Forecasting of the seasonal rainfall is very important to semi-arid area
like Khorasan Province in Northeast of Iran. Artificial neural network
is an innovative approach to construct computationally intelligent systems
that are supposed to possess humanlike expertise within a specific domain,
adapt themselves and learn to do better in changing environments and explain
how they make decisions. Considering the significance of rainfall in many
decision making processes such as water resource management and agriculture,
the present study aims to find out the relationship between large-scale
climatic signals and regional rainfall using artificial neural network.
McCullagh et al. (1995) investigated the use
of an Artificial Neural Network (ANN) to estimate the 6 h rainfall over the
South-east coast of Tasmania. The results confirm that ANNs have the potential
for successful application to the problem of rainfall estimation. Karamouz
et al. (2004) have used from ANN, fuzzy logic and time series for
seasonal rainfall forecasting in the western regions of Iran. In this research,
The ANN model displayed a better performance compared to the other models. Nazemosadat
and Cordery (2000) used Sea Surface Temperature (SST) for seasonal rainfall
variability in the winter (Jan. to Mar.) in the southern regions of Iran. Results
showed that rainfall is inversely proportion with Sea Surface Temperature (SST)
of the Persian Gulf. Khoshakhlagh (1998) used from correlation
between ENSO and rainfall in Iran. Results showed that the Iranian rainfall
predictors are strongly related with ENSO.
Suwardi et al. (2006) have used of a neuro-fuzzy
system for modeling wet season tropical rainfall. The models resulted low values
of the RMSE indicated that the prediction models are reliable in representing
the recent inter-annual variation of the wet season tropical rainfall.
Lee et al. (1998) used RBF (Radial Basis Function)
networks and linear regression based on the locational information only for
the daily rainfall prediction at 367 locations in Switzerland. Comparison with
the observed data revealed that RBF networks produced good predictions while
the linear models poor predictions. Maria et al. (2005)
used ANN and linear regression for rainfall forecasting in Sao Paulo State,
Brazil. The results show that ANN forecasts were superior to the ones obtained
by the linear regression model thus revealing a great potential for an operational
suite. Cavazos (2000) used self organization map and
ANN for daily rainfall forecasting in Bucharest. Liu and
Lee (1999) also used from ANN for short-term rainfall forecasting in Hong
Kong region. This neural-based rainfall forecasting system is useful and parallel
to traditional forecasts from the Hong Kong observatory. Wong
et al. (2003) constructed fuzzy rule bases with the aid of SOM (Self-Organization
Map) and back propagation neural networks and then with the help of the rule
base developed predictive model for rainfall over Switzerland using spatial
interpolation. Matayo et al. (2000) used Empirical
Orthogonal Function (EOF) and simple correlation analysis for seasonal rainfall
anomalies over East Africa. Pozo-Vazquez et al. (2001)
the association among ENSO (El Nino-Southern Oscillation), the Northern Hemisphere
Sea Level Pressure (SLP) and temperatures in Europe has been analyzed during
the period 1873-1995. Surajit and Manojit (2007) used
artificial neural network as a Soft Computing technique to anticipate the average
monsoon rainfall over India and results have been compared with those obtained
through conventional techniques.
MATERIALS AND METHODS
Study area: The area of this study is Khorasan Province in North
east of Iran in Fig. 1. Total precipitation from December
to May over a period of 33 yeas (1970-2002) was selected as data of our
interest in this research.
Data of 37 stations including four Synoptic, five climatology and 28
rain gauges (all belong to Iranian Meteorological organization) were selected
for each year.
Calculating of local average rainfall: Digitizes Elevation model
is used to get the amount of local average rainfall. The following steps
were taken to obtain the time series of average regional rainfall:
||Making input files for the Arc GIS software
||Obtaining the relation between rainfall and elevation using regression
||Obtaining digitized map of area under study
||Analyzing and drawing annual spatial changes of rainfall in the
||Obtaining the values of annual average rainfall in the region under
||Making time series of rainfall in the region under study
|| Map of area of study with selected stations
|| Name and properties of selected stations in the area
Data collection: The data used in this study are:
||Thirty seven rainfall station data for the seasonal
rainfall (Dec.-May) were obtained from Iranian Meteorological organization.
All of these stations are in the Eastern north region of Iran. Properties
of these stations have shown in Table 1
||Large-scale ocean and atmospheric circulation variables such as
Sea Surface Temperature (SST), Sea Level Pressure (SLP), the difference
sea level pressure, the difference sea surface temperature between
surface and 1000 hPa level, relative humid at 300 hPa level, geopotential
height at 500 hPa level, air temperature at 850 hPa level during months
(June-Nov.). These data were obtained from NCEP/NCAR Re-analysis data.
These data sets span the period of 1948-current, covering the globe
on a 2.5x2.5 grid and available at http://www.cdc.noaa.gov
National Oceanic and Atmospheric Administration (NOAA) website
||Standard ENSO indices: NINO3, NINO1+2, Southern Oscillation
Index (SOI) Available at http://www.cdc.noaa.gov
||Indian Ocean Dipole (IOD) index (Saji et al.,
1999). This is an index based on SST anomaly difference between the
Eastern and Western tropical Indian Ocean. The index, its impact on the
adjoining continental rainfall, interactions with El Nino Southern Oscillation
(ENSO) and teleconnections can all be obtained from the IOD home page http://www.jamstec.go.jp/frsgc/research/d1/iod/
Identification of predictors: The aim of identification of predictors
is to identify predictors for Khorasan Province seasonal rainfall, which
can then be used in forecast models. The two main requirements for any
useful predictors are good relationship with the seasonal rainfall and
reasonable lead-time (i.e., months to season). The earlier research indicated
that seasonal rainfall in the region is strongly correlated with predictors.
So, the first step is to look for relationship with standardized predictors
during the season (June-Nov.) and follow up with correlations between
the rainfall and large-scale ocean-atmospheric variables (SST, SLP and
so on). This approach of correlation with large-scale ocean-atmospheric
circulation variables used to identify predictors for seasonal rainfall
forecasting in the Northern East of Iran.
Correlation with large-scale variables: The predictors large-scale
aspects and also the seasonal rainfall correlation with predictors such
as SST, SLP and etc. were checked during pre-season rainfall (June-Nov).
In this research, the correlations that are significant at 95% confidence
level have been selected. Figure 2 show properties of
selected points which have used for relation between rainfall and remote
Predictor selection: Based on the correlations with indices and
the correlation with large-scale variables, predictors with high correlations
to the seasonal rainfall were identified. With this criterion, the selected
predictors parameters are:
Standardized pressure of Aden gulf (x1), South of Persian
Gulf (x2), North of Red Sea (x3), South of Red Sea
(x4), The difference of pressure standardized between Adriatic
Sea and south of Persian Gulf (x5). Aral Lake and North of
Caspian Sea (x6), South of Persian Gulf and Arab Sea (x7),
Oman Sea and South of Persian Gulf (x8), South of Persian Gulf
and South of Red sea (x9), The standardized sea surface temperature
of Siberian network (x10), The difference of temperature standardized
between sea surface and the 1000 mb level of the Island network (x11).
|| Name and coordinates that have used for relation between
rainfall and remote linkage controlling
The factor analysis of the relative humidity in the index area of factor
1 in a 5x5 degree networks (x12). These regions have shown
in Fig. 3.
Time series of rainfall and selected predictors that are used for relation
between rainfall and Remote linkage controlling that mentioned above have
shown in Table 2.
THE STRUCTURE OF ARTIFICIAL NEURAL NETWORKS
Artificial neural networks were first introduced in 1943 by McCulloch et
al. (1995). Later, with the development of back propagation algorithm for
feed forward networks the application of neural network entered a new stage
Like natural neural networks, artificial neural networks are made up of parts
called neural cells. As in natural neural networks where some cells are responsible
for receiving the external stimulus, some for processing and some for the transfer
of response to the intended part, in artificial neural networks, too, some cells
receive the data of the problem, some process the data and some provide the
solution to the problem. Thus, every neural network is made up of the input
layer, the hidden layer and the output layer, with the three layers connected
by means of connectors of different weights. In all neural networks, there is
one input layer, one output layer and several hidden layers (Mahdizadeh,
2004). Figure 4 shows the structure of one kind of such
networks (Mohammadi and Misaghi, 2003).
Figure 5 shows the model of a multi-input neuron (Christodoulou
and Georgiopoulos, 2000). The three elements of a multi-input neuron are
||The set of synapses each specified with a certain weight.
As it is shown in the Fig. 5, the neuron k with
the output xk is connected to the intended neuron j through
a proper weight connector called wjk. The effect of the
neuron k on the neuron j is calculated through xk.wjk.
If the neuron k is active and wjk is positive (excitatory
synapse), the neuron k will have a positive effect on the neuron j.
On the other hand, if the neuron k is active, but wjk is
negative (inhibitory synapse), the neuron k will have a negative effect
on the neuron j. Special attention should be paid to the written form
and the subtitle of the weight of the synapse wjk. The
first subtitle belongs to the target neuron and the second to the
source neuron of the intended synapse.
|| The detected areas of relative humidity at 300 mb level
in networks of 5x5 degree
|| Time series of rainfall and selected predictors
||The overall structure of feed forward monolayer neural
|| A model of a multi-input neuron
||A capacitor for collecting the incoming signals which
are weighted by the synapses of the neuron. The accumulating effect
of all neurons which are connected to the intended neuron (Neuron
j) is calculated by adding up all the effects of individual neurons
on the neuron j.
||A function of activity is used to limit the output range of the
neuron. Activity function is considered as a constraining function
in which the eligible changes of the range of output signals is restricted
to some finite values. The net input and the output y are calculated
by Eq. 1 and 2.
In the Equation x1, x2,
, xk represent
the incoming signals, wj1, wj2,
stand for synaptic weights accumulating in a neuron, net is the accumulated
effect of all the neurons connected to the neuron j and the internal threshold
of the neuron j. g is activity function and yj is the output
signal of the neuron.
To assess the accuracy of the model, the index of Root Mean Square Error
(RMSE) has used which is calculated by the following formula:
In the above Equation, RMSE is Root Mean Square Error, oi
and ei are the observed and predicted value of the variable,
respectively in the point i and n number of observations.
Research methodology: The methodology described is used as a diagnostic
tool to derive anomalous atmospheric patterns characteristic of rainfall
events and to derive seasonal rainfall at the area and local scales. The
methodology consists of three main steps: (1) classification of the atmospheric
controls into different climate signals (i.e., weather types), (2) derivation
of large-scale climate anomalies associated with rainfall events and (3)
derivation of empirical transfer functions between the atmospheric controls
and seasonal rainfall at the area and local scales.
The statistical methods are used to discover patterns; in this case,
the correlation between rainfall and climate signals is supposed to find
significant features that characterize the seasonal atmospheric circulation
and the atmospheric control fields (e.g., humidity) during June to November
over the study area.
The purpose of this study is to improve our understanding of the physical and
remote linkages associated with rainfall events over the Iran region. This is
accomplished by exploring the climate anomalies characteristic of seasonal rainfall
events. As explained earlier, the factor analysis classes are used to composite
upper troposphere moisture, temperature and mid troposphere geopotential height
patterns. These mean fields and their anomalies are chosen to explore the physical
characteristics of the atmosphere during seasonal rainfall events in Khorasan
Province in Iran during the 1970-2002 period.
Experimentation setup for training and performance evaluation:
The data was obtained for the period from 1970 to 2002. The rainfall data
was standardized and divided from 1970-1992 as training set and the data
from 1993-2002 as test set. The operation was started with a network having
12 input nodes. Further experimentation showed that it was not necessary
to include information corresponding to the whole year, but 6 month information
centered over the predicted month of the 33 years in the time series would
give good generalization properties because in the area under study, 80%
total of the rainfall fall from Dec. to May and predictors of the pre-season
(from June to Nov.) is suitable to predict amount of the rainfall from
Dec. to May.
Thus, based on the information from the earlier years, the network would
predict the amount of rain to be expected in each 6 month of the each
year. The training was terminated after 1000 epochs. Experiments were
carried out on a machine and the model was executed using Neuro Solution
software. In this research, tangent hyperbolic axon transfer function
was used in the hidden layer and linear tangent hyperbolic axon transfer
function in the output layer. Test data was presented to the network and
the output from the network was compared with the actual data in the time
series. Following are the details of network training.
ANN training: For neural networks using momentum algorithm, 1
input layer, 1 hidden layers and an output layer were used. Input layer
consists of 12 neurons corresponding to the input variables and the hidden
layer consists of 10 neurons and the output layer consists of 1 neuron
(rainfall) [12-10- 1].
RESULTS AND DISCUSSION
After conducting various tests to test the network and the number of
neurons of the hidden layer and different functions of activity in the
hidden and output layers, as mentioned above, the final model with one
input layer, one hidden layer and one output layer (rainfall from Dec.
to May), had the least error, so in this research, used it as the main
Table 3 shows the training results for ANN model. As
an evident from Table 3, Mean Square Error (MSE), Normalized
Mean Square Error (NMSE), Mean Absolute Error (MAE), Minimum Absolute
Error (Min. Abs Error), Maximum Absolute Error (Max. Abs. Error) was obtained
1755.43, 0.9642, 36.21, 6.51, 73.009 and 0.3, respectively. Table
4 shows the comparative performance of between the actual rainfall
and the predicted rainfall by using ANN. Figure 6 also
shows actual data versus predicted data. Table 5 shows
the characteristic of ANN structure. Root mean square error for the model
was obtained 41 mm.
The investigation of the model results showed that the difference between
actual rainfall (mm) and predicted rainfall (mm) is acceptable and the
model can predict the amount of the rainfall in the most years. The root
mean square error for the model was obtained 41.5 mm. As a result, the
entered variables in the model can successfully explain the rainfall distribution
and dispersal pattern of the seasonal rainfall in the study area.
|| Training results for ANN
|| Seasonal rainfall prediction (6 months) using ANN
|| Properties of ANN network
||Comparative of observed rainfall and predicted rainfall
using ANN model
object has the important role in the planning and agricultural water management.
The results comparative of this research and other researches such as Karamouz
et al. (2004) and Maria et al. (2005)
showed that ANN techniques are efficiencies in the rainfall prediction and they
can successfully predict amount of the rainfall. Then, the results of this research
support the other researches in the study area and with considering of these
predictions; we can planning future politicals for the maximum operation.
In this study, it has been attempted to evaluate amount of the rainfall (six
month ahead) based on Artificial Neural Network (ANN). As the RMSE values on
test data are comparatively less, the prediction model was reliable. There have
been few deviations of the predicted rainfall value from the actual. As climate
and rainfall predication involves tremendous amount of imprecision and uncertainty.
The proposed prediction model based on soft computing on the other hand is easy
to implement and produces desirable mapping function by training on the given
data set. A network requires information only on the input variables for generating
forecasts. In these experiments, only 23 years training data were used to evaluate
the learning capability. Network performance could have been further improved
by providing more training data. Moreover, the considered connectionist models
are very robust, capable of handling the noisy and approximate data that are
typical in weather data and therefore should be more reliable in worst situations.
Choosing suitable parameters for the soft computing models is more or less a
trial and error approach. Optimal results will depend on the selection of parameters.
Selection of optimal parameters may be formulated as an evolutionary search
(Fogel, 1999) to make ANN models fully adaptable and optimal
according to the requirement.