediment rating curve

(SRC) models to predict the suspended sediment load. The actual time series of river flow

discharge and sediment load were decomposed to sub time series at different scales by discrete
24

wavelet transform. After the decomposition, the effective sub-series was summed together and

added to the input of the neuro-fuzzy method to predict the daily sediment load. The results

illustrated that the performance of the WNF model was significantly better than the results

obtained from the NF, MLR and SRC models.

Shiri and Kişi (2012) estimated the suspended sediment load by using two different

model techniques. They employed GEP, NF and ANN techniques as a single model and another

type by using conjunction models (WGEP, WNF, and WANN) which are a combination of

wavelet and single models. Daily river flow and sediment load data were used as input variables

to estimate the sediment load. Comparison of the predicted output show that GEP is capable of

simulating suspended sediment values reasonably well and also performed better than other

single models in estimating the sediment load. The application of DWT with single AI models

significantly increases their accuracies in suspended sediment estimation.

Liu et al. (2013) designed a complementary model of the wavelet artificial neural

network to predict the suspended sediment concentrations (SSC). Daily SSC and water discharge

data of Yellow River catchments of China were used from 1967 to 1972. The discrete wavelet

transform and Daubechies wavelet of order 4 (Db4) were employed to decompose the data in

seven sub-series. From these sub-series, only the effective sub-series were chosen by using

cross-correlation. The selected subseries were obtained to be input variables to the feed-forward

back-propagation artificial neural network (BP ANN) for a 1 day ahead SSC prediction. The

selection of sub-series performed by cross-correlation analysis was summed to reconstruct noise-

free discharge and SSC time series. Comparison of the results obtained from WANN, ANN and

SRC found that the WANN model could predict the highly nonlinear and non-stationary SSC
25

time series more precisely than ANN or SRC. Removal of noise from the data by WANN led to

a noticeable improvement in the fit of the predicted and measured SSC values.

Haji et al. (2013) used a Wavelet-Support Vector Machine (WSVM) and SVM to predict

the daily suspended sediment load. Input parameters of stream flow and suspended sediment load

time series were decomposed into several multi-frequency time series by the wavelet transform.

The autocorrelation and cross-correlation were used to assess the best input combinations. The

WSVM model with a combination of two antecedent values of flow and suspended sediment

showed better results than other combinations of antecedent values. The study supports the usage

of WSVM and SVM models as a more accurate alternative to the traditional SRC.

Nourani and Andalib (2015a) examined the performance of a wavelet-based model in the

suspended sediment load prediction. They used two types of wavelet complementary models, i.e.

the wavelet based Least Square Support Vector Machine (WLSSVM) and Wavelet based

Artificial Neural Network (WANN). Flow discharge and suspended sediment load were

decomposed into a sub-series by wavelet and then summed together which was used as input for

both SVM and ANN models to predict the monthly suspended sediment load. Even though the

single model of ANN performed better than SVM, the integrated model of WLSSVM provided

more accurate results in comparison with WANN.

The major advantage of the wavelet over the traditional Fourier transform is that it allows

for representation of the signal in the base of wavelet functions. Each wavelet function is located

at a different spot along the time axis and its range depends on the range of applied scales.

However, Wavelet faces a major challenge regarding the choice of the threshold values that

should be first selected. This value is critical in identifying the level of de-noising in time series
26

data and has a strong influence on the final accuracy of the model. The hard threshold method

excludes all smaller values than the threshold, leaving other details unchanged. This point does

not receive sufficient discussion in the proposed models in the sediment model, which is

considered a vital issue in the prediction model utilizing wavelet as preprocessing method for the

time series.

6 Evaluation and assessment

The current review article presented the-state-of-the-art artificial intelligence approaches applied

in sediment transport modelling and prediction. The authors discussed the utility of the machine

learning techniques (i.e. ANN, SVM, ANFIS, EC and coupled AI models) and overviewed their

advantages and disadvantages. In addition, the simulation accuracy of the sediment predictive

model, which is based on the hydrological input variables, was investigated. We found that the

application of AI methods has increased since 2001 until recently, as shown in Figure 1. Studies

undertaken during these two decades of modelling enhancement have contributed to the

improvement of prediction performance. In light of the above literature, the following

conclusions are reached:

i. Theoretically, artificial neural network can be considered the top option, if not the

most popular method in comparison to other techniques. Although the ANN method

has some drawbacks and limitations (e.g. over-fitting, slow learning speed, local

minima, and difficulty capturing the high complexity, non-stationarity, dynamism and

non-linearity of time series), the performance of ANN in terms of modelling was

consistently satisfactory. We predict that with further studies to overcome the


27

mentioned drawbacks and limitation, the efficiency of ANN in the future is likely to

be improved.

ii. In the last five years, several studies effectively modeled the sediment transport

prediction using fuzzy logic and evolutionary computing techniques. Many authors

have studied these two approaches and remarkably established that their models

outperformed the conventional and ANN methods (Cobaner et al., 2009; El-Shafie et

al., 2011a; El-Shafie et al., 2006; Hipni et al., 2013; Kisi et al., 2012; Kisi and Shiri,

2012). In fact, fuzzy set is characterized by its capability to capture the non-linearity

relationship between the input and output hydrological parameters. Furthermore,

ANFIS has been found to sufficiently handle the uncertainties and vagueness of the

sediment transport in rivers. Whereas, the evolutionary computing methods have the

advantages of their skill to evolve solutions for the problem instead of giving a direct

fixed solution.

iii. Of all the time scales used in the reviewed papers, daily sediment transport in river

was the most commonly used parameter. Indeed, this time horizon is more applicable

than weekly or monthly time horizons due to the fact that daily sediment pattern

provides comprehensive information about the phenomena. Obviously, the

probability of accurate modelling is enhanced.

iv. Generally, suspended sediment prediction studies could be categorized into two major

groups, namely, time based (time domain) and case base studies. Actually, the

development of prediction models based time domain category relatively difficult to

be generalized and successfully applied in other case river case study, on the other
28

hand, the potential for generalization is higher for those models under case base

category. Indeed, this review research focused on the models under the category

of time domain studies where these studies are purposely developed for certain river

only utilizing its particular sediment time series data, such category of models could

be successful for similar rivers. While, the models under the case base category, the

prediction of suspended sediment is developed based on the utilization

of physical data from several rivers that has not time series data (e.g. Pektas, 2015;

Petals et al., 2015; Tayfur et al., 2013). Therefore, suspended sediment load (SSL)

and suspended sediment concentration (SSC) have been modelled and predicted

based on several types of hydrological parameters. Basically, the hydrological

variables are categorized into two types: time domain data, e.g. SSL, SSC, discharge,

rainfall, water level and water temperature; and non-time domain data which is used

in case base studies, e.g. flow velocity, flow depth, water surface slope, river width,

flow area, hydraulic radius and others. The majority of the studies utilized the time

series data of discharge and SSL in their modeling. This is expected, as discussed in

the hydrological concept, since streamflow is the most associated variable which

effects the amount and concentration of sediment in rivers. Even so, relying solely on

parameter discharge is not sufficient to guarantee accurate modelling and the

inclusion of more variables is needed to contemplate the complexity of flow-sediment

behaviour.

v. Most recently, there is an effective method that impressively deals with the multi-

scale and non-stationarity pattern of the time series, which is the wavelet transform

analysis. This technique solves the problem of dynamical behaviour of sediment


29

transformation in the river. It has the ability to de-noise the time series data in order

to improve the results of estimation. Throughout the review of the current research,

paired data-driven models with a wavelet preprocessing technique showed

outstanding results compared with the individual data-driven models (Kişi, 2009;

Nourani and Andalib, 2015a; Rajaee et al., 2010b; Shiri and Kişi, 2012). In addition,

the revision of wavelet transform application in sediment transport reveals that all

studies employed a discrete wavelet transform (DWT) method, while continuous

wavelet transform (CWT) has not yet been applied and needs to be investigated.

vi. Among the surveyed articles, it was found that models were evaluated based on

several evaluation indicators (i.e., R, RMSE, MSE, MAE, RE, NS and E). Between

all the mentioned indicators, R and RMSE were used as fundamental performance

criteria to assess the models. The researchers highlighted that “goodness-of-fit” and

one absolute error measures are recommended to be carried out for the hydrological

modelling and specifically for predictive models.

7 Recommendations for future research

From a thorough review of the literature, it can be seen that an enhancement for sediment

modelling is needed. Here, we suggest some recommendations and suggestions to develop a

complementary AI sediment model as shown in Figure (3):

1- After collecting the time series data for sediment or discharge, it is necessary to

preprocess the data before creating any estimation model. The purpose of preprocessing

is to remove unwanted variation. There are many common preprocessing methods, but

most of them involve the possibility of losing important data when deleting some of the
30

variation and this makes the problem more complicated. The two most recommended

methods in preprocessing data are the Orthogonal partial least squares (OPLS) and the

O2PLS (bidirectional OPLS). These two methods are novel extensions of the PLS

method (Trygg and Wold, 2002). The structure of O2PLS is a bidirectional model, and

this type of structure improves understanding of the studied data sets (Gabrielsson et al.,

2006). Other common alternative methods that can be used for preprocessing are

principal component regression (PCR), partial least squares (PLS), multiple linear

regression (MLR), ridge regression (RR), multiplicative signal correction (MSC),

standard normal variation (SNV), and orthogonal signal correction (OSC), to name a few.

2- Discharge is one of the most important parameters which has a direct effect on the

behaviour of transported sediment in the river. Due to the dynamism of the sediment

behaviour based on the changing of discharge in high and low discharges, it is

recommended to create two sub models based on the discharge values. Classifying the

data into two groups will keep the data value close enough and improve the accuracy of

prediction.

3- Denoising data is an important issue. Wavelet transform is one of the most popular

processing data tools, and many studies have utilized this tool to enhance the

performance of the AI model in the sediment transport estimation. The non-

stationary and complexity in the time series is considered as the major obstacles that face

the modelling utilizing soft computing methods for suspended sediment prediction. A

hybrid ANN-wavelet model has been proven that the decomposition of time series could

be consider as one the optimal way to deal with the non-stationary and complex time

series data and then enhance the performance of model in terms of the prediction
31

accuracy. One of the best alternatives and recommended methods for de-noising is fast

orthogonal search (FOS). FOS is a signal processing tool developed to provide a high-

resolution spectral estimation. In addition, it is a general non-linear modeling technique

that finds functional expansions using an arbitrary set of non-orthogonal candidate

functions. The proposed method has two significant advantages: firstly, FOS yields a

parsimonious sinusoidal series representation by firstly selecting the most significant

sinusoidal components, and secondly, the frequencies of the sinusoids selected need not

be commensurate nor integral multiples of the fundamental frequency corresponding to

the record length (Korenberg, 1989).

4- ANN is one of the best techniques that has been used widely in the sediment transport

estimation and accomplished good results compared with other classical methods such as

SRC. Employing the ANN method in the suggested complementary AI sediment model

will increase the performance of the model. Indeed, an improvement of back propagation

in ANN is needed. The issue with the gradient decent in ANN is that it sticks with local

minima or maxima due to the difficulty of searching for the optimal solution. This issue

can be solved by utilizing new evolutionary algorithms (EAs) that have the ability to

search for global minima or maxima. Numerous high potential algorithms such as

Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), Artificial Bee

Colony algorithm (ABC), Ant Lion Optimizer (ALO), Bacterial Colony Optimization

(BCO) and Gravitational Search Algorithm (GSA) are available, although their efficiency

in a sediment related modelling type is yet to be asserted with confidence.

(Figure 3)
32

8 Conclusions

In this paper, a thorough review of all the AI approaches, specifically applied in the estimation of

sediment transport, particularly suspended sediment, was conducted. AI methods have

demonstrated successful results to estimate the phenomena of sediment transport in rivers. In

recent years, the Artificial Intelligence and computation methods have shown the ability to tackle

the complexity of estimating suspended sediment. The artificial neural network, neuro-fuzzy,

Wavelet based model, and some evolutionary algorithms (EAs) are the most efficient methods

that have been applied in the sediment field. The uninterrupted development in AI, soft

computing and its application has increased competition between researchers to develop an

accurate model to overcome the complexity, non-stationarity and dynamism of the sediment

transport behaviour in rivers. Therefore, it is necessary to develop a structure of a

complementary model consisting of the most advanced and recent AI methods as recommended

in this paper.

Acknowledgement

This research was partially supported by Universiti Kebangsaan Malaysia, under the research

grant for the second author DIP-2015-012 and for the third author DIP-2015-006.

References

Ab. Ghani, A., Azamathulla, H.M., 2012. Development of GEP-based functional relationship for
sediment transport in tropical rivers. Neural Computing and Applications, 24(2): 271-
276. DOI:10.1007/s00521-012-1222-9
Abrahart, R., Kneale, P.E., See, L.M., 2004. Neural networks for hydrological modeling. CRC
Press.
33

Afan, H.A. et al., 2014. ANN Based Sediment Prediction Model Utilizing Different Input
Scenarios. Water Resources Management, 29(4): 1231-1245. DOI:10.1007/s11269-014-
0870-1
Agarwal, A., Singh, R., Mishra, S., Bhunya, P., 2005. ANN-based sediment yield models for
Vamsadhara river basin (India). Water SA, 31(1): p. 95-100.
Alp, M., Cigizoglu, H., 2007. Suspended sediment load simulation by two artificial neural
network methods using hydrometeorological data. Environmental Modelling & Software,
22(1): 2-13. DOI:10.1016/j.envsoft.2005.09.009
Ardıçlıoğlu, M., Kişi, Ö., Haktanır, T., 2007. Suspended sediment prediction using two different
feed-forward back-propagation algorithms. Canadian Journal of Civil Engineering, 34(1):
120-125. DOI:10.1139/l06-111
Bayram, A., Kankal, M., Onsoy, H., 2012. Estimation of suspended sediment concentration from
turbidity measurements using artificial neural networks. Environmental monitoring and
assessment, 184(7): 4355-65. DOI:10.1007/s10661-011-2269-2
Bayram, A., Kankal, M., Tayfur, G., Önsoy, H., 2013. Prediction of suspended sediment
concentration from water quality variables. Neural Computing and Applications, 24(5):
1079-1087. DOI:10.1007/s00521-012-1333-3
Campolo, M., Soldati, A., Andreussi, P., 2003. Artificial neural network approach to flood
forecasting in the River Arno. Hydrological Sciences Journal, 48(3): 381-398.
Chang, F.-J., Chen, P.-A., Lu, Y.-R., Huang, E., Chang, K.-Y., 2014a. Real-time multi-step-
ahead water level forecasting by recurrent neural networks for urban flood control.
Journal of Hydrology, 517: 836-846.
Chang, F.-J. et al., 2014b. Watershed rainfall forecasting using neuro-fuzzy networks with the
assimilation of multi-sensor information. Journal of Hydrology, 508: 374-384.
Cigizoglu, H.K., 2004. Estimation and forecasting of daily suspended sediment data by multi-
layer perceptrons. Advances in Water Resources, 27(2): 185-195.
DOI:10.1016/j.advwatres.2003.10.003
Cigizoglu, H.K., Alp, M., 2006. Generalized regression neural network in modelling river
sediment yield. Advances in Engineering Software, 37(2): 63-68.
DOI:10.1016/j.advengsoft.2005.05.002
Cobaner, M., Unal, B., Kisi, O., 2009. Suspended sediment concentration estimation by an
adaptive neuro-fuzzy and neural network approaches using hydro-meteorological data.
Journal of Hydrology, 367(1-2): 52-61. DOI:10.1016/j.jhydrol.2008.12.024
Daliakopoulos, I.N., Coulibaly, P., Tsanis, I.K., 2005. Groundwater level forecasting using
artificial neural networks. Journal of Hydrology, 309(1): 229-240.
Demirci, M., Baltaci, A., 2012. Prediction of suspended sediment in river using fuzzy logic and
multilinear regression approaches. Neural Computing and Applications, 23(S1): 145-151.
DOI:10.1007/s00521-012-1280-z
Doğan, E., Yüksel, İ., Kişi, Ö., 2007. Estimation of total sediment load concentration obtained by
experimental study using artificial neural networks. Environmental fluid mechanics, 7(4):
271-288.
El-Shafie, A., Jaafer, O., Akrami, S.A., 2011a. Adaptive neuro-fuzzy inference system based
model for rainfall forecasting in Klang River, Malaysia.
El-Shafie, A., Mukhlisin, M., Najah, A.A., Taha, M., 2011b. Performance of artificial neural
network and regression techniques for rainfall-runoff prediction. International Journal of
the Physical Sciences, 6(8): 1997-2003.
34

El-Shafie, A., Noureldin, A., Taha, M., Hussain, A., Mukhlisin, M., 2012. Dynamic versus static
neural network model for rainfall forecasting at Klang River Basin, Malaysia. Hydrology
and Earth System Sciences, 16(4): 1151-1169. DOI:10.5194/hess-16-1151-2012
El-Shafie, A., Taha, M.R., Noureldin, A., 2006. A neuro-fuzzy model for inflow forecasting of
the Nile river at Aswan high dam. Water Resources Management, 21(3): 533-556.
DOI:10.1007/s11269-006-9027-1
Firat, M., Güngör, M., 2009. Monthly total sediment forecasting using adaptive neuro fuzzy
inference system. Stochastic Environmental Research and Risk Assessment, 24(2): 259-
270. DOI:10.1007/s00477-009-0315-1
Firat, M., Turan, M.E., Yurdusev, M.A., 2009. Comparative analysis of fuzzy inference systems
for water consumption time series prediction. Journal of hydrology, 374(3): 235-241.
Gabrielsson, J. et al., 2006. The OPLS methodology for analysis of multi‐block batch process
data. Journal of chemometrics, 20(8‐10): 362-369.
Govindaraju, R.S., 2000. Artificial neural networks in hydrology. I: Preliminary concepts.
Journal of Hydrologic Engineering, 5(2): 115-123.
Goyal, M.K., 2014. Modeling of Sediment Yield Prediction Using M5 Model Tree Algorithm
and Wavelet Regression. Water Resources Management, 28(7): 1991-2003.
DOI:10.1007/s11269-014-0590-6
Guven, A., Kişi, Ö., 2010. Estimation of Suspended Sediment Yield in Natural Rivers Using
Machine-coded Linear Genetic Programming. Water Resources Management, 25(2): 691-
704. DOI:10.1007/s11269-010-9721-x
Haji, M.S., Mirbagheri, S., Javid, A., Khezri, M., Najafpour, G., 2013. A wavelet support vector
machine combination model for daily suspended sediment forecasting. International
Journal of Engineering-Transactions C: Aspects, 27(6): 855.
Haykin, S., Lippmann, R., 1994. Neural Networks, A Comprehensive Foundation. International
Journal of Neural Systems, 5(4): 363-364.
Heng, S., Suetsugi, T., 2013. Using Artificial Neural Network to Estimate Sediment Load in
Ungauged Catchments of the Tonle Sap River Basin, Cambodia. Journal of Water
Resource and Protection, 05(02): 111-123. DOI:10.4236/jwarp.2013.52013
Hipni, A. et al., 2013. Daily forecasting of dam water levels: comparing a support vector
machine (SVM) model with adaptive neuro fuzzy inference system (ANFIS). Water
resources management, 27(10): 3803-3823.
Hsu, K.l., Gupta, H.V., Sorooshian, S., 1995. Artificial Neural Network Modeling of the
Rainfall‐Runoff Process. Water resources research, 31(10): 2517-2530.
Isa, M.M.M., 2014. Comparative Study of MLP and RBF Neural Networks for Estimation of
Suspended Sediments in Pari River, Perak.
Jain, S.K., 2001. Development of integrated sediment rating curves using ANNs. Journal of
hydraulic engineering, 127(1): 30-37.
Jang, J.-S.R., 1993. ANFIS: adaptive-network-based fuzzy inference system. Systems, Man and
Cybernetics, IEEE Transactions on, 23(3): 665-685.
Jothiprakash, V., Garg, V., 2009. Reservoir sedimentation estimation using artificial neural
network. Journal of Hydrologic Engineering, 14(9): 1035-1040.
Kakaei Lafdani, E., Moghaddam Nia, A., Ahmadi, A., 2013. Daily suspended sediment load
prediction using artificial neural networks and support vector machines. Journal of
Hydrology, 478: 50-62. DOI:10.1016/j.jhydrol.2012.11.048
35

Kerem Cigizoglu, H., Kisi, Ö., 2006. Methods to improve the neural network performance in
suspended sediment estimation. Journal of Hydrology, 317(3-4): 221-238.
DOI:10.1016/j.jhydrol.2005.05.019
Kisi, O., 2005. Suspended sediment estimation using neuro-fuzzy and neural network
approaches/Estimation des matières en suspension par des approches neurofloues et à
base de réseau de neurones. Hydrological Sciences Journal, 50(4).
DOI:10.1623/hysj.2005.50.4.683
Kisi, O., 2012. Modeling discharge-suspended sediment relationship using least square support
vector machine. Journal of Hydrology, 456-457: 110-120.
DOI:10.1016/j.jhydrol.2012.06.019
Kisi, Ö., 2004. Multi-layer perceptrons with Levenberg-Marquardt training algorithm for
suspended sediment concentration prediction and estimation/Prévision et estimation de la
concentration en matières en suspension avec des perceptrons multi-couches et
l’algorithme d’apprentissage de Levenberg-Marquardt. Hydrological Sciences Journal,
49(6).
Kisi, Ö., 2008. Constructing neural network sediment estimation models using a data-driven
algorithm. Mathematics and Computers in Simulation, 79(1): 94-103.
DOI:10.1016/j.matcom.2007.10.005
Kişi, Ö., 2009. Daily suspended sediment estimation using neuro-wavelet models. International
Journal of Earth Sciences, 99(6): 1471-1482. DOI:10.1007/s00531-009-0460-2
Kişi, Ö., 2010. River suspended sediment concentration modeling using a neural differential
evolution approach. Journal of Hydrology, 389(1-2): 227-235.
DOI:10.1016/j.jhydrol.2010.06.003
Kisi, O., Dailr, A.H., Cimen, M., Shiri, J., 2012. Suspended sediment modeling using genetic
programming and soft computing techniques. Journal of Hydrology, 450-451: 48-58.
DOI:10.1016/j.jhydrol.2012.05.031
Kisi, O., Shiri, J., 2012. River suspended sediment estimation by climatic variables implication:
Comparative study among soft computing techniques. Computers & Geosciences, 43: 73-
82. DOI:10.1016/j.cageo.2012.02.007
Kisi, O., Yuksel, I., Dogan, E., 2008. Modelling daily suspended sediment of rivers in Turkey
using several data-driven techniques / Modélisation de la charge journalière en matières
en suspension dans des rivières turques à l'aide de plusieurs techniques empiriques.
Hydrological Sciences Journal, 53(6): 1270-1285. DOI:10.1623/hysj.53.6.1270
Korenberg, M.J., 1989. A robust orthogonal algorithm for system identification and time-series
analysis. Biol Cybern, 60(4): 267-76.
Koza, J.R., 1992. Genetic programming: on the programming of computers by means of natural
selection, 1. MIT press.
Liu, Q.-J., Shi, Z.-H., Fang, N.-F., Zhu, H.-D., Ai, L., 2013. Modeling the daily suspended
sediment concentration in a hyperconcentrated river on the Loess Plateau, China, using
the Wavelet–ANN approach. Geomorphology, 186: 181-190.
DOI:10.1016/j.geomorph.2013.01.012
Lohani, A.K., Goel, N., Bhatia, K., 2014. Improving real time flood forecasting using fuzzy
inference system. Journal of Hydrology, 509: 25-41.
Lohani, A.K., Goel, N.K., Bhatia, K.K.S., 2007. Deriving stage–discharge–sediment
concentration relationships using fuzzy logic. Hydrological Sciences Journal, 52(4): 793-
807. DOI:10.1623/hysj.52.4.793
36

Melesse, A.M., Ahmad, S., McClain, M.E., Wang, X., Lim, Y.H., 2011. Suspended sediment
load prediction of river systems: An artificial neural network approach. Agricultural
Water Management, 98(5): 855-866. DOI:10.1016/j.agwat.2010.12.012
Mirbagheri, S.A., Nourani, V., Rajaee, T., Alikhani, A., 2010. Neuro-fuzzy models employing
wavelet analysis for suspended sediment concentration prediction in rivers. Hydrological
Sciences Journal, 55(7): 1175-1189. DOI:10.1080/02626667.2010.508871
Mustafa, M.R., Rezaur, R.B., Saiedi, S., Isa, M.H., 2012. River Suspended Sediment Prediction
Using Various Multilayer Perceptron Neural Network Training Algorithms—A Case
Study in Malaysia. Water Resources Management, 26(7): 1879-1897.
DOI:10.1007/s11269-012-9992-5
Nagy, H., Watanabe, K., Hirano, M., 2002. Prediction of sediment load concentration in rivers
using artificial neural network model. Journal of Hydraulic Engineering, 128(6): 588-
595.
Noori, N., Kalin, L., 2016. Coupling SWAT and ANN models for enhanced daily streamflow
prediction. Journal of Hydrology, 533: 141-151.
Noori, R., Safavi, S., Shahrokni, S.A.N., 2013. A reduced-order adaptive neuro-fuzzy inference
system model as a software sensor for rapid estimation of five-day biochemical oxygen
demand. Journal of Hydrology, 495: 175-185.
Nourani, V., Andalib, G., 2015a. Daily and monthly suspended sediment load predictions using
wavelet based artificial intelligence approaches. Journal of Mountain Science, 12(1): 85-
100. DOI:10.1007/s11629-014-3121-2
Nourani, V., Andalib, G., 2015b. Wavelet Based Artificial Intelligence Approaches for
Prediction of Hydrological Time Series, Artificial Life and Computational Intelligence.
Springer, pp. 422-435.
Nourani, V., Hosseini Baghanam, A., Adamowski, J., Kisi, O., 2014. Applications of hybrid
wavelet–Artificial Intelligence models in hydrology: A review. Journal of Hydrology,
514: 358-377. DOI:10.1016/j.jhydrol.2014.03.057
Nourani, V., Kalantari, O., Baghanam, A.H., 2012. Two semidistributed ANN-based models for
estimation of suspended sediment load. Journal of Hydrologic Engineering, 17(12): 1368-
1380.
Pan, T.-Y. et al., 2013. Improvement of watershed flood forecasting by typhoon rainfall climate
model with an ANN-based southwest monsoon rainfall enhancement. Journal of
Hydrology, 506: 90-100.
Partal, T., Cigizoglu, H.K., 2008. Estimation and forecasting of daily suspended sediment data
using wavelet–neural networks. Journal of Hydrology, 358(3-4): 317-331.
DOI:10.1016/j.jhydrol.2008.06.013
Pektaş, A.O., 2015. Determining the essential parameters of bed load and suspended sediment
load. International Journal of Global Warming, 8(3): 335-359.
Pektaş, A.O., Doğan, E., 2015. Prediction of bed load via suspended sediment load using soft
computing methods. Geofizika, 32.
Polikar, R., 1999. The story of wavelets. Physics and modern topics in mechanical and electrical
engineering: 192-197.
Raghuwanshi, N., Singh, R., Reddy, L., 2006. Runoff and sediment yield modeling using
artificial neural networks: Upper Siwane River, India. Journal of Hydrologic
Engineering, 11(1): 71-79.
37

Rai, R.K., Mathur, B.S., 2007. Event-based Sediment Yield Modeling using Artificial Neural
Network. Water Resources Management, 22(4): 423-441. DOI:10.1007/s11269-007-
9170-3
Rajaee, T., 2010. Wavelet and Neuro-fuzzy Conjunction Approach for Suspended Sediment
Prediction. CLEAN - Soil, Air, Water, 38(3): 275-286. DOI:10.1002/clen.200900191
Rajaee, T., Mirbagheri, S., Nourani, V., Alikhani, A., 2010a. Prediction of daily suspended
sediment load using wavelet and neurofuzzy combined model. International Journal of
Environmental Science & Technology, 7(1): 93-110.
Rajaee, T., Mirbagheri, S.A., Zounemat-Kermani, M., Nourani, V., 2009. Daily suspended
sediment concentration simulation using ANN and neuro-fuzzy models. The Science of
the total environment, 407(17): 4916-27. DOI:10.1016/j.scitotenv.2009.05.016
Rajaee, T., Nourani, V., Zounemat-Kermani, M., Kisi, O., 2010b. River suspended sediment load
prediction: Application of ANN and wavelet conjunction model. Journal of Hydrologic
Engineering, 16(8): 613-627.
Ramezani, F., Nikoo, M., Nikoo, M., 2014. Artificial neural network weights optimization based
on social-based algorithm to realize sediment over the river. Soft Computing, 19(2): 375-
387. DOI:10.1007/s00500-014-1258-0
Sadeghpour Haji, M., Mirbagheri, S.A., Javid, A.H., Khezri, M., Najafpour, G.D., 2014.
Suspended sediment modelling by SVM and wavelet. Građevinar, 66(03.): 211-223.
Shiri, J., Kişi, Ö., 2012. Estimation of Daily Suspended Sediment Load by Using Wavelet
Conjunction Models. Journal of Hydrologic Engineering, 17(9): 986-1000.
DOI:10.1061/(asce)he.1943-5584.0000535
Si, J. et al., 2015. Modeling soil water content in extreme arid area using an adaptive neuro-fuzzy
inference system. Journal of Hydrology, 527: 679-687.
Singh, A., Imtiyaz, M., Isaac, R.K., Denis, D.M., 2013. Comparison of Artificial Neural Network
Models for Sediment Yield Prediction at Single Gauging Station of Watershed in Eastern
India. Journal of Hydrologic Engineering, 18(1): 115-120. DOI:10.1061/(asce)he.1943-
5584.0000601
Sivakumar, B., Wallender, W.W., 2005. Predictability of river flow and suspended sediment
transport in the Mississippi River basin: a non-linear deterministic approach. Earth
Surface Processes and Landforms, 30(6): 665-677. DOI:10.1002/esp.1167
Solomatine, D.P., 2005. Data‐Driven Modeling and Computational Intelligence Methods in
Hydrology. Encyclopedia of hydrological sciences.
Takagi, T., Sugeno, M., 1985. Fuzzy identification of systems and its applications to modeling
and control. Systems, Man and Cybernetics, IEEE Transactions on(1): 116-132.
Talebizadeh, M., Morid, S., Ayyoubzadeh, S.A., Ghasemzadeh, M., 2009. Uncertainty Analysis
in Sediment Load Modeling Using ANN and SWAT Model. Water Resources
Management, 24(9): 1747-1761. DOI:10.1007/s11269-009-9522-2
Tayfur, G., 2002. Artificial neural networks for sheet sediment transport. Hydrological Sciences
Journal, 47(6): 879-892.
Tayfur, G., Karimi, Y., Singh, V.P., 2013. Principle component analysis in conjuction with data
driven methods for sediment load prediction. Water resources management, 27(7): 2541-
2554.
Trygg, J., Wold, S., 2002. Orthogonal projections to latent structures (O‐PLS). Journal of
chemometrics, 16(3): 119-128.
38

van Maanen, B., Coco, G., Bryan, K.R., Ruessink, B.G., 2010. The use of artificial neural
networks to analyze and predict alongshore sediment transport. Nonlinear Processes in
Geophysics, 17(5): 395-404. DOI:10.5194/npg-17-395-2010
Vapnik, V.N., Vapnik, V., 1998. Statistical learning theory, 1. Wiley New York.
Yaseen, Z.M. et al., RBFNN versus FFNN for daily river flow forecasting at Johor River,
Malaysia. Neural Computing and Applications: 1-10.
Yurdusev, M.A., Firat, M., 2009. Adaptive neuro fuzzy inference system approach for municipal
water consumption modeling: An application to Izmir, Turkey. Journal of hydrology,
365(3): 225-234.
39

12

10
Number of papers

0
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
year of publication

Figure 1: Number of AI publication in sediment estimation since 2001 until 2014 (indexed in
Scopus)

Figure 2: Types of input and output variables for ANN sediment transport model
40

Figure 3: The stages of complementary suggested model with some recommended AI methods
41

Table 1 Details of the ANN reviewed papers, including authors, type of ANN, input variables,

time scale, and performance criteria

Author ANN type / Input variables Time scale Performance


compared with criteria
Jain (2001) FFNN/ Conventional Water level, Daily Correlation,
curves discharge, SSC SSE
Cigizoglu MLP/ SRC Discharge, SSL Daily MSE, R2
(2004)
Agarwal et al. FFNN/- Rainfall, runoff , Daily, RMSE,
(2005) SSC weekly, ten- CC,CE
day, monthly
Cigizoglu and FFNN,GRNN/ MLR, Discharge, SSL Daily RSME, R2
Alp (2006) SRC
Alp and FFNN/RBF Rainfall, SSL, Daily MSE, R2
Cigizoglu discharge
(2007)
Rai and Mathur FFNN/- Rainfall, runoff, Daily RMSE, CC,
(2007) SSL CE
Kisi et al. MLP Discharge, SSL Daily R2,CC,
(2008) MARE
Jothiprakash MLP Rainfall, inflow, Yearly RMSE,
and Garg reservoir capacity, MAE, E,
(2009) sediment volume AARE
Melesse et al. FFNN/ MLR, SSL, rainfall, Daily, Weekly R2 ,RMSE,
(2011) ARIMA, MNLR discharge MAPE, E
Mustafa et al. MLP/- Discharge, SSL Daily RMSE,
(2012) MARE, R2
Singh et al. SBP / RBNN Rainfall, discharge Monthly R2 , RMSE
(2013)
Afan et al. FFNN/RBF SSL, discharge Daily R2, RMSE,
(2014) MAE
(Sum Square Error (SSE), Mean Square Error (MSE), Root Mean Square Error (RMSE),
coefficient of determination (R2), Nash efficiency (CE), Correlation Coefficient (CC),
42

Mean Absolute Relative Error (MARE), Efficiency (E), Mean Absolute Percent Error
(MAPE)).

Table 2 The application of Fuzzy Logic in sediment transport modeling.

Author Variables Time scale Performance criteria

Kisi (2005) Sediment load, discharge Daily RMSE, R2

Lohani et al. (2007) SSC, discharge, water level Daily RMSE, R

Kisi et al. (2008) Sediment load, discharge Daily RMSE, R2

Firat and Güngör (2009) Sediment load, discharge Monthly RMSE, E, CC

Rajaee et al. (2009) SSC, discharge Daily RMSE, R2

Cobaner et al. (2009) SSC, discharge, rainfall Daily RMSE, MAE, R2 ,


Relative Error %
Demirci and Baltaci SSC, discharge, mean water Daily MSE, MAE, R
(2012) temperature
43

Table 3 The application of SVM and evolutionary methods in sediment transport estimation

Author AI type Variables Time

scale

Guven and Kişi (2010) LGP, GEP SSL, discharge Daily

Kişi (2010) NDE SSC, discharge Daily

Kisi (2012) LSSVM SSC, discharge Daily

Kisi et al. (2012) GP SSL, discharge Daily

Kisi and Shiri (2012) GEP SSC, discharge, rainfall Daily

Goyal (2014) M5 SSL, discharge, rainfall Monthly


44

Highlights

 The authors introduced a comprehensive state-of-the-art for AI-based sediment model.

 The authors defined each data-driven of AI and AI-complementary model.

 An assessment and evaluation have been carried out for the AI sediment models.

 The comparison proved that AI-complementary model provided an accurate modelling.

 Several recommended advanced techniques have been proposed for future research.

You might also like