FB

TD

a

ARRA

KAWGSMECS

1

ldncduoc[bltpwgmt

0h

Applied Ocean Research 53 (2015) 208–217

Contents lists available at ScienceDirect

Applied Ocean Research

journal homepage: www.elsevier.com/locate/apor

orecasting of chlorophyll-a concentrations in South San Franciscoay using five different models

aher Rajaee, Amir Boroumand ∗

epartment of Civil Engineering, University of Qom, Qom, Iran

r t i c l e i n f o

rticle history:eceived 8 January 2015eceived in revised form 31 August 2015ccepted 1 September 2015

eywords:NNavelet transforms

AVRultiple linear regressions

a b s t r a c t

Accurate and reliable eutrophication level forecasting models are necessary for characterizing compli-cated water quality processes in bays. In this study, the ability of coupled discrete wavelet transform(DWT) with artificial neural network (ANN) and multi linear regression (MLR) (WANN and WMLR),ANN, MLR and genetic algorithm-support vector regression (GA-SVR) models for chlorophyll-a level fore-casting applications were considered. The data used to develop and validate the models were monthlychlorophyll-a (Chl-a) data recorded from January 1994 to December 2013 were obtained from the NO.36station located in the South San Francisco bay, USA. In the proposed WANN and WMLR models, theobserved time series of Chl-a were decomposed to sub time series at different scales by DWT. After-wards, the sub time series were used as input data to the ANN and MLR systems to predict the 1 monthahead Chl-a. Also the genetic algorithm was linked to SVR models to search for the optimal SVR param-

utrophicationhlorophyll-aan Francisco Bay

eters. The relative performance of the proposed models was compared together and the results showedthat the WANN models were found to provide more accurate monthly Chl-a forecasts compared to theother models. The determination coefficient was 0.87, −0.04, 0.31, −2.36 and 0.24 for the WANN, WMLR,ANN, MLR and GA-SVR models, respectively. In addition, the WANN model predicted extreme Chl-a val-ues precisely. The results indicate that the WANN models are a promising new method for eutrophication

uch a

level forecasting in bays s

. Introduction

Eutrophication has been one of the major water quality prob-ems in estuaries and coastal waters in many countries. In recentecades, human activities have considerably increased the ofutrients delivery to many estuarine and coastal areas. Eutrophiconditions, which include low dissolved oxygen concentrations,eclining sea grasses and harmful algal blooms, may impact theses of estuarine and coastal resources by reducing the successf commercial and sport fisheries, fouling swimming beaches, andausing other problems due to the decay of excess amounts of algae1,2]. Chlorophyll-a concentrations may be used to determine aay’s trophic status. Chlorophyll is the green pigment in plants’

eaves that allows them to create energy light through photosyn-hesis. By measuring chlorophyll, the amount of photosynthesizinglants is indirectly measured. In a bay water sample, these plantsould be algae or phytoplankton. Chlorophyll is a measure of all

reen pigments whether they are alive or dead. Chlorophyll-a is aeasure of the portion of the pigment that is still alive. Sunlight,

emperature, nutrients, and wind all affect both algae numbers and

∗ Corresponding author. Tel.: +98 9125853026.E-mail address: a.broumand@stu.qom.ac.ir (A. Boroumand).

141-1187/$ – see front matter © 2015 Elsevier Ltd. All rights reserved.ttp://dx.doi.org/10.1016/j.apor.2015.09.001

s those found in South San Francisco Bay.© 2015 Elsevier Ltd. All rights reserved.

Chlorophyll-a concentration. During the spring when water beginsto warm, the days are sunnier, and nutrients are still plentiful, thefirst outbreak or “bloom” of algae may occur. As the days becomeincreasingly warmer and sunnier, algae will continue to grow more;however, they may soon outgrow the available supply of nutrients.Consequently, the total amount of algae growth may be limited. Assummer turns to fall and temperature and sunlight decrease, algaeconcentrations will decrease as well. Same pattern of variation canbe seen at San-Francisco Bay.

In recent years, eutrophic condition has been monitored at thebays, in terms of both temporal and spatial variation. At San Fran-cisco Bay, they run a long-term program to monitor chlorophyll-a.The measurements are done monthly at fixed stations in the bay.

Understanding and modeling the level of eutrophication(Chlorophyll-a) can be helpful to estuaries ecosystem management.In this regard, water quality and environmental models have beenused widely to assist water resources’ managers developing controlstrategies for estuarine water quality management.

Nutrients’ load of phosphate and nitrate plays a key role inoutbreak and growth of algal blooms. In other words, phosphate

and nitrate are the indices that control this process. Managementof bays upstream catchment land use, implementation of totalmaximum daily load plan and or nutrients’ load allocation plans(increasing or decreasing the consumption of the fertilizers and

dx.doi.org/10.1016/j.apor.2015.09.001http://www.sciencedirect.com/science/journal/01411187http://www.elsevier.com/locate/aporhttp://crossmark.crossref.org/dialog/?doi=10.1016/j.apor.2015.09.001&domain=pdfmailto:a.broumand@stu.qom.ac.irdx.doi.org/10.1016/j.apor.2015.09.001

d Ocea

ca

doeib

atshemmara(ntsdaio

alpesctAtv(TtpdpistbeBbenpCttfDqo

rssb

T. Rajaee, A. Boroumand / Applie

hemical pesticides in upstream farms) may control eutrophicationt the bays [3].

So catchments nutrient management is an important factor toecrease eutrophication process. The management incompetencyccurs when there is no link between practical catchment mod-ls such as SWAT and water body simulation models includingntelligent models. As a result linking between these models cane assisted for bay water quality managers.

Traditional models used for water bodies’ qualitative modelingre complex models. Putting together the assumption and limita-ions of the models with data uncertainty leads to a very complexystems that make the modeling outcomes uncertain. On the otherand application of these models require sufficient expertise andxperience and users’ skills to calibrate, validate and verify theodels affects the results directly. Hence, the limitation of theseodels on accurate forecasting the level of the algal blooms, and

lso the nonlinear relationship between the water quality and envi-onmental indices and that of the level of Chlorophyll-a necessitate

new method using intelligent models. Artificial neural networksANNs) are the most used intelligent ones. Neural networks offer aumber of advantages, including requiring less formal statisticalraining, ability to implicitly detect complex nonlinear relation-hips between dependent and independent variables, ability toetect all possible interactions between predictor variables, and thevailability of multiple training algorithms. Disadvantages includets “black box” nature, greater computational burden, proneness tover fitting, and the empirical nature of model development [4].

The last decade has seen a tremendous growth in interest in thepplication of ANNs to water resources and environmental prob-ems. In recent years ANNs have found a number of applications inrediction of eutrophic conditions. Lee et al. [5] used ANN to mod-ling of coastal algal blooms of Hong Kong coastal waters. Theyhowed that the algal concentration in the eutrophic sub-tropicaloastal water is mainly dependent on the antecedent algal concen-rations in the previous 1–2 weeks. Their study also showed that anNN model with a small number of input variables is able to capture

rends of algal dynamics, but data with a minimum sampling inter-al of 1 week is necessary. Kuo et al. [6] applied a back-propagationBP) ANN to reservoir eutrophication prediction in central Taiwan.hey showed that the ANN is able to predict water quality indica-ors with reasonable accuracy. Melesse et al. [7] used a multilayererceptron-back propagation (MLP-BP) algorithm of ANN to pre-ict the level of eutrophication (chlorophyll-a) from water qualityarameters monitored at two Florida Bay water quality monitor-

ng stations (FLAB03 and FLAB14). They studied seven input datacenarios, and compared the models performance. They showedhat the prediction with antecedent chlorophyll-a alone gave a sta-le result with smaller error and higher performance attributed toasier and more efficient training. They also showed that the MLP-P technique is applicable to the monitoring and prediction of algallooms and will be crucial to coastal watershed management. Huot al. [8] used ANN model to relate the key factors that influence aumber of water quality indicators such as dissolved oxygen, totalhosphorus, chlorophyll-a and secchi disk depth in Lake Fuxian,hina. Their results indicated that the ANN model performs well inen months prediction and the ANN is able to predict eutrophica-ion indicators with reasonable accuracy. Saghiet al. [9] developed aeed forward ANN model to analysis trophic state index (TSI) in theez Dam reservoir; Iran. They showed that ANN is a suitable tool foruality modeling of dam reservoir and increment and decrementf nutrients in trend of eutrophication.

In recent years, wavelet analysis and ANNs was used for water

esources and environmental engineering problems. Kisi [10] con-idered the accuracy of WANN and single ANN models in monthlytream flow prediction and resulted that the WANN performs muchetter than single ANN. Partal and Cigizoglu [11] applied WANN for

n Research 53 (2015) 208–217 209

prediction of suspended sediment load in rivers. Rajaee et al. [12,13]proposed a model by combining wavelet analysis and the neuro-fuzzy (NF) approach to predict daily suspended sediment load (S). Inthe developed WNF models, the daily observed time series of riverdischarge and S were decomposed into several sub time series atdifferent scales using discrete wavelet transform. The results indi-cated that the WNF model more efficient than the NF and sedimentrating curve models in the prediction of S. In another study, Rajaeeet al. [14] applied a WANN model to Iowa River suspended sedimentload prediction, USA. They showed that wavelet-transformed dataimprove the ability of predicting model by capturing useful infor-mation on various resolution levels. SatyajiRao et al. [15] proposeda WANN model for the prediction of daily runoff in the West flowingRivers of India. The results of daily runoff time series modeling illus-trated that the performances of WANN models are more effectivethan the ANN models.

A number of studies have applied wavelet analysis and ANNsfor eutrophication modeling in lakes. Kim et al. [16] proposed aWANN model to chlorophyll-a concentration forecasting 1, 3, and 7days ahead, in Korea. They showed that WANN models constitute apromising new method for short-term chlorophyll-a concentrationforecasting in large lakes. Wang et al. [17] proposed a hybrid WANNmethod for chlorophyll-a simulation in the lake Baiyangdian, NorthChina. They compared the performance of the proposed WANNmodel for monthly chlorophyll-a simulation in the lake ecosys-tem with a multiple stepwise linear regression (MSLR) model, anautoregressive integrated moving average (ARIMA) model and aregular ANN model. Their results showed that the WANN modelwas suitable for chlorophyll-a simulation providing a more accu-rate performance than the MSLR, ARIMA, and ANN models. Nouraniet al. [18] reviewed applications of hybrid wavelet–artificial intel-ligence models in hydrology.

In recent years, support vector regression (SVR) was used forwater resources and environmental engineering problems. Manyapplications SVR can be found in water resource and environmentalengineering literature. Some of these studies in the field of waterquality modeling are presented in Table 1 [19–23].

In this research, WANN, ANN, WMLR, MLR, and GA-SVR modelswere developed for one month ahead forecasting of eutrophicationin San Francisco Bay gauging station in the USA, and were com-pared together. Based on the author’s findings, this paper is thefirst application of the WANN, WMLR, and GA-SVR models for bayeutrophication prediction.

This paper is organized as follows: in Section 2, a brief descrip-tion about wavelet, ANN, GA, SVR, MLR models and modelsperformance criteria’s are presented, respectively. Gauging stationand data analysis are presented in Section 3. Section 4 representsthe proposed models. The models application for chlorophyll-a pre-diction and results are summarized in Section 5. The concludingremarks will be the last section.

2. Methods

2.1. Wavelet analysis

A wavelet system is a set of building blocks to construct or repre-sent a signal or function. It has become a popular analytical tool dueto its ability to simultaneously elucidate both spectral and temporalinformation within the signal. This property overcomes the basicshortcoming of Fourier analysis, which is that the Fourier spectrumcontains only globally averaged data. Wavelet transforms, which

provide information in both the time and frequency domains ofa signal, give considerable information about the physical struc-ture of data. It provides a time-frequency representation of a signalin many different periods of the time domain [24]. The time-scale

210 T. Rajaee, A. Boroumand / Applied Ocean Research 53 (2015) 208–217

Table 1Some illustrative reviews of SVR application in the field of water quality modeling.

Researchers Proposed models Predictive performance

Malek et al. [19] SVM for dissolved oxygen prediction Relatively good performanceGarcíaNieto et al. [20] GA–SVR model for cyanotoxins presence forecasting High performance

predr qualality p

wa

T

wtTam

x

tocpdottbdf

g

wasoac

g

b

T

wsfiNwid[n

tion and the output vector is computed during the employment ofa nonlinear function called the activation function. The feed for-ward operation computes an output for each input vector and then

Inpu t 2

Inpu t 1

Inpu t layer Hidden Layer Out put Lay er

Output

Xie et al. [21] SVM for freshwater algal bloomLiu et al. [22] RGA–SVR for aquaculture wateDu and Tao [23] IRSVR for aquaculture water qu

avelet which transforms a continuous time signal, x(t), is defineds:

(a, b) = 1√a

+∞∫−∞

g∗(

t − ba

)x(t)dt (1)

here the asterisk (*) corresponds to the complex conjugate, andhe function g(t) is called the wavelet function or mother wavelet.he parameter a acts as a dilation factor, while b corresponds to

temporal translation of the function g(t) [25]. The original signalay be reconstructed using the inverse wavelet transform:

(t) = 1cg

+∞∫−∞

∞∫0

1√a

a(

t − ba

)T(a, b)

dadba2

(2)

For practical applications, a hydrologist has access to a discreteime signal, rather than a continuous time signal. A discretizationf Eq. (1) based on the trapezoidal rule maybe is the simplest dis-retization of the continuous wavelet transform. This transformroduces N2 coefficients from a data set of length N. Hence, redun-ant information is locked up within the coefficients, which mayr may not be a desirable property [26]. To overcome the men-ioned redundancy, logarithmic uniform spacing can be used forhe scale discretization with considerably coarser resolution of the

locations, which allows N transform coefficients to completelyescribe a signal of length N. A discrete wavelet of this type has theollowing form:

m,n(t) = 1√am0

g

(t − nboam0

am0

)(3)

here m and n are integer values, b0 is the location parameter,nd a0 is a specific calculated dilation step. The most common andimplest choice for the parameters are a0 = 2 and b0 = 1. This powerf two logarithmic scaling of the translation and dilation is knowns the dyadic grid arrangement [27]. The dyadic wavelet can beondensed as:

m,n(t) = 2m/2g(2−mt − n) (4)

For a discrete time series xi, the dyadic wavelet transformecomes:

m,n = 2m/2n−1∑t=0

g(2−mi − n)xi (5)

here Tm,n is the wavelet coefficient for the discrete wavelet ofcale a = 2m and location b = 2mn. The terms in Eq. (5) considers anite time series, xi, i = 1, 2, . . ., N − 1 and N is an integer power of 2,

= 2M. At the largest wavelet scale (i.e., 2m, where m = M), just oneavelet is needed to cover the time interval and only one coefficient

s created. Therefore, the total number of wavelet coefficients for aiscrete time series of length N = 2M is 1 + 2 +4 + 8 + · · · +2M−1 = N − 126]. A signal smoothed component T̄ is left, which is the sig-al mean. Therefore, a time series of length N is broken into N

iction Relatively good performanceity prediction High performancerediction Good performance

components, with zero redundancy. The inverse discrete transformis given by:

xi = T̄ +M∑

m=1

2M−n−1∑n=0

Tm.n(2−m/2i − n) (6)

or in a simple format as:

xi = T̄(t) +M∑

m=1Wm(t) (7)

in which T̄(t) is called the approximation sub signal at level M,and Wm(t) is the details of sub signals at levels m = 1, 2, . . ., M.The wavelet coefficients, Wm(t), m = 1, 2, . . ., M, provide the detailsignals, which can capture small features of interpretational val-ues in the data. The residual term, T̄(t), represents the backgroundinformation of the data. Because of the simplicity of W1(t), W2(t),. . ., WM(t), T̄(t), some important characteristics can be determinedeasily by means of these components. Recommended literature forthe wavelet beginner is contained in Rao and Bopardikar [28].

2.2. Artificial neural network

The ANN architecture is a massive parallel distributedinformation-processing system that has certain performance char-acteristics resembling biological networks of the human brain. Backpropagation (BP) is a type of ANN among the most researchedand widely-used structures in hydrology and water-resource prob-lems. Some ANN structures are dependent on the learning strategyadopted. This section briefly explains the ANN employed in thispaper, namely the multilayer perceptron, usually referred to as thefeed-forward neural network (FFNN) with BP algorithm. This ANNtype can map input real valued vectors into an output of real val-ued vectors. Fig. 1 shows the standard structure of an FFNN. ThisANN type contains three layers: input and hidden layers, contain-ing neurons, and an output layer. In the ANN, the inputs presentedto the neurons in an input layer are propagated in a forward direc-

Inpu t n

Fig. 1. Standard FFNN with one hidden layer.

d Ocea

ca

E

wtvM

2

oopriutmaacstcota

2

dstvpmttiilE

f

wvwstb

L

R

T. Rajaee, A. Boroumand / Applie

ompares it with the actual value. The total error is calculated forll training sets.

=M∑

m=1Em =

m∑m=1

n∑k=1

(Tmk − Omk)2 (8)

here Em = error for the mth input vector; M = number of input vec-ors; N = number of outputs; and Omk, Tmk = observed and predictedalues, respectively. A detailed description of ANN can be found inasters [29].

.3. Genetic algorithms

GA, as powerful and broadly applicable stochastic search andptimization techniques, are perhaps the most widely known typesf evolutionary computation methods today. The GA maintains aopulation of individuals for each generation. Each individual rep-esents a potential solution to the problem at hand. Each individuals evaluated to give some measure of its fitness. Some individualsndergo stochastic transformations by means of genetic operationso form new individuals. There are two types of transformation:

utation which creates new individuals by making changes in single individual, and crossover which creates new individu-ls by combining parts from two individuals. New individuals,alled offspring, are then evaluated. A new population is formed byelecting the more fit individuals from the parent population andhe offspring population. After several generations, the algorithmonverges to the best individual, which hopefully represents anptimal or suboptimal solution to the problem. A detailed descrip-ion of GA and its application in engineering can be found in Gennd Cheng [30].

.4. Support vector regression

As a prediction algorithm, the foundations of SVM were firstlyeveloped by Vapnik [31] and are an effective tool for data clas-ification and regression. A SVM uses a linear model to separatehe sample data through some nonlinear mapping from the inputectors into the high-dimensional feature space. In regressionroblems, a non-linear function is learned by a linear learningachine in a kernel induced feature space, while the capacity of

he system is controlled by a parameter that does not depend onhe dimensionality of the feature space [32]. The process of apply-ng SVMs in regression problems is referred to SVR. SVR maps thenput space x to the high dimensional feature space g(x) in a non-inear manner. This relationship in the feature space is shown inq. (9):

(x, w) =n∑

j=1wigi(x) + b (9)

here gi(x), j = 1, . . ., n are a set of nonlinear functions map the inputectors into a high-dimensional feature space. W represents theeighting vector and b is the bias. The quality of estimation is mea-

ured by the loss function L(y, f (x, w)). SVM regression uses a newype of loss function called the insensitive loss function proposedy Vapnik [31,33].

ε(y, f (x, w)) ={

0 if∣∣y − f (x, w)∣∣ ≤ ε∣∣y − f (x, w)∣∣ − ε otherwise (10)

The empirical risk is

emp(w) = 1m

m∑i=1

Lε(yi, f (xi, w)) (11)

n Research 53 (2015) 208–217 211

SVR performs linear regression in the high-dimension featurespace using ε insensitive loss and, at the same time, tries to reducemodel complexity by minimizing ||w||2 . This can be describedby introducing (non-negative) slack variables, �i, �∗i = 1, ..., m tomeasure the deviation of training samples outside the ε insensitivezone. Thus, SVR is formulated as the minimization of the followingfunction:

Min12

||w||2 + Cm∑

i=1

(�i + �∗i

)

Such that

⎧⎪⎪⎨⎪⎪⎩

yi − f (xi, w) ≤ ε + �∗i

f (xi, w) − yi ≤ ε + �i�i, �

∗i

≥ 0, i = 1, ..., m

(12)

where C > 0 represents the penalty on samples out of error ε and itis used to control the medium solution between model complexityand error. This primal optimization problem can be converted intothe following dual problem:

f (x) =nsv∑i=1

(˛i − ˛∗i

)k(xi, x)

Subject to 0 ≤ ˛∗i

≤ C, 0 ≤ ˛i ≤ C(13)

where nsv is the number of support vectors (SVs) and the k (xi, x) isthe kernel function.

This optimization model can be solved using the Lagrangianmethod, which is almost equivalent to the method used to solvethe optimization problem in the separable case.

Accordingly, the coefficients i can be found by solving the fol-lowing convex quadratic programming problem.

The kernel function is formulated as

k(x, xi) =n∑

j=1gj(x)gj(xi) (14)

It is well known that SVM generalization performance dependson a good setting of parameters C and ε and the kernel parameters.The choices of C and ε control the prediction (regression) modelcomplexity. Kernel functions are used to change the dimension-ality of the input space to perform the regression task with moreconfidence.

2.5. Multiple linear regressions

MLR method is a statistical technique that is used to predicta dependent variable from several independent variables. In MLRmodel development, it is assumed that the relationship betweendependent and independent variables is linear, the variables havenormal distribution, and the values of variables are measured pre-cisely. MLR methods have dominated many areas of time seriesprediction. The model for MLR, given n observations, is:

yi = a0 + a1xi,1 + a2xi,2 + · · · + anxi,n + ei (15)where yi = prediction on month i; xi,n = value of the nth predictoron month i; a0 = regression constant; an = coefficient on the nthpredictor; n = total number of predictors; and ei = error term. Rec-ommended literature for details of the MLR includes the study ofSnedecor and Cochran [34].

2.6. Models performance comparison

The performance of different forecasting models can be assessedusing several statistical tests that describe the errors associatedwith the models. After each of the models is calibrated using the

2 d Ocean Research 53 (2015) 208–217

tiavm

m

E

wcdaop

s

R

Rcw

3

3

Bta4csBsBTn

3

J

Table 2Statistics analysis for training, testing and all data set.

Data set Max Min Mean Sd

Training set 144.20 0.95 11.73 20.95

and antecedent Chl-at and despite the fact that the combinationC3 shows comparatively higher correlation, due to the negligi-ble discrepancy between the correlation coefficients of C2 and C3

Table 3R2 between Chl-at+1 and Chl-at , Chl-at−1, Chl-at−2, Chl-at−5, Chl-at−8, and Chl-at−11.

Combination R2

C1 Chl-at+1 (Chl-at) 0.071C2 Chl-at+1 (Chl-at , Chl-at−1) 0.097C3 Chl-at+1 (Chl-at , Chl-at−1, Chl-at−2) 0.098

12 T. Rajaee, A. Boroumand / Applie

raining/validation data set, the performance can then be assessedn terms of statistical measures of goodness of fit. In order to providen indication of goodness of fit between the observed and predictedalues the Nash–Sutcliffe model efficiency coefficient (E) and root-ean-squared error (RMSE) were used in this research.The E is used to assess the predictive power of hydrological

odels [35]:

= 1 −∑i=1

n (Chl-ameasured − Chl-apredicted)2

∑i=1n (Chl-ameasured − Chl-ameasured Ave)

2(16)

here n is the number of data in time series. An efficiency of oneorresponds to a perfect match of forecasted data to the observedata. An efficiency of zero indicates that the model predictions ares accurate as the mean of the observed data, whereas an efficiencyf less than zero (E < 0) occurs when the observed mean is a betterredictor than the model [34].

The RMSE evaluates the variance of errors independently of theample size, and is given by [36]:

MSE =

√∑i=1n (Chl-ameasured − Chl-apredicted)

2

n(17)

MSE indicates the discrepancy between the observed and fore-asted values. A perfect fit between observed and forecasted valuesould have an RMSE of 0.

. Study areas and data

.1. Study area

San Francisco Bay is a part of the more complex San Franciscoay estuary system, which includes San Pablo Bay and Suisun Bay,he Carquinez Strait, the tidal marshes surrounding these waters,nd river tributaries. San Francisco Bay estuary, which consists of80 square miles, 12 islands, and two trillion gallons of salt water,an be thought of as two separate areas: the northern, which passesouth and westward from the Delta through Suisun and San Pabloays, and the southern (also called the South Bay) which extendsouth eastward toward San Jose. These two areas join in the Centralay near the Golden Gate Bridge and flow out to the Pacific Ocean.he entire bay is relatively shallow, with narrow, deep channelsear the Golden Gate.

.2. Data

The chlorophyll-a (Chl-a) data used in this study were fromanuary 1994 to December 2013 collected from the NO. 36

Fig. 2. Map of San Francisco Bay (California) and po

Testing set 34.30 2.25 7.04 5.49All data set 144.20 0.95 10.66 17.13

sampling station (Fig. 2) by the United State Geological Survey(USGS) (http://sfbay.wr.usgs.gov/access/wqdata). The samplingdepth is 2 m below water surface and the monitoring frequency isonce a month. Modeling has been done monthly due to the restric-tion in sampling intervals and chlorophyll-a seasonally variation inmacroscopic scale.

For all the WANN, ANN, WMLR, MLR and the GA-SVR models,the data series were divided into a training/calibration set (60% ofthe data), a validation set (20% of the data) and a testing set (theremaining 20% of the data). Brief statistics of the studied data arepresented in Table 2 which includes the mean, standard deviation(Sd), minimum and maximum.

4. Model development

To achieve the best combination of input data driven from timeseries, correlation coefficients (R2) between Chl-at+1 with Chl-at,Chl-at−1, Chl-at−2, . . ., Chl-at−i time series were computed and pre-sented in Table 3. R2 measures the degree of correlation among theChl-at+1 and antecedent Chl-at. R2 was used to judge how manytime steps into the past would allow the best efficiency. The com-bined data with 3, 6 and 12 months delay was used to investigatingthe effects of seasonal variation of the Chl-a on the value of Chl-at+1. Table 3 totally shows the low correlation between the Chl-at+1

C4 Chl-at+1 (Chl-at , Chl-at−2, Chl-at−5) 0.095C5 Chl-at+1 (Chl-at , Chl-at−5, Chl-at−11) 0.081C6 Chl-at+1 (Chl-at−5, Chl-at−11) 0.002C7 Chl-at+1 (Chl-at−2, Chl-at−5, Chl-at−11) 0.013C8 Chl-at+1 (Chl-at−2, Chl-at−5, Chl-at−8, Chl-at−11) 0.027

sition of NO.36 water qualiy sampling station.

http://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdatahttp://sfbay.wr.usgs.gov/access/wqdata

T. Rajaee, A. Boroumand / Applied Ocean Research 53 (2015) 208–217 213

opose

cs

4

tAoAnpwtacR

4

ufauaCrAtsoW

4

koTtpS

Fig. 3. Construction of the pr

ombinations, in order to simplify the models, C2 combination con-idered as input data for the models.

.1. Development of ANN models

In this research we used one of the most common ANN archi-ectures, the feed forward-back propagation neural network. ThisNN architecture is very popular, because it can be used to vari-us fields. For the regular ANNs used in this research, three layerNN models consisting of an input layer composed of two inputeurons, one single hidden layer with 1–15 neurons, and one out-ut layer consisting of one neuron denoting the predicted Chl-a,ere developed. Levenberg–Marquardt algorithm was used to train

he networks and the Tansig and Logsig activation functions werepplied in hidden layer nodes. 30 ANN models were developed andompared using two statistical measures of goodness of fit (E andMSE) and optimum architecture of the ANN were reported.

.2. Development of coupled WANN and WMLR models

The WANN and WMLR models are ANN and MLR models whichse, as inputs, sub-time series which are derived by wavelet trans-orms. In this research, the original Chl-a data was decomposed into

sub-time series of Chl-ad1(t), Chl-ad2(t), . . ., Ch-adi(t) and Chl-aa(t),sing discrete wavelet transform (DWT), for which Chl-ad1(t), Chl-d2(t), . . ., Chl-adi(t) and Chl-aa(t) are the details and approximationhl-a time series, respectively. These sub-time series play variousoles in the original time series and the behavior of each is distinct.ll sub-series were then used as input data to the ANN and MLR sys-

ems to predict the one month ahead Chl-a, because all sub-timeeries coefficients are important and contain information about theriginal time series. Fig. 3 shows the construction of the proposedANN and WMLR models.

.3. Development of GA-SVR model

In this research, radial base function is used in the SVR. In thisind of SVR C, � and ε are user-determined parameters. To constructf an effective SVR model, the SVR parameters must beset precisely.

his study combined the SVR with genetic algorithm to search forhe optimal SVR parameters using genetic algorithm optimizationrogram, and then applied the optimal parameters to construct theVR model.

d WANN and WMLR models.

4.4. Development of MLR model

The MLR model was trained by using the same calibration datasets used for the WANN, WMLR, ANN, and GA-SVR models, toenable a direct comparison. The efficiency of the models was alsotested with the same data set used to validate the WANN, WMLR,ANN and GA-SVR models, thus making the result comparable.

5. Results and discussion

5.1. WANN, WMLR, ANN, MLR, and GA-SVR models performance

For the WANN models, ANN networks consisting of an inputlayer with 4–12 input neurons, one single hidden layer with 1–15neurons and one output layer consisting of one node denoting theChl ai+1 were developed. The Tansig and Logsig activation functionswere applied in hidden layer nodes and the process of calibratingthe ANN inputs and outputs was carried out with BP algorithm.The Levenberg–Marquardt algorithm was used to train the ANNmodels. In this study measured Chl-a time series was decomposedto different scales from 1 to 5 by twenty one different kind wavelettransform functions (haar, daubechies2-8, coeiflet1-5 and symlet1-8). The original Chl-a time series and the generated sub time seriesusing db5 wavelet transform function (Chl-ad1(t), Chl-ad2(t) andChl-aa(t)) are shown in Fig. 4.

3150 different models were tested in this study and the bestWANN models are shown in Table 4.

Table 4 shows that using db5 wavelet transform function withsecond level decomposition was yielded the best prediction.

To direct comparison between the WANN and WMLR models,the WMLR model used the same wavelet transform function anddecomposition level with the WANN model and these two modelswere compared to the best ANN, MLR and GA-SVR models. GA wasused to optimize the SVR parameters and GA-SVR model was devel-oped using the optimum parameters (C, � , ε) = (4.08244, 3.54397,1.53998) which is then used to forecast the one month ahead Chl-ausing the training data and test data.

Table 5 shows the WANN, WMLR, ANN, MLR and GA-SVR mod-els performance statistics (E and RMSE). According to Table 5, the Eand RMSE for WANN model was 0.87 and 1.58. In the WMLR, ANNand MLR models, the parameters were −0.04, 5.53, 0.31, 4.51 and−2.36, 9.94 respectively, and in the GA-SVR model, the parameters

were 0.24, and 4.72, respectively. The time series of the measuredand predicted Chl-a using WANN, WMLR, ANN, MLR and GA-SVRmodels for the verification period were plotted in Figs. 5–9. Further-more, the scatter plots of Chl-a predictions using the models were

214 T. Rajaee, A. Boroumand / Applied Ocean Research 53 (2015) 208–217

020406080

100120140160

Jan-9

Chl-

a (t

)

-60-40-20

0204060

Jan

Chl-

a d1(

t)

93 Jan-98

Ti

n-93 Jan-98

T

Jan-03 Jan-

me (Month)

Jan-03 Jan

Time (Month)

-08 Jan-13

n-08 Jan-13

)

Chl-a

a (t

)

--4-

4

Chl-

a d2(

t)

-100

102030405060

Jan-93 Jan-

604020

0204060

Jan-93 Jan-9

98 Jan-03 J

Time (Mont

8 Jan-03

Time (Mon t

an-08 Jan-13

h)

Jan-08 Jan-13

h)

Fig. 4. Original and decompos

Table 4E and RMSE in Chl-a prediction by differrent WANN models in testing period.

Wavelet type Decomposition level ANN structure E RMSE

haar 3 8-13-1 0.51 3.80db2 2 6-5-1 0.73 2.55db3 3 8-8-1 0.77 2.29db4 3 8-13-1 0.83 1.61db5 2 6-14-1 0.87 1.58db6 4 10-7-1 0.83 1.71db7 2 6-13-1 0.81 1.93db8 3 8-13-1 0.77 2.24coif1 4 10-13-1 0.81 1.95coif2 4 10-9-1 0.83 1.74coif3 2 6-14-1 0.81 1.90coif4 2 6-13-1 0.82 1.85coif5 4 10-13-1 0.83 1.75sym1 3 8-14-1 0.49 3.91sym2 4 10-3-1 0.69 2.85sym3 2 6-10-1 0.77 2.23sym4 2 6-5-1 0.77 2.30sym5 4 10-12-1 0.76 2.38sym6 2 6-9-1 0.76 2.35

pWtcSccfFm

TE

sym7 3 8-7-1 0.79 2.06sym8 2 6-15-1 0.81 1.93

lotted against the observed Chl-a. It is found that none of the ANN,MLR, MLR and GA-SVR models didn’t provide accurate predic-

ions in the studied sampling station, although the ANN model hadomparatively good predictions than the WMLR, MLR and the GA-VR models. Also, it can be seen that the WANN model, because ofapability of feature extraction and considering unknown physics

haracteristics of the Chl-a phenomenon, provided more accurateorecasting results than the regular artificial neural network model.urthermore, WMLR model resulted better predictions than MLRodel. Fig. 5 shows a good agreement between the observed and

able 5 and RMSE values in Chl-a prediction by all predictive models for verification period.

Model Wavelet transform function Decomposition level

WANN db5 2 WMLR db5 2 ANN – – MLR – – GA-SVR – –

ed Chl a sub time series.

predicted values obtained by the WANN model. The magnitudes oflow, medium, and high Chl-a predictions using the proposed WANNmodel were closer to the measured values than the other models.On the other hand, the WANN results were closer to the 45◦ straightline in the scatter plots than the other models. The WANN modelunderestimated the Chl-a values.

5.2. Prediction of high value

The value of 34.3 which was the highest value of studied dataset was laid in the verification set. The WANN, WMLR, ANN, MLRand GA-SVR models forecasted values were 32.87, 27.11, 22.96,16.66, and 10.22, respectively. The WANN error percentage (4%)was less than the errors for the other models, and the GA-SVR modelwith 70% error presented very poor efficiency. The WANN modelforecasted this value precisely.

5.3. Prediction of Chl-a values at the station No. 18

The developed models were used to predict Chl-a values at thestation No. 18 located in central bay in order to indicate the appli-cability of them. The performance statistics were calculated andthe E and RMSE for WANN model were 0.62 and 1.40. In the WMLR,ANN and MLR models, the parameters were −0.04, 2.32, −0.32, 2.62and −1.15, 4.58, respectively, and in the GA-SVR model, the param-eters were −3.41, and 4.80, respectively. Although WANN resultsshowed reasonable agreement with the actual Chl-a data but theresults of other models indicated low accuracy. The results express

that developed models for the station No. 36 did not yield accuratepredictions at the station No. 18. It is the nature of the intelligentmodels that if they were developed for the prediction of the speci-fied time series, the accurate results couldn’t be derived necessarily

ANN function AAN structure E RMSE

Tansig 6-14-1 0.87 1.58– – −0.04 5.53Tansig 2-11-1 0.31 4.51– – −2.36 9.94– – 0.24 4.72

T. Rajaee, A. Boroumand / Applied Ocean Research 53 (2015) 208–217 215

0

10

20

30

40

Feb-10 Feb-11 Feb-12 Feb-13 Feb-14Ch

l-a

Time (Month)

WANN Observed

0

10

20

30

40

0 10 20 30 40

Pred

icte

d (C

hl-a

)

Observed (Chl-a)

Fig. 5. Time series comparison of measured and WANN predicted monthly Chl-a for verification period.

0

10

20

30

40

Feb-10 Feb-11 Feb-12 Feb-13 Feb-14

Chl-

a

Time (Month )

ANN Observed

0

10

20

30

40

0 10 20 30 40

Pred

icte

d (C

hl-a

)

Observed (Chl-a)

Fig. 6. Time series comparison of measured and ANN predicted monthly Chl-a for verification period.

0

10

20

30

40

1 13 25 37 49

Chl-

a

Time (Month )

WML R Obs erved

0

10

20

30

40

0 20 40

Pred

icte

d (C

hl-a

)

Observed (Chl -a)

Fig. 7. Time series comparison of measured and WMLR predicted monthly Chl-a for verification period.

0

10

20

30

40

Feb-10 Feb-11 Feb-12 Feb-13 Feb-14

Chl-a

Time (Month )

MLR Observed

0

10

20

30

40

0 10 20 30 40

Pred

icte

d (C

hl-a

)

Observed (Chl-a)

Fig. 8. Time series comparison of measured and MLR predicted monthly Chl-a for verification period.

0

10

20

30

40

Feb-10 Feb-11 Feb-12 Feb-13 Feb-14

Chl-a

Time (Month )

GA-SVR Obs erved

0

10

20

30

40

0 20 40

Pred

icte

d (C

hl-a

)

Fig. 9. Time series comparison of measured and GA-SVR

Observed (Chl -a)

predicted monthly Chl-a for verification period.

216 T. Rajaee, A. Boroumand / Applied Oce

Table 6Comparison between E and RMSE values in Chl-a prediction by using large and smalldatasets.

Model 240 Months Dataset 60 Months Dataset

E RMSE E RMSE

WANN 0.87 1.58 0.08 4.82WMLR −0.04 5.53 0.05 3.98ANN 0.31 4.51 −2.53 9.46

itim

5

fmsdduo

6

esbobaacot

fafdpthfiwpmceaincon

frrf

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[[

[

MLR −2.36 9.94 −0.26 5.66GA-SVR 0.24 4.72 −0.23 5.59

n the similar ones. The results obtained in this section approvedhis feature. But the major advantage of intelligent models usedn the study is the non-liner and complicated Chl-a phenomena

odeling.

.4. Prediction of Chl-a values by using small dataset

The Chl-a data recorded from January 2009 to December 2013rom the NO. 36 station were used to develop and validate the

odels in order to illustrate the capability of modeling by usingmall datasets. The results obtained by using these datasets (smallataset) were compared to the results of sub section No. 5.1 (largeataset) in Table 6. This table obviously demonstrates that thatsing small datasets could not yield accurate predictions becausef inadequate training sets.

. Conclusion

The potential of WANN, WMLR, ANN, MLR and GA-SVR mod-ls for Chl-a forecasting 1 month ahead was investigated in thistudy for gauging station in the San Francisco Bay- USA. In theest of the authors’ knowledge, this paper is the first applicationf WANN, WMLR and GA-SVR models for prediction of Chl-a inays. In this study, all proposed models were compared togethernd resulted that the best WANN model was substantially moreccurate than the other models. Since the temporal and seasonalhanges existed at the time series used in this study, the resultsbtained from the WANN model showed that this model forecastshese changes accurately.

The WANN model was more accurate because wavelet trans-orms provide useful decompositions of the observed time series,nd the decomposed data improved the performance of the ANNorecasting model by analyzing useful information on variousecomposition levels. Non stationary time series wavelet decom-osition into different subseries prepares an interpretation of theime series structure and extracts important information about itsistory by using several coefficients. WANN implements a single-tting procedure to nonlinear and complex Chl-a phenomenonithout the need for establishing a formal model. The ANN modelrovided some reasonable results, but in comparison to the WANNodel, it was not highly accurate. The errors demonstrate that some

ontributions of the physics are disguised. WANN and WMLR mod-ls involve this unknown physics and thus improve the predictionccuracy. The performance of the MLR linear model, due to thenherent nonlinearity and complexity of Chl-a phenomenon wasot suitable. In the WANN and ANN models, nonlinear propertiesan help the models to detect and capture the nonlinear featuresf the Chl-a phenomenon, although the GA-SVR model despite ofonlinearity was not yielded the good predictions.

The results indicate that if an intelligent model was developed

or the prediction of the determined time series, the exquisiteesults couldn’t be derived necessarily in the similar ones. Also theesults showed that using small datasets couldn’t lead to accurateorecasts because of inadequate training sets.

[

[

an Research 53 (2015) 208–217

As a result, highly accurate Chl-a forecasting models such as theWANN model that was developed in this study are useful tools inbays water quality management. In order to complete the currentresearch, it is suggested to model the Chl-a variations by consider-ing other variables (e.g. PO4 or NO3) and use the proposed model topredicting Chl-a in the second, third, or following months. Further-more, the writers suggested that the proposed WANN model couldbe used to other bays to reconfirm the effectiveness of the model.

References

[1] National Research Council. Clean coastal waters: understanding and reducingthe effects of nutrient pollution. Washington, DC: National Academy Press;2000. p. 405.

[2] Duda AM. Municipal point source and agricultural nonpoint source contribu-tions to coastal eutrophication. Water Resour Bull 1982;18:397–407.

[3] Tu JV. Advantages and disadvantages of using artificial neural networksversus logistic regression for predicting medical outcomes. J Clin Epidemiol1996;11(49):1225–31. Elsevier.

[4] Afshar A, Saadatpour M. Reservoir eutrophication modeling, sensitivity anal-ysis, and assessment: application to Karkheh Reservoir, Iran. Environ Eng Sci2009;26(7):1227–38.

[5] Lee JHW, Huang Y, Dickman M, Jayawardena AW. Neural network modellingof coastal algal blooms. Ecol Model 2003;159:179–201.

[6] Kuo JT, Hsieh MH, Lung WS, She N. Using artificial neural network for reservoireutrophication prediction. Ecol Model 2007;200:171–7.

[7] Melesse AM, Krishnaswamy J, Zhang K. Modeling coastal eutrophication atFlorida Bay using neural networks. J Coast Res 2008;24(2B):190–6.

[8] Huo S, He Z, Su J, Xi B, Zhu C. Using artificial neural network models for eutroph-ication prediction. Procedia Environ Sci 2013;18(2013):310–6.

[9] Saghi H, Karimi L, Javid AH. Investigation on trophic state index by artifi-cial neural networks (case study: Dez Dam of Iran). Appl Water Sci 2014,http://dx.doi.org/10.1007/s13201-014-0161-2.

10] Kisi O. Stream flow forecasting using neuro-wavelet technique. Hydrol Process2008;22(20):4142–52.

11] Partal T, Cigizoglu HK. Estimation and forecasting of the daily sus-pended sediment data using wavelet-neural networks. J Hydrol (Amsterdam)2008;358(3–4):317–31.

12] Rajaee T, Mirbagheri SA, Zounemat-Kermani M, Nourani V. Daily suspendedsediment concentration simulation using ANN and neuro-fuzzy models. SciTotal Environ 2009;407(17):4916–27.

13] Rajaee T. Wavelet and neuro-fuzzy conjunction approach for suspended sedi-ment prediction. Clean Soil Air Water 2010;38(3):275–86.

14] Rajaee T, Nourani V, Zounemat-Kermani M, Kisi O. River suspended sedimentload prediction: application of ANN and wavelet conjunction model. J HydrolEng ASCE 2011:613–27.

15] SatyajiRao YR, Krishna B, Venkatesh B. Wavelet based neural networks for dailystream flow forecasting. Int J Emerg Technol Adv Eng 2014;4(1):307–17.

16] Kim ME, Shon TS, Shin HS. Forecasting algal bloom (Chl-a) on the basis of cou-pled wavelet transform and artificial neural networks at a large lake. DesalinWater Treat 2012;51(2013):4118–28.

17] Wang F, Wang X, Chen B, Zhao Y, Yang Z. Chlorophyll a simulation in a lakeecosystem using a model with wavelet analysis and artificial neural network.Environ Manag 2013;51(5):1044.

18] Nourani V, Baghanam A, Adamowski J, Kisi O. Applications of hybrid wavelet-artificial intelligence models in hydrology: a review. J Hydrol 2014;514:358–77.

19] Malek S, Mosleh M, M- Syed SH. Dissolved oxygen prediction using supportvector machine. Int J Comput Inf Syst Control Eng 2014;8(1):46–50.

20] GarcíaNieto PJ, AlonsoFernández JR, de Cos Juez FJ, SánchezLasheras F,DíazMuñiz C. Hybrid modelling based on support vector regression withgenetic algorithms in forecasting the cyanotoxins presence in the Trasonareservoir (Northern Spain). Environmental 2013;122(2013):1–10.

21] Xie Z, Lou I, Ung WK, Mok KM. Freshwater algal bloom prediction by supportvector machine in Macau storage reservoirs. Math Probl Eng 2012;2012:1–12.

22] Liu S, Tai H, Ding Q, Li D, Xu L, Wei Y. A hybrid approach of support vectorregression with genetic algorithm optimization for aquaculture water qualityprediction. Math Comput Model 2011;58(2013):458–65.

23] Du J, Tao T. An improved recurrent support vector regression algorithm forwater quality prediction. J Comput Inf Syst 2011;7(12):4455–62.

24] Daubechies I. The wavelet transform, time-frequency localization and signalanalysis. IEEE Trans Inform Theor 1990;36(5):961–1005.

25] Mallat SG. A wavelet tour of signal processing. San Diego, CA: Academic; 1998.26] Addison PS, Murrary KB, Watson JN. Wavelet transform analysis of open chan-

nel wake flows. J Eng Mech 2001;127(1):58–70.27] Szilagyi J. The local effect of intermittency on the inertial subrange

energy spectrum of the atmospheric surface layer. Boundary-Layer Meteorol

1996;79:35–50.

28] Rao RM, Bopardikar AJ. Wavelet transforms. In: Introduction to theory andapplications. Reading: Addison-Wesley; 1998.

29] Masters T. Practical neural network recipes in C++. San Diego: Academic Press;1993.

http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0185http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0190http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0195http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0200http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0205http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0210http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0215http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0220dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2dx.doi.org/10.1007/s13201-014-0161-2http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0230http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0235http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0240http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0245http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0250http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0255http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0260http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0265http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0270http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0275http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0280http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0285http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0290http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0295http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0300http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0305http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0310http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0315http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0320http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0325

d Ocea

[

[

[

[

[

[

T. Rajaee, A. Boroumand / Applie

30] Gen M, Cheng R. Genetic algorithms and engineering optimization. Wiley;2000.

31] Vapnik V. The nature of statistical learning theory. New York: Springer-Verlag;

1995.

32] Water Issues Study Group, School of Oriental, African Studies (SOAS). Waterdemand management (WMD): a case study from South Africa. Technical Report1999.

33] Vapnik V. Statistical learning theory. NY: Springer; 1998.

[

n Research 53 (2015) 208–217 217

34] Snedecor GW, Cochran WG. Statistical methods. 7th ed. Ames, IA: Iowa StateUniversity Press; 1981.

35] Pulido-Calvo I, Gutierrez-Estrada JC. Improved irrigation water demand fore-

casting using a soft-computing hybrid model. Biosyst Eng 2009;102(2):202–18.

36] Sreekanth P, Geethanjali DN, Sreedevi PD, Ahmed S, Kumar NR, JayanthiPDK. Forecasting groundwater level using artificial neural networks. Curr Sci2009;96(7):933–9.

http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0330http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0330http://refhub.elsevier.com/S0141-1187(15)00115-7/sbref0330http://refhub.elsevier.com/S0141-1187(