You are on page 1of 17

Journal of Hydrology 577 (2019) 123952

Contents lists available at ScienceDirect

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

Research papers

A novel approach for the prediction of the incipient motion of sediments T


under smooth, transitional and rough flow conditions using Geno-Fuzzy
Inference System model

Hussein Bizimana , Abdüsselam Altunkaynak
Hydraulics Division, Faculty of Civil Engineering, Istanbul Technical University, Maslak 34469, Istanbul, Turkey

A R T I C LE I N FO A B S T R A C T

This manuscript was handled by A. Bardossy, In this present study, a novel approach is introduced for accurate prediction of the incipient motion of uniform
Editor-in-Chief. grain particles in sand and gravel-bedded open channels under unidirectional flow by improving Sugeno Fuzzy
Keywords: Inference System (Sugeno FIS). The Adaptive Neural Fuzzy Inference System (ANFIS) tool is based on Sugeno
Fuzzy logic FIS. Consequent part of the ANFIS tool is limited to either as a constant or a linear function. This means that not
Shields entrainment function only a non-linear function is available for the consequent part of the model but also it cannot be represented by
Shear Reynolds number the constant and linear functions, at the same time. ANFIS tool optimizes antecedent parameters (fuzzy sets) and
Genetic algorithm consequent parameters (constant or linear functions) by utilizing neural and least square methods, respectively.
Sediments
In this study, a novel hybrid model named as Geno-Fuzzy Inference System (GENOFIS) is introduced by in-
Artificial neural network
tegrating improved Sugeno FIS and Genetic Algorithms (GAs) that refer to where, the antecedent (fuzzy sets) and
consequent (constant, linear and non-linear functions) parameters are optimized using Genetic Algorithms (GAs)
tool. A quantitative comparison is implemented between the ANFIS and GENOFIS models using root mean
square error (RMSE) and the Nash-Sutcliffe coefficient of efficiency (CE) as evaluation criteria by using three
different types of incipient motion of sediment measurements that exist in the literature as reference, visual and
development of competence functions. The results of this present study demonstrated that the novel GENOFIS
model provided more accurate prediction results in comparisons with the ANFIS model results for three different
types of incipient motion of sediment.

1. Introduction related to the bed surface morphology, grain particle size, protrusion to
the flow and packing orders (Fenton and Abbott, 1977; Giménez-curto
1.1. Incipient motion and sediment transport overview and Corniero, 2009), most notably the recurring behavioural aspect of
turbulence in near-bed flow (Sutherland, 1967; Dey et al., 2012).
Following an important increase in the volume of flow pressing Fig. 1a illustrates a description of a bed grain particle that is targeted at
against stationary sediment particles, at a given level which particles the incipience of sediment entrainment by hydraulically rough wall-
cannot possibly sustain undergoing pressure acting upon them caused shear flow conditions. The catalyst forces for the grain particle trans-
by the hydrodynamic forces (drag force and lift force), suddenly dis- port are the lift force FL and drag force FD. Elseways, the force brings the
location will happen and movement will initiate (Buffington and stabilizing condition to the particle is the submerged weight FG of the
Montgomery, 1997; Andrews, 1994). The instantaneous changes from a actual particle. The targeted bed particle can be possibly under these
no motion state to initiation of sediment motion is defined as the following entrainment modes, rolling, sliding, and lifting. Fig. 1b,
threshold of initial motion or threshold of incipient motion which refers Fig. 1c, and Fig. 1d illustrate all those transport modes. The first mode
to the earliest stage of sediment transport. The study of the incipient of entrainment is rolling and is depicted in Fig. 1b, it happens when the
motion of sediment transport is very crucial in various areas such as destabilizing moment MD around the pivoting point exceeds the stabi-
channel design, river morphology and restoration, pollutant transport, lizing moment MS around the same point (Ali and Dey, 2016; Dey and
sediment mobility, and their transport, etc. (Paphitis, 2001; Ali and Ali, 2018). Elseways, Fig. 1c illustrates the sliding mode where the drag
Dey, 2016; Dey and Ali, 2018). The sediment transport is extremely force FD surpasses the frictional resistance FR at the contact points;


Corresponding author at: Istanbul Technical University, Maslak 34469, Istanbul, Turkey.
E-mail address: mutembealhussein@gmail.com (H. Bizimana).

https://doi.org/10.1016/j.jhydrol.2019.123952
Received 27 February 2019; Received in revised form 10 July 2019; Accepted 12 July 2019
Available online 15 July 2019
0022-1694/ © 2019 Elsevier B.V. All rights reserved.
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 1. Targeted bed grain particle at the incipience of entrainment by a rough wall-shear flow condition.

moreover, in lifting mode Fig. 1d, the lift force FL surpasses the sub- where a1, a2, a3 and an are N basic quantities of the set ai (i = 1,2,…,n).
merged weight FG of the grain particle. Hence, all entrainment modes of These N basic quantities are combined with each one of the remaining
a grain particle entrainment are as follows (Dey and Ali, 2018): n-3 power products X1, X2,…,XN formed by n characteristic parameters
ai as shown by eq.1, as the dimensionless variables of the two-phase
Rolling mode: MD ≥ MS → FDlv + FLlh ≥ FGlh , (1a)
phenomenon. mj,αj,βj and γj must be selected so that each of the power
products above (Eq. (2)) becomes dimensionless (Yalin, 1977). In this
Sliding mode: FD ≥ FR → FD ≥ kfr (FD − FR) (1b)
present study the set ρ, µ, ρs, υ, D, h, u*, γs is considered for the two-
Lifting mode: FL ≥ FG phase phenomenon in sediment transport. Furthermore, selection of the
(1c)
parameters D,ρ,u*, which have independent dimensions as basic quan-
where lv and lh define the vertical and the horizontal lever arms, re- tities is performed. Moreover, the application of the procedure implied
spectively, and kfr defines the coefficient of friction. kfr = tanØ, if Ø is by Eq.2 permits to determine the following 8–3 = 5 power products as
the internal friction angle. The equal sign presented in Eq. (1) for a the dimensionless variables of the two-phase phenomenon. In this study
specific grain particle entrainment mode represents the incipience of the incipient motion of sediment transport or Shields entrainment
entrainment mode conditions, whilst the greater than sign means a function is defined by two dimensionless characteristic parameters, this
transport or entrainment mode (Elhakeem et al., 2017). In order to is because of data used in this study from previous researches that were
measure the incipient motion of sediment transport accurately; some defined by only two dimensionless characteristic parameters as follows:
dimensionless variables and functions of the two-phase phenomenon u*cr D
are used. The two-phase phenomenon is herein defined by the fluid, X1 = X cr = D1ρ1u 1cr μ−1 =
* υ (3)
cohesionless granular medium and flow characteristics such as density
(ρ), viscosity (µ), density of the granular material (ρs), diameter of the ρu 2cr
X2 = Ycr = D−1 ρ1 u 2cr γs−1 = *
granular material (D), average depth of flow (h), slope (J) and force of * γs D (4)
gravity (g) (Yalin 1977). According to Yalin (1977), when defining the
where X1 and X2 in Eqs. (3) and (4) stand for shear Reynolds number
two-phase motion, it should always be preferred to use the shear ve-
(Re*) and Shield entrainment function (Ycr), respectively, D defines
locity (u*) and specific density (ρs) rather than using the slope (J) and
particle diameter, ρ defines the fluid density, u* defines the bed shear
force of gravity (g). In addition, basic quantities can be defined as
velocity,µ defines the fluid dynamic viscosity, υ defines the fluid kine-
characteristic parameters that have independent dimensions. The pro-
matic viscosity and γs defines particle specific weight. For the calcula-
cedure to transform the characteristic parameters into dimensionless
tion of Re*, the roughness length is taken as D by considering uniform
variables was defined by Yalin (1977) as follows:
granular material, otherwise, Nikuradse’s equivalent sand roughness
X1 = a1α1 a2β1 a3γ1 a4m1 ⎫ should be used.

X2 = a1α 2 a2β2 a3γ 2 a5m2 (with N= n− 3)
-------------------- ⎬
XN=a1αN a2βN a3γN anmN ⎪
⎭ (2)

2
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

1.2. Incipient motion modeling overview entrainment. Therefore, Competence and reference-based approaches
can be considered the best alternatives while representing reach-
Using purely theoretical reasoning, it has been established that the average bed sediment motion. Unlikeness, visual based approaches
initiation of sediment transport on a plane mobile bed is given by a specifically focus on to record local incipience of sediment entrainment
certain condition that must be fulfilled by the dimensionless variables and are the best fit for application to entrainment studies of distinct bed
shear Reynolds number (Xcr) and Shields entrainment function (Ycr) surface textural blotches (Carson and Griffiths, 1985; Wilcock, 1988;
(Yalin, 1977). Shields introduced to plot the experimental values Ycr Buffington and Montgomery, 1997; Dey and Ali, 2018). Moreover, the
versus Xcr with the intention of producing a curve determining the in- approaches used to analyze the incipience of sediment entrainment,
itiation of sediment transport for the first time (Shields, 1936). There- carry additionally considerable drawbacks such as many iterations,
after, in recent years some implicit and explicit mathematical ap- time-consuming nature and requiring expensive experiments and un-
proaches have been used to study the threshold of incipient motion of fortunately the results found still carry considerable errors. Un-
sediments (Shields, 1936; Yalin and Karahan, 1979; Paphitis, 2001;, doubtedly, the accurate measurement of the incipient (initiation) of
Yalin and da Silva, 2001, Cao et al., 2006, Kitsikoudis et al., 2016; motion could contribute great valuable information to water resources
Sipiliotis et al., 2018; Dey and Ali, 2018;). Due to the complexity and engineering planners and designers. Recently, in order to bring a new
high variability in defining hydrodynamics of sediments; defining the approach to the accurate measurement of the initiation of sediment
threshold of incipient motion of sediments under a natural environment transport; fuzzy approach has been used in rough flow conditions
also becomes very challenging (Shields, 1936; Komar and Clemens, (Spiliotis et al., 2018; Kitsikoudis et al., 2016). Spiliotis et al. (2018)
1986; Paphitis, 2001; Vollmer and Kleinhans, 2007; Elhakeem et al., applied the fuzzy set knowledge to the Zanke’s equation (Dey and Ali,
2017). In order to overcome this challenge, numerous empirical 2018) in rough flow conditions. Spiliotis et al. (2018) opted not to
threshold curves representing the threshold (critical) condition for the consider the application in smooth and transitional flow conditions
incipient (initial) motion of sediments or Shields entrainment function because of higher influence of viscous forces present. Previously, the
as discussed herein, have been represented (Hjulstrom, 1935; Shields, same group of researchers (Kitsikoudis et al., 2016) has applied the
1936; Inman, 1949; Sundborg, 1956; Komar and Clemens, 1986; fuzzy regression analysis combined with goal programming to model
Paphitis, 2001). The most commonly used approach to defining the the incipient motion under rough flow conditions. Obviously, a need to
threshold of initiation of sediment transport up today is still regarded as extend this approach for all hydrodynamic flow conditions to con-
the Shields approach (Shields 1936) even though; details from Shields tribute to the current knowledge is present. Knowing that Neuro-Fuzzy
results remain extremely vague (Kennedy, 1995; Buffington, 1999). approach which is a powerful combination of Artificial Neural Net-
This same fact was supported by Shvidchenko and Pender (2000) by works (ANN) and Adaptive Neural Fuzzy Inference System (ANFIS) has
showing how the vagueness persists in defining what actually con- been successfully used in hydrology, hydraulics and other water re-
stitutes the incipient of motion. Furthermore, Vanoni (2006) concluded sources fields such as rainfall-runoff modeling (Şen and Altunkaynak,
that the different definitions given to the initiation of motion by dif- 2006; Nasr and Bruen, 2008; Remesan et al., 2009), flood hazard risk
ferent researchers highly generate a bias to the published data. As it can mapping ( Sonmez and Bizimana, 2018), modeling of lake water level
be noted in previously several approaches to defining the initiation of fluctuations (Altunkaynak and Şen, 2007; Bizimana et al., 2016),
motion for sediment transport have been undertaken (Shields, 1936; stream flow prediction (Özger, 2009), model for equilibrium scour in
Yalin and Karahan, 1979; Yalin and da Silva, 2001). In the actual lit- sediment transport (Uyumaz et al., 2006). Moreover, Tütmez et al.
erature, the incipience of sediment transport is commonly defined using (2007) who compared the spatial interpolation of mechanical proper-
three different approaches, namely reference, visual and competence- ties of rocks by using Mamdani and Sugeno fuzzy models, concluded
based approaches (Buffington and Montgomery, 1997; Buffington and that Sugeno fuzzy model was better than Mamdani fuzzy model in in-
Montgomery, 1998; Buffington, 1999, Cao et al., 2006). The critical vestigating the randomness presented in mechanical properties of rocks
shear stress defined using reference-based approach is determined with under nonlinear distribution. Furthermore, Ying (1999), in their re-
either a zero or low reference entrainment rate extrapolated to from search concluded that Sugeno fuzzy approach is very economical while
joined shear stress and bed sediment entrainment analysis (Ashworth defining the number of inputs fuzzy sets compared to Mamdani fuzzy
and Ferguson, 1989; Ashworth et al., 1992). Visual observation ap- approach when nontrapezoidal/nontriangular inputs fuzzy sets are
proach is direct but can carry high subjectivity depending on re- applied. Consequently, the advantages found in literature of Sugeno
searcher’s definition of how much sediment entrainment defines the fuzzy approach over Mamdani fuzzy approach and successful applica-
incipience of sediment motion (Kramer, 1935; Neill and Yalin, 1969; tion of Sugeno ANFIS approach in previous sediment transport based
Yalin and Karahan, 1979; Wilcock, 1988; Turowski et al. (2011); studies such as in evaluation of sediment entrainment in sand-bed rivers
Schiereck, 2017). The competence functions are highly linked to the (Azamathulla et al., 2009; Kitsikoudis et al., 2015), computation of bed
size and efficiency of the sediment catch, sample size, presence of load in gravel-bedded river channels (Kitsikoudis et al.,2014;Zounemat-
coarse sediment sizes and definitely fitting approach (Wilcock, 1988, Kermani et al., 2018) andcomputation of sediment entrainment in
Wathen et al., 1995). Even though all methods used to define the in- sewer pipe systems (Azamathulla et al., 2012), increased the motivation
cipience of sediment transport carry methodological bias that can be to apply machine learning or soft computation techniques, specifically
the cause of subjectivities and inappropriateness while deciding which Sugeno ANFIS to determine the incipient motion of sediment in all
research method is the best while defining the Shields entrainment hydrodynamic flow conditions namely hydrodynamically smooth,
function value (Ycr). None of these research methods used to define the transitional and hydrodynamically rough. The neuro-fuzzy method
incipience of sediment motion is considered superior (Buffington and clearly makes use of a combination of Artificial Neural Network (ANN)
Montgomery, 1997, Dey and Ali, 2018). Each and every research architecture and Fuzzy Logic (FL) reasoning ability. These two ap-
method has relatively its weakness and strength. Yet, every method is proaches fall into soft computation techniques which are very im-
more competitive in a particular case than others (Carson and Griffiths, portant when encountered with systems that are hard to be described
1985). For instance, since competence and reference-based values of Ycr analytically by conventional physical and mathematical approaches.
are computed from bedload entrainment estimations, they can be ac- ANN can be described as a black box modeling tool which is able to
cepted as the best research approaches that can be used to compute obtain accurate results matching with observed data. According to
bedload transport. Following the bed load and the investigative tech- Jovanovic et al. (2004), the mechanism of ANN learning is free of
nique applied, competence and reference-based approaches can in- human expertise. Fuzzy Inference System (FIS) has recently become a
tegrate differential bed mobility that comes from bed surface textural popular computing approach with its concept of fuzzy set theory. Fuzzy
blotches and reach-scale separation of shear stress and sediment sets, fuzzy operators and expert base are building components of FIS.

3
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

The expert base of a FIS together with its rule base made of a number of function. The function defines the level to which X (or Y ) meets the
fuzzy IF-THEN rules and corresponding membership functions of the quantifier A, (or Bi-2) and also called as the membership function of Ai or
fuzzy sets considered in the fuzzy rules; can utilize expert knowledge. (Bi-2). The FIS provides tremendous freedom in defining the type of
Although FIS is very easy to use when expressing the knowledge with a membership functions in relation to our needs in terms of speed, sim-
set of if-then rules, still there is no methodical approach to make this plicity, convenience, and efficiency. The only condition that should be
transformation (Jang, 1992). satisfied is that a membership function must definitely vary between 0
Consequently, the combination of ANN and FIS generates a mu- and 1 (Tagaki and Sugeno, 1985). Fig. 2 depicts the ANFIS architecture
tually beneficial relationship that tackles previously mentioned draw- suggested by Jang (1992).
backs of each individual method. This has meshed through integrated However, the membership function is a function of its parameters,
Neuro-Fuzzy (NF) building blocks that share data structures and therefore changing its parameters alters the membership function
knowledge representation (Jang, 1992; Jang, 1993). ANFIS is a dif- shape. Parameters found in the first layer refer to antecedent (premise)
ferent type of NF building architectures which belongs to the category parameters.
of full fused NF systems (Jovanovic et al., 2004). Compared to other Layer 2: Every node in this layer multiplies the incoming signals
most used NF architectures such as Evolving Fuzzy Neural Network and also sends the product which is defined as the firing strength of a
(Kasabov, 1999), Fuzzy Inference Software (Tano et Al., 1996) and rule.
many more, ANFIS architecture provides the lowest root mean square
Oi2 = Wi = μAi (X )xμBi (Y ), i = 1, 2 (6)
error (RMSE) when used for prediction of any chaotic data set
(Abraham and Nath, 2000; Abraham, 2001). Layer 3: The name given to the outputs in this layer are called
normalized firings strengths because each i-th node in this particular
1.3. Research objectives layer calculates the ratio of the i-th strength to the total summation of
all rules of the firing strengths.
The objectives of this present study are i) improve the widely used − Wi
Adaptive Neural Fuzzy Inference System (ANFIS) tool, which is based Oi3 = Wi = , i = 1, 2
∑i Wi (7)
on the Sugeno Fuzzy Inference System (Sugeno FIS) approach. The
consequent part of ANFIS is allowed to be represented either as a Layer 4: Every node i in this particular layer is represented by a
constant or a linear function. This means that not only a non-linear square node with a node function.
function is available for the consequent part of the model but also it − −
Oi4 = Wi = Wi (pi x + qi y + ri ), i = 1, 2 (8)
cannot be represented by the constant and linear functions, at the same

time. ii) develop a novel method to overcome the shortcomings of where Wi is the output characteristic of layer 3, and (p1, q1, r1) is a set of
ANFIS model with the combination of improved Sugeno FIS and parameters referred to consequent parameters. It should be noticed that
Genetic Algorithms (GAs) that has skills to represent the consequent in case of pi = 0 and qi = 0 (i = 1,2) a Sugeno FIS which is named as
part of the model simultaneously by a constant, linear and non-linear zero-order is implemented, otherwise it is the first order Sugeno FIS
function or a combination of them. In addition, the newly developed that is incorporated (Sugeno, 1985; Tagaki and Sugeno, 1985).
model optimizes the antecedent parameters (fuzzy sets) and consequent Layer 5: The computation of the overall output as the summation of
parameters (constant, linear and non-linear functions) via utilizing the all signals received is made available by the single node in this layer.
GAs tool. iii) develop a model for sediment entrainment for smooth,
transitional and rough flow conditions. iv) investigate and compare the − ∑i wi fi
Oi5 = ∑ wifi =
performances of ANFIS and newly developed model which is named as i
∑i wi (9)
Geno-Fuzzy Inference System (GENOFIS) in predicting the inci-
pient motion of sediments under smooth, transitional and rough flow In order to reach its best performance, ANFIS makes use of the least
conditions. In the following, Section 2 represents the theory establish- mean square optimization method in order to determine the consequent
ment in this study, section 3 represents the materials and methods parameters and backpropagation to obtain the antecedent (premise)
applied in this study and section 4 represents the results and discussion parameters. Every step in the learning process is composed of two steps:
of this study whilst section 5 qualitatively summarizes the main con- In the first step training data set is used as the input, the antecedent or
clusions of this study. boundary parameters are considered as fixed values and the optimal
consequent parameters are obtained by an iterative (epoch) least mean
square method. In the second step, patterns are once again propagated,
2. Theory establishment but herein this step the consequent parameters are supposed to be
fastened and backpropagation is utilized to alter the antecedent para-
2.1. Adaptive Neural fuzzy Inference System meters. But unfortunately, ANFIS does represent the optimal con-
sequent output as a constant or linear function whereas many problems
Adaptive Neural Fuzzy Inference System (ANFIS) is based on behave strongly non-linear. Therefore, this present study introduces a
Takagi-Sugeno FIS (Sugeno and Kang, 1988). The Neuro-fuzzy model is novel approach named as GENOFIS model which is an integration of the
introduced by R.Jang, (1992) and with the name of ANFIS by Jang improved SUGENO FIS and Genetic Algorithms (GAs). Moreover,
(1992), Jang (1993) that used two inputs and generated one output by GENOFIS allows the representation of the consequent part as constant,
using two fuzzy if-then rules. linear and non-linear functions, at the same time as well as optimize
Rule 1: If × is A1 and y is B1 then fi = p1x + q1y + r1 antecedent and consequent parameters utilizing GAs.
Rule 2: If × is A2 and y is B2 then f2 = p2x + q2y + r2
As depicted in Fig. 2, five structure layers of the ANFIS tool are
2.2. Geno fuzzy Inference System
defined as follows:
Layer 1: Every node available in this layer has a node function
Practical assumptions in determination of the suitable Sugeno FIS
OÄ1° = μ Ai (X ) for i = 1, 2 or;
architecture for any practical social or engineering problem are a
Oi1 = μ Bi − 2 (Y ) for i = 3, 4 compromise between the accuracy and the computational entangle-
(5)
ment. For the conventional ANFIS architecture, for n inputs, f mem-
where X or Y is the input to node i, and Ai or (Bi-2) represents the bership functions and k parameters that are given to each input and
linguistic label (big, small, high, close, etc.) in association with the node membership function, respectively need the total number of fitting

4
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 2. ANFIS with the rule base containing two IF-THEN rules.

parameters equal to F(n,f,k) = nf.k + fn. (n + 1). It should be noticed parameters b3, c3, d3, and e3 by using its inherent properties. In order to
that the time needed for computation during the training process in an successfully achieve this objective, in the beginning, chromosome po-
ANFIS, directly rely on this number; therefore, caution needs to be pulation belonging to firmness variables is picked out randomly. In
taken when increasing it. It should also be kept in mind that the re- most cases, GAs is coded in binary (0, 1) basis in order to simplify the
lationship between computational entanglement and ANFIS di- calculation. These digital series are called chromosomes. Foregoing and
mensionality is one of the drawbacks of the ordinary fuzzy approach. present state variables connecting the weights Rules’ of chromosomes
This is frequently referred to as the ‘curse of dimensionality' (Maguire formation which is in binary basis digital system is depicted in Fig. 3.
et al., 1998). Furthermore, as mentioned before, ANFIS generally can In the chromosome structure depicted in Fig. 3, a member of state
only define the consequent part either as a constant or linear function, transition rules for Rule 1 (a1,b1,c1,d1,e1), Rule 2 (a2,b2,c2,d2,e2) and
at the same time. Consequently, the present study combines the ap- Rule 3 (a3,b3,c3,d3,e3) were presented by 10 bits (for each member)
proach defined by Jovanovic et al. (2004) proposed that in order to respectively. After being coded in this way, the activeness of each
enhance the accuracy of ANFIS and at the same time reduce the com- chromosome in the community was calculated separately. The roulette
putational entanglement, the number of if-then rules in FIS should be wheel approach was used to determine the activeness of each chro-
represented by the number of input fuzzy sets. Under this presumption mosome. The activeness criterion of the chromosomes was determined
the total number of fitting parameters is given by G (n, f, according to the difference between the observed data and predicted
k) = n × f × k + n × f (n + 1) = n × f × (n + k + 1). This approach values.
considerably increases the accuracy as well as reduces the computation The global optimum solution point was reached at the point for
entanglement in linear problems. Unfortunately, it cannot solve non- which the root-mean-square error (RMSE) value becomes the smallest.
linear problems. Consequently, as mentioned before in section 2; this The activeness of each chromosome defines whether a chromosome has
present study introduces the GENOFIS, a novel approach that permits to stay in or leave the population. Cross-over transaction at GAs pos-
the representation of the consequent part with a non-linear function sesses an important function in shaping the members with more ac-
where antecedent and consequent parameters are optimized by the GAs tiveness at the community. The high activeness of any member defines
tool. Hence, the present novel method allows presenting the consequent its survival or otherwise its dismissal. The existing generation is varied
part of the introduced novel method as non-linear functions for the by using the operators of GAs; cross-over and mutation. According to
particular Shields entrainment function (Ycr) prediction of uniform se- the literature it is likely regarded that after the new member joins the
diment grain materials, which highly behaves in a non-linear manner population, it becomes closer to the solution in order to attain the ob-
(Shields, 1936; Yalin and Karahan, 1979; Yalin and da Silva, 2001; jective (Altunkaynak, 2008). In this present study, each chromosome in
Paphitis, 2001; Kitsikoudis et al., 2016). The general rule bases for the the population was coded with a randomly performed genetic digital
novel GENOFIS model used three approaches (reference, visual and series and a cross-over between the genes. It is called random cross-
competence) used in the literature to define the incipient motion of over. The randomness was also used to define the cross-over’s length.
sediments under smooth, transitional and rough flow conditions as Unlike many other approaches such as the least square method (LSM),
follows: GA generates a vast number of solution points rather than only one
Rule 1: If Xcr is Laminar then solution point. Therefore, GAs can reach the global optimum solution
Ycr = a1 (Re *)n + b1 (Re *)n − 1 + c1 (Re *)n − 2 + d1 (Re *)n − n + e1 among all solution points (Altunkaynak, 2009; Altunkaynak, 2013,
Rule 2: If Xcr is Transitional then Hadi and Tombul 2018). In table 1 a comparison between Fuzzy Logic
Ycr = a2 (Re *)n + b2 (Re *)n − 1 + c2 (Re *)n − 2 + d2 (Re *)n − n + e2 (FL) and Genetic Algorithms (GAs) for their differences is exhibited.
Rule 3: If Xcr is Turbulent then
Ycr = a3 (Re *)n + b3 (Re *)n − 1 + c3 (Re *)n − 2 + d3 (Re *)n − n + e3 where Ycr 2.3. Geno-Fuzzy Inference System versus Adaptive Neural Fuzzy Inference
defines the Shields entrainment function and Re* is the shear Reynolds System
Number.
a1,b1,c1,d1,e1,a2,b2,c2,d3,e3,a3,b3,c3,d3, and e3 are consequent Adaptive Neural Fuzzy Inference System (ANFIS) is a technique that
parameters which are optimized using Genetic Algorithms (GAs) tool holds the capacity to self-organize its network system and to train the
which is a method that searches the optimum solution by generating parameters of the system. The ANFIS tool uses the advantage of
random number in a defined interval. The GAs goal is to determine a1 Artificial Neural Network (ANN) learning algorithms and Fuzzy
which combines parameters b1, c1, d1, and e1 in Rule 1, a2 which Inference Systems (FIS). In other words, ANFIS merges the verbal ca-
combines parameters b2, c2, d2 and e2, and a3 which combines pacity of FIS and the numerical capacity of the ANN. An ANFIS is

5
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 3. Chromosomes structure.

Table 1 details by Altunkaynak (2010, 2013) and also by Ebtehaj and Bonakdari
Comparison between fuzzy logic and Genetic Algorithm. (2014) and Najafzadeh et al. (2017) and Ouyang (2017). The novel
Issues of comparison Fuzzy logic Genetic algorithm
GENOFIS application allows that drawbacks presented by the regular
Sugeno approach are eliminated by a modified FIS on the consequent
Soft computing Yes Yes part and at the same time solves the entanglement problems resulting
Mathematical Yes No from ANFIS approach process. The entanglement problems are in terms
complexity
of lack of simplicity and time effectiveness. The novel GENOFIS solves
Problem solving Depends on probability Depends on fitness/
and input parameters Objective function the aforementioned entanglement problems by considering equal the
IF-THEN rules Yes No number of antecedents (premises) and consequents. Therefore, the
Learning scheme Supervised Non-supervised novel GENOFIS reduces the number of IF-THEN rules and time used to
Processing time Little High
process the final output. Finally, in this study, the Shields entrainment
CH electing Distributed algorithm Centralized algorithm
function (Ycr) of uniform bed grain particles under unidirectional flow
CH: Cluster head. conditions was predicted by using the observed shear Reynolds number
(Re*). Additionally, a comparative study between novel GENOFIS and
defined as a hybrid-learning algorithm that utilizes gradient-descent ANFIS is accomplished. Eventually, the novel GENOFIS model is pro-
and least-squares (LS) methods to improve antecedent (premise) and posed as a novel and practical alternative to the ANFIS model for
consequent parameters, respectively. Though, the LS method utilizes modeling the Shields entrainment function (Ycr) of uniform bed grain
derivatives to arrive at an optimal solution. This indicates that used particles in unidirectional flow under smooth, transitional and rough
functions should always be continuous, otherwise optimization proce- flow conditions. In this study, in order to accurately split the data space
dure fails. Therefore, ANFIS works definitely under restrictive pre- into accurate fuzzy sets, a fuzzy logic clustering method is used. In the
sumptions such as linearity, normality, and constant variance (Özger, literature, the most cited fuzzy logic clustering approaches are the
2009; Akrami et al., 2013; Kitsikoudis et al., 2016; Spiliotis et al., Gustafson Kessel (GK) approach (Gustafson and Kessel, 1979) and
2018). Hence it can be clearly seen that ANFIS is inaccurate to deal with Subtractive clustering (SC) approach (Chiu, 1996). The GK approach is
highly non-linear problems. Moreover, the ANFIS model parameter based on iterative minimization of the objective functions whereas SC is
estimations are reached by considering the gradient-descent and LS based on non-iterative minimization of the objective functions. In the
methods (Jang, 1993). Therefore, the basic ANFIS approach demands present study, the SC approach is used to compute the Membership
differentiable and continuous linear target function at each and every Functions (MFs) of the Fuzzy Inference Systems (FIS) by using MATLAB
point in the system structure. In case of a nonlinear target function, it commercial software together with Genetic Algorithm optimization
definitely becomes difficult to take derivative and find an optimal so- operation for both ANFIS and GENOFIS. The selection of the best MFs
lution. On the other hand, in order to bring a novel solution to the was made after developing numerous models with different MF struc-
aforementioned problems resulting from ANFIS approach; the novel tures such as Gaussian (Bell shape), trapezoidal, triangular and finding
GENOFIS model was developed by improving consequent part of the that as also supported by Moosavi et al. (2013) and Gholami et al.
Sugeno consequent part so that it can be defined at the same time if (2017), the Gaussian shape is the best to be used. Fig. 4 depicts the
needed as constant, linear and non-linear function options. To optimize flowchart of calculation for both ANFIS and novel GENOFIS.
the novel GENOFIS model parameters, Genetic Algorithms (GAs) are
used. The GAs advantages and previous applications are discussed in

6
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 4. Flowchart of calculation of ANFIS and developed novel GENOFIS for the prediction of Shields entrainment function values.

2.4. Performance evaluation criteria 0.85, the performance is very good.

Solomatine and Shrestha (2009) and others have demonstrated that


among many other available model evaluation criteria in the literature, 3. Materials and methods
root-mean-square error (RMSE) and the Nash-Sutcliffe coefficient of
efficiency (CE) are most widely utilized and accepted (Nash and 3.1. Observed data and data analysis
Sutcliffe (1970); Bowden et al., 2013) also mentioned that RMSE is a
very good indicator of the degree of accuracy of a model. Hence, RMSE Both ANFIS and novel GENOFIS models were implemented to model
and CE are utilized to evaluate the performances of ANFIS and novel the Shields entrainment function. As indicated in the “Theory estab-
GENOFIS models in this present study. The following equation is used lishment” section, ANFIS and GENOFIS models were developed to
to calculate the root mean square error (RMSE): predict the Shields entrainment function for three different Shields
entrainment function modeling approaches, extrapolation base tech-
N
1 nique (Shields, 1936); which involves determination of critical shear
RMSE =
N
∑ (Ycrpi−Ycrmi )2 stress values that result from a low or completely a zero reference
i=1 (10)
transport frequency from jointed shear stress and bed load entrainment
where N is the total number of measurements; Ycrpi and Ycrmi are the measurements. The visual base technique (Yalin and Karahan, 1979) is
predicted and modeled Shields entrainment values, respectively. For a a direct visual observation of sediment entrainment from a given bed
perfect model RMSE equals to 0 and as the perfection of a model re- shear stress and development of competence functions (Wilcock, 1988;
duces the inconsistency between the actual data and predicted model Yalin and da Silva, 2001; Ali and Dey, 2016; Dey and Ali, 2018). The
values increases, RMSE value also increases. As regards the calculation development of competence functions technique relates the bed shear
of efficiency (CE) following equation is used: stress to the entrainment of the biggest sediment particle. The results of
these predicting models were then evaluated in terms of their perfor-
N
⎡ ∑ (Ycrpi − Ycrmi )2 ⎤ mance in predicting the Shields entrainment function or dimensionless
CE = ⎢1 − iN= 1 ⎥
∑i = 1 (Ycrmi − Ycrm )2 ⎦ shear stress.
⎣ (11)
The data used herein was collected from three different previous
where Ycrm is the mean predicted Shield entrainment value. CE of a most cited researches in terms of sediment incipient motion. Exploited
model is considered acceptable when its values are greater than 0.5 data are from Yalin and Karahan (1979), Buffington and Montgomery
(Moriasi et al., 2007; Altunkaynak and Nigussie, 2015). Moreover, (1997) and Yalin and da Silva (2001). Those data were collected and
Donigian and Love (2003), Wang and Altunkaynak (2011) and digitized using the software webPlotDigitizer-3.8. Fig. 5 depicts plots
Altunkaynak and Wang (2012) categorized the performance of a pre- from which the data was collected. Fig. 5b depicts the plot for visual-
dictive model as follows: If CE values are in the range 0.65–0.75, the based data (Yalin and Karahan, 1979; Dey and Ali, 2018), Fig. 5a de-
performance of the model is fair, if CE values are in the range picts the plot for reference-based data (Buffington and Montgomery,
0.75–0.85, the performance is good and if CE results are greater than 1997) and Fig. 5c depicts the plot for competence-based data (Yalin and

7
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 5. Data plots for reference (a), visual (b) and competence (c) based incipient motion approaches.

Table 2 Kennard and Stone (1969) have described the SSMD. The SSMD offers
Statistical measurements of employed experimental data (Buffington and the possibility to extract data subsets in a way that the extracted subsets
Montgomery, 1997). form the original data set contains minimum similarities with max-
Statistical measures Ycr Re* imum dissimilarities assessed between included members. Moreover,
the use of SSMD offers the unique possibility to increase the generality
Minimum value 0.010 0.035 in a dataset by extracting data in a way that do not centralize to the
Maximum value 0.078 363.1
special space; instead, it incorporates the data from the center to the
Mean value 0.036 76.1
Standard deviation 0.020 107.6 defined dataset boundary. Lajiness and Watson (2008) also stressed that
Skewness coefficient 0.539 1.306 by using the SSMD to divide the original dataset, provides highest
dissimilarities between subsets that cover the original dataset. There-
*
Both Re* and Ycr are dimensionless. fore, 60% and 40% (remaining) of observed data were used for cali-
bration (training) and to validate the performance of the model, re-
da Silva, 2001; Hutter and Wang, 2016). According to Altunkaynak and spectively, in this present study. The available data is very diverse as
Şen (2007) and Özger (2009), each data set should be divided into two they were collected from different experimental results that used dif-
groups before the development of the model. In order to divide the ferent techniques to define the threshold of the incipient motion of
original dataset into train and test data sets, the subset selection of sediment. A reference based technique utilizes statistical values rather
maximum dissimilarity method (SSMD) is applied. The original datasets than full obtained values and thus data collected from this technique
are divided into 60% for train dataset and 40% for the test dataset. demonstrates smaller shear Reynolds numbers related to the Shields

8
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Table 3 Table 5
Statistical measures of employed experimental data (Yalin and Karahan, 1979). ANFIS approach fuzzy rule bases.
Statistical measures Ycr Re* Rule no. Antecedent Consequent
Re* Ycr
Minimum value 0.029 0.57
Maximum value 0.187 1173.75 ANFIS (reference) 1 L1 Ycr = 0.22(Re*)-0.12
Mean value 0.069 184.55 2 L2 Ycr = 0.04(Re*)-0.08
Standard deviation 0.044 286.23 3 L3 Ycr = 0.12(Re*)-0.04
Skewness coefficient 1.355 2.08 4 T1 Ycr = 0.02(Re*)-0.004
5 T2 Ycr = 0.0035(Re*)-0.02
* Both Re* and Ycr are dimensionless. 6 T3 Ycr = 0.048(Re*)-0.001
7 R1 Ycr = 0.05(Re*)-0.002
8 R2 Ycr = 0.06(Re*)-0.003
Table 4
9 R3 Ycr = 0.11(Re*)-0.014
Statistical measures of employed experimental data (Yalin and da Silva, 2001).
ANFIS (visual) 1 L1 Ycr = 0.025(Re*)-0.11
Statistical measures Ycr Re* 2 L2 Ycr = 0.16(Re*)-0.10
3 L3 Ycr = 0.18(Re*)-0.04
Minimum value 0.03 2.775 4 T1 Ycr = 0.06(Re*)-0.004
Maximum value 0.099 19020.9 5 R1 Ycr = 0.064(Re*)-0.004
Mean value 0.049 1975.9
Standard deviation 0.0168 4113.6 ANFIS (competence) 1 L1 Ycr = 0.14(Re*)-0.038
Skewness coefficient 1.554 2.795 2 L2 Ycr = -0.17(Re*) + 0.51
3 L3 Ycr = -0.12(Re*)-0.15
* Both Re* and Ycr are dimensionless. 4 T1 Ycr = -0.20(Re*)-0.13
5 T2 Ycr = 0.025(Re*)-0.0065
6 T3 Ycr = 0.0068(Re*)-0.008
entrainment function values or the dimensionless shear stress than in 7 R1 Ycr = − 0.086(Re*) + 0.046
e.g. visual technique that was used by Yalin and Karahan (1979); it can 8 R2 Ycr = 0.018(Re*)-0.06
be seen that shear Reynolds numbers in Table 3 that are higher than 9 R3 Ycr = 1.6x10-7(Re*) + 4.9.10-7
those presented in Table 2. Furthermore, Table 4 contains data obtained
* Both Re* and Ycr are dimensionless.
from a development of competence function based technique used by
Yalin and da Silva (2001). As can be seen in Table 4, shear Reynolds
and Montgomery (1997), Yalin and Karahan (1979) and Yalin and da
values are even higher than of those in Table 2 and Table 3 with respect
Silva (2001), in which only shear Reynolds number was used to predict
to Shields entrainment function, which is because the technique used to
the Shields entrainment function. Table 5 and Table 6 present the de-
define the incipient motion relies on selective transport (Wilcock 1988).
cision matrix (fuzzy rules) at the end of the calibration phase for both
The selective transport relates the bed shear stress to the largest mobile
ANFIS and GENOFIS models, respectively. As it can be seen in Table 5
grain size (Jiang and Haff, 1993). The data collected from the research
ANFIS tool has a drawback of representing the consequent part of the
of Yalin and da Silva (2001), are exposed in Table 4. Furthermore,
model either as a constant or a linear function, which means that it does
statistically, an anticipated observation from the collected data, shows
not allow representing the consequent part of the model as a constant,
that the ANFIS and GENOFIS model results obtained using data from
linear and non-linear or integration of them, simultaneously. Because of
Table 2 will likely give higher prediction error due to relatively bigger
this disadvantage, the ANFIS model can yield inaccurate results when it
standard deviation considering the available maximum and minimum
comes to non-linear problems such as the definition of the Shields en-
Shield entrainment function (Ycr) values and shear Reynolds number
trainment function. However, by reconsidering the same MFs as used
(Re*). All data sets were collected from uniform grain material and non-
for the ANFIS, but this time redefining the parameters of the MFs using
viscous fluid base models. Additionally, these data were obtained from
GAs optimization technique, the developed novel GENOFIS provides
laboratory experiments performed for sand bedded open channels
the ability to represent the consequent part of the Sugeno FIS as any
under unidirectional flows. Some of the statistical parameters of the
function such as polynomial, logarithmic, Gaussian, etc. or combination
data used herein are presented in Table 2, Table 3 and Table 4 for re-
of them, simultaneously. After the determination of FIS using GAs op-
ference, visual and development of competence functions based tech-
timization technique, the centroid approach for defuzzification process
nique, respectively.
is performed to define the final Ycr optimized values.
In this present study, the consequent part of the newly GENOFIS
3.2. Application of ANFIS versus GENOFIS model is defined as polynomial functions for sediment incipient motion
or threshold of sediment entrainment under smooth, transitional and
It is well known that the ANFIS tool allows representing consequent rough flow conditions (see Table 6). Fuzzy sets applied herein to de-
part either as a constant or a linear function. This implies that not only a velop both ANFIS and GENOFIS models for reference, visual and
non-linear function is not allowed but also the consequent part of the competence techniques are depicted in Figs. 6, 7 and 8, respectively.
model cannot be represented with both the constant and linear func- Model parameters of Gaussian-type membership functions for the
tion, simultaneously. In this present study, to overcome the short- GENOFIS approach are presented in Table 7 and 8. The GAs tool was
comings of ANFIS, a new model is developed with the combination of used to optimize the antecedent and consequent parameters. For both
modified Sugeno FIS and GAs and named as GENOFIS which is capable ANFIS and GENOFIS models the same number of membership functions
of representing the consequent part of the model with constant linear or was used. Nine fuzzy sets are depicted as Laminar 1 (L1), Laminar 2
non-linear functions or even with a combination of them, at the same (L2), Laminar 3 (L3), Transitional 1 (T1), Transitional 2 (T2), Transi-
time. Furthermore, antecedent and consequent parameters are opti- tional 3 (T3), Rough 1 (R1), Rough 2 (R2) and Rough 3 (R3). In Tur-
mized by the GAs tool. The ANFIS and GENOFIS model results are bulent flow, fuzzy sets were separated as follows; Rough 1 (R1), Rough
evaluated quantitatively based on RMSE and CE values. These models 2 (R2) and Rough 3 (R3)) for reference-based approach), Laminar 1
are used to define the Shields Entrainment function as the consequent (L1), Laminar 2 (L2), Laminar 3 (L3), Transitional 1 (T1) and Rough 1
part of the model by using the shear Reynolds number which is selected (R1) for visual-based approach. Additionally, nine fuzzy sets were used
as the sole antecedent parts of both ANFIS and GENOFIS models. Only for competence-based approach and in the latter; fuzzy sets were di-
the shear Reynolds number is considered as the sole antecedent value in vided as similarly as in the reference-based approach. For ANFIS nine
both models because the models are trained by using data of Buffington

9
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Table 6
GENOFIS approach fuzzy rule bases.
Rule no. Antecedent Consequent
Re* Ycr

GENOFIS (reference) 1 L(1,2,3) Ycr = a1(Re*)6 + b1(Re*)5 + c1(Re*)4 + d1(Re*)3 + e1(Re*)2 + f1(Re*) + g1
2 T(1,2,3) Ycr = a2(Re*)6 + b2(Re*)5 + c2(Re*)4 + d2(Re*)3 + e2(Re*)2 + f2(Re*) + g2
3 R(1,2,3) Ycr = a3(Re*)6 + b3(Re*)5 + c3(Re*)4 + d3(Re*)3 + e3(Re*)2 + f3(Re*) + g3
GENOFIS (visual) 1 L(1,2,3) Ycr = b1(Re*)5 + c1(Re*)4 + d1(Re*)3 + e1(Re*)2 + f1(Re*) + g1
2 T(1) Ycr = a2(Re*)6 + b2(Re*)5 + c2(Re*)4 + d2(Re*)3 + e2(Re*)2 + f2(Re*) + g2
3 R(1) Ycr = a3(Re*)6 + b3(Re*)5 + c3(Re*)4 + d3(Re*)3 + e3(Re*)2 + f3(Re*) + g3
GENOFIS (competence) 1 L(1,2,3) Ycr = f1Re* + g1
2 T(1,2,3) Ycr = e2(Re*)2 + f2Re* + g2
3 R(1,2,3) Ycr = d3(Re*)3 + e3(Re*)2 + f3Re* + g3

* Both Re* and Ycr are dimensionless.

consequent linear outputs with nine fuzzy rules were used whereas for neural network (ANN) and least square method, respectively whereas
the GENOFIS model all laminar flow sub-parts were assigned to one GENOFIS model trains antecedent parameters (fuzzy sets) and con-
consequent output and the same procedure was performed for the sequent parameters only by the GAs tool. The consequent parameters of
transitional flow and turbulent flow parts by assigning each of them to ANFIS and GENOFIS models for reference, competence, and visual-
one consequent output as polynomial functions (see Table 6). based techniques are presented in Table 5 and Table 6, respectively. In
Table 6, the equations chosen to assume the processes in the consequent
part of the GENOFIS model are presented for three different techniques;
4. Results and discussion these functions were chosen after the optimization of the generalized
consequent part of the GENOFIS presented in Theory establishment
Due to the extremely complex interrelation between the hydro- section, Geno-fuzzy inference system subsection. Artificial intelligence
dynamic forces caused by the fluid and the bed sediment particles with methods such as Artificial Neural Network (ANN), Fuzzy Logic (FL) and
different packing conditions and high turbulence that occurs near the Genetic Programming (GP) are black-box modeling tools. Likewise, the
bed; the developed analytical approaches cannot accurately capture the proposed novel GENOFIS and ANFIS models are also black-box methods
interrelationship between bed-sediment-particle and fluid in motion which transform the input data to output data mathematically (not
easily (Dey and Ali, 2018). The accuracy of Adaptive Neural Fuzzy using physical concepts). In this study, the consequent functions are
Inference System (ANFIS) and novel Geno Fuzzy Inference System generated by published data which was divided into two parts for
(GENOFIS) models in predicting Shields entrainment function (Ycr) is training (calibration) and validation (prediction) phases. Moreover,
investigated by utilizing shear Reynolds number (Re*) as input for three antecedent parameters (fuzzy sets) and consequent parameters (con-
different techniques; namely reference, visual and competence based sequent functions) are optimized with training (calibration) data by
techniques. 123, 73 and 50 data points were used for reference, com- utilizing the GAs tool. Subsequently, GENOFIS model is validated with
petence, and visual-based techniques, respectively. For the develop- remaining (testing) data. Furthermore, knowing that the sediment en-
ment of the model, 72, 43 and 30 data points were used for training trainment behave in a different way following the type of flow condi-
(calibration) phase and remaining 51, 30 and 20 data points were used tions (Ebtehaj et al., 2017; Spiliotis et al., 2018; Wan Mohtar et al.,
for testing phase (prediction) for reference, competence and visual 2018); the novel GENOFIS model trained with the GA and compared to
based techniques, respectively. ANFIS and newly developed GENOFIS the published experimental data (Yalin and Karahan, 1979; Buffington
models were simulated and fuzzy sets of the models are obtained after and Montgomery, 1997; Yalin and da Silva, 2001) generated the best
the training phase for ANFIS and developed GENOFIS models which are optimal functions for the consequent part of the improved SUGENO
depicted in Figs. 6, 7, and 8. Also, the decision matrix (rule bases) is fuzzy inference system as presented in Table 6. Consequently, the novel
demonstrated in Tables 5 and 6. In the ANFIS model, the antecedent GENOFIS model offers a practical solution to mathematically represent
(fuzzy sets) and consequent parameters are trained by an artificial

Fig. 6. ANFIS and GENOFIS models fuzzy sets for Re* under the reference-based approach.

10
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 7. ANFIS and GENOFIS models fuzzy sets for Re* under the visual-based approach.

the behavior of sediment entrainment in all types of individual observed outcome. According to Pianosi and Wagener (2015), any
boundaries and flow conditions by offering different types of functions model input parameter is defined as sensitive to the model output when
at the same time, unlike ANFIS that offers only linear or constant an R2 is higher than 60 (R2>60). R2 is defined as follows,
functions and not both in the same model. The constant parameters of
R2 = 1 − VAR(e)/VAR(q) (12)
functions (consequent outputs) of GENOFIS model for laminar, transi-
tional and turbulent flows with reference, competence, and visual- where VAR (e) defines the variance of training stochastic terms or er-
based techniques after GA training and optimization processes, are rors and VAR (q) defines the variance of observed Shields entrainment
presented in Tables 7 and 8, respectively. In this study, we applied the values. The confidence ranges were defined. Moreover, for the con-
PAWN approach to investigate the sensitivity for Shear Reynolds fidence level definition; in this study, an approach discussed by Brase
number (Re*) as a sole input for all performed models in all hydro- and Brase (2009) was applied. The confidence band for the dataset
dynamic flow conditions. PAWN approach was explained in details by means μ is defined as,
Pianosi and Wagener (2015). PAWN is an abbreviation derived from − s − s
Pianosi and Wagener. PAWN approach allows investigating the pro- [x − tc , x + tc
n n (13)
pagation of Shear Reynolds number (Re*) uncertainty, and sensitivity
when used a sole input to predict the Shields entrainment function where tc defines the critical values for the data distribution for con-

under hydrodynamically smooth, transitional and rough flow condi- fidence level c. x defines the mean value, s is the sample standard
tions. We, therefore, consider the model output which is the Shields deviation, n define the number of sampled data. The validation of the
entrainment function when the predicted Ycr values are highly corre- confidence interval is qualified as a dynamic approach in the categor-
lated to the observed Ycr values under feasible ranges or confidence ization process because it requires executing the model as also stressed
ranges of variation of Re* as given in Table 9. The confidence ranges are by Banks (1998) and Petty (2012). The validation process involves a
computed and defined in a way that model training with the defined comparative approach (Petty, 2012); during this process, an observed
range would exhibit a sensible output performance. The PAWN ap- Shields entrainment function value (Ycr) is compared to the Shield en-
proach to defining the sensibility of model parameters uses a coefficient trainment value (Ycr) obtained when the model is executed for the
of determination (R2) that is evaluated between computed and specific confidence range or band and the value performance is re-
corded. Moreover, R2is calculated for the computed and observed

Fig. 8. ANFIS and GENOFIS models fuzzy sets for Re* under the competence-based approach.

11
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Table 7
Gaussian-type MF parameters for the GENOFIS model.
Incipient motion approach Laminar Transitional

a1 b1 c1 d1 e1 f1 g1 a2 b2 c2 d2 e2 f2 g2

Competence approach 0 0 0 0 0 −0.06 0.13 0 0 0 0 0.06 −0.21 0.21


Visual approach 0 −0.06 0.01 0.11 −0.03 0.1 0.13 −74.2 773.4 −3346.7 7696.7 9921.7 6796.7 −1932.9
Reference approach −0.011 9.10−5 0.04 0.01 −0.04 −0.03 −0.03 18.14 −147.09 492.52 −871.8 860.59 −449.27 96.94

outcome with an estimate of a confidence limit, separately for in- Table 10. Consequently; this present study has chosen GENOFIS as the
dividual boundaries generated from the confidence bands and flow best model following its smaller RMSE and the larger CE values as
conditions presented in this study. The results are presented in Table 9 aforementioned and presented in Table 10. Based on these indicator
accordingly. Pandis (2015) discussed in detail how to calculate con- criteria, the prediction accuracy of the Shields entrainment function by
fidence limits. The parametric outcomes described in Table 9 for the GENOFIS model is much better than the ANFIS model for reference,
sensitivity analysis by using the PAWN approach, indicate an increase competence, and visual-based techniques. According to the calculated
of the sensitivity for the shear Reynolds number (Re*) as an input to- CE values from prediction values of ANFIS and developed GENOFIS
wards the output which is the Shields entrainment function (Ycr) when models, both models have exhibited very good performance in terms of
GENOFIS was performed. This means that the GENOFIS model has Shields entrainment function prediction. However, GENOFIS model
successfully better trained even data samples that remain very distant performed superior to the ANFIS model for all Shields entrainment
from the mean of the sampled data than the ANFIS model. Performance functions (Ycr). Furthermore, RMSE values calculated from GENOFIS
of these two models was evaluated using RMSE and CE as indicator model results are lower than those calculated from the ANFIS model
criteria. Good prediction accuracy was yielded by the ANFIS and results for reference, visual and competence based techniques. RMSE
GENOFIS approaches where solely shear Reynolds number was used as values of ANFIS and GENOFIS models obtained from the reference
input into the models. based data are higher than in visual and competence based models
Results of RMSE and CE values are illustrated in Table 10. It can be because of previous discussed high fluctuations available in Buffington
seen that RMSE values of ANFIS model for reference, competence, and and Montgomery data as it can be seen from the standard deviation in
visual-based techniques are found to be Table 2.
6.84 × 10−3, 4.77 × 10−3, 2.19 × 10−3 , whereas 2.33 × 10−3, This definitely indicates that the predicted Ycr values by novel
1.93 × 10−3, 9.35 × 10−4 are the RMSE values of GENOFIS model for GENOFIS model are significantly better than those of Ycr values pre-
reference, competence, and visual-based techniques are found to be, dicted by the ANFIS model. By referring to Donigian and Love (2003)
respectively. Obviously, the GENOFIS model provided relatively and Altunkaynak and Nigussie (2015) and taking evaluation criteria
smaller RMSE values than the ANFIS model for reference, competence, RMSE and CE into consideration; the performance of ANFIS model for
and visual-based techniques. This implies that the GENOFIS model prediction of Shields entrainment function under reference, competence
yielded better prediction accuracy of Shields entrainment function than and visual based techniques is good whereas the prediction perfor-
the ANFIS model for all techniques. For the ANFIS model, CE values mance of GENOFIS model is very good.
under reference, competence, and visual-based techniques are calcu- Additionally, the predictions of Shields entrainment function (Ycr)
lated as 0.95, 0.98 and 0.98, and for the GENOFIS model CE values are by ANFIS and GENOFIS and corresponding observed values of Shields
found as 0.98, 0.99 and 0.99, respectively. The GENOFIS model has entrainment function versus shear Reynolds number (Re*) values are
relatively smaller RMSE values and relatively bigger CE values com- plotted in a logarithmic scale as shown in Figs. 9, 10 and 11, respec-
pared to the ANFIS model. Moreover, a number of inputs, processing tively. Considering Re* and referring to Shields (1936), Yalin and
time and IF-THEN rules number are presented in Table 10 for com- Karahan (1979) and many other researchers; in the present study ANFIS
parative purposes. Both ANFIS and novel GENOFIS used one single and novel GENOFIS results show that for small Re* numbers (Re*≤ 1.5),
input, however in order to reach the final results, 9, 5 and 9 IF-THEN Shields entrainment values are strongly dependent on shear Reynolds
rules have been applied for ANFIS based models for reference, visual number (Re*). For larger values of shear Reynolds number, Shields
and competence-based approach respectively whereas, 3,3 and 3 IF- entrainment values become independent from shear Reynolds numbers
THEN rules have been applied for GENOFIS based models for reference, and approximate to a constant value for Re* ≥ 3. According to Fig. 9 the
visual and competence-based approach respectively. This shows that predicted Shields entrainment values by GENOFIS model tend to follow
GENOFIS model has reduced significantly the entanglement that results exactly the observed Shields entrainment data of reference approach
from using more IF-THEN rules in a given fuzzy model (Jovanovic et al., while the ANFIS model prediction values also captured the observed
2004; Akrami et al., 2013). In Table 4, also the ANFIS and GENOFIS Shields entrainment data. Clearly, the GENOFIS model results follow
processing time are presented for reference, visual and competence- the corresponding observed Shields entrainment data of reference ap-
based approaches. Slightly ANFIS based models have exhibited shorter proach closer than the ANFIS model results. As regards the visual ap-
periods for the processing of the final results compared to the novel proach, Fig. 10 revealed that the prediction values of GENOFIS model
GENOFIS but overall the novel GENOFIS based model exhibited better quiet captured the observed Shields entrainment data better than the
performance than ANFIS following the evaluation criteria shown in ANFIS model results. For the competence approach, Fig. 11 shows that

Table 8
Gaussian-type MF parameters for the GENOFIS model.
Incipient motion approach Turbulent

a3 b3 c3 d3 e3 f3 g3

Competence approach 0 0 0 0.004 −0.043 0.145 −0.13


Visual approach −0.07 1.11 −6.79 22.03 39.97 38.43 −15.25
Reference approach −3481 45,504 −2.4.105 7.1.105 -106 106 −3.6.105

12
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Table 9
Model parameters of the ANFIS and GENOFIS models, their units of measurements and confidence ranges and levels.
Model type Flow conditions Input parameters uom Sensitivity (PAWN) Individual boundary Confidence limit (%)
ANFIS GENOFIS ANFIS GENOFIS

Smooth Re* Dimensionless 0.97 0.98 [-1.3, 1.2] 95% 95%


Reference based model Transitional Re* Dimensionless 0.77 0.85 [1.1, 1.7] 75% 75%
Rough Re* Dimensionless 0.86 0.88 [1.65, 2.7] 85% 85%
Visual based model Smooth Re* Dimensionless 0.96 0.98 [-1.2, 1.2] 95% 95%
Transitional Re* Dimensionless 0.95 0.98 [1.15, 1.94] 95% 95%
Rough Re* Dimensionless 0.84 0.89 [1.8, 2.8] 80% 80%
Competence based model Smooth Re* Dimensionless 0.98 1 [0.4, 1.2] 95% 95%
Transitional Re* Dimensionless 0.94 1 [1.15, 1.92] 92% 92%
Smooth Re* Dimensionless 0.96 1 [1.85, 4.28] 95% 95%

Re* is dimensionless, PAWN is defined by R2 and is dimensionless and uom is unity of measurement.

the GENOFIS model results are exactly matched with the corresponding encapsulate the findings of Yalin and Karahan (1979), Miller et al.
observed Shields entrainment data of competence approach under (1977) and Yalin and da Silva (2001) where Shields entrainment values
smooth, transitional and rough flow conditions. Whereas the ANFIS under rough flow conditions were found to be 0.045, 0.047 and 0.043,
model values follow the observed data closely for smooth flow condi- respectively. On top of that the novel GENOFIS model offers the ability
tion but particularly for transitional condition over predicted the ob- to define the Shields entrainment function (Ycr) that initiates the sedi-
served data and for rough flow condition under predicted observed ment entrainment for each and every defined corresponding shear
data. This means that the prediction performance of the ANFIS model is Reynolds number (Re*) in all flow conditions, unlike previous re-
particularly fair for competence approach under transitional and rough searchers who defined the critical dimensionless shear stress in turbu-
flow conditions. However, it has been observed that the performance of lent flows only (Shields, 1936; Yalin and Karahan, 1979; Yalin and da
GENOFIS model is enhanced with relatively larger sets of experimental Silva, 2001; Cao et al., 2006; Tamburrino et al., 2016). The results of
data. Besides, GENOFIS performed very well for visual and reference this present study revealed that developed GENOFIS model out-
based models compared to reference-based data. This is because the performed the ANFIS model in predicting Shields entrainment data in
used data of Yalin and Karahan (1979) and Yalin and da Silva (2001) all techniques for smooth, transitional and rough flow conditions. The
(visual and reference based models) was collected under deterministic observed Shields’s entrainment function versus predicted Shields’ en-
approaches whereas reference-based data used herein was collected by trainment function by ANFIS and GENOFIS models for the reference,
using a deterministic approach considering only statistical values. competence, and visual-based approaches are plotted with the exact
Therefore, it was not possible for the reference-based approach to re- line (1:1) in Fig. 12. The visual investigation of these figures indicates a
present the extreme and lowest crisp values which eventually help to strong consistency exhibited between predicted Shields’ entrainment
define fuzzy sets and allow capturing the available initial and extreme function values by ANFIS and GENOFIS models and corresponding
Shield entrainment conditions in fuzzification process. The high viscous observed Shields entrainment function data as shown in Fig. 12. As can
forces present in smooth flow condition were overcome by assigning be seen from Fig. 12a, the prediction results of the GENOFIS model for
higher fuzzy membership degrees to lower shear Reynolds numbers in the reference based approach quietly scattered around the exact line
order to make them be part of the fuzzy Rough 3 (R3) set which cor- while the scattering is slightly more for the predicted values by ANFIS
responds to extreme Shield entrainment values. Gaussian MF types around exact line compared to GENOFIS model results. This means that
were used in all hydrodynamic flow conditions because of the ad- GENOFIS model has a very good performance in terms of prediction of
vantage of Gaussian distribution in which variables tend towards a Shields entrainment function for the reference-based approach. Like-
normal distribution. The latter was therefore very useful because of the wise, Fig. 12b illustrates the results of ANFIS and GENOFIS models and
available experimental data used herein that had considerable fluc- corresponding observed data of the competence-based approach dis-
tuations (Buffington and Montgomery, 1997). Using developed GEN- tributed around the exact line. The predicted values of GENOFIS mode
OFIS model for high shear Reynolds values in gravel-bedded rivers matched almost exactly with the corresponding observed data of
(gravel-bedded rivers which were used by Buffington and Montgomery, competence-based approach; while the ANFIS model results exhibit a
1997), Shield entrainment values varied in range 0.056–0.076 and good agreement with observed data which implies that, GENOFIS
completely stayed in the range described by Buffington and model has relatively better accuracy than ANFIS model for competency-
Montgomery (1997) in gravel-bedded river under rough flow condi- based approach. This is to say that the developed GENOFIS model
tions. Additionally, GENOFIS model performed very well for visual and outperformed the ANFIS model for prediction of Shields’ entrainment
competence-based models where Shields entrainment values for rough function using observed data of the reference and competence-based
flow conditions varied between 0.042 and 0.048 and 0.041–0.046 for approaches. Similarly, the predicted values by GENOFIS model re-
visual and competence, respectively. These results completely markably follow the data of visual-based approach as depicted in

Table 10
Performance Comparison of ANFIS and GENOFIS models in terms of the RMSE and CE values.
Reference based approach Visual based approach Competence based approach

Evaluation criterio ANFIS GENOFIS ANFIS GENOFIS ANFIS GENOFIS

Number of inputs 1 1 1 1 1 1
Processing time 586 936 476 742 425 695
IF-THEN rules 9 3 5 3 9 3
RMSE 6.84 × 10−3 2.33 × 10−3 4.77 × 10−3 1.93 × 10−3 2.19 × 10−3 9.35 × 10−4
CE 0.95 0.98 0.98 0.99 0.98 0.99

*
RMSE is dimensionless and *Processing time is in seconds.

13
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 9. Shields Entrainment Function of observed and predicted dimensionless shear stress of the Buffington and Montgomery collective data for reference approach.

Fig. 6c. On the other hand, the results of the ANFIS model is slightly time. In general, the training process using Genetic Algorithms (GAs)
scattered around the exact line referring to the relatively lower per- for GENOFIS manifested relatively longer output processing time while
formance of the ANFIS model for visual-based approach data. Conse- converging to the final solution compared to Artificial Neural Networks
quently, the results of this present study indicate that the incorporation based ANFIS.
of modified Sugeno FIS approach and GAs tool further improved the
accuracy of GENOFIS model significantly for the reference, compe-
5. Conclusions
tence, and visual-based approaches. It is to note that the developed
GENOFIS model can accurately predict Shields entrainment function
Adaptive Neural Fuzzy Inference System (ANFIS) tool is based on
better than the ANFIS model without expensive computation effort for
Sugeno Fuzzy Inference System (Sugeno FIS). However, the ANFIS
all three different techniques that were used to define the threshold of
model has a shortcoming in defining the consequent part of the model.
the sediment incipient motion under smooth, transitional and rough
That is, it could only be represented either as a constant or a linear
flow conditions. The GENOFIS and ANFIS model results are compared
function. This implies that ANFIS is not capable of describing con-
with experimental data quantitatively and qualitatively and close si-
sequent part of the model simultaneously as a constant, linear, non-
milarity is found under all hydrodynamic flow conditions. Additionally,
linear or combination of them. In this present study, Geno-Fuzzy
the processing time of a model is so important in order to be more
Inferences System (GENOFIS) is introduced with the incorporation of
useful nowadays. In this study, the processing time used for the training
improved Sugeno FIS and Genetic Algorithms (GAs) tool, which im-
process of both ANFIS and GENOFIS for 1000 iterations was recorded.
proved the accuracy significantly, in predicting the Shields’ entrain-
586 s, 476 s, and 425 s were recorded for the training process of re-
ment values. The ANFIS and developed GENOFIS models are im-
ference, visual and competence based ANFIS models whereas 936 s,
plemented to predict sediment incipient motion by utilizing the
742 s and 695 s were recorded for the training process of reference,
reference, competence, and visual-based data under smooth, transi-
visual and competence based GENOFIS models. Consequently, compe-
tional and rough flow conditions, respectively. The prediction perfor-
tence based ANFIS model has the least output processing time whereas
mance of the ANFIS and GENOFIS models is evaluated quantitatively by
Competence-based GENOFIS model has the highest output processing
taking calculated root-mean-square error (RMSE) and coefficient of

Fig. 10. Shields Entrainment Function of observed and predicted dimensionless shear stress of the Yalin and Karahan collective data for visual approach.

14
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Fig. 11. Shields Entrainment Function of observed and predicted dimensionless shear stress of the Yalin and da Silva data for competence approach.

Fig. 12. Comparison of Shields entrainment function results predicted by ANFIS and GENOFIS and observed Shields entrainment function data for three different
approaches for reference-based approach (a), competence based approach (b) and visual-based approach (c).

15
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

efficiency (CE) values into consideration. Clearly, developed GENOFIS Banks, J., 1998. Principles. Handbook of simulation: principles, methodology, advances, ap-
model outperformed the ANFIS model for the reference, competence, plications, and practice. John Wiley & Sons, Atlanta, Georgia.
Bowden, J.H., Nolte, C.G., Otte, T.L., 2013. Simulating the impact of the large-scale
and visual-based data under smooth, transitional and rough flow con- circulation on the 2-m temperature and precipitation climatology. Clim. Dyn. 40
ditions based on RMSE and CE values as performance criteria. (7–8), 1903–1920.
Moreover, the prediction performance of the ANFIS and GENOFIS Brase H., C., Brase P., C., 2009. Concepts and Methods. Understandable Statistics, ninth
ed. Houghton Mifflin, Boston, MA.
models is investigated in terms of visual analysis. The ANFIS and novel Buffington, J.M., 1999. The legend of AF Shields. J. Hydraul. Eng. 125 (4), 376–387.
GENOFIS performances were evaluated in predicting the instantaneous Buffington, J.M., Montgomery, D.R., 1998. Correction to “A systematic analysis of eight
Shields’ entrainment values (Ycr), which is defined in terms of shear decades of incipient motion studies, with special reference to gravel-bedded rivers”.
Water Resour. Res. 34 (1), 157.
Reynolds number (Re*) in sand-bedded open channels under unidirec- Buffington, J.M., Montgomery, D.R., 1997. A systematic analysis of eight decades of in-
tional flow regime. The GENOFIS exhibited higher performance than cipient motion studies, with special reference to Gravel-bedded Rivers. Water Resour.
conventional ANFIS model in predicting the Shields entrainment values Res. 33 (8), 1993–2029.
Cao, Z., Pender, G., Meng, J., 2006. Explicit formulation of the Shields diagram for in-
for all considered data under smooth, transitional and rough flow
cipient motion of sediment. J. Hydraul. Eng. 132 (10), 1097–1099.
conditions. In this present study, sand beds with uniform packing and Carson, M.A., Griffiths, G.A., 1985. Tractive stress and the onset of bed particle movement
direction under unidirectional flow were considered. These considera- in gravel stream channels: Different equations for different purposes. J. Hydrol. 79
tions helped the fuzzy expert system to define the Shields’ entrainment (3–4), 375–388.
Chiu, S., 1996. Method and software for extracting fuzzy classification rules by sub-
values for the initiation of motion from weak to general motion of se- tractive clustering. In: Proceedings of North American Fuzzy Information Processing.
diments under all hydrodynamic flow conditions and to define the fuzzy IEEE, pp. 461–465.
transition of motion that varies from weak to general motion. The fuzzy Dey, S., Ali, S.Z., 2018. Advances in modeling of bed particle entrainment sheared by
turbulent flow. Phys. Fluids 30 (6), 061301.
transition was defined with Shields entrainment values in terms of Dey, S., Das, R., Gaudio, R., Bose, S.K., 2012. Turbulence in mobile-bed streams. Acta
fuzzy sets. As a result, developed GENOFIS is recommended to be used Geophys. 60 (6), 1547–1588.
in predicting the Shields entrainment values in gravel and sand bedded Donigian, A.S., Love, J.T., 2003. Sediment calibration procedures and guidelines for
watershed modeling. In: Proceedings of the Water Environment Federation, 2003(4),
open channels with uniform packing and direction of sediment particles 728–747.
under unidirectional flow and in all hydrodynamic flow conditions. Ebtehaj, I., Bonakdari, H., 2014. Comparison of genetic algorithm and imperialist com-
Declaration of Competing Interest petitive algorithms in predicting bed load transport in clean pipe. Water Sci. Technol.
70 (10), 1695–1701.
The authors declare that there is no conflict of interests regarding Ebtehaj, I., Bonakdari, H., Khoshbin, F., Bong, C.H.J., Ab Ghani, A., 2017. Development of
publishing this article. group method of data handling based on genetic algorithm to predict incipient mo-
tion in rigid rectangular storm water channel. Sci. Iran. Trans. A Civ. Eng. 24 (3),
1000.
Acknowledgment
Elhakeem, M., Papanicolaou, A.T., Tsakiris, A.G., 2017. A probabilistic model for sedi-
ment entrainment: The role of bed irregularity. Int. J. Sedim. Res. 32 (2), 137–148.
The authors are very thankful to the Editor-in-Chief and all Fenton, J.D., Abbott, J.E., 1977. Initial movement of grains on a stream bed: The effect of
Reviewers for their insightful comments that helped to improve the relative protrusion. Proc. Royal Soc. Lond. A. Math. Phys. Sci. 352 (1671), 523–537.
Gholami, A., Bonakdari, H., Ebtehaj, I., Akhtari, A.A., 2017. Design of an adaptive neuro-
clarity of the final paper. fuzzy computing technique for predicting flow variables in a 90° sharp bend. J.
Hydroinform. https://doi.org/10.2166/hydro.2017.200. jh2017200.
References Giménez-curto, L.A., Corniero, M.A., 2009. Entrainment threshold of cohesionless sedi-
ment grains under steady flow of air and water. Sedimentology 56 (2), 493–509.
Gustafson, D.E., Kessel, W.C., 1979. Fuzzy clustering with a fuzzy covariance matrix. In:
Abraham, A., 2001. Neuro-fuzzy systems: State-of-the-art modeling techniques. In: 1978 IEEE conference on decision and control including the 17th symposium on
International Work-Conference on Artificial Neural Networks. Springer, Berlin, adaptive processes. IEEE, pp. 761–766.
Heidelberg, pp. 269–276. Hadi, S.J., Tombul, M., 2018. Monthly streamflow forecasting using continuous wavelet
Abraham, A., Nath, B., 2000. Designing optimal neuro-fuzzy architectures for intelligent and multi-gene genetic programming combination. J. Hydrol. 561, 674–687.
control. In 6th International Conference on Control, Automation, Robotics, and Hjulstrom, F., 1935. Studies of the morphological activity of rivers as illustrated by the
Vision. River Fyris, Bulletin. Geol. Inst. Upsalsa 25, 221–527.
Akrami, S.A., El-Shafie, A., Jaafar, O., 2013. Improving rainfall forecasting efficiency Hutter, K., Wang, Y., 2016. Dimensional analysis, similitude and physical experiments at
using modified adaptive neuro-fuzzy inference system (MANFIS). Water Resour. laboratory scale. In: Fluid and Thermodynamics. Springer, Cham, pp. 537–607.
Manage. 27 (9), 3507–3523. Inman, D.L., 1949. Sorting of sediments in the light of fluid mechanics. J. Sediment. Res.
Ali, S.Z., Dey, S., 2016. Hydrodynamics of sediment threshold. Phys. Fluids 28 (7), 19 (2), 51–70.
075103. Jang, J.S., 1992. Neuro-Fuzzy Modeling: Architectures, Analyses, and Applications.
Altunkaynak, A., 2010. A predictive model for well loss using fuzzy logic approach. University of California, Berkeley.
Hydrol. Process. 24 (17), 2400–2404. Jang, J.S., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst.
Altunkaynak, A., 2013. Prediction of significant wave height using geno-multilayer per- Man Cybern. 23 (3), 665–685.
ceptron. Ocean Eng. 58, 144–153. Jiang, Z., Haff, P.K., 1993. Multiparticle simulation methods applied to the micro-
Altunkaynak, A., 2009. Sediment load prediction by genetic algorithms. Adv. Eng. Softw. mechanics of bed load transport. Water Resour. Res. 29 (2), 399–412.
40 (9), 928–934. Jovanovic, B.B., Reljin, I.S., Reljin, B.D., 2004. Modified ANFIS architecture-improving
Altunkaynak, A., 2008. Adaptive estimation of wave parameters by Geno-Kalman fil- efficiency of ANFIS technique. In: 7th Seminar on Neural Network Applications in
tering. Ocean Eng. 35 (11–12), 1245–1251. Electrical Engineering, 2004. NEUREL 2004. IEEE, pp. 215–220.
Altunkaynak, A., Şen, Z., 2007. Fuzzy logic model of lake water level fluctuations in Lake Kasabov, N., 1999. Evolving connectionist and fuzzy connectionist systems–theory and
Van, Turkey. Theor. Appl. Climatol. 90 (3–4), 227–233. applications for adaptive, on-line intelligent systems. Neuro-Fuzzy Techniques for
Altunkaynak, A., Nigussie, T.A., 2015. Prediction of daily rainfall by a hybrid wavelet- Intelligent Information Processing. Physica Verlag.
season-neuro technique. J. Hydrol. 529, 287–301. Kennard, R.W., Stone, L.A., 1969. Computer aided design of experiments. Technometrics
Altunkaynak, A., Wang, K.H., 2012. Estimation of significant wave height in shallow lakes 11 (1), 137–148. https://doi.org/10.1080/00401706.1969.10490666.
using the expert system techniques. Expert Syst. Appl. 39 (3), 2549–2559. Kitsikoudis, V., Sidiropoulos, E., Hrissanthou, V., 2015. Assessment of sediment transport
Andrews, E.D., 1994. Marginal bed load transport in a gravel bed stream, Sagehen Creek, approaches for sand-bed rivers by means of machine learning. Hydrol. Sci. J. 60 (9),
California. Water Resourc. Res. 30 (7), 2241–2250. 1566–1586.
Ashworth, P.J., Ferguson, R.I., 1989. Size-selective entrainment of bed load in gravel bed Kennedy A., A., 1995. The Albert Shields story. J. Hydraul. Eng. 121 (11), 766–772.
streams. Water Resour. Res. 25 (4), 627–634. Kitsikoudis, V., Sidiropoulos, E., Hrissanthou, V., 2014. Machine learning utilization for
Ashworth, P.J., Ferguson, R.I., Ashmore, P.E., Paola, C., Powell, D.M., Prestegaards, K.L., bed load transport in gravel-bed rivers. Water Resour. Manage. 28 (11), 3727–3743.
1992. Measurements in a braided river chute and lobe: 2. Sorting of bed load during Kitsikoudis, V., Spiliotis, M., Hrissanthou, V., 2016. Fuzzy regression analysis for sedi-
entrainment, transport, and deposition. Water Resour. Res. 28 (7), 1887–1896. ment incipient motion under turbulent flow conditions. Environ. Process. 3 (3),
Azamathulla, H.M., Ghani, A.A., Fei, S.Y., 2012. ANFIS-based approach for predicting 663–679.
sediment transport in clean sewer. Appl. Soft Comput. 12 (3), 1227–1230. Komar, P.D., Clemens, K.E., 1986. The relationship between a grain's settling velocity and
Azamathulla, H.M., Chang, C.K., Ghani, A.A., Ariffin, J., Zakaria, N.A., Hasan, Z.A., 2009. threshold of motion under unidirectional currents. J. Sediment. Res. 56 (2), 258–266.
An ANFIS-based approach for predicting the bed load for moderately sized rivers. J. Kramer, H., 1935. Sand mixtures and sand movement in fluvial model. Trans. Am. Soc.
Hydro-environ. Res. 3 (1), 35–44. Civil Eng. 100 (1), 798–838.
Bizimana H., Demir F., Sönmez O., 2016. Modeling of water level changing of Yuvacik Lajiness, M., Watson, I., 2008. Dissimilarity-based approaches to compound acquisition.
dam with fuzzy logic. 4th Int.Sy. On innovative technologies in Engineering and Curr. Opin. Chem. Biol. 12 (3), 366–371. https://doi.org/10.1016/j.cbpa.2008.03.
Science, 3-4 November 2016, Alanya, Turkey. 010.

16
H. Bizimana and A. Altunkaynak Journal of Hydrology 577 (2019) 123952

Maguire, L.P., Roche, B., McGinnity, T.M., McDaid, L.J., 1998. Predicting a chaotic time using machine-learning techniques. Water Resour. Res. 45 (12).
series using a fuzzy neural network. Inf. Sci. 112 (1–4), 125–136. Spiliotis, M., Kitsikoudis, V., Kirca, V.O., Hrissanthou, V., 2018. Fuzzy threshold for the
Miller, M.C., McCave, I.N., Komar, P., 1977. Threshold of sediment motion under uni- initiation of sediment motion. Appl. Soft Comput. 72, 312–320.
directional currents. Sedimentology 24 (4), 507–527. Sugeno, M., 1985. An introductory survey of fuzzy control. Inf. Sci. 36 (1–2), 59–83.
Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L., Sugeno, M., Kang, G.T., 1988. Structure identification of fuzzy model. Fuzzy Sets Syst. 28
2007. Model evaluation guidelines for systematic quantification of accuracy in wa- (1), 15–33.
tershed simulations. Trans. ASABE 50 (3), 885–900. Sundborg, Å., 1956. The river Klarälven a study of fluvial processes. Geogr. Ann. 38 (2–3),
Moosavi, V., Vafakhah, M., Shirmohammadi, B., Behnia, N., 2013. A wavelet-ANFIS hy- 125–316.
brid model for groundwater level forecasting for different prediction periods. Water Sutherland, A.J., 1967. Proposed mechanism for sediment entrainment by turbulent
Resour. Manage. 27 (5), 1301–1321. flows. J. Geophys. Res. 72 (24), 6183–6194.
Najafzadeh, M., Tafarojnoruz, A., Lim, S.Y., 2017. Prediction of local scour depth Tagaki, T., Sugeno, M., 1985. Fuzzy identification of systems and its application to
downstream of sluice gates using data-driven models. ISH J. Hydraul. Eng. 23 (2), modelling and control. IEEE Trans. Syst. Man and Cybernetics 15 (1), 116–132.
195–202. Tamburrino, A., Carrillo, D., Negrete, F., Ihle, C.F., 2016. Critical shear stress for incipient
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models part motion of non-cohesive particles in open-channel flows of pseudoplastic fluids. Can.
I—A discussion of principles. J. Hydrol. 10 (3), 282–290. J. Chem. Eng. 94 (6), 1084–1091.
Nasr, A., Bruen, M., 2008. Development of neuro-fuzzy models to account for temporal Tano, S.I., Oyama, T., Arnould, T., 1996. Deep combination of fuzzy inference and neural
and spatial variations in a lumped rainfall–runoff model. J. Hydrol. 349 (3–4), network in fuzzy inference software—FINEST. Fuzzy Sets Syst. 82 (2), 151–160.
277–290. Tütmez, B., Dağ, A., Tercan, A.E., Kaymak, U., 2007. Lignite thickness estimation via
Neill, C.R., Yalin, M.S., 1969. Quantitative definition of beginning of bed movement. J. adaptive fuzzy-neural network. Mining Congress Turkey 151.
Hydraul. Div. 95 (1), 585–588. Turowski, J.M., Badoux, A., Rickenmann, D., 2011. Start and end of bedload transport in
Özger, M., 2009. Comparison of fuzzy inference systems for streamflow prediction. gravel-bed streams. Geophys. Res. Lett. 38 (4).
Hydrol. Sci. J. 54 (2), 261–273. Uyumaz, A., Altunkaynak, A., Özger, M., 2006. Fuzzy logic model for equilibrium scour
Ouyang, H.T., 2017. Optimization of autoregressive, exogenous inputs-based typhoon downstream of a dam's vertical gate. J. Hydraul. Eng. 132 (10), 1069–1075.
inundation forecasting models using a multi-objective genetic algorithm. Eng. Optim. Vanoni, V.A., 2006. ASCE Manuals and Reports on Engineering No. 54, Sedimentation
49 (7), 1211–1225. Engineering. ASCE, Virginia, USA.
Paphitis, D., 2001. Sediment movement under unidirectional flows: an assessment of Vollmer, S., Kleinhans, M.G., 2007. Predicting incipient motion, including the effect of
empirical threshold curves. Coast. Eng. 43 (3–4), 227–245. turbulent pressure fluctuations in the bed. Water Resour. Res. 43 (5).
Pandis, N., 2015. Statistical inference with confidence intervals. Am. J. Orthod. Dentofac. Wang, K.H., Altunkaynak, A., 2011. Comparative case study of rainfall-runoff modeling
Orthop. 147 (5), 632–634. between SWMM and fuzzy logic approach. J. Hydrol. Eng. 17 (2), 283–291.
Petty D., M., 2012. Calculating and using confidence intervals for model validation. In: In Wan Mohtar, W.H.M., Afan, H., El-Shafie, A., Bong, C.H.J., Ab Ghani, A., 2018. Influence
Proceedings of the Fall 2012 Simulation Interoperability Workshop, pp. 10–14. of bed deposit in the prediction of incipient sediment motion in sewers using artificial
Pianosi, F., Wagener, T., 2015. A simple and efficient method for global sensitivity ana- neural networks. Urban Water J. 15 (4), 296–302.
lysis based on cumulative distribution functions. Environ. Model. Softw. 67, 1–11. Wathen, S.J., Ferguson, R.I., Hoey, T.B., Werritty, A., 1995. Unequal mobility of gravel
Remesan, R., Shamim, M.A., Han, D., Mathew, J., 2009. Runoff prediction using an in- and sand in weakly bimodal river sediments. Water Resour. Res. 31 (8), 2087–2096.
tegrated hybrid-modelling scheme. J. Hydrol. 372 (1–4), 48–60. Wilcock, P.R., 1988. Methods for estimating the critical shear stress of individual fractions
Schiereck, G.J., 2017. Introduction to Bed, Bank and Shore Protection. CRC Press. in mixed-size sediment. Water Resour. Res. 24 (7), 1127–1135.
Şen, Z., Altunkaynak, A., 2006. A comparative fuzzy logic approach to runoff coefficient Yalin, M.S., 1977. Mechanics of Sediment Transport, second ed. Pergamon Press.
and runoff estimation. Hydrol. Process. Int. J. 20 (9), 1993–2009. Yalin, M.S., Da Silva, A.M.F., 2001. Fluvial processes. IAHR monograph. International
Shields, A., 1936. Anwendung der Ahnlichkeistmechanik und der turbulenzforschung auf Association for Hydraulic Research, Delft.
die geschiebebewegung, Mitteilungen der Preuss. Versuchsanst. f. Wasserbau u. Yalin, M.S., Karahan, E., 1979. Inception of sediment transport. J. Hydraul. Div. 105 (11),
Schiffhau, Berlin, Heft 26. (In German). 1433–1443.
Shvidchenko, A.B., Pender, G., 2000. Flume study of the effect of relative depth on the Ying, H., 1999. Analytical analysis and feedback linearization tracking control of the
incipient motion of coarse uniform sediments. Water Resour. Res. 36 (2), 619–628. general Takagi-Sugeno fuzzy dynamic systems. IEEE Trans. Syst. Man Cyber. Part C
Sonmez, O., Bizimana, H., 2018. Flood hazard risk evaluation using fuzzy logic and (Applications and Reviews) 29 (2), 290–298.
weightage based combination methods in Geographic Information System (GIS). Sci. Zounemat-Kermani, M., Meymand, A.M., Ahmadipour, M., 2018. Estimating incipient
Iran. motion velocity of bed sediments using different data-driven methods. Appl. Soft
Solomatine, D.P., Shrestha, D.L., 2009. A novel method to estimate model uncertainty Comput. 69, 165–176.

17

You might also like