Typhoon Damage Forecasting with Self-Organizing Maps,

Multiple Regression and Decision Trees

Kazuhiro Kohara and Ryo Hasegawa

Department of Electrical, Electronics and Computer Engineering

Chiba Institute of Technology, 2-17-1, Tsudanuma, Narashino, Chiba, 275-0016, Japan

Abstract. Damage caused by typhoons to both people and structures has de-

creased in Japan due to improvements of countermeasures against natural disas-

ters, however, such damage still occurs. A typhoon warning that represents the

risk posed by a typhoon with high accuracy should be issued appropriately.

Thus, we propose a new typhoon warning system which forecasts the likely ex-

tent of damage associated with a typhoon towards humans and buildings. The

relation between typhoon data and damage data is investigated and typhoon

damage is forecast using typhoon data. Self-organizing maps (SOM), multiple

regression analysis and decision trees were used for typhoon damage forecast-

ing. We consider two types of forecasting: two-class (yes or no) and three-class

(small, medium or large scale) damage forecasting. Experimental results on ac-

curacy of two-class and three-class forecasting with SOM were 93.3% and

96.8%, respectively. The accuracy with SOM was much better than that with

multiple regression and decision trees. We recommend a new typhoon damage

forecasting method based on these results.

1 Introduction

Intelligent techniques such as back-propagation neural networks (BPNN) [1], self-

organizing maps (SOM) [2], decision trees [3] and Bayesian networks [4] have been

extensively investigated, and various attempts have been made to apply them to iden-

tification, prediction and control [e.g., 1-10]. Harada et al. applied BPNN to forecast-

ing typhoon course [8], Takada et al. applied BPNN to forecasting typhoon damage

of electric power systems [9] and Udagawa et al. applied Bayesian networks to rain

prediction [10]. This paper applies intelligent techniques to forecasting typhoon dam-

age to human and buildings.

Damage caused by typhoons to both people and structures has decreased in Japan

due to improvements of countermeasures against natural disasters, however, such

damage still occurs [11, 12]. A typhoon warning that represents typhoon menace with

high accuracy should be issued appropriately. A typical typhoon warning currently

issued may be “This typhoon is large and very strong”. We propose a new typhoon

warning which forecasts the risk of damage scale to both human and buildings. We

investigate relation between typhoon data and damage data and forecast typhoon

damage using typhoon data. The typhoon data includes the month when the typhoon

Kohara K. and Hasegawa R. (2009).

Typhoon Damage Forecasting with Self-Organizing Maps, Multiple Regression and Decision Trees.

In Proceedings of the 5th International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing, pages 106-111

DOI: 10.5220/0002254601060111

 SciTePress

was born, latitude and longitude where the typhoon was born, lowest atmospheric

pressure, maximum wind speed and total precipitation. Damage data includes human

damage data such as number of fatalities and injured persons and building damage

data such as number of completely destroyed houses and number of houses under

water.

We use SOM, multiple regression analysis and decision trees for typhoon damage

forecasting. SOM [2] are neural networks which consist of two layers: input layer and

map layer. As an interesting feature of SOM, teaching vectors are not required and

input vectors are automatically classified in accordance with similarity, updating the

weight of the winning neuron and the neighbor neurons. After trained by SOM algo-

rithm, the weight vectors of the neurons form the cluster of input vectors. A decision

tree [3] is an inductive learning algorithm. In a decision tree algorithm, an explicit

decision boundary is extracted from the training data, and an example E is classified

into class c if E falls into the decision area corresponding to c. Viscovery SOMine 4.0

was used as SOM software and See5 release 1.19 was used as decision tree software

with default parameter values.

2 Forecasting Damage Data using Typhoon Data

139 data records of typhoon data and damage data from June 1981 to September 1999

were collected from the typhoon database [13, 14]. The types of typhoon and damage

data are shown in Table 1. There are nine types of typhoon data and nine types of

damage data, divided into three types of human damage and six types of building

damage. We used 111 data records (to September 1995) for learning and 28 data

records (from July 1996) for testing.

Table 1. Types of typhoon data and damage data used in this study.

Typhoon data

Month when the typhoon was born,

Latitude and longitude where the typhoon was born,

Lowest atmospheric pressure,

Maximum wind speed,

Total, one-hour and twenty-four-hour precipitation,

Life span

Damage data

Number of fatalities,

Number of injured persons,

Number of dead and injured persons

Number of completely destroyed houses,

Number of half destroyed houses,

Number of partially destroyed houses,

Total number of damaged houses,

Number of houses under water,

Total number of destroyed non-house structures

The average and maximum of every damage type are shown in Table 2. The min-

imum of every damage type was zero.

107

Table 2. Average and maximum of every type of damage data.

Data type Average Maximum

Number of fatalities 5.5 100

Number of injured persons 39.2 1499

Number of dead and injured persons 44.8 1561

Number of completely destroyed houses 21.9 541

Number of half destroyed houses 1839.8 169877

Number of partially destroyed houses 1051.7 85989

Total number of damaged houses 2913.4 170418

Number of houses under water 7829.6 174124

Total number of destroyed non-house structures 163.6 15840

In this study, we consider two types of typhoon damage forecasting: two-class

(yes or no) and three-class (small, medium or large scale) damage forecasting.

3 Two-class (Yes or No) Damage Forecasting

In two-class damage forecasting, a predictor is trained by two values (0 and 1). In this

case, 0 means that the damage is zero (no) and 1 means that the damage is not zero

(yes). Experiments were made with nine types of continuous typhoon data as inputs

and one damage data (two values) as an output. Here, we expect that typhoon data

such as lowest atmospheric pressure, maximum wind speed and precipitation can be

forecast with high accuracy by a weather forecasting system such as Japanese

SYNFOS [15] and hence actual typhoon data was used as inputs.

Table 3. Average accuracy of two-class (yes or no) damage forecasting.

Method Learning data Test data

Self-organizing maps (SOM) 100% 93.3%

Multiple regression (MR) 70.9% 70.2%

Decision trees (DT) 77.7% 63.9%

Table 4. Accuracy of two-class (yes or no) damage forecasting for test data.

Damage type SOM MR DT

No. fatalities 92.9% 57.1% 50.0%

No. injured persons 89.3% 75.0% 75.0%

No. dead and injured persons 96.4% 89.3% 85.7%

No. completely destroyed houses 92.9% 57.1% 57.1%

No. half destroyed houses 89.3% 71.4% 71.4%

No. partially destroyed houses 92.9% 67.9% 60.7%

Total no. of damaged houses 96.4% 75.0% 64.3%

No. houses under water 96.4% 85.7% 78.6%

Total no. destroyed non-house structures 92.9% 53.6% 32.1%

Average 93.3% 70.2% 63.9%

SOM: self-organizing maps, MR: multiple regression, DT: decision trees

The average accuracy of two-class (yes or no) damage forecasting for the three in-

telligent methods is shown in Table 3. Here, average accuracy means the average of

108

the accuracy of nine damage data. The average accuracy of the learning and test data

using SOM was 100% and 93.3%, respectively. This experiment confirmed that dam-

age data are well related with typhoon data and that SOM learned the nonlinear rela-

tion very well. The accuracy for each damage test data is shown in Table 4. Each

damage data was forecast very well by SOM. The accuracy with SOM was much

better than that with multiple regression and decision trees.

4 Three-class (Small, Medium or Large Scale) Damage

Forecasting

In three-class damage forecasting, two experiments were made with nine types of

continuous typhoon data as inputs and one damage data as an output. In the first ex-

periment, a predictor is trained by continuous damage data. As this is a regression

problem, decision trees were not used. The average of each damage data was calcu-

lated as shown in Table 2. Small scale corresponds to under half of the average, me-

dium scale corresponds to between half of the average and the average, and large

scale corresponds to over the average, respectively. The prediction was considered

accurate when both the predicted value and the actual value correspond to the same

size. The average accuracy of three-class damage forecasting when trained by conti-

nuous damage data is shown in Table 5. The average accuracy of the learning and test

data with SOM was 100% and 78.6%, respectively. The accuracy for each damage

type is shown in Table 6. The accuracy with SOM was much better than that with

multiple regression, however, each damage data was not always forecast very well by

SOM. For example, accuracy for number of fatalities and number of partially de-

stroyed houses was 67.9% and 92.9%, respectively.

Table 5. Average accuracy of three-class (small, medium or large scale) damage forecasting

when trained by continuous damage data.

Method Learning data Test data

Self-organizing maps (SOM) 100% 78.6%

Multiple regression (MR) 52.8% 43.7%

Table 6. Accuracy of three-class (small, medium or large scale) damage forecasting for test

data when trained by continuous damage data.

Damage type SOM MR

No. fatalities 67.9% 35.7%

No. injured persons 75.0% 46.4%

No. dead and injured persons 75.0% 42.9%

No. completely destroyed houses 60.7% 46.4%

No. half destroyed houses 89.3% 42.9%

No. partially destroyed houses 92.9% 42.9%

Total no. of damaged houses 89.3% 39.3%

No. houses under water 82.1% 39.3%

Total no. destroyed non-house structures 75.0% 57.1%

Average 78.6% 43.7%

109

In the second experiment, a predictor is trained by three values (0, 1 and 2). As

this is a classification problem, decision trees were used. In the learning data, 0 means

that the damage is small scale, 1 means the damage is medium scale and 2 means the

damage is large scale. The prediction was considered accurate when the predicted

size was equal to the actual size. The average accuracy of three-class damage fore-

casting when trained by three values is shown in Table 7. The average accuracy of the

learning and test data with SOM was 100% and 96.8%, respectively. This also con-

firmed that damage data are also well related with typhoon data. The accuracy for

each damage type is shown in Table 8. Each damage type was also forecast very well

by SOM. For example, accuracy for number of fatalities and number of partially

destroyed houses was 85.7% and 100%, respectively. The accuracy with SOM was

also much better than that with multiple regression and decision trees.

Table 7. Average accuracy of three-class (small, medium or large scale) damage forecasting

when trained by three values.

Method Learning data Test data

Self-organizing maps (SOM) 100% 96.8%

Multiple regression (MR) 77.5% 65.1%

Decision trees (DT) 90.1% 78.6%

Table 8. Accuracy of three-class (small, medium or large scale) damage forecasting for test

data when trained by three values.

Damage type SOM MR DT

No. fatalities 85.7% 42.9% 78.6%

No. injured persons 100% 53.6% 71.4%

No. dead and injured persons 100% 53.6% 71.4%

No. completely destroyed houses 92.9% 39.3% 53.6%

No. half destroyed houses 100% 85.7% 96.4%

No. partially destroyed houses 100% 85.7% 85.7%

Total no. of damaged houses 100% 85.7% 92.9%

No. houses under water 92.9% 50.0% 60.7%

Total no. destroyed non-house structures 100% 89.3% 96.4%

Average 96.8% 65.1% 78.6%

5 Conclusions

We investigated typhoon damage forecasting with intelligent techniques. Using nine

types of typhoon data as inputs to SOM, experimental results on the average accuracy

of two-class (yes or no) and three-class (small, medium or large scale) damage fore-

casting were 93.3% and 96.8%, respectively. The accuracy with SOM was much

better than that with multiple regression and decision trees. As a result, a typhoon

forecasting method is proposed as follows: 1) Evaluate two-class damage forecasting,

2) When two-class forecasting result is yes, evaluate three-class damage forecasting,

3) Issue a typhoon warning based on above three-class damage forecasting. For ex-

ample, such a warning may be issued, “According to the Japanese typhoon database,

we forecast that the coming typhoon has a risk of causing both large scale human and

building damage. Please take care.” In further research, we will consider more de-

110

tailed damage forecasting and use other predictors such as support vector machines.

References

1. Rumelhart, D., Hinton, G., Williams, R.: Learning internal representations by error propa-

gation. In: Rumelhart, D., McClelland, J., the PDP Research Group (eds.): Parallel Distri-

buted Processing, Vol. 1. MIT Press, Cambridge, MA (1986).

2. Kohonen, T.: Self-Organizing Maps. Springer (1995).

3. Quinlan, J.: C4.5: Programs for Machine Learning. Morgan Kaufmann (1993).

4. Jensen, F.: Bayesian Networks and Decision Graphs. Springer (2001).

5. Pham, D., Liu, X.: Neural Networks for Identification, Prediction and Control. Springer

(1995).

6. Kohara, K.: Neural networks for economic forecasting problems. In: Leondes, C. T.: Expert

Systems - The Technology of Knowledge Management and Decision Making for 21st Cen-

tury -. Academic Press (2002).

7. Kohara, K.: Combining selective-presentation and selective-learning-rate approaches for

neural network forecasting of stock markets, Proceedings of International Workshop on Ar-

tificial Neural Networks and Intelligent Information Processing. Madeira (2008) Pp 3-9.

8. Harada, H., Momma, E., Ishii, H., Ono, T.: Forecast of typhoon course using multi-layered

neural network (III), Proceedings of National Convention of the Institute of Electrical En-

gineers of Japan, Toyama (2007) Vol. 3, 111.

9. Takata, H., Kawaji, S., Ha, T.: Study on a Prediction Method of Typhoon Damage of Elec-

tric Power Systems in each District on the Main Island in Kagoshima Prefecture, Technical

Report 48, Faculty of Engineering, Kagoshima University (2006).

10. Udagawa, S., Nishio, S., Kimura, M.: Rain prediction by the Bayesian network, Proceed-

ings of National Convention of the Information Processing Society of Japan (2005) Vol.3,

237-238.

11. Murayama, K.: Introduction to Typhoon Study. Yama-Kei Publishers (2006).

12. Nyoumura, Y.: Weather Damage Prediction and Countermeasure. Ohmsha (2002).

13. National Research Institute for Earth Science and Disaster Prevention (2008).

http://www.bosai.go.jp/index.html

14. National Institute of Informatics (2008). http://agora.ex.nii.ac.jp/digital-typhoon

15. Japan Weather Association (2008). http://www.jwa.or.jp/synfos/

111