Você está na página 1de 14
Acta Geophysica vol. 55, no. 3, pp. 369-382 DOI 10.2478/s11600-007-0020-8 Feed forward Artificial Neural Network model

Acta Geophysica

vol. 55, no. 3, pp. 369-382 DOI 10.2478/s11600-007-0020-8

Feed forward Artificial Neural Network model to predict the average summer-monsoon rainfall in India

Surajit CHATTOPADHYAY

Pailan College of Management and Technology, Kolkata, India e-mail: surajit_2008@yahoo.co.in

Abstract

In the present research, possibility of predicting average summer-monsoon rainfall over India has been analyzed through Artificial Neural Network model. In formulating the ANN – based predictive model, three-layer network has been con- structed with sigmoid non-linearity. The monthly summer monsoon rainfall totals, tropical rainfall indices and sea surface temperature anomalies have been consid- ered as predictors while generating the input matrix for the ANN. The data pertain- ing to the years 1950-1995 have been explored to develop the predictive model. Fi- nally, the prediction performance of neural net has been compared with persistence forecast and Multiple Linear Regression forecast and the supremacy of the ANN has been established over the other processes.

Key words: summer-monsoon rainfall, prediction of monsoon rainfall, Artificial Neural Network model, Multiple Linear Regression forecast.

1.

INTRODUCTION

Weather forecasting is one of the most imperative and demanding operational respon- sibilities carried out by meteorological services all over the world. It is a complicated procedure that includes numerous specialized fields of know-how (Guhathakurata 2006). The task is complicated because in the field of meteorology all decisions are to be taken in the visage of uncertainty. Several authors (i.e., Brown and Murphy 1988, Handerson and Wells 1988, Wilks 1991, Elsner and Tsonis 1992, Jacovides et al. 1994, Kondratyev and Varotsos 2001a, b, Cartalis and Varotsos 1994, Varotsos et al. 2001) have discussed the uncertainty associated with the weather systems. Chaotic features associated with the atmospheric phenomena also have attracted the attention

© 2007 Institute of Geophysics, Polish Academy of Sciences

370

S. CHATTOPADHYAY

of the modern scientists (Varotsos 2005, Varotsos and Krik-Davidoff 2006, Sivakumar 2000, 2001, Sivakumar et al. 1999, Men et al. 2004). Different scientists over the globe have developed stochastic weather models which are basically statistical models that can be used as random number generators whose output resembles the weather data to which they have been fit (Wilks 1999). Statistical models have the drawback that in most of the cases they depend upon several tacit assumptions regarding the sys- tem. But, a chaotic system like the atmosphere cannot be bound by any postulation. The numerical models are based on the scheme of nonlinear operator equations pre- vailing in the atmospheric system. However, in the absence of any analogue solution of this system of operator equations, numerical solutions based on different assump- tions are the only alternative. Furthermore, the chaotic behaviors of these nonlinear equations sensitive to initial conditions make it more intricate to solve these equations (Guhathakurta 2006). As a result, flawed forecast comes out.

Amongst all weather happenings, rainfall plays the most imperative role in hu- man life. Human civilization to a great extent depends upon its frequency and amount to various scales. Several stochastic models have been attempted to forecast the occur- rence of rainfall, to investigate its seasonal variability, to forecast monthly/yearly rain- fall over some given geographical area. Daily precipitation occurrence has been viewed through Markov chain by Chin (1977). Gregory et al. (1993) applied a chain- dependent stochastic model, known as Markov chain model to investigate interannual variability of area average total precipitation. Wilks (1998) applied mixed exponential distribution to simulate precipitation amount at multiple sites exhibiting realistic spa- tial correlation.

The Asian monsoon circulation influences most of the tropics and subtropics of the Eastern Hemisphere and a major portion of the Earth’s population. Accurate, long lead prediction of monsoon rainfall can improve planning to mitigate the adverse im- pacts of monsoon variability and to take advantage of beneficial conditions (Reddy and Salvekar 2003). The southwest (summer) and the northeast (winter) monsoons in- fluence weather and climate between 30N and 30S over the African, Indian and Asian land masses (Reddy and Salvekar 2003 and references therein). The variability in the monsoon rainfall depends heavily upon the sea surface temperature anomaly over the Indian Ocean (Clark et al. 2000). As the extra tropical circulation anomalies display energy dispersion away from the region of anomalous tropical convection, they have been interpreted as a Rossby wave respons e to the latent heat release associated with the tropical convection (Ferranti et al. 1990). In regions of anomalous tropical heating, there is a dynamical response with anomalous large-scale ascent and upper tropospheric divergence, which acts as a Rossby wave source (Sardeshmukh and Hoskins 1988) for extratropical waves. Conversely, in regions of reduced convection and anomalous cooling, the tropical response is one of anomalous descent and upper- tropospheric convergent inflow (Matthews et al . 2004). Positive (negative) sea sur- face temperature (SST) anomalies lead the enhanced (suppressed) MJO convection by approximately 10-12 days (a quarter cycle), consistent with the atmosphere respond- ing to the ocean forcing (Matthews 2004). The SST anomalies themselves have been

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

371

simulated in thermodynamical ocean models as the response to the observed anoma- lous surface fluxes of latent heat and shortwave radiation (Matthews 2004, Shinoda and Hendon 1998) without the need to a mass flux approach and closed on buoyancy. SST anomalies have been addressed by Woolnough et al. (2001), who examined the equilibrium response to an idealized, eastward-propagating equatorial SST dipole anomaly in an ‘‘aquaplanet’’ atmospheric general circulation model (AGCM).

The present paper endeavors to develop an Artificial Neural Network (ANN) model to forecast average rainfall during summer-monsoon in India. The term “mon- soon” seems to have been derived either from the Arabic mausin or from the Malayan monsin. It was first applied to southern Asia and the adjacent waters, where it referred to the seasonal surface air streams which reverse their directions between winter and summer, southwest in summer and northeast in winter in this area. In 1686 Halley ex- plained the Asiatic monsoon as resulting from thermal contrasts between the continent and oceans. During the summer, the continent is heated, leading to rising motion and lower pressure. This induces airflow from sea to land at low elevations.

Eastern and southern Asia has the Earth’s largest and best-developed monsoon circulations. The tropical monsoon circulation of southern Asia, including India- Pakistan and southeast Asia, differs significantly from east Asia monsoon. The Indian monsoon is effectively separated from that of China by the Himalayan-Tibet system. In summer, a deep and widespread surface pressure trough extends across northern In- dia-Pakistan into southeast Asia. This is part of the planetary intertropical convergence zone, which here reaches its maximum poleward displacement. To the south of the trough is a deep current of maritime tropical air called the southwest monsoon. This current appears to originate in the southeast trades of the Indian Ocean east of Africa. As this stream of air approaches and crosses the equator its direction becomes south- erly and then southwesterly. Along the Somali coast of Africa the flow becomes espe- cially strong, taking the form of a low-level jet. In crossing the Arabian Sea the southwesterly current gains considerable moisture and becomes less stable. This un- stable southwesterly current crosses India, continues eastward over the Indochina pen- insula, and then move northward over much of eastern Asia. It is a great moisture source for most of southern Asia.

Indian economy is standing on Indian summer monsoon. So the prediction of In- dian summer monsoon is a challenging topic to Indian atmospheric scientists. Hasten- rath (1988) developed statistical model using regression method to predict Indian summer monsoon rainfall anomaly. Rajeevan (2001) discussed the problems and prospects in prediction of Indian summer monsoon and revealed that Indian summer monsoon predictability exhibits epochal variations. Gadgil et al. (2005) investigated the causes of failure in prediction of Indian summer monsoon and expected the dy- namical models to generate better prediction only after the problem of simulating year-to-year variation of monsoon is addressed. Kishtawal et al. (2003) assessed the feasibility of a nonlinear technique based on genetic algorithm, an artificial intelli- gence technique for the prediction of summer rainfall over India. Guhathakurta (2006) implemented ANN technique to predict rainfall over a state (Kerala) of India, and to

372

S. CHATTOPADHYAY

the best of the knowledge and belief of the authors of the present study, Guhathakurta (2006) is the first ever attempt to implement ANN to forecast the summer-monsoon over India. However, Guhathakurta (2006) confined his study within a state of India. The present contribution deviates from the study of Guhathakurta (2006) in the sense that instead of choosing a particular state, the authors implement backpropagation ANN to forecast the average summer-monsoon rainfall over the whole country and aroma of newness further lies in the fact that here various multilayer ANN models are attempted to find out the best fit.

  • 2. ANN IN RAINFALL PREDICTION – A LITERATURE SURVEY

Hu (1964) initiated the implementation of ANN, an important soft computing method- ology in weather forecasting. Since the last few decades, a voluminous development in the application field of ANN has opened up new avenues to the forecasting task in- volving atmosphere related phenomena (Gardner and Dorling 1998, Hsieh and Tang 1998). Michaelides et al. (1995) compared the performance of ANN with multiple lin- ear regressions in estimating missing rainfall data over Cyprus. Kalogirou et al. (1997) implemented ANN to reconstruct the rainfall time series over Cyprus. Lee et al. (1998) applied ANN in rainfall prediction by splitting the available data into homogeneous subpopulations. Wong et al. (1999) constructed fuzzy rule bases with the aid of Self Organizing Map (SOM) and backpropagation neural networks and then with the help of the rule base developed predictive model for rainfall over Switzerland using spatial interpolation.

  • 3. MATERIALS AND METHOD OF DEVELOPING THE PREDICTIVE MODEL

ANNs have recently become an important alternative tool to conventional methods in modelling complex non-linear relationships. In the recent past, the ANN has been ap- plied to model large data with large dimensionality (i.e. Gevrey et al. 2003, Nagendra and Khare 2006). Most of the ANN studies related to the problem allied with pattern recognition, forecasting and comparison of the neural network with other traditional approaches in ecological and atmospheric sciences. However, the step-by-step proce- dure involved in development of ANN-based models is less discussed (Nagendra and Khare 2006). This paper develops ANN model step-by-step to predict the average rainfall over India during summer-monsoon by exploring the India rainfall data avail- able at the website http://www.tropmet.res.in/ published by Indian Institute of Tropi- cal Meteorology. In this paper, tropical rainfall indices have been used as one of the predictors, and the data have been collected from http://jisao.washington.edu/ data_sets/global_sstanomts/. From the same website the data of Tropical-average (30N-30S) sea surface temperature anomalies have been collected to develop the pre- dictive model. The time series made available from the said site captures both year-to- year ENSO variability and more decadal time scale changes (Smith et al. 1996).

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

373

The ANN approach has several advantages over conventional phenomenological or semi-empirical models, since they require known input data set without any as- sumptions (Gardner and Dorling 1998, Nagendra and Khare 2006). It exhibits rapid information processing and is able to develop a mapping of the input and output vari- ables. Such a mapping can subsequently be used to predict desired outputs as a func- tion of suitable inputs (Nagendra and Khare 2006). A multilayer neural network can approximate any smooth, measurable function between input and output vectors by se- lecting a suitable set of connecting weights and transfer functions or activation func- tion (Gardner and Dorling 1998, Kartalopoulos 1996, Nagendra and Khare 2006).

ANN based prediction of summer-monsoon rainfall in India

The model building process consists of the four sequential steps:

(i) Selection of the input and output for the supervised backpropagation learning. (ii) Selection of the activation function. (iii) Training and testing of the model. (iv) Testing the goodness of fit of the model.

Selection of the input and output for the supervised backpropagation learning

In India, the months June, July, and August are identified as the summer-monsoon months. In the present paper, data pertaining to the years 1950 through 1995 are ex- plored to develop the predictive model. The inputs to the proposed model are summa- rized below:

– Average monthly Indian rainfall (in mm) of the months of June, July, and August.

– Tropical-average (30N-30S) sea surface temperature anomalies (in °C) in the months of June, July, and August. – Tropical rainfall indices in the months of June, July, and August. – The predictand is the average Indian summer monsoon rainfall.

The physical relationships between the predictand and the predictors have al- ready been discussed in Chapter 1. It is interesting to see that the time series of Indian summer monsoon rainfall are not pair wise correlated. The mutual Pearson correlation values are: –0.06 (June–July), –0.01 (June–August), and –0.01(July–August). Thus, all the correlation values are too small, indicating that the relationships are highly non- linear. The Pearson correlation between the average rainfall and the predictors are nei- ther too low nor significantly high. The correlation values between the average sum- mer monsoon rainfall and the sea surface temperature anomalies are: –0.145, –0.256, and –0.291 in the months of June, July, and August, respectively. Again, the correla- tion between the tropical rainfall indices and the average summer monsoon rainfall are:

0.294, 0.206, and 0.334, respectively. Thus, the necessity of implementing ANN in the prediction problem is felt highly relevant. Furthermore, the autocorrelations in the rainfall time series for each month are found to be significantly small (Fig. 1). This in-

374

S. CHATTOPADHYAY

0.8 June 0.6 July 0.4 August 0.2 0 -0.2 lag1 lag2 lag3 lag4 lag5
0.8
June
0.6
July
0.4
August
0.2
0
-0.2
lag1
lag2
lag3
lag4
lag5

Fig. 1. Schematic showing the autocorrelation function corresponding to the rainfall amounts in the summer monsoon months.

dicates that the data exhibit no serial correlation or persistence. The aim of this paper is to develop a multilayer feed forward ANN model so that the average summer- monsoon rainfall of a given year can be predicted using the rainfall data of the sum- mer-monsoon months of the immediately previous year. Thus, the input matrix would consist of four columns of which the first three columns would correspond to the summer-monsoon months’ rainfall of year n and the fourth column would correspond to the average summer-monsoon rainfall of the year (n + 1). Basically, the fourth col- umn would correspond to the ‘desired output’ in the supervised backpropagation learning (Kartalopoulos 1996) procedure. The first 75% data are taken as the training set and the remaining 25% data are taken as the test set or validation set.

To avoid the asymptotic effect the raw data are scaled according to

z

i

=

0.1

+

0.8

x

i

x

min

x

max

x

min

,

(1)

where, z i denotes the transformed appearance of the raw data x i . After the modelling is completed, the scaled data are reverse scaled according to

Px

=+

i

min

1

0.8

⎡ ( ⎣
(

y

i

0.1

)(

x

max

x

min

)

,

(2)

where P i denotes the prediction in original scale.

Activation function

The advent of backpropagation algorithm (BP), the adaptation of steepest descent method, opened up new avenues of application of multilayered ANN for many prob- lems of practical interest (see e.g., Perez and Reyes 2001, Kamarthi and Pittner 1999,

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

375

Sejnowski and Rosenberg 1987). A multilayer ANN contains three basic types of layer: input layer, hidden layer(s), and output layer. Basically the backpropagation learning involves propagation of error backwards from the output layers to the hidden layers in order to determine the update for the weights leading to the units in the hid- den layer(s). The methodology is detailed in next section.

The non-linear relationship between input and output parameters in any network requires a function, which can appropriately connect and/or relate the corresponding parameters (Nagendra and Khare 2006). In the present paper, backpropagation learn- ing of ANN would be adopted with steepest descent (Kartalopoulos 1996). Thus, the sigmoidal function is taken as the ideal activation mathematically defined as (Karta- lopoulos 1996)

f

(

z

) =

1

1

+

exp(

z

)

.

Training and testing of the model

(3)

The proposed ANN model is basically a three layered ANN with backpropagation learning. Before going into the implementation details, the backpropagation algorithm for multilayered ANN is briefly discussed.

Backpropagation learning is exactly the same as delta learning at the output layer and is similar to the delta learning with the propagated error in the hidden layer(s), and thus it is called generalized delta rule (Yegnanarayana 2000). In this algorithm, at first step, the input and desired output are identified. Then an arbitrary weight vector w 0 is initialized. Then the feed forward neural network is iteratively adopted according to the recursion (Kamarthi and Pittner 1999)

w

k

+

1

=

wd

+

η

kk

,

(4)

where w l denotes the weight matrix at epoch l; the positive constant η, which is se- lected by the user, is called the learning rate.

The direction vector d k is negative of the gradient of the output error function E

d

k

= −∇Ew .

k

(

)

(5)

There are two standard learning schemes for the BP algorithm: on-line learning and batch learning. In on-line learning, the weights of the network are updated imme- diately after the presentation of each pair of input and target patterns. In batch learning all the pairs of patterns in the training sets are treated as a batch, and the network is updated after processing of all training patterns in the batch. In either case, the vector w k contains the weights computed during kth iteration, and the output error function E is a multivariate function of the weights in the network (Kamarthi and Pittner 1999)

E

(

w k
w
k

)

E

p

= ⎨

p

(

w

k

)

[

on-line

]

E

p

(

w

k

)

[

batch

]

(6)

376

S. CHATTOPADHYAY

where E p (w k ) denotes the half-sum-of-squares error functions of the network output

for a certain input pattern p. The purpose of the supervised learning (or training) is to

find out a set of weight that can minimize the error E over the complete set of training

pair. Every cycle in which each one of the training patterns is presented once to the

neural network is called an epoch.

The direction vector d k , expressed in terms of error gradient, depends upon the

choice of activation function. When the sigmoid function (as defined by eq. (3)) is

adopted, the BP algorithm becomes ‘back propagation for the sigmoid adaline’

(Widrow and Lehr 1990). In this method the input matrix is multiplied by the weight

matrix and the product is used as the variable for the sigmoid activation function. For

example, at epoch k, the sigmoid non-linearity is produced as

where

W

k

=

[

ww 2 ...

1

w n
w
n

]

f

(

WX

k

k

)

1

=

1

+

exp

i

w x

i

i

,

(7)

and

X

k

=

[

xx

1

2

...

x

n

]

T

are the weight matrix and the

transpose of the input matrix, respectively, at epoch k.

After training or learning the ANN with BP algorithm with sigmoid non-

linearity, an ultimate weight matrix is obtained. This weight matrix is applied to an-

other set of independent inputs to examine the efficiency of the model. This phase is

called the testing or validation phase.

Testing the goodness of fit of the model

After developing the model through training and testing, goodness of fit of the model

is examined statistically. The overall prediction error (PE) is measured as (Perez and

Reyes 2001)

PE =

y − y predicted actual y actual
y
y
predicted
actual
y
actual

,

where

376 S. CHATTOPADHYAY where E ( w ) denotes the half-sum-of-squares e rror functions of the

implies the average over the whole test set.

(8)

The predictive model is identified as a good one if the PE is sufficiently small,

i.e., close to 0. The model with minimum PE is identified as the best prediction model.

4.

IMPLEMENTATION DETAILS AND COMPARISON WITH OTHER MODELS

Details of the input and output variables are presented in Section 3. The learning rate η

(see eq. (4)) is taken to be 0.9. A three-layered feed forward neural net is now de-

signed. The problem is to find out the number of hidden nodes producing the best

model. Since the number of adjustable parameters in a one hidden-layer feed forward

neural network with n i input units, n 0 output units, and n h hidden units is

n + nn ++

0

h

(

i

n

0

1

)

(Perez et al. 2000), for n i = 3, n 0 = 1 and with 23 training

cases, it is not possible to use an

n

h

greater than 2.

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

377

Now, three-layered feed forward ANN models with n h = 2 and η = 0.9 are gen-

erated. The backpropagation learning using the methodology explained in the previous

chapter is applied. In the ANN models the initial weights are chosen randomly from

–0.5 to +0.5 (Pal and Mitra 1999). After each training iterations/epochs, the network is

tested for its performance on validation data set. The training process is stopped when

the performance reaches the maximum on validation data set (Haykin 2001, Sarle

1997, Gardner and Dorling 1998, Nagendra and Khare 2006). The network compo-

nents are presented in Table 1.

Table 1

Basic network components of the ANN model

 

Network Architecture

 

Number of inputs

9

Number of outputs

1

Number of hidden

1

   

layers

   

Hidden layer sizes

2

Learning parameter

0.2

Initial weight range (0±w)

0.5

Momentum

0.9

   

Training options

     

Total rows

45

No of training cycles

500

in your data

   

Training mode

On-line

Save network

With least

   

weights

training error

Training/validation

set

Partition data

   

into training/

validation set

After 500 epochs, the final weight matrix is found to be:

Hdn1_bias

0.0000

0.0000

0.0000

 
  • 0.0000 0.0000

 
  • 0.0000 0.0000

0.0000

0.0000

0.0000

Hdn1_Nrn1

0.0200

4.6640

6.0560

–1.9587

–2.7452

  • 3.9752 –3.6464

–3.5230

–2.7460

–4.0021

Hdn1_Nrn2

–1.2040

–0.7148

0.3910

  • 1.6052 0.5995

  • 0.2009 0.2943

–4.6899

–4.3097

–1.6299

After proper training and validation with the ANN with backpropagation learn-

ing, the model outputs are compared with other statistical model outputs to measure its

potential over the other models. Using eq. (8), the overall prediction error over the

whole test set is found to be approximately 0.15. The prediction made by the ANN is

made visible in Fig. 2, where the actual yearly average summer monsoon rainfall is

plotted against those predicted through the neural net.

In the next phase, the prediction made by the neural net is compared with the fol-

lowing statistical methods:

– Persistence forecast,

– Multiple Linear Regression (MLR) forecast.

378

S. CHATTOPADHYAY

1983 1984 1995 1988 1978 1987 1977 1974 1991 1992 1979 1980 1981 1982 1985 1986
1983
1984
1995
1988
1978
1987
1977
1974
1991
1992
1979
1980
1981
1982
1985
1986
1975
1976
1989
1990
0
1993
1994
Average summer monsoon rainfall [mm]
100
150
200
250
300
ANN
Average predicted by
Actual average
50

Test cases [years]

Fig. 2. Actual versus predicted average monsoon rainfall in the test cases. The prediction is

made by Artificial Neural Network (ANN).

Persistence forecast is made by shifting the predicted time series back one year.

In this method, the value of the time series at time step t is assumed to be in the same

state, which means that the predicted value at time t + 1 is the actual value at time t.

In the present modeling, 1974 to 1995 are taken as test years. Thus, the overall predic-

tion error (PE) should always be calculated using these years’ actual and predicted

values. The PE, as calculated using eq. (8) is found to be approximately equal to 0.18,

which is greater than that produced by neural net model.

A Multiple Linear Regression equation is now fit to the set of 9 predictors and 1

predictand. Just like neural net model, first 75% data are used for training, and the last

25% data are used for testing. The predictive MLR equation is then fitted using the

method of least squares and the equation comes out to be

yx ˆ =+++++++++ 0.28 0.14 x 0.15 x 0.11x 0.03 x 0.05 x 0.28 x 0.13
yx
ˆ =+++++++++
0.28
0.14 x
0.15 x
0.11x
0.03 x
0.05 x
0.28 x
0.13 x
0.18 x
171.12 .
(9)
123456789
0.4
0.3
0.2
0.1
0
PE (Artificial Neural
Network)
PE (Persistence
Forecast)
PE (Multiple Linear
Regression)
Value of PE
0.150775108
0.175310325
0.383413897

Fig. 3. This figure compares the prediction errors produced by three competitive models in

predicting average summer monsoon rainfall over India. The computation is made over the test

cases.

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

379

The first three variables in the right hand side imply the total rainfall amounts in the

months of June, July, and August of the year n, the second three variables imply

Tropical Rainfall Indices in the months of June, July, and August of the year n, the last

three variables imply Sea Surface Temperature anomalies in the months of June, July,

and August of the year n, and the left hand side implies the average summer monsoon

rainfall predicted for the year (n + 1). The PE produced by the MLR model for the test

years is found to be approximately equal to 0.29 which is significantly greater than

those produced by other two models. All the prediction errors, PE, are pictorially pre-

sented in Fig. 3.

5.

CONCLUSIONS

In the present paper, Artificial Neural Network with backpropagation learning has

been implemented to predict average summer monsoon rainfall over India. Nine pre-

dictors have been explored to generate the input matrix for the neural net. After 500

epochs, the ANN has been found to produce a forecast with small prediction error.

Finally, the model performance has been compared with two customary statistical

models and ANN has been found to produce lower prediction error than persistence

forecast and Multiple Linear Regression forecast. The study, therefore, establishes

supremacy of ANN over customary processes in predicting average summer-monsoon

rainfall over India.

References

Brown, B.G., and A.H. Murphy, 1988, The economic value of weather forecasts in wildfire

suppression mobilization decisions, Canadian J. Forest Res. 18, 1641-1649.

Cartalis, C., and C. Varotsos, 1994, Surface ozone in Athens, Greece, at the beginning and at

the end of the 20th-century, Atmos. Environ. 28, 3-8.

Chin, E.H., 1977, Modeling daily precipitation occurrence process with Markov chain, Water

Resour. Res. 13, 949-956.

Clark, O.C., J.E. Cole and P.J. Webster, 2000, Indian Ocean SST and Indian summer monsoon

rainfall: Predictive relationships and their decadal variability, J. Climate 14, 2503-

2519.

Elsner, J.B., and A.A. Tsonis, 1992, Non-linear prediction, chaos, and noise, Bull. Am. Me-

teor. Soc. 73, 49-60.

Ferranti, L., T.N. Palmer, F. Molteni and E. Klinker, 1990, Tropical-extratropical interaction

associated with the 30-60 day oscillation and its impact on medium and extended

range prediction , J. Atmos. Sci. 47 , 2177-2199.

Gadgil, S., M. Rajeevan and R. Nanjundiah, 2005, Monsoon prediction – Why yet another fail-

ure?, Current Science 88,1389-1499.

Gardner, M.W., and S.R. Dorling, 1998, Artificial Neural Network (Multilayer Perceptron) –

A review of applications in atmospheric sciences, Atmos. Environ. 32, 2627-2636.

380

S. CHATTOPADHYAY

Gevrey, M., I. Dimopoulos and S. Lek, 2003, Review and comparison of methods to study the

contribution of variables in artificial neural network models, Ecological Modeling

160, 249–264.

Gregory, J.M., T.M.L. Wigley and P.D. Jones, 1993, Application of Markov models to area av-

erage daily precipitation series and inter annual variability in seasonal totals, Cli-

mate Dynamics 8, 299-310.

Guhathakurta, P., 2006, Long-range monsoon rainfall prediction of 2005 for the districts and

sub-division Kerala with artificial neural network, Current Science 90, 773-779.

Haykin, S., 2001, Neural Networks: A Comprehensive Foundation, 2 ed., Pearson Education

Inc., New Delhi, India.

Hsieh, W.W., and T. Tang, 1998, Applying neural network models to prediction and data

analysis in meteorology and oceanography, Bull. Am. Meteor. Soc. 79, 1855-1869.

Hu, M.J.C., 1964, Application of ADALINE system to weather forecasting, Technical Report,

Stanford Electron, Stanford, CA.

Jacovides, C.P., C. Varotsos, N.A. Kaltsounides, M. Petrakis and D.P. Lalas, 1994, Atmos-

pheric turbidity parameters in the highly polluted site of Athens basin, Renewable En-

ergy 4, 5, 465-470.

Kalogirou, S.A., C.N. Constantinos, S.C. Michaelides and C.N. Schizas, 1997, A time series

construction of precipitation records using Artificial Neural Networks, EUFIT ’97,

September 8-11, 2409-2413.

Kamarthi, S.V., and S. Pittner, 1999, Accelerating neural network training using weight ex-

trapolation, Neural Networks 12, 1285-1299.

Kartalopoulos, S.V., 1996, Understanding Neural Networks and Fuzzy Logic – Basic Concepts

and Applications, Prentice Hall, New-Delhi.

Kishtawal, C.M., S. Basu, F. Patadia and P.K. Thapliyal, 2003, Forecasting summer rainfall

over India using Genetic Algorithm, Geophys. Res. Lett. 30, doi: 10.1029/2003GL

018504.

Kondratyev, K.Y., and C.A. Varotsos, 2001a, Global tropospheric ozone dynamics – Part I:

Tropospheric ozone precursors, Environ. Sci. Pollution Res. 8, 1, 57-62.

Kondratyev, K.Y., and C.A. Varotsos, 2001b, Global tropospheric ozone dynamics – Part II:

Numerical modelling of tropospheric ozone variability, Environ. Sci. Pollution Res. 8,

2, 113-119.

Lee, S., S. Cho and P.M. Wong, 1998, Rainfall prediction using Artificial Neural Network,

J. Geograph. Inform. Decision Anal. 2, 233-242.

Matthews, A.J., 2004, Atmospheric response to observed intraseasonal tropical sea surface

temperature anomalies, Geophys. Res. Lett. 31, doi:10.1029/2004GL020474.

Matthews, A.J., B.J. Hoskins and M. Masutani, 2004, The global response to tropical heating

in the madden-julian oscillation during northern winter, Quart. J. Roy. Met. Soc.

130 , 1-20.

Men, B., Z. Xiejing and C. Liang, 2004, Chaotic analysis on monthly precipitation on Hills

Region in Middle Sichuan of China, Nature and Science 2, 45-51.

Michaelides, S.C., C.C. Neocleous and C.N. Schizas, 1995, Artificial Neural Networks and

multiple linear regression in estimating missing rainfall data, Proc. DSP95 Intern.

Confer. “Digital Signal Processing”, Limassol, Cyprus. 668-673.

ANN MODEL TO PREDICT SUMMER-MONSOON RAINFALL IN INDIA

381

Nagendra, S.M.S., and M. Khare, 2006, Artificial neural network approach for modelling ni-

trogen dioxide dispersion from vehicular exhaust emissions, Ecological Modeling

190, 99-115.

Pal, S.K., and S. Mitra, 1999, Neuro-Fuzzy Pattern Recognition: Methods in Soft Computing,

Willey Intersc. Publ., New York.

Perez, P., and J. Reyes, 2001, Prediction of particulate air pollution using neural techniques,

Neural Computing and Application 10, 165-171.

Perez, P., A. Trier and J. Reyes, 2000, Prediction of PM2.5 concentrations several hours in ad-

vance using neural networks in Santiago, Chile, Atmosph. Environ. 34, 1189-1196.

Reddy, P.R.C., and P.S. Salvekar, 2003, Equatorial East Indian Ocean sea surface tempera-

ture: A new predictor for seasonal and annual rainfall, Current Science 85, 1600-

1604.

Sardeshmukh, P.D., and B.J. Hoskins, 1988, The generation of global rotational flow by steady

idealized tropical divergence , J. Atmos. Sci. 45 , 1228-1251.

Sarle, W., 1997, Neural network frequently asked questions, ftp://ftp.sas.com/pub/neural/FAQ.

html.

Sejnowski, T.J., and C.R. Rosenberg, 1987, Parallel networks that learn to pronounce English

text, Complex Systems 1, 145-168.

Shinoda, T., and H.H. Hendon, 1998, Mixed layer modeling of intraseasonal variability in the

tropical Pacific and Indian Oceans, J. Clim. 11, 2668-2685.

Sivakumar, B., 2000, Chaos theory in hydrology: important issues and interpretations, J. Hy-

drology 227, 1-20.

Sivakumar, B., 2001, Rainfall dynamics in different temporal scales: A chaotic perspective,

Hydrology and Earth System Sci. 5, 645-651.

Sivakumar, B., S.Y. Liong, C.Y. Liaw and K.K. Phoon, 1999, Singapore rainfall behavior:

Chaotic?, J. Hydrol. Eng., ASCE 4, 38-48.

Smith, T.M., R.W. Reynolds, R.E. Livezey and D.C. Stokes, 1996, Reconstruction of historical

sea surface temperatures using empirical orthogonal functions, J. Climate 9, 1403-

1420.

Varotsos, C., 2005, Power-law correlations in column ozone over Antarctica, Intern. J. Remote

Sensing 26, 3333-3342.

Varotsos, C., and D. Krik-Davidoff, 2006, Long-memory processes in ozone and temperature

variations at the region 60 degrees S-60 degrees N, Atmos. Chem. Phys. 6, 4093-

4100.

Varotsos, C., K.Y. Kondratyev and M. Efstathiou, 2001, On the seasonal variation of the sur-

face ozone in Athens, Greece, Atmos. Environ. 35, 2, 315-320.

Widrow, B., and M.A. Lehr, 1990, 30 years of Adoptive Neural Netwoks; Perceptron,

Madaline, and Back propagation, Proc. IEEE 78, 1415-1442.

Wilks, D.S., 1991, Representing serial correlation of meteorological events and forecasts in

dynamic decision-analytic models, Month. Weather Rev. 119, 1640-1662.

Wilks, D.S., 1998, Multisite generalization of a daily stochastic precipitation generation

model, J. Hydrology 210, 178-191.

382

S. CHATTOPADHYAY

Wong, K.W., P.M. Wong, T.D. Gedeon and C.C. Fung, 1999, Rainfall prediction using neural

fuzzy technique, URL: www.it.murdoch.edu.au/~wong/publications/SIC97.pdf, 213-

221.

Woolnough, S.J., J.M. Slingo and B.J. Hoskins, 2000, The relationship between convection

and sea surface temperature on intraseasonal timescales, J. Climate 13, 2086-2104.

Yegnanarayana, B., 2000, Artificial Neural Networks, Prentice-Hall of India Pvt Ltd., New

Delhi, India.

Zhang, M., and A.R. Scofield, 1994, Artificial Neural Network techniques for estimating rain-

fall and recognizing cloud merger from satellite data , Intern. J. Remote Sensing 16,

3241-3262.

Received 7 December 2006

Accepted in revised form 5 May 2007