Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.

28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

ARTIFICIAL NEURAL NETWORK AND SUPPORT VECTOR

MACHINE IN FLOOD FORECASTING: A REVIEW
Azizah Suliman1, Nursyazana Nazri1, Marini Othman1, Marlinda Abdul
Malek1, and Ku Ruhana Ku-Mahamud2
1
Universiti Tenaga Nasional (UNITEN),Malaysia, azizah@uniten.edu.my,nursyazana@uniten.edu.my,
marini@uniten.edu.my; marlinda@uniten.edu.my
2
Universiti Utara Malaysia (UUM), Malaysia, ruhana@uum.edu.my

ABSTRACT. Flood is a natural phenomenon that can cause havocs and

deaths. Although flood is sometimes unavoidable, early flood forecasting
can be helpful for people to take precaution. In the past decades, researchers
have been working on flood forecasting models using artificial intelligence
(AI). AI models such as Artificial Neural Network (ANN) and Support
Vector Machine (SVM) have been developed and implemented in different
locations to help in weather forecasting over the past years. This paper
reviews both methods and compares their experimental results.

Keywords: Artificial Neural Network, Support Vector Machine, Flood

forecasting, flood

INTRODUCTION

In China, millions of people are affected each year and are forced to evacuate promptly
leaving their belongings behind. While in Kuala Lumpur, thousands of people are stranded in
the middle of the city center, patiently waiting in their vehicles hoping and praying for it to
subside soon. Two different scenarios, but are caused by one same thing - flood. Definition
given by Oxford English Dictionary for flood is an overflow of a large amount of water
beyond its normal limits. (Abhas et al., 2012) generally characterized flood into fluvial (or
river) floods, pluvial (or overland) floods, coastal floods, groundwater floods or the failure of
artificial water systems. What causes flood can vary from heavy downpour to sea level rise. It
can last for a few hours to days, or even a longer period depending on the cause. The deadliest
flood in China that occurred in 1931, also known as 1931 Central China Flood killed 3.7
million people – recorded as the worst case ever. Perhaps, this is the worst natural disaster of
20th century. As time goes by, researchers started to take precaution by developing flood
forecasting model in order to give early warning to citizens in order to avoid catastrophe.

FLOOD FORECASTING MODEL

Over the past decades, researchers have shown interest in developing flood forecasting
model. Weerts and Beckers (2009) from Netherlands have constructed a framework named
Uncertainty Framework for flood and storm surge forecasting. It is built around procedural
and operational constraints. The framework is said to help in deciding which method, and in
which part of the model chain, it is most suitable to increase the accuracy or quantifying the
(predictive) uncertainty of the flood forecast. Figure 1 shows the uncertainty framework that
offers a structured approach to reduce the predictive uncertainty.

327
 Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.
28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

Uncertainty is divided into three parts in model used for flood forecasting. a) Input
boundary conditions for the prediction. b) Initial conditions of the area or model. c) Behavior
of the model during the prediction phase.

Although there are other models as stated

in the figure that use different algorithms,
the application of AI in handling and
reasoning under uncertainty has been used
widely in diverse areas (Levitt, 1988). The
provision of making inferences with
uncertainty and the availability of learning
mechanism in AI techniques makes it a very
useful tool in making prediction and
forecasting. Among the common AI method
used in flood forecasting are ANN and
SVM. This paper focuses solely on models
using artificial intelligence and is divided
into two parts: (1) Flood forecasting models
using Artificial Neural Network; and (2)
Flood forecasting models using Support
Vector Machine. This paper will further
Figure 1. Uncertainty Framework discuss on ANN and SVM in flood
forecasting domain in Discussion section.

Artificial Neural Network

ANNs were first introduced to water resources research for their use to predict monthly
water consumption and to estimate occurrences of flood. Since then, ANNs have been used
for a number of different water resource applications which include time-series prediction for
rainfall forecasting, rainfall-runoff processes and river salinity. ANNs have also been used for
modeling soil and water table fluctuations, pesticide movement in soils, water table
management and water quality management (Parson, 1999).

Models of Artificial Neural Network

Mandal et al. (2005) employed ANN model, namely Multi-layer Perceptron (MLP) using
back-propagation network technique and used delta rule for training. Environmental
parameters used for this research are temperature, humidity, underground water level,
precipitation and wind speed. It is found that underground water level is the most significant
parameter for the prediction model. Simulation runs for this model using NeuroSolutions
v4.10 has resulted in 97.33% of given overall prediction accuracy.

Ayalew et al. (2007) adopted three-layer back-propagation ANN model for real-time flood
forecasting in Omo River, Ethiopia. Floods in Omo river are sudden, non-linear and of short
duration. ANN models are best suited for forecasting such types of floods. This research uses
sigmoid function which is commonly used for hydrological studies. Two important
parameters in this research are magnitude and time-to-peak discharge. Comparisons of
observed and forecasted runoff values for training and testing for all models showed little
discrepancies.

328
 Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.
28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

Tan et al. (2008) combined two models of ANN and SVM in order to come out with a new
model called Reward Learning Ensemble (RLEnsemble). One model will learn the problem
while the other will learn from the error of its counterpart. SVM is the first model,
subsequently followed by ANN using MLP. Error produced by SVM will be the input for
MLP. Output produced from MLP will be taken as final prediction. RLEnsemble is the one
with highest accuracy in predicting the rainfall pattern in Singapore.

Pang et al. (2011) developed a non-linear perturbation model adopting ANN (NLPM-
ANN) and the results are compared to ANN and also linear perturbation model (LPM). In this
model, it is recognized that seasonal hydrological behavior, as incorporated in the model is a
very important source of information in flood forecasting. It is shown that the NLPM-ANN
obtains better simulation results than ANN by 2.7%, while results compared to LPM is higher
by 6.32%.

Support Vector Machine

The idea of Support Vector Machine was initially developed in Russia in the 60's by
Vapnik and Lerner. Vapnik further developed the field and wrote the definitive book on the
subject. A SVM consists of a set of support vectors and a kernel function. The support vectors
are a set of vectors from the training data. The support vectors together with the kernel create
the function approximation.

Models of Support Vector Machine

Han et al. (2007) is an example that employed Sequential Minimal Optimization algorithm
(SMO) for their SVM model. On top of that, they also incorporated an algorithm called
SVMLight. The data used for model training is from October 1955 to September 1963, while
the testing data is from November 1972 to November 1974. Data are from the catchment in
Bird Creek, Oklahoma, USA. Tools used for this research are LIBSVM, coupled with Gunn‘s
Toolbox for data normalization. A comparison with some benchmarking models has been
made and it demonstrates that SVM is able to surpass all of them in the test data series, at the
expense of a huge amount of time and effort.

Wiriyarattanakul et al. (2008) used fuzzy support vector machine regression (FSVMR) to
predict the runoff of Yom River at Sukhotai province, Thailand. They selected runoff data
from June until October, between 1995-2000 and 2002-2004. The data are compared using
FSVMR and standard SVMR. Average MAE of the best FSVMR model is 3.627 m3/s and
7.728 m3/s in the training and testing data set, respectively. While the average MAE of the
best SVMR model is 3.954 m3/s and 8.041 m3/s in the training and testing data set,
respectively. The MAE of the blind test data set from the best FSVMR model and best SVMR
model are 7.8588 m3/s and 9.0895 m3/s, respectively. This shows that the FSVMR is more
effective and efficient in forecasting runoff than the standard SVMR.

Hu et al. (2011) adopted SVM model which provided higher runoff forecast accuracy
compared to the forecasts of the ANN model for monthly runoff in the upstream of the Fenhe
River. It used a hybrid forecasting technique of support vector regression and its applications
for rainfall-runoff forecasting in order to investigate its feasibility in forecasting runoff
amounts. Various SVM models were trained to simulate monthly and daily rainfall-runoff
relationships and compared with the ANN model. The results show that the SVM model has
higher nonlinear mapping capabilities and thus can more easily capture runoff data patterns
than can the ANN models.

329
 Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.
28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

Bell et al. (2012) adopted SVM for river runoff forecasting, with Smola/Scholkopf's
Sequential Minimal Optimization algorithm for training a SVM with a RBF kernel. They used
monthly precipitation and snow data gathered from 10 precipitation monitoring stations and
28 snow monitoring stations located in the American River basin. The calculations were made
using WEKA v3.6 and the results using SMOreg with a RBF kernel yield a relative absolute
error 48.65% versus 63.82% for the human ensemble forecast.

SUMMARY

All the previous works by researchers are summarized as shown in the Table 1 below for
easy comparison. Even though a direct comparison might not seem fair as the parameter used
differ, it is highly noticeable that SVM does give a better accuracy. (Han et al., 2007,
Wiriyarattanakul et al., 2008, Hu et al., 2011, Bell et al., 2012).

Table 1. Summary of ANN and SVM in flood forecasting models

Method
Location Technique Parameters Tools Outcome
ANN SVM
i. Temp
erature
ii. Humi
dity Water level is the
Multi-Layer
India iii. Unde key parameter
Perceptron NeuroSoluti
 (Mandal et rground related to flood.
(delta rule for ons v4.10
al., 2005) water level Overall prediction
training)
iv. Preci accuracy is 97.33%.
pitation
v. Wind
speed
Comparisons of
Omo 3 layers back- i. Magn
observed and
River, propagation itude
forecasted runoff
 Ethiopia ANN ii. Time -
values for all
(Ayalew et (sigmoid -to-peak
models showed
al., 2007) function) discharge
little discrepancies.
RLEnsemble is the
Error
RLEnsemble one with highest
Singapore produced by
(combination accuracy in
  (Tan et al., SVM will be -
of ANN and predicting rainfall
2008) the input for
SVM) pattern in
ANN.
Singapore.
Lower
NLPM-ANN
Yellow Non-linear
obtains better
River, Perturbation Discharge
 - simulation results
China Model time series
than the APM and
(Pang et adopting ANN
ANN.
al., 2011)
LIBSVM,
It demonstrates that
Oklahoma, Gunn‘s
SVM is able to
USA SMO, Toolbox (for
 River flow surpass all of the
(Han et al., SVMLight data
compared models in
2007) normalizatio
the test data series.
n)
Yom Fuzzy Support Average error of
 River, Vector Runoff - FSVMR model is
Thailand Machine lower than SVMR

330
 Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.
28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

(Wiriyaratt Regression models.

anakul et
al., 2008)
The results show
Fenhe
SVM model has
River, Support
Precipitation- higher non-linear
 China Vector -
Runoff mapping
(Hu et al., Regression
capabilities than
2011)
ANN model.
Results using
American Smola/Scholk SMOreg yield a
River, opf‘s Monthly Machine relative absolute
 California Sequential precipitation Learning error of 48.65%
(Bell et al., Minimal and snow data Tool WEKA versus 63.82% for
2012) Optimization the human
ensemble forecast.

DISCUSSION

G. Zhang et al. (1998) stated reasons why ANN is highly used for forecasting. ANNs are
well suited for problems whose solutions require knowledge that is difficult to specify but for
which there are enough data or observations. Second, ANNs can generalize. As forecasting is
performed via prediction of future behavior from examples of past behavior, it is suitable to
be applied in forecasting flood. Records of rainfall in past years can be trained to see the trend
and eventually a prediction can be made. Third, ANNs are nonlinear. ANN, which are
nonlinear data-driven approaches are capable of performing nonlinear modeling without
knowledge about the relationships between input and output variables. Thus, they are a more
general and flexible modeling tool for forecasting.

In choosing suitable AI models for forecasting model – not limited to flood, it is always
crucial to question ourselves of how well will the model make predictions for events that are
not in the training set. As for ANN model, when a little modification is done to be NLPM-
ANN model, it becomes a flexible tool for flood forecasting, especially in the area without
detailed hydrometer data, a common situation particularly in developing countries. On the
other hand, although we can see SVM has been increasingly used in recent hydrological
modeling research, it still has its limitations such as poor performance in skewed dataset. Q.
Li et al. (2007) stated that SVM is highly dependent on its parameters and the kernel
parameters. The inference process of SVM may become time-consuming and computationally
expensive due to the large number of support vectors. Looking at the results of previous
researches, it is highly recommended to further explore SVM in building flood forecasting
model.

CONCLUSION

In reducing side effects of high computation time of SVM, it is also recommended that the
use of parallel SVM be investigated. With the availability of GPU and multicore processors
on current machines, that would be the best direction to take in the flood forecasting model
development. A preliminary work has been done to develop a flood forecasting model for a
selected river in Malaysia using parallel SVM on GPU. Further results of this will be
discussed in next publication.

331
 Proceedings of the 4th International Conference on Computing and Informatics, ICOCI 2013 Paper No.
28-30 August, 2013 Sarawak, Malaysia. Universiti Utara Malaysia (http://www.uum.edu.my ) 030

ACKNOWLEDGMENT

The authors wish to thank the Ministry of Higher Education Malaysia for funding this
study under Long Term Research Grant Scheme (LRGS/b-u/2012/UUM/Teknologi
Komunikasi dan Informasi).

REFERENCES
Abhas, K. J., Bloch, R. & Lamond, J. (2012). Cities and Flooding - A Guide to Integrated Urban Flood
Risk Management for the 21st Century.

Ayalew, L., Möller, D. P. F. & Reik, G. (2007). Using ANN for Real-Time Flood Forecasting. Summer
Computer Simulation Conference (SCSC).

Mandal, S., Saha, D., & Banerjee T. (2005). A Neural Network Based Prediction Model for Flood in A
Disaster Management System with Sensor Networks. International Conference on Intelligent
Sensing and Information Processing (ICISIP).

Tan, T. Z., Lee, G. K. K., Liong, S. Y., Lim, T.K., Chu, J. & Hung. T. (2008). Rainfall Intensity
Prediction by a Spatial-Temporal Ensemble. International Symposium on Neural Networks
(ISNN), pp. 1721-1727.

Hu, C., Wu, Z., Wang, J. & Liu L. (2011). Application of the SVM on Precipitation-Runoff Modeling
in Fenhe River. International Symposium on Water Resource and Environmental Protection
(ISWREP).

Bell, B., Wallace, B. & Du. Z. (2012). Forecasting River Runoff through Support Vector Machines.
Proc. 11th IEEE International Conference on Cognitive Informatics & Cognitive Computing
(ICCI*CC‘12).

Li, Q., Jiao, L. & Hao Y. (2007). Adaptive Simplification of Solution for Support Vector Machine.
Pattern Recognition, vol. 40, 972-980.

Wiriyarattanakul, S., Auephanwiriyakul, & Umpon, S. N. T. (2009). Runoff Forecasting Using Fuzzy
Support Vector Regression. IEEE International Symposium on Intelligent Signal Processing and
Communication Systems (ISPACS).

Parson, C. S. (1999). Use of Artificial Neural Networks to Facilitate Hydrological Modeling on the
Internet. PhD thesis

Zurada (1992). Artificial Neural Systems. West Publishing Company.

Weerts, A. & Beckers, J. (2009), Uncertainty Framework for Operational Flood and Storm Surge
Forecasting. Research for Flood Control 2015 Program.

Pang, B. & Liang, Y. (2011), Application of an Improved Neural Network to Flood Forecasting of the
Lower Yellow River. 4th International Symposium on Computational Intelligence and Design
(ISCID).

Han, D., Chan, & Zhu, L. N. (2007), Flood Forecasting Using Support Vector Machines. Journal of
Hydroinformatics, IWA Publishing.

Levitt, T. S. (1988), Uncertainty in Artificial Intelligence. AI Magazine, vol. 9, Number 4, pp. 77-78.

332