DOI: https://doi.org/10.4316/fens.2022.011

107

Journal homepage: www.fia.usv.ro/fiajournal

Journal of Faculty of Food Engineering,

Ştefan cel Mare University of Suceava, Romania

Volume XXI, Issue 2 - 2022, pag. 107 - 115

USE OF ARTIFICIAL NEURAL NETWORKS FOR MODELING INFLOW AND

OUTFLOW AND SALINITY OF LAKE FETZARA IN THE REGION-ANNABA

(NE ALGERIA)

*Zahra BOUHALI
1
, Larbi DJABRI

1
, Hamza BOUGUERRA

1
, Fatma TRABELSI

2
,

Azzedine HANI
1
, Hicham CHAFFAI

1 Faculty of earth sciences, Badji Mokhtar University, Annaba, Algeria. Zara.bou5@gmail.com

2 Medjez Elbab Engineering School, Djendouba University. Tunisia

*Corresponding author

Received 8th December 2021, accepted 25th June 2022

Abstract: Lake Fetzara is one of the important lakes in the northeast of Algeria; the water supplying

this lake comes from different precipitation and wadis. Moreover, Meboudja wadi constitutes the

drainage channel. The water of the lake and the underlying groundwater is exposed to excessive

overuse; which seriously threatens the hydrological and ecological balance. The overexploitation is
explained by the increase in water mineralization, which poses a risk of soil salinization. To this end,

this article deals with the subject of current salinity and predict its evolution over time by means of the

modeling of the artificial neural network (ANN), according to the period of low water and the period
of high water. The ANN were trained using three different algorithms: the Scaled Conjugate Gradient

back propagation (SCG) algorithm and One Step Secant back propagation (OSS) algorithm and

Quasi-Newton algorithm (BFGS). The performance results indicate that the three algorithms provided
satisfactory simulations according to the determination coefficient (R2) and the performance criteria

of the mean square error (RMSE), with priority to the BFGS algorithm; where the coefficient of

determination using the BFGS algorithm varies between 69.5% and 95.3%. The BFGS method

presents better results in order to design appropriate institutional mechanisms, capable of leading to
the protection of the quality of these resources essential to the promotion of sustainable development.

Keywords: Inflow, Outflow, Drainage, Salinity, Oued, Water table, Artificial neural network.

1. Introduction

Water is necessary element for the life of

human beings, fauna and flora. The quality

of groundwater has been degraded due to

climate change, the intensive use of

fertilizers, domestic wastewater and

various discharges from factories rejected

directly into the streams; particularly in the

wadis feeding the lake Fetzara. [1]

The hydrographic network of the lake

Fetzara is composed of four streams,

which are: the Mellah stream, with a length

of nearly 8 km, the El Hout stream, with a

length of 10 km, the Zied stream, with a

length of about 10.5 km and finally the

Bou Messous stream which has a length of

9 km. All these streams flow into Fetzara

Lake with a very irregular regime,

torrential in winter and dry in summer. [2]

The degradation of quality and quantity

threatens hydrological and ecological

balance. The increase of water

mineralization leads to the risk of soil

salinization. [2]

To this end, this article deals with the

subject of current salinity and predict its

evolution over time by means of the

modeling of the artificial neural network

(ANN) according to the period of low and

http://www.fia.usv.ro/fiajournal
mailto:Zara.bou5@gmail.com

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

Zahra BOUHALI , Larbi DJABRI, Hamza BOUGUERRA, Fatma TRABELSI, Azzedine HANI, Hicham CHAFFAI,
modeling inflow and outflow and salinity of lake Fetzara (NE Algeria),Food and Environment Safety, Volume XXI, Issue 2 -
– 2022, pag. 107 – 115

108

high water and weekly measurement of the

flow of wadis along two hydrological

cycles.

Regarding the provision of water resources

and the management of application, it is

undeniable to predict the quality of

groundwater, various factors contributes to

the variation in water quality. Their

uncertainly inherent in weight, as more

than one variable effects water quality. The

inhomogeneity the environment reflect the

prediction of quality and approaches into

complexity. The salinity parameter is one

of the fundamental parameters for potable

and agricultural water. [3] This is directly

related to the conductivity of the water and

its dissolved salt (Sodium Chloride,

Magnesium Chloride, Magnesium Sulfate,

etc.) and the quality of potable water.

Salinity was examined in this study as such

a suitable approach to examine the

behavior of groundwater by model system;

it is necessary to know the mechanism of

quality fluctuations over time in order to

have an idea about the situation of the

water table and the quantity of accessible

water. [4]

Artificial neural network are “the black

box” non-linear mathematical models

capable of establishing relationships

between inputs and outputs of a system.[5]

ANN is considered practical substitutes for

regression and empirical models to predict

the behavior of water ressources; due to

their temporal reliability and adaptability

to unforeseen changes. These are applied

not only in qualitative forecasting, but also

in forecasting the situation and the volume

of groundwater, much of the modeling has

been done in this regard. [6] As for

prediction based on the neural network, [6]

2. Materials and methods

1.2. Description of study area

Lake Fetzara is cosedered as wetland. Its

location is about 18 km southwest from the

chief state of Annaba (Daïra and

municipality of Berrahal), it covers

18.600hectares. [7] Several small tows

border it: to the north, the municipality of

Berrahal, to the south the municipalities of

ElEulma (crosseb by ElHout wadi) and

Cheurfa, to the east the small villages of

ElGantra and Zied wadi. (Figure 1)

Lake Fetzara has been listed on the

RAMSAR list of wetlands of international

importance since 2002. Its large extent and

relatively temporary nature make it a

representative wetland of the

Mediterranean region [7]. The

hydrological function of the lake

contributes to flood control, as well as to

the recharge of the water table.

Lake Fetzara stretches 17km length and

13km width. The open water area, which

depends almost exclusively on the

intensity of rainy season, occupies an area

of more than 5800hectares. In addition,

there are several thousand hectares of flood

plains forming vast wet meadows. [7]

2.2. Analytical method

To ensure rigorous monitoring of the

quality and quantity of water entering and

leaving the lake, weekly monitoring of the

flow of the wadis during two hydrological

cycles (2016/17) and (2017/18) as well as

seasonal piezometric campaigns has been

realized. For sampling, the sites chosen are

in near-urban areas and nearby the

industrial zone of the municipality of

Berrahal. The number of the wells selected

is 39 wells. In-situ parameters were

measured using multi-parameter band

HANNA instruments INC (made in

Romania). The main elements presents in

the water analyzed using a Metrohm 883

basic IC plus ion chromatography

apparatus (made in Courtaboeuf, France)

with a conductometric detector. Prior to

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

109

their chemical analysis. The water samples

are filtered through a nylon membrane

(0.22 µm).

Fig. 1. Geographic location of Fetzara lake in the wilaya of Annaba. Algeria.

(Design by P. Pentsch) [8]

The mobile phase used is mixture of

NaHCO3 (168mg) and Na2CO3 (678mg)

dissolved in two liters of ultra-pure water.

The injection rate is 0.7ml/min and the

volume injected is 20 µL. The

chromatograph was calibrated using five

standard solutions containing a

concentration of cations or anions ranging

from 1 to 50mg/l. The prediction of

salinity of groundwater is done using three

different artificial network algorithms.

2.3. Data analysis

The data set is arranged into two layers:

- The first one is the training part
(72% of the data)

- The second one is used for model
validation (28% of data).

Several tests were made to ensure that all

training and validation modules were

examined in a trial and error manner, in

order to obtain better results. In order to

detect model perturbation, a portion of data

is used for training and another portion is

reserved for testing the performance of the

model in order to decide to stop training

with optimal hidden nodes [9]. This stop is

made when the model validation error

starts to increase. [10]

The conductivity is directly proportional to

the water mineralization; [11] for this

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

110

purpose, we have seven descriptor

parameters: Conductivity, Chloride,

Sulphates, Bicarbonates, Calcium,

Magnesium and Sodium. The sampling

date is from the 08/11/2016 to 15/05/2018.

Simple descriptive statistic of the raw data

is shown in (table 1).

The correlation corresponding to the

statistical data collected from the sampling

of 39 wells in the region surrounding lake

Fetzara and the wadis that feed it, shows

that the most significant correlation with

the conductivity parameter is observed

among the other parameters, in which the

correlation was established through the

software statistica 8 (table 2)

Table 1. Statistical description of parameters.

Variables Training dataset Validation dataset

Min Max Mean SD CV Min Max Mean SD CV

EC [μs/cm] 281 6800 1789.75 1225.67 0.68 276 2440 1093.43 621.71 0.57

Cl
-
[mg/l] 85.2 2996.2 675.02 523.60 0.78 35.5 1065 402.02 268.04 0.67

SO4
2-

[mg/l] 19.2 725.3 231.11 133.25 0.58 19.2 528 171.29 106.88 0.62

HCO3
-
[mg/l] 23.18 1284.53 357.88 209.51 0.59 34.77 614.88 301.55 140.96 0.47

Ca
2+

[mg/l] 33.6 672 167.33 106.71 0.64 43.2 348.8 139.36 77.89 0.56

Mg
2+

[mg/l] 1.92 886.4 63.87 89.85 1.41 4.8 176.64 32.38 30.24 0.93

Na
+
[mg/l] 46 2622 382.33 390.66 1.02 23 575 196.02 126.12 0.64

Selection of the input combination:

Table 2. Correlation matrix (Pearson) between variables.

Variables EC Cl
-
SO4

2-
HCO3

-
Ca

2+
Mg

2+
Na

EC 1 0.941 0.474 0.500 0.748 0.465 0.789

Cl
-
0.941 1 0.409 0.407 0.746 0.464 0.776

SO4
2-

0.474 0.409 1 0.216 0.359 0.296 0.425

HCO3
-
0.500 0.407 0.216 1 0.274 0.331 0.504

Ca
2+

0.748 0.746 0.359 0.274 1 0.250 0.492

Mg
2+

0.465 0.464 0.296 0.331 0.250 1 0.419

Na
+
0.789 0.776 0.425 0.504 0.492 0.419 1

Different combination of inputs used in the modeling:
No. Combination

01 Cl-

02 Cl- & Na+

03 Cl-, Na+ & Ca2+

04 Cl-, Na+, Ca2+ & HCO3
-

05 Cl-, Na+, Ca2+, HCO3
- & SO4

06 Cl-, Na+, Ca2+, HCO3
-, SO4

2- & Mg2+

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

111

Training algorithms

The performance of the neural artificial

network was constructed in order to

determine which algorithm would be the

most effective in predicting the salinity of

groundwater in the region by the following

algorithms:

- BFGS: quasi-Newton back propagation

(Brayden 1970; Fletcher 1970; Goldfarb

1970; Shanno 1970).

- SCG: Scaled Conjugate Gradient back

propagation

- OSS: One-Step Secant back propagation

3. Results and discussion

In this research, we created a model to

predict the total variables of groundwater

salinity based on different water quality

states of lake Fetzara and the underlying

aquifer. The artificial neural network is

used to estimate the relationship between

different variables. The results of the

prediction show that the neural network

approach has good and wide applicability

for modeling the salinity of waters of the

region. (Figure 2) presents the calibration

of predicted salinity according to the

BFGS, SCG and OSS algorithms with the

observed salinity by ANNO5.

(2A)

(2B)

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

112

Fig. 2C.

Fig. 2. Observed and predicted salinity ''EC'' according to BFGS (2A), SCG (2B) and OSS (2C)

algorithms using ANN05 in the validation phase.

The (figure 3) presents the calibration of

the values of observed salinity simulated

by the BFGS, SCG and OSS algorithms

and those of the predicted salinity by

ANNO6 for each water point in the

validation phase. The error graph indicates

the observations of the observed salinity

values with the predicted salinity values in

the trainig phase (figure 4). Note that R2 is

a measure of the rate of explanation of

phenomenal reality by the model adopted.

The predicted salinity versus the observed

salinity with the best improvement of the

region’s water is well presented in

ANNO6. Note that R2 is around 0.9 for

calibration and validation, which proves

that the model retains the statistical

characteristics linked to the values

observed with almost unchanged

coefficients of variation.

By comparing between the model’s values

by ANN with experimental data (Table 3)

reveals that the MLP model (BFGS)

(provides values: R2=0.99, RMSE= 114.98

for the training phase and R2=0.95,

RMSE=138.41 for the validation phase) is

the best model.

In this study we use the artificial neural

network to estimate groundwater salinity

based on other quality analisis parameters.

4. Conclusion

Lake Fetzara is the main source of

irrigation for the surrounding town. [13]

The deterioration of the supply and the

quality of its waters can cause irreparable

damages to the environment. Monitoring to

predict future conditions needs precise

models. These tasks are the foundation for

planning to conserve and manage critical

resources.

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

113

Fig. 3. Observed and predicted salinity ''EC'' according to BFGS, SCG and OSS algorithms using ANN06

in the validation phase.

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

114

Fig. 4. Observed versus predicted salinity by BFGS, SCG and OSS algorithms using ANN05 and ANN06
models for training phase.

This study consists of the use of three

autonomous hybrid algorithms to

determine and predict the salinity of the

lake and the water of the underlying

aquifer.

We used the method of artificial neural

network: multilayer perceptron (MLP)

with three algorithms (BFGS, SCG and

OSS). The collected weekly flow

measurement and physico-chemical

analyzes of the water over two

hydrological cycles were used to construct

several sets of combinations of variables as

inputs for the model construction and

evaluation. Qualitative and qualitative

criteria were applied to validate and

compare the models. The results reveal

that the most suitable model for modeling

the salinity of groundwater in a future

studies is the MLP (BFGS), where the

RMSE=114.98 and the coefficient of

determination R2=0.99 for the training

phase and RMSE=138.41, R2=0.95 for the

validation.

The analysis of the data acquired and the

impact of the environmental factors show

the qualitative and quantitative degradation

of the inputs water of the lake. To this end,

it is recommended to carry out geophysical

studies to determine the extension of the

underlying aquifer and to carry out regular

measurement campaigns in order to

encourage organic farming in vulnerable

sectors.

5. Acknowledgment

We would like to thank Dr Lamine

SAYAD monitor at BADJI Mokhtar

University, Annaba, Algeria, for his

support and guide.

Food and Environment Safety - Journal of Faculty of Food Engineering, Ştefan cel Mare University - Suceava

Volume XXI, Issue 2 – 2022

115

6. References

[1]. HABES S, DJABRI L, HANI A,
BOUHSINA S & MUDRY J. Quantification of

inputs and dewatering of a lake: case of the Fetzara

lake. North-east of Algeria, Rev. Sci. Technol.,

Synthèse (2011)

[2]. HABES S, DJABRI LARBI., Chemical
characteristics of a lake belonging to the humid

ecosystems of the north Algeria, case of the Fetzara
lake. North-east of Algeria. PhD thesis in

Hydrogeology, faculty of earth sciences,

department of geology, Badji Mokhtar University,

Annaba. Algeria. (2013)

[3]. BOUDJEMAA S., Mapping of soil-
vegetation relationships in a salty environment

(Fetzara lake), Magister thesis, faculty of sciences,

department of Biology. Badji Mokhtar University.

Annaba. Algeria, (2010)

[4]. HAKIM KHELFAOUI., HICHAM
CHAFFAI., AZZEDINE HANI., RABAH
LAOUAR., Impacts des rejets industriels sur les

eaux de la région de Berrahal (Nord Est Algérien),

Rev. Sci. Technol., Synthèse. (2012)

[5]. P. ABBASI MAEDEH, N. MEHRDADI,
G.R. NABI BIDHENDI, H. ZARE ABYANEH,

Application of Artificial Neural Network to Predict

Total Dissolved Solids Variations in Groundwater

of Tehran Plain, Iran. , International Journal of

Environment and Sustainability, (2013)

[6]. HANI A., LALLAHEM S., MANIA J.,
DJABRI L, The use of finite difference and neural
network models to evaluate the impact of

underground water overexploitation”. Hydrol

Process 20:4381– 4390. doi:10.1002/hyp.6173.

(2006)

[7]. BOUMEZBEUR A., Fact sheet on
RAMSAR wetlands, chott Zaherz Chergui General

directorate of forests. (Algeria), (2003)

[8]. http://journals.openedition.org/mediterrane
e/docannexe/image/8077/img-2.jpg

[9]. BROWN R. et al., Transitioning to water
sensitive cities: Historical, current and future

transition states, 11th International conference on

urban drainage, United Kingdom. (2008)

[10]. BRADDOCK RD., KREMMER ML.,
SANZOGNI L., Feed forward artificial neural

network model for forecasting rainfall,

Environmetrics journal. (1997)

[11]. BOUGUERRA HAMZA, TACHI
SALAH-EDDINE, DERDOUS OUSSAMA,
BOUANANI ABDERRAZAK, KHANCHOUL

KAMEL. Suspended sediment discharge modeling

during flood events using two different artificial

neural network algorithms, Springer International

Publishing, (2019)

[12]. KAVEH OSTAD-ALI-ASKARI.,
MOHAMMAD SHAYANNEJAD., HOSSEIN

GHORBANIZADEH-KHARAZI., “Artificial

Neural Network for Modeling Nitrate Pollution of

Groundwater in Marginal Area of Zayandeh-rood

River, Isfahan, Iran “, KSCE Journal of Civil
Engineering. (2017)

[13]. MELLOUK K, AROUA N., The Fetzara
lake, a fragile wetland threatened by the urban

growth of Annaba city (north-east Algeria) (2015)

[14]. KAHOUL M., DERBAL N., ALIOUA A.,
AYAD W., Assessment of the physical-chemical

quality of well water in the region of Berrahal

(Algeria), Larhyss Journal. (2014)

[15]. BACHIR SAKAA., HICHAM
CHAFFAI., AZZEDINE HANI., ANN’s approach

to identify water demand drivers for Saf-Saf river
basin, Journal of Applied Water Engineering and

Research. (2020)

[16]. SALIM HEDDAM & OZGUR KISI,
Extreme learning machines: a new approach for

modeling dissolved oxygen (DO) concentration

with and without water quality variables as

predictors, Springer. (2017)

javascript:;
javascript:;
javascript:;
javascript:;
javascript:;

1. Introduction