Alexandria Engineering Journal (2014) 53, 655–662

H O S T E D BY
Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

REVIEW

Artificial Neural Networks (ANNs) for flood

forecasting at Dongola Station in the River Nile,
Sudan
Sulafa Hag Elsafi *

Faculty of Arts, Department of Geography, King Saud University, Ryiad, Saudi Arabia

Received 23 May 2012; revised 22 March 2014; accepted 22 June 2014

Available online 22 July 2014

KEYWORDS Abstract Heavy seasonal rains cause the River Nile in Sudan to overflow and flood the surroundings
River Nile; areas. The floods destroy houses, crops, roads, and basic infrastructure, resulting in the displacement
Dongola; of people. This study aimed to forecast the River Nile flow at Dongola Station in Sudan using an
Artificial neural network; Artificial Neural Network (ANN) as a modeling tool and validated the accuracy of the model
Flood forecasting against actual flow. The ANN model was formulated to simulate flows at a certain location in the
river reach, based on flow at upstream locations. Different procedures were applied to predict
flooding by the ANN. Readings from stations along the Blue Nile, White Nile, Main Nile, and River
Atbara between 1965 and 2003 were used to predict the likelihood of flooding at Dongola Station. The
analysis indicated that the ANN provides a reliable means of detecting the flood hazard in the River
Nile.
ª 2014 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria
University.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
2. Study area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
3. Artificial neural networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
4. Artificial neural network models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
4.1. Modeling a neuron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
4.2. Training an artificial neural network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
5. Application of the neural network method in flood forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
5.1. ANN model results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662

* Tel.: +966 4808155.

E-mail address: sulafa007@gmail.com.
Peer review under responsibility of Faculty of Engineering, Alexandria
University.
http://dx.doi.org/10.1016/j.aej.2014.06.010
1110-0168 ª 2014 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria University.
656
6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1. Introduction
Flooding leads to numerous hazards, with consequences
including risk to human life, disturbance of transport and
communication networks, damage to buildings and infrastruc-
ture, and the loss of agricultural crops. Therefore, prevention
and protection policies are required that aim to reduce the
vulnerability of people and public and private property. Many
solutions for flood mitigation and prevention have been
suggested however, a vast amount of data and knowledge
are required about the causes and influencing factors of floods
and their resulting damage.
Flood forecasting and prediction capabilities evolved
slowly during the 1970s and 1980s. However, recent technolog-
ical advances have had a major impact on forecasting
methodologies. For instance, hydrological models use physical
detection systems to forecast flood conditions based on pre-
dicted and/or measured parameters [9]. River flow models
are used as components in actual flood forecasting schemes,
where forecasts are required to issue warnings and to permit
the evacuation of populations threatened by rising water levels.
The basis of such forecasts is invariably observation and/or
predictions of rainfall in the upper catchment area and/or river
flows at upstream points along main rivers or tributaries.
Forecasts about the discharge are obtained in real-time, by
using the model to transform the input functions into a
corresponding discharge function time.
At present, two main approaches are employed in hydro-
logical forecasting. The first approach is based on mathemati-
cal modeling. It models the physical dynamics between the
principal interacting components of the hydrological system.
In general, a rainfall-runoff model is used to transform the
point values of rainfall, evaporation, and flow data into
hydrograph predictions by considering the spatial variation
in storage capacity. A hydraulic channel flow routing model
is then used to calculate flow. An example of this type of deter-
ministic modeling is River Flow Forecasting (RFFS), which is a
large-scale operational system currently employed by the Outer
River Catchment [8]. The second approach is based on model-
ing the statistical relationship between the hydrologic input
and output, without explicitly considering the relationships that
exist among the involved physical processes. Examples of
stochastic models used in hydrology are autoregressive moving
average models (ARMA) [2] and the Markov method [13].
Models that provide a physically sound description about
the hydrological processes that occur in a basin are expected
to have significant advantages over purely empirical models.
The main advantages of these models are their accuracy
and the potential for performing comprehensive sensitivity
analyses. The parameters of these models have direct physical
interpretation, and their values might be established through
field or laboratory investigations [9].
Stream flow modeling is a key tool in water resources man-
agement, early warning for flood hazards, and related impacts.
Many advanced types of models exist, but they have been
S.H. Elsafi
662
662
662
developed for a diverse range of climatic regions. Hence, these
models must be adapted to the local situation of the Nile Basin
prior to application. The ability to simulate river flow quickly
and accurately is of crucial importance in flood forecasting
operations. Hydrodynamic models provide a sound physical
basis for this purpose, and have the capability of simulating
a wide range of flow situations. However, these models require
accurate river geometric data, which may not be available at
many locations. It is also not possible to integrate observed
data directly at desired locations to improve model results [9].
Thus, the Artificial Neural Network (ANN) provides a
quick and flexible approach for data integration and model
development. Therefore, this research used ANN models to
forecast flooding along the River Nile. It is anticipated that
this work will provide baseline information toward the estab-
lishment of a flood warning system for certain sections of the
River Nile.
2. Study area
Sudan has been subjected to a number of major floods along
the Nile and its tributaries, as well as along seasonally flowing
riverbeds (termed wadis). Examples of major floods occurred
in 1946, 1988, 1994, 1996, 1998, 1999, 2000, 2001, 2006, and
2007. Such floods normally cause the loss of life and massive
scale damage to the agricultural sector and properties in the
vicinity of the rivers.
The study area was located in Dongola town (Fig. 1). This area
is located down-stream of the junction of the main tributaries
to the Nile; including the White Nile, Blue Nile, and River
Atbara. This region has suffered a severe flood in 1998.
3. Artificial neural networks
An alternative approach to flow forecasting has been devel-
oped in the recent years, which is based on the ANN [3].
Recent studies have reported that ANN may offer a promising
alternative for the hydrological forecasting of stream flow [7].
The ANN is a computer program that is designed to model the
human brain and its ability to learn tasks [4]. An ANN differs
to other forms of computer intelligence in that it is not rule-
based, as in an expert system. An ANN is trained to recognize
and generalize the relationship between a set of inputs and
outputs.
Early artificial neural networks were inspired by percep-
tions of how the human brain operates. In the recent years,
ANN technological developments have made it more of an
applied mathematical technique with some similarities to the
human brain. ANNs retain two characteristics of the brain
as primary features: the ability to (1) ‘learn’ and (2) generalize
from limited information [5].
Both biological and artificial neural networks employ mas-
sive, interconnected simple processing elements, or neurons.
The knowledge stored as the strength of the interconnecting
Artificial Neural Networks (ANNs) for flood forecasting at Dongola Station in the River Nile, Sudan 657

Main Nile
Atbara River

Blue Nile
White Nile

Figure 1 River Nile and its tributaries in Sudan.

(i.e., back-propagation) is used to modify the weights in the

Figure 2 Example of a simple feed forward network.
weights (a numeric parameter) in ANNs is modified through a
process called learning, using a learning algorithm. This
algorithmic function, in conjunction with a learning rule,
network in an orderly fashion.
Unlike most computer applications, an ANN is not ‘‘pro-
grammed,’’ rather it is ‘‘taught’’ to give an acceptable answer
to a particular problem. Input and output values are sent to
the ANN, initial weights to the connections in the architecture
of the ANN are assigned, and the ANN repeatedly adjusts
these interconnecting weights until it successfully produces
output values that match the original values. This weighted
matrix of interconnections allows the neural network to learn
and remember [10].
When using an ANN to solve a problem, the first step is to
train the ANN to ‘‘learn’’ the relationship between the input
and outputs. This action is accomplished by presenting the
network with examples of known inputs and outputs, in
conjunction with a learning rule. The ANN maps the relation-
ship between the inputs and outputs, and then modifies its
internal functions to determine the best relationship that is
be represented by the ANN.
The inner workings and processing of an ANN are often
thought of as a ‘‘black box’’ with inputs and outputs. One use-
658 S.H. Elsafi

The input value received by a neuron is calculated by

summing the weighted input values from its input links.
That is:
X
ini ¼ j
Wj ; i aj

An activation function takes the neuron input value and

Figure 3 Neuron Model, in which aj equals the activation value
of unit j, wj,i equals the weight on the link from unit j to unit i, ini
equals the weighted sum of inputs to unit i, ai equals the activation
value of unit i (also known as the output value), and g equals the
activation function.
produces a value that becomes the output value of the neuron.
This value is passed to other neurons in the network. This
process is summarized in the schematic presented in Fig. 3.
A neuron is connected to other neurons via its input and
output links. Each incoming neuron has an activation value
and each connection has a weight associated with it. The
neuron sums the incoming weighted values, representing the
input from the neuron [6].

4.2. Training an artificial neural network

ful analogy that helps to understand the mechanism occurring
inside the black box is to consider the neural network as a
super-form of multiple regression. Like linear regression,
which finds the relationship that {y} = f{x}, the neural net-
work finds some function f{x} when trained. However, the
neural network is not limited to linear functions. It finds its
own best function to the best of its ability, given the complex-
ity used in the network, and without the constraint of linearity
(Hewitson and Crane [5]).
The most common type of artificial neural network consists
of three groups, or layers, of units: (1) a layer of ‘‘input’’ units
is connected to (2) a layer of ‘‘hidden’’ units, which is con-
nected to (3) a layer of ‘‘output’’ units (Fig. 2).
The activity of the input units represents the raw informa-
tion that is fed into the network. The activity of each hidden
unit is determined by the activities of the input units and the
weights on the connections between the input units and the
hidden units. The behavior of the output units depends on
the activity of the hidden units and the weights between the
hidden units and output units [12].
4. Artificial neural network models
Once a network has been structured for a particular applica-
tion, it is ready to be trained. To start this process, the initial
weights are chosen at random. Then, the training, or learning,
begins.
There are two approaches to training: supervised and
unsupervised. Supervised training involves a mechanism that
provides the network with the desired output, either by manually
‘‘grading’’ the network’s performance or by providing the desired
outputs with the inputs. Unsupervised training is where the net-
work has to make sense of the inputs without outside help [1].
The computational complexity of a network and its gener-
alization capability directly depends on network topology. A
larger-than-necessary network tends to over fit the training
data, leading to poor generalization performance. In contrast,
a smaller-than-necessary network faces difficulty in learning
the training data. The standard method of trial and error is
used to design the network. In each case of network design,
a large number of networks, differing in the number of hidden
nodes and memory depth, are trained, with the smallest net-
work that learns the training data with the best generalization
performance being selected [11].

4.1. Modeling a neuron 5. Application of the neural network method in flood forecasting

To model a neuron, each neuron performs a simple computa- To observe the modeling performance of ANNs, different
tion. It receives signals from its input links and uses these stations across the Nile were used in this study. All of the data
values to compute the activation level (or output) for the neu- were obtained from the Sudan Ministry of Irrigation as
ron. This value is passed to other neurons via its output links. daily readings. Because of the continuity and availability of

Table 1 Four selected scenarios and model efficiency results.

Scenario Stations used as input Strength of station Slab structure and scale function Model efficiency
Calibration period Verification period
First Tamaniat 0.34672 2,58,1 R2 = 94.32% R2 = 91.59%
Atbara 0.65328 Linear(1, 1), Logistic, Logistic RMSE = 0.338 RMS = 0.374
Second Eddeim 0.23007 3,58,1 R2 = 94.51% R2 = 91.62%
Tamaniat 0.31869 Linear (1,1), Logistic, Logistic RMS = 0.333 RMS = 0.374
Atbara 0.45123
Third Hassanab 0.30906 2,58,1 R2 = 91.85% R2 = 70.29%
AtbaraK3 0.69094 Linear(1, 1), Logistic, Logistic RMS = 0.439 RMS = 0.640
Fourth Tamaniat 0.48644 2,58,1 R2 = 91.31% R2 = 67.92%
AtbaraK3 0.51356 Linear (1, 1), Logistic, Logistic RMS = 0.453 RMS = 0.669
 Artificial Neural Networks (ANNs) for flood forecasting at Dongola Station in the River Nile, Sudan
Table 2 ANN models for the ‘‘Eddeim, Tamaniat, Atbara’’ stations as input and the ‘‘Dongola’’ Station as output for the calibration period (1970–1985) and the verification period
(1986–1987).
ANN model used Slab structure Scale function Model efficiency
Calibration period Verification period
3 Layers, Standard Connections 3, 58, 1 Linear (1, 1), Logistic, Logistic R2 = 94.51% R2 = 91.62%
RMSE = 0.333 RMSE = 0.374
3, 30, 1 Linear (1, 1), Logistic, Logistic R2 = 94.49% R2 = 91.73%
RMSE = 0.333 RMSE = 0.371
4 Layers, Standard Connections 3, 29, 29, 1 Linear (1, 1), Logistic, Logistic, Logistic R2 = 94.47% R2 = 91.75%
RMSE = 0.333 RMSE = 0.371
3, 12, 12, 1 Linear (1, 1), Logistic, Logistic, Logistic R2 = 94.52% R2 = 91.72%
RMSE = 0.333 RMSE = 0.371
5 Layers, Standard Connections 3, 19, 19, 19, 1 Linear (1, 1), Logistic, Logistic, Logistic, Logistic R2 = 94.49% R2 = 91.71%
RMSE = 0.491 RMSE = 0.445
3 Layers, Jump Connections 3, 58, 1 Linear (1, 1), Logistic, Logistic R2 = 94.35% R2 = 91.32%
RMSE = 0.338 RMSE = 0.381
4 Layers, Jump Connections 3, 29. 29, 1 Linear (1, 1), Logistic, Logistic, Logistic R2 = 94.40% R2 = 91.28%
RMSE = 0.336 RMSE = 0.382
5 Layers, Jump Connections 3, 30, 30, 30, 1 Linear (1, 1), Logistic, Logistic, Logistic, Logistic R2 = 94.43% R2 = 91.24%
RMSE = 0.335 RMSE = 0.382
3, 10, 10, 10, 1 Linear (1, 1), Logistic, Logistic, Logistic, Logistic R2 = 94.39% R2 = 91.38%
RMSE = 0.336 RMSE = 0.379
Input Layer, Dampened Feed Back Link 3, 58, 1, 3 Linear (1, 1), Logistic, Logistic, Logistic R2 = 96.56% R2 = 96.13%
RMSE = 0.263 RMSE = 0.255
Hidden Layer Dampened Feed Back Link 3, 58, 1, 58 Linear (1, 1), Logistic, Logistic R2 = 97.37% R2 = 97.05%
RMSE = 0.230 RMSE = 0.221
Output Layer Dampened Feed Back Link 3, 58, 1, 58 Linear (1, 1) Logistic, Logistic R2 = 92.22% R2 = 93.90%
RMSE = 0.396 RMSE = 0.319
Two Hidden Slabs with Different Activation Function 3, 29, 29, 1 Linear (1, 1), Gaussian, Gaussian Camp, Logistic R2 = 94.57% R2 = 91.22%
RMSE = 0.330 RMSE = 0.382
Three Hidden Slabs, Different Activation Function 3, 19, 19, 19, 1 Linear (1, 1), Gaussian, Tanh, Gaussian Camp, Logistic R2 = 94.51% R2 = 91.19%
RMSE = 0.333 RMSE = 0.383
Two Hidden Slabs, Different Activation Function, Jump Connection: 3, 29, 29, 1 Linear (1, 1), Gaussian, Gaussian Camp, Logistic R2 = 94.59% R2 = 91.33%
RMSE = 0.330 RMSE = 0.380

659
660 S.H. Elsafi

Table 3 Predicted stage for Dongola during August and September (flood season) of 1998.
Date Tamaniat Atbara Observed Dongola Predicted Error Water level
Days Gauge (m) Gauge (m) Gauge (m) (m) (m) (m)
8/1/1998 14.30 14.02 12.70 13.15211 0.45211 225.18
8/2/1998 14.42 14.22 12.98 13.33011 0.35011 225.36
8/3/1998 14.58 14.365 13.32 13.46949 0.14949 225.50
8/4/1998 14.60 14.51 13.36 13.58692 0.22692 225.62
8/5/1998 14.69 14.60 13.56 13.67016 0.11016 225.70
8/6/1998 14.70 14.65 13.56 13.71087 0.15087 225.74
8/7/1998 14.68 14.72 13.60 13.7635 0.1635 225.79
8/8/1998 14.86 14.80 13.70 13.84758 0.14758 225.88
8/9/1998 15.14 14.86 13.75 13.92208 0.17208 225.95
8/10/1998 15.58 14.98 13.80 14.04664 0.24664 226.08
8/11/1998 15.86 15.16 13.95 14.19609 0.24609 226.23
8/12/1998 16.16 15.28 14.08 14.29524 0.21524 226.33
8/13/1998 16.10 15.38 14.20 14.36589 0.16589 226.40
8/14/1998 16.10 15.41 14.24 14.38748 0.14748 226.42
8/15/1998 16.08 15.46 14.26 14.42266 0.16266 226.45
8/16/1998 16.10 15.56 14.29 14.49409 0.20409 226.52
8/17/1998 16.18 15.58 14.50 14.51027 0.01027 226.54
8/18/1998 16.26 15.70 14.61 14.59528 0.014724 226.63
8/19/1998 16.38 15.80 14.74 14.66544 0.074562 226.70
8/20/1998 16.45 15.65 14.89 14.5644 0.325603 226.59
8/21/1998 16.48 15.60 15.03 14.53023 0.499772 226.56
8/22/1998 16.60 15.63 15.07 14.55258 0.517422 226.58
8/23/1998 16.68 15.70 15.05 14.60143 0.448572 226.63
8/24/1998 16.75 15.78 15.15 14.65614 0.493859 226.69
8/25/1998 16.80 15.80 15.23 14.66983 0.560172 226.70
8/26/1998 16.88 15.85 15.30 14.70344 0.596564 226.73
8/27/1998 16.93 15.92 15.38 14.74963 0.630369 226.78
8/28/1998 16.96 15.96 15.45 14.7757 0.674296 226.81
8/29/1998 16.96 16.05 15.58 14.83352 0.746475 226.86
8/30/1998 16.90 16.10 15.6 14.8651 0.734899 226.90
8/31/1998 16.86 16.10 15.65 14.86502 0.784977 226.90
9/1/1998 16.82 16.14 15.68 14.88993 0.790071 226.92
9/2/1998 16.76 16.17 15.74 14.90828 0.831717 226.94
9/3/1998 16.65 16.28 15.76 14.97457 0.78543 227.00
9/4/1998 16.50 16.33 15.80 15.00284 0.797162 227.03
9/5/1998 16.52 16.34 15.82 15.00899 0.811011 227.04
9/6/1998 16.51 16.32 15.86 14.99703 0.862966 227.03
9/7/1998 16.60 16.00 15.88 14.79988 1.080116 226.83
9/8/1998 16.69 16.17 15.88 14.90786 0.972136 226.94
9/9/1998 16.80 16.17 15.87 14.90846 0.961535 226.94
9/10/1998 16.90 16.28 15.90 14.9758 0.924203 227.01
9/11/1998 16.98 16.33 15.91 15.00572 0.90428 227.04
9/12/1998 17.00 16.30 15.91 14.98785 0.922149 227.02
9/13/1998 17.06 16.20 15.88 14.92718 0.952822 226.96
9/14/1998 17.10 16.12 15.86 14.87748 0.982523 226.91
9/15/1998 17.13 16.12 15.83 14.87737 0.952627 226.91
9/16/1998 17.16 16.12 15.82 14.87725 0.942751 226.91
9/17/1998 17.10 16.11 15.80 14.87121 0.92879 226.90
9/18/1998 17.00 16.00 15.79 14.80153 0.988475 226.83
9/19/1998 16.95 15.77 15.76 14.65007 1.109925 226.68
9/20/1998 16.90 15.77 15.65 14.65 0.999995 226.68
9/21/1998 16.80 15.91 15.48 14.74276 0.737243 226.77
9/22/1998 16.75 15.87 15.44 14.71617 0.723826 226.75
9/23/1998 16.7 15.78 15.42 14.65581 0.764191 226.69
9/24/1998 16.65 15.63 15.41 14.55309 0.856906 226.58
9/25/1998 16.80 15.64 15.16 14.5611 0.598895 226.59
9/26/1998 16.83 15.56 14.94 14.50585 0.434149 226.54
9/27/1998 16.68 15.65 14.96 14.56715 0.392845 226.60
9/28/1998 16.60 15.60 14.96 14.5318 0.428201 226.56
9/29/1998 16.40 15.60 14.90 14.52893 0.371072 226.56
9/30/1998 16.28 15.26 14.70 14.28399 0.416008 226.31
Artificial Neural Networks (ANNs) for flood forecasting at Dongola Station in the River Nile, Sudan 661

Figure 4 Maximum observed and predicted floodplain in 1998.

data, 1970–1985 was selected as the training (calibration) per-
iod, while 1986–1987 was selected as the verification period.
Four scenarios were examined, as shown in Table 1. Different
stations were used to provide input data, while Dongola Sta-
tion was used as the output. A three layer standard connection
model was applied to the four scenarios. To evaluate the
performance of the models in forecasting flows, the root mean
square error (RMSE) was used.
Model performance varied with the structure of the model
and scenarios used. The second model obtained the highest
efficiency, in which ‘‘Eddeim, Tamaniat, Atbara’’ as inputs
and ‘‘Dongola’’ as the output. This result indicates that
ANN models are viable for forecasting. Of note, there is no
rule for selecting the number of intermediate layers or the
number of neurons in the network. The procedure is based on
trial and error. Thus, different numbers of intermediate layers
and neurons are tried and assessed. The performance of each
model is checked, and the structure with the minimum number
of intermediate layers and neurons is normally selected.
Therefore, in the following steps, different model structures
are checked using ‘‘Eddeim, Tamaniat, Atbara’’ as the input
and ‘‘Dongola’’ as the output to select the best model. Table 2
662
shows 12 different ANN models using ‘‘Eddeim, Tamaniat,
Atbara’’ flood levels as input and ‘‘Dongola’’ flood levels as
output. Most models produced relatively good results, with
some performing better than others. The table shows the
model that was used in the first column, the slabs structure
used, the model efficiency R2 for the calibration period, and
the verification period and the Root Mean Square Error
(RMSE).
In all of the models the Learning rate = 0.1, Momen-
tum = 0.1, and Initial weights = 0.3.
5.1. ANN model results
For the flood disaster of 1998 at Dongola, the performance of
the ANN model was compared with the actual readings from
the River Nile. The ANN model was run to predict the flood
level at Dongola Station. The water surface level equaled the
predicted value at station plus the zero gauge of this station.
Therefore, the water surface level at Dongola equaled the
predicted stage using ANNs + 212.03 (m) Table 3.
The maximum flood level was recorded on September 5,
1998. On this date, the predicted gauge was 15.01 m. Thus
the predicted water surface level was 227.04 m. The expected
flooded area is presented in Fig. 4.
On the same day, the observed gauge was 15.91 and the
water surface level was 227.94 m. This output is also shown
in Fig. 4 to illustrate the actual maximum floodplain of this
year.
6. Conclusions
In conclusion, this research aimed to predict flooding levels
along the River Nile using ANNs. This method is advanta-
geous because only one variable is required, whereas other
models require several variables to produce accurate predic-
tions. Because of the lack of data and difficulty of acquisition,
this study provides a brief overview and of the steps taken
toward using ANN models for the River Nile.
This study facilitated the production of rapid and repeated
analytical testing. The first assessment was undertaken using
several different stations to check the performance of the
models. Four scenarios were examined, using the same output
station ‘‘Dongola’’, but different inputs: (1) ‘‘Tamaniat,
Atbara,’’ (2) ‘‘Eddeim, Tamaniat, Atbara,’’ (3) ‘‘Hassanab,
Atbara K3,’’ and (4) ‘‘Tamaniat, Atbara K3.’’
The second assessment was based on the second scenario
(inputs: Eddeim, Tamaniat, Atbara; output: Dongola), which
produced the best results in the first assessment. Then, 12
different model options were checked. There is no rule for select-
ing the number of intermediate layers and the number of neurons
in the ANNs. The best way to define these numbers is by testing
different model structures with different numbers of intermediate
S.H. Elsafi
layers and neurons. The performance of each scenario was
checked, and the structure with the minimum number of layers
and neurons was selected to avoid any redundancy.
ANNs have the advantage of simplicity when compared to
other more sophisticated models. Therefore, in situations
where information is lacking or difficult to obtain, the ANN
method provides the most viable option for flood forecasting.
Neural networks (ANNs) offer a means of reducing the
analytical costs of topographical and hydrological information
by reducing the amount of time spent analyzing the data.
Acknowledgments
This is a research project that was supported by a grant from
the Research Centre for the Humanities, Deanship of Scientific
Research at King Saud University.
References
[1] Anderson, Dave, McNeil, George. Artificial neural networks
technology. Data and analysis Centre for Software, Rome,
August 1992 <http://www.dtic.mil>.
[2] G.E.P Box, G.M. Jenkins, Time Series Analysis: Forecasting
control, Holden – Day, Oakland, California, 1976.
[3] Govindoraju, R.S., Rao, A.R., Artificial neural networks in
hydrology. Netherlands, 2000.
[4] Haykin, S., Neural networks. http://www.cul.salk.edu/.tewon/
ICA/teaching-KAIST/references.htmc/1994.
[5] B.C Hewiston, Crane, Precipitation Controls in Southern
Mexico, in Neural Nets, Kluwer Academic Publisher, 1994.
[6] Kendall, Graham. Introduction to Artificial Intelligence, Course
in University of Nottingham, School of Computer Science and
IT <http://www.cs.nott.ac.uk>, 2001.
[7] Özgür. Kisi, A combined generalized regression neural network
wavelet model for monthly stream flow prediction, KSCE J.
Civil Eng. 15 (8) (2011) 1469–1479.
[8] Moore, R.J., Jones, D.A., Black, K.B., Austin, R.M.,
Carrington, D.S., Tinnion, M., Akhondi, A., 1994. RFFS and
HYRAD: Integrated System for Rainfall and River Flow
Forecasting in Real-Time and their Application in Yorkshire.
BHS Ossasional paper No. 4, 12.
[9] Nile Basin Capacity Building Network (NBCBN). Flood and
Drought Forecasting and Early Warning Program, 2005.
[10] Obermeier and Barron. Integration of GIS and Artificial Neural
Networks for Natural Resources Applications. <http://
gis.esri.com/library/userconf/proc96/To150/pAp126/
P126.HTM>, 1989.
[11] Roy, Parthajit, Choudhury, P.S., Saharia, Manabendra.
Dynamic ANN modeling for flood forecasting in a river
network. In: International Conference on Modeling,
Optimization, and Computing (ICMOC), 2010.
[12] Christos Stergiou, Dimitrios Siganous, Neural Networks,
Imperial College, London, 1998.
[13] S.J. Yakowitz, Markov flow models and the flood warning
problem, Water Res Res 21 (1985) 81–88.