Você está na página 1de 12

Nationwide Prediction of Drought Conditions

in Iran Based on Remote Sensing Data


Mahdi Jalili, Member, IEEE, Joobin Gharibshah, Seyed Morsal Ghavami,
Mohammadreza Beheshtifar, and Reza Farshi
AbstractIran is a country in a dry part of the world and extensively suffers from drought. Drought is a natural, temporary, and
iterative phenomenon that is caused by shortage in rainfall, which affects peoples health and well-being adversely as well as impacting
the societys economy and politics with far-reaching consequences. Information on intensity, duration, and spatial coverage of drought
can help decision makers to reduce the vulnerability of the drought-affected areas, and therefore, lessen the risks associated with
drought episodes. One of the major challenges of modeling drought (and short-term forecasting) in Iran is unavailability of long-term
meteorological data for many parts of the country. Satellite-based remote sensing datathat are freely availablegive information on
vegetation conditions and land cover. In this paper, we constructed artificial neural network to model (and forecast) drought conditions
based on satellite imagery. To this end, standardized precipitation index (SPI) was used as a measure of drought severity. A number of
features including normalized difference vegetation index (NDVI), vegetation condition index (VCI), and temperature condition index
(TCI) were extracted from NOAA-AVHRR images. The model received these features as input and outputted the SPI value (or drought
condition). Applying the model to the data of stations for which the precipitation data were available, we showed that it could forecast
the drought condition with an accuracy of up to 90 percent. Furthermore, TCI was found to be the best marker of drought conditions
among satellite-based features. We also found multilayer perceptron better than radial basis function networks and support vector
machines forecasting drought conditions.
Index TermsModeling, prediction, neural networks, data mining, performance measures

1 INTRODUCTION
D
ROUGHT is a serious climate condition affecting many
places of world and is indeed the most complex
natural hazard. Severity of drought is often measured using
a combination of variables including rainfall, temperature,
humidity, wind, soil moisture and stream flow. Since
drought spreads over wide areas, develops slowly, has
complex dynamics, and lacks a universal definition, its
monitoring is difficult. Iran has suffered from severe
drought conditions in the last decades and is predicted to
be a victim of drought in future years. A very recent
drought happened in 2008 in many areas of Iran resulted in
lots of damage to the national economy. The frequent
occurrence of drought in Iran is mainly due to lack of
enough rainfall, and lack of proper water management
system. Abnormal high temperature (especially a long-
lasting one), while the rainfall is normal, can also cause
drought conditions. The drought risks can be minimized by,
first, properly modeling the drought, and then, trying to
manage the available resources.
Drought is often defined in three types: meteorological,
agricultural, and hydrological. Meteorological drought is
defined as a situation when the normal precipitation in an
area significantly deceases in a period of time. There are
various methods for quantifying severity of drought [1] of
which palmer drought severity index [2] and standardized
precipitation index (SPI) [3] are the most commonly used
ones. SPI globally accepted as a good metric quantifying
severity of drought. SPI is difference of precipitation from
the mean value over a specified period, where the mean and
standard deviation are calculated from climatological data.
Since the precipitation data do not have normal distribu-
tion, a proper transformation (e.g., Gamma function) is
applied to the data to obtain normally distributed data. SPI
is advantageous over other drought metrics for that it has
variable time scale and uses only rainfall records. It is often
used for real-time monitoring and analysis of drought.
Satellite-based remote sensing data can be effectively
used for monitoring and forecasting environmental issues
[4]. They are increasingly used for monitoring drought-
related vegetation conditions all over the world [5], [6],
[7], [8], [9]. Since early 1980s, the advanced very high
resolution radiometer (AVHRR) on the National Oceanic
and Atmospheric Administration (NOAA)s satellite has
been used for extracting moisture-related vegetation indices
such as normalized difference vegetation index (NDVI),
vegetation condition index (VCI) and temperature condi-
tion index (TCI). These indices are then used for monitoring
the vegetation conditions (and somehow drought). Since
AVHRRs imagery data have high temporal resolution (on a
daily basis) with large area coverage, it is a major tool for
90 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
. M. Jalili, J. Gharibshah, and R. Farshi are with the Department of
Computer Engineering, Sharif University of Technology, Tehran, Iran.
E-mail: mjalili@sharif.edu, {gharibshah, rezafarshi}@gmail.com.
. S.M. Ghavami is with the Department of Geodesi and Geomatics
Engineering, Faculty of Geodesy and Geomatics Engineering, Khajeh
Nasir University of Technology, No 1346, Valiasr Street, Mirdamad Cross,
Tehran, Iran. E-mail: mghavami@mail.kntu.ac.ir.
. M. Beheshtifar is with the Iran Remote Sensing Center, ISA, Saye Alley,
Vali-Asr Avenue, Tehran, Iran. E-mail: beheshtifar@gmail.com.
Manuscript received 13 Sept. 2012; revised 20 Feb. 2013; accepted 12 May
2013; published online 21 May 2013.
For information on obtaining reprints of this article, please send e-mail to:
tc@computer.org, and reference IEEECS Log Number TCSI-2012-09-0663.
Digital Object Identifier no. 10.1109/TC.2013.118.
0018-9340/14/$31.00 2014 IEEE Published by the IEEE Computer Society
mapping the time series of vegetation conditions at global
or regional levels. Among these metrics, NDVI has been
extensively applied to vegetation and drought monitoring,
which is mainly due to strong relationship between
vegetation vigor and soil moisture [6], [10], [11], [12], [13],
[14], [15], [16].
In this paper, we developed a neural network model for
predicting severity of drought based on satellite images.
The models takes a number of features based on satellite
images (such as NDVI, TCI and VCI) as input and outputs
SPI as an indicator of drought severity. To this end, we
extracted the SPI time series for the period 1978-2008 and
those of NDVI, TCI, and VCI from the launch of NOAA-
AVHRR to 2008. We used a number of neural network
architectures (including multilayer perceptron (MLP) and
radial-basis function (RBF) networks) to perform the
modeling task. The results showed good performance of
the model in modeling and predicting drought conditions.
Here, we aimed at identifying the best performing satellite-
based vegetation index to model drought conditions
expressed by SPI. We considered three neural networks
models (multilayer perceptron, radial basis function net-
works, and support vector machines (SVMs)) and com-
pared their performance.
2 RELATED WORKS
A number of scholars studied the relation between
drought severity expressed by SPI and AVHRR-based
vegetation features such as NDVI and VCI. Song et al. [17]
developed a tool for early detection of drought in East
Asia based on NOAA-AVHRR NDVI data. Jin and Peters
[18] correlated the NDVI time series with those of SPI
obtained for US Great Plains and found the best correlation
value for 3-month SPI in the middle of growing season.
They suggested that by taking into consideration the
seasonal timing, NDVI is an effective metric for vegeta-
tion-moisture condition [18]. Jain et al. [12] considered
NDVI and VCI obtained from NOAA-AVHRR for some
western provinces of India and found significant correla-
tions between the time series of these indices and those of
SPI. Quiring and Ganesh [19] studied the relationship
between VCI and SPI using data obtained from Texas
counties during 18 growing seasons and found that the
growing-season VCI is correlated to prolonged moisture,
while it was less sensitive to short-term precipitation
deficiencies. Strong agreement between VCI and SPI was
found for Tigra, Northern Ethiopia, which allowed produ-
cing multitemporal drought maps [20]. Moreover, Gebre-
hiwot et al. [20] showed that climate region is an important
factor determining the nature of relationship between VCI
and drought indices. These studies clearly show that
integrated analysis of satellite-based remote sensing data
and ground measured data have great potential in drought
monitoring, modeling and forecasting.
In this work, we used neural network architectures to
apply to the classification task (i.e., prediction of drought
conditions). Artificial neural networks have been previously
applied to drought prediction [21], [22]. Linear stochastic
models and recursive multistep neural networks were
applied to predict SPI values for Kansabati River Basin
and resulted in good performance [21]. Dastorani et al. [23]
used neural networks for predicting SPI values (where both
the input and the output of the model were SPI) in a station
located in central Iran. Employing an aggregate drought
index, Barua et al. [24] developed neural networks for
predicting drought severity. However, these works were
only based on SPI and did not use any satellite-based data,
and thus, their results cannot be directly compared to ours.
All previous works on satellite-based drought prediction
have been applied on small regions with small number of
stations. In this work, we considered all stations in Iran.
These stations are distributed in different climate zone.
We trained a separate model for each climate zone.
Furthermore, we used sophisticated neural network models
and created a decision-making tool for policy makers.
3 DATA PREPARATION
3.1 Region of Study
Iran is located in Southwest Asia and shares border
with Turkmenistan, Armenia, Azerbaijan, Afghanistan,
Iraq, Pakistan, and Turkey. Iran covers an area of about
1.64 million km
2
and extends between latitudes 25

and
40

N and longitudes 44

and 63

E. It is surrounded by two
mountain ranges, namely Alborz to the north and Zagros to
the west, both with important role in the spatial and
temporal distribution of precipitation over the country.
These mountains avoid Mediterranean moisture bearing
systems by crossing through the region to the east [25].
Irans climate ranges from arid or semiarid to subtropical
along the Caspian coast and the northern forests. On the
northern edge of the country (the Caspian coastal plain),
temperature rarely falls below freezing and the area is
humid in most of the time. Annual precipitation is 680 mm
in the eastern part of the plain and more than 1700 mm in
the western part. To the west, settlements in the Zagros
basin experience lower temperatures, severe winters with
below zero average daily temperatures and heavy snowfall.
The eastern and central basins are arid with less than
200 mm of rain and have occasional desert. The coastal
plains of the Persian Gulf and Gulf of Oman in southern
Iran have mild winters and very humid and hot summers.
The annual precipitation ranges from 135 to 355 mm [26].
3.2 Meteorological Data
The metrological data used in this study are the monthly
rainfall and temperature data gathered from Iran Metrolo-
gical Organization. The precipitation data are for 1,767 rain
gauge stations between years 1978 and 2008 (Fig. 1). Before
using the data for extracting SPI indices, we statistically
checked data reliability and removed those stations with
considerable lack of suitable data. However, we could
reconstruct the values for some of the stations. Finally, we
identified 701 stations with complete (or reconstructed)
data to be used for extracting SPI time series.
The consistency of the rainfall data can be verified using a
number of methods such as double mass curve analysis, Von
Neumann ratio test, and likelihood ratio test [27]. In this
study, double mass curve analysis [28] was used to verify
the consistency among adjacent stations. As a result, those
stations showing inconsistency were removed from further
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 91
analysis. A number of stations had some missing data that
we had to reconstruct them before using in calculation of
SPI. There are some algorithms for reconstructing missing
data; we used linear regression method in this work [29]. We
also used runs test [30] to evaluate the correctness of the data
that indeed checks the randomness of data series. All
statistical tests were performed in SPSS. It must be taken into
consideration that these verification methods must be
applied to the data before any other processing steps such
as calculating climate zones of each station.
There are a number of indices for quantifying drought
severity [1]. Standard precipitation index is one of the
widely used metrics for drought modeling and forecasting
[31]. SPI not only captures information about the amount of
rainfall, but also provides a measure representing rainfall
conditions against a long-term mean. It describes how
precipitation is more (or less) than normal in a period of
time in an observation point. SPI is an invariant drought
index that can be calculated for any desired time. To this
end, a probability density function should be determined
that describes precipitation time series. It was shown that
Gamma distribution provides a good fit for the precipita-
tion data [32], and therefore, in this work, we fitted a
Gamma distribution to the precipitation time series. Then, a
cumulative probability of the precipitation was calculated
and a normal Gaussian function with mean zero and
variance one was applied to it. This makes the mean SPI for
the desired location and time to be zero. We finally obtained
an index for which the positive value illustrates greater than
median precipitation and the negative ones shows it is
lower than the median [31]. Table 1 shows correspondence
with the SPI values and drought severity conditions [33].
We also classified the stations into different climate
zones. It has been shown that De Marton index is the best
for Iran [34]. We used the data of precipitation and
temperature to calculate the De Marton indices as
1
1
T 10
. 1
where T is the annual mean temperature in Celsius and 1 is
the annual mean precipitation in millimeters. Table 2 shows
correspondence between the De Morton coefficient and the
class of the climate zone [34]. These classifications were
used in the modeling process. Table 3 shows the number of
stations in these six climate zones. As it is seen, many of the
stations have been classified as arid (29 percent) or semiarid
(54 percent), indicating that many parts of Iran are dry.
3.3 AVHRR-NOAA Data
Drought monitoring has been successfully carried out in
many areas using indices derived from optical remote
sensing data. The sensors installed in earth observation
satellites are able to quantify the capacity which the
presence of vegetation covers in the earth surface possesses
for initiating the process of photosynthesis. When drought
exists, due to considerable reduction of the rainfall, the
capacity to carry out the chlorophyllian function on the part
of the vegetation is notably reduced. This occurrence is
92 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
TABLE 1
Correspondence between SPI Values
and Drought Severity Conditions
TABLE 2
Correspondence between De Marton Index and Climate Zone
Fig. 1. Distribution of meteorological stations in Iran. Different colors
show the zonings based on the De Morgan method for the precipitation
data collected for the period 1978-2008.
TABLE 3
Number of Stations in Each Climate Zone (Based on Table 2)
Zone 1 is sever humid (shown in dark blue in Fig. 1), Zone 2 is humid
climate zone (shown in blue in Fig. 1), Zone 3 is semihumid (shown in
dark green in Fig. 1), Zone 4 is Mediterranean (shown in green in Fig. 1),
Zone 5 is semiarid (shown in yellow in Fig. 1), and Zone 6 is arid (shown
in orange in Fig. 1).
demonstrated by the spectral response; the affected
vegetation covers provide the response of the green
vegetation (in a good physiological and healthy state) that
is characterized by a substantial absorption in the red
region and a large reflection in the infrared region near the
electromagnetic spectrum [35]. Having this characteriza-
tion, some remote sensing indices have been introduced for
drought monitoring. In this work, we used a number of
well-known vegetation indices, which are described in the
following.
3.3.1 Normalized Difference Vegetation Index
NDVI is the most commonly used vegetation index account-
ing the amount of vegetation cover in the land. NDVI was
first suggestedby Rouse et al. [36] andTucker [37] as anindex
of vegetation health and density. It is calculated as
`
/
`11
/
111
/
`11
/
111
. 2
where ` is the NDVI and /
NIR
and /
RED
are the reflectance
in the NIR and red bands, respectively.
3.3.2 NDVI-DEV
NDVI itself does not reflect drought or nondrought
conditions. However, the severity of a drought may be
captured by deviation of NDVI from its long-term mean.
This deviation is calculated as the difference between the
NDVI for the current time step and a long-term mean NDVI
for that month as
`
dc.
`
i
`
i.icoi
. 3
where `
i
is the NDVI value for month i and `
i.icoi
is the
long-term mean NDVI for month i over the period that
NDVI values have been calculated. Negative values of `
dc.
indicate the below-normal vegetation condition/health, and
therefore, suggest a prevailing drought situation. Similarly,
positive values of `
dc.
correspond to above-normal vegeta-
tion conditions in the area. However, this index has its own
limitation and drawbacks; the deviation from the mean
does not take into account the standard deviation, and
hence, it can be misinterpreted when the variability in
vegetation conditions in a region is very high in a given
year [38].
3.3.3 Vegetation Condition Index
VCI was suggested by Kogan [39], which shows how close
the NDVI of the current month is to the minimum NDVI
calculated from the long-term record. It is calculated as
\
,

`
,
`
min
`
max
`
min

100. 4
where \
,
is the VCI value of month ,, and `
max
and `
min
are, respectively, the maximum and the minimum values of
NDVI that are calculated from a long-term record for that
month (or week) and , is the index of the current month
(week). The condition/health of the ground vegetation
presented by VCI is measured in percent and may serve as
an approximate measure of how dry the current month is.
3.3.4 Temperature Condition Index
TCI is calculated similarly to VCI. However, in contrast to
VCI, TCI includes the deviation of the current months
value from the recorded maximum, as
T
,

T1
max
T1
,
T1
max
T1
min

100. 5
where T
,
is the TCI value of month ,, and TB, T1
max
and
T1
min
are, respectively, the absolute maximum and mini-
mum smoothed monthly (weekly) brightness temperature
[40]. TCI provides opportunity to identify subtle changes
in vegetation health due to thermal effect as drought
proliferates when moisture shortage is accompanied by
high temperature.
To calculate NDVI, NDVI-DEV, VCI, and TCI, NOAA-
AVHRR data with spatial resolution of 1.1 km 1.1 km
were used for January 1981 to January 2009 period. NOAA-
AVHRR data have some distinct advantages. High tempor-
al resolution, low data volume, and low cost compared to
other high-resolution satellite images are considered as
advantages in using this data source. However, it might
have some limitations such as its coarse resolution. The
spatial resolution of the AVHRR data is 1.1 km at nadir;
however, it decreases with increase in the view angle off
nadir. The maximum off-nadir resolution along track is
2.4 km and across track is 6.9 km.
Iran has a receiving station for downloading NOAA-
AVHRR data allowing us have access to near-nadir images.
Thus, we could obtain images with a resolution of 1.1 km
for all parts of Iran. The data were in high-resolution picture
transmission (hrpt) format. We applied standard proce-
dures for radiometric and geometric correction on the
images. Then, to remove cloud effect, a number of cloud
detection methods were applied and their performance was
compared. Threshold tests are often used for cloud masking
[41], which are based on a rather simple principle, that is if
the measured brightness temperature in one of AVHRRs
infrared window channels is smaller than a predefined
threshold, or if the measured reflectance in one of AVHRRs
visible channels is higher than a predefined threshold, the
pixel is regarded as cloud contaminated. We adopted a
modified version of the algorithm suggested in [41], which
is as follows:
/
4
qt 265 and /
1
/
2
lt 70
/
1
/
2
lt 90
/
4
gt 285.
8
<
:
6
where /
1
, /
2
, and /
4
are the corresponding AVHRR bands
and gt and lt denote greater than and less than,
respectively. The second row of (6) is Reflectance test that is
if reflectance of a pixel in visible bands is lower than the
threshold value, then the pixel is masked as a cloudy one.
The third row of (6) is Brightness temperature test that is
if the brightness temperature (band 4) of a pixel is higher
than the threshold value, then the pixel is masked as a
cloudy one. The first row of (6) is Reflectance and
Brightness test. Some cloud types such as thin cirrus, low
stratus at night, and small cumulus are difficult to detect
because of insufficient contrast with the surface radiance.
By using the first row of (6), we can detect such clouds.
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 93
The GPS coordinates of the meteorological stations were
co-registered with the NDVI imagery data corresponding to
each site. Finally, we constructed two types of time series
for each station; one for SPI and one for satellite data
(NDVI, NDVI-DEV, VCI, and TCI).
4 MODELING/FORECASTING BASED ON NEURAL
NETWORKS
Our aim in this manuscript was to use the power of
satellite images to model and predict the severity of
drought in Iran. Therefore, we considered a number of
features based on satellite images (including NDVI, NDVI-
DEV, TCI, and VCI) as inputs to the model and the drought
condition based on SPI as its output. Fig. 2 shows the
general architecture of the model.
4.1 Neural Network Models
Various architectures of artificial neural networks have
been applied on many real-world applications such as
modeling, classification, smoothing, filtering, prediction,
function approximation, and optimization [42], [43], [44].
Their major advantage is that they are often based on
parallel computations using individual neurons enabling
them to properly model complex tasks. We used a number
of neural network architectures for predicting drought
conditions based on features extracted from NOAA-
AVHRR imagery. These models included multilayer per-
ceptron and radial basis function networks. It has been
proven that networks such as MLPs and RBFs are universal
approximators meaning that they can approximate any
function with a desired precision [44]. Neural networks are
robust against failure in their components, and if trained
properly, they demonstrate high generalizability.
In previous studies, linear models such as regression and
adaptive linear model have been applied to link SPI to
vegetation indices [12], [17], [18], [19]. The free parameters
of these models can be easily determined using methods
such as Delta rule or least mean square. In this work, we
started the modeling task by applying such linear models.
However, due to complexity of the data (for the whole Iran),
linear models did not result in satisfactory performance,
and in some cases, they did not converge to the optimal
parameters. Although being simple in training, linear
models are incapable of capturing interrelations in highly
nonlinear systems (such the case in our data). Therefore, we
used proper neural network structures with nonlinear
modeling elements. MLP and RBF networkswhich are
used hereuse such nonlinear elements and have been
extensively applied for modeling of nonlinear systems.
Fig. 3 shows general architecture of MLP networks. They
consist of an input layer and an output layer. Furthermore,
they may consist of a number of hidden layers that include
the computational elements. These networks can model
nonlinear multi-input multi-output systems. Each neuron in
the hidden and output layers has a (nonlinear) activation
functionoften taken as Heaviside or tangent-hyperbolic
functionsrelating its output to its input. Let us consider
neuron , in the output layer. The output of this neuron is
calculated as
y
,
c
X
/
n
/,
y
/
!
. 7
where y
/
is the output of /th neuron (in the hidden layer
just before the output layer) and n
/,
is the weight of the link
from that neuron to neuron ,. is the activation function of
the neurons. The outputs of the neurons in the hidden
layers can also be computed similarly, based on the outputs
of their preceding layers got practical applications after
invention of the backpropagation (BP) learning algorithm
[45]. MLPs with only one layer often do not results in
accepted performance and one has to increase the number
of hidden layers to get good performance. However,
usually two hidden layer is enough and such architectures
result in acceptable performance. BP is a simple learning
algorithm based on the steepest descent optimization
method, which has two steps. In the forward step, the
output of the neurons in different layers is computed, and
finally, the error signal is obtained. The error signal for
neuron , in the output layer is obtained as
c
,
d
,
y
,
. 8
94 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
Fig. 2. The model that is constructed in this work. It takes the features based on NOAA-AVHRR (e.g., TCI) as input and outputs the drought condition
based on SPI.
Fig. 3. Architecture of an MLP networks with input, hidden, and
output layers.
where d
,
is the desired output for that neuron. As the errors
are computed, in the backward step of BP algorithm, these
error signals are backpropagated to the hidden layers.
Finally, the weights are updated as
n
/,
new n
/,
old cy
/
c
,
. 9
where n
/,
is the link from neuron / to neuron ,. c is the
learning rate, y
/
is the output neuron / (presynaptic
neuron), and c
,
is the local gradient of neuron , (post-
synaptic neuron). Local gradient of neuron , in the output
layer is computed as
c
,
c
,
c
0
.
,
10
and that of any neuron / in the hidden layers is computed as
c
/
c
0
.
/

X
,
n
/,
c
,
. 11
where .
/
is the input to neuron /, which is obtained by
summing all the inputs neuron / receives from those of the
previous layer.
BP algorithm is a powerful tool for performing the
optimization task in MLP and has been used in many
applications.
Other commonly used neural network approach for
modeling nonlinear tasks is RBF [46]. RBFs are fundamen-
tally different from MLP; however, they are, like MLPs,
universal approximators. RBF networks typically consist of
three layers: an input layer, a hidden layer, and an output
layer. Unlike MLPs for which the neurons in the output
and hidden layers are similar, they are different in RBF.
While the output neurons are linear, those in the hidden
layer are nonlinear and have essential role in the nonlinear
modeling task. The output of neuron , in the output layer is
calculated as
y
,

X
/
n
/,
expkr c
/
k. 12
where n
/,
is the weight from neuron / in the hidden layer to
neuron , in the output layer, r is the input vector to the
network, and c
/
is the center of the basis function of
neuron / in the hidden layer. In its simplest form, an RBF
can be learned linearly [46]; however, more complex
learning algorithms may also apply.
SVMs have also been shown good performance in
classification tasks. SVM is designed such that the distance
between the data points in each class from the classifying
boundary is maximized. In this work, we used SVMs with
nonlinear (Gaussian) kernels.
There is no concrete way for designing the optimal
structure of neural networks. Here, first, we tried MLPs
with only single hidden layer. However, regardless of the
number of neurons in this layer, it did not result in a good
performance. Therefore, we considered a network with two
hidden layers. After some try and error, we used MLPs with
two hidden layers of 25 and 15 neurons in each. We also
used the standard BP learning algorithm for optimizing the
free parameters. The activation function of the neurons
was identical and considered to be hyperbolic tangent
sigmoid functionthat has been shown to perform well for
many applications. For RBF networks, we considered
35 computational neurons in the layer and adopted multi-
variate Gaussian functions as basis function.
4.2 Prediction Model
We considered standard autoregressive moving average
(ARMA) model to construct a model with SPI (or the class
for which the SPI value belongs to) as output and NDVI,
NDVI-DEV, VCI, or TCI as input. Let us suppose that TCI is
the input to the model. The general relation between SPI
and TCI is
St j )Tt. Tt 1. Tt 2. . . . . Tt . 13
where o indicates the SPI value and t denotes the current
time and the values of SPI are predicted j steps ahead.
indicates how many past samples of input is used in the
model. ). is the function that is indeed determined by
MLP and RBF.
Here, we will show the results for one step ahead
prediction that is when j 1 (we also computed two and
three step ahead prediction and the results were very
similar to those of one step ahead prediction). There is no
concrete way to determine the optimal for the model, and
it is often tuned by trial and error. As increase, the
accuracy of the model also increases with the price of
increase in computation time. However, in practice, there is
a value for such that further increasing it does not have
considerable influence on the outcome. For our data, we
found 12 to be a good parameter resulting in good
performance. In other words, to predict the SPI value (or
drought condition) in the current month, the TCI values for
the past 12 months were fed into the model as inputs. To
obtain the values of NDVI, NDVI-DEV, TCI, and VCI for
the pixels where the meteorological stations are located (for
which we had the SPI values), we considered a window of
3 3 pixel size with the station in the central pixel. We
then obtained the values for the central pixels (e.g., the
TCI values) by making average over these nine values. The
stations might be placed in locations without enough
vegetation, while their surrounding lands have a vegeta-
tion cover. Averaging the values over the neighboring
pixels with the station in the center makes it possible to
have a smooth vegetation index. Such setting has also been
used in previous works [11], [12].
As indicated in previous sections, Iran had different
climate zones (six climate zones based on the measure that
has been used in this work). It has been shown that the
relation between SPI and satellite-based vegetation index
depend on the specific climate zone the station is placed.
Thus, we used a separate model for each climate zone
leading to six models for the whole stations. It is worth
mentioning that the architecture of these six models is the
same (e.g., the number of hidden layers and neurons);
however, since data of different stations are used to train
each of them, the optimized parameter values are different.
Indeed, we train one neural network model for arid regions,
one for semiarid regions, and so on.
To assess the effectiveness of the models, one may use a
number of metrics such as mean absolute error and
accuracy. Since our task is a classification task (predicting
the class of droughtas indicated by drought conditions
shown in Table 1rather than the real SPI value), we used
the accuracy measure. Accuracy of the classification task is
defined as
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 95
Accuracy
No
c
No
. 14
where No and No
c
are the total number of satiation and
those with correctly predicted drought conditions, respec-
tively. The above metric indicates the portion of stations
for which the model was successful in predicting the
drought condition.
5 RESULTS AND DISCUSSIONS
We used MLP and RBF networks to capture the essential
nonlinearity between SPI and satellite-based vegetation
indices. To this end, in each specific climate zone, the
SPI values were classified into the corresponding drought
conditions (Table 1). We then applied MLP/RBF networks
to solve the classification task as follows: Given the value of
NDVI, NDVI-DEV, VCI, or TCI in the previous months, we
would like to forecast the correct drought condition (class).
To verify the effectiveness of the model, we applied the
trained networks to forecast the drought condition in all
12 months of years 2006, 2007, and 2008. More precisely, to
forecast the drought condition in a specific month of a
particular year (e.g., June 2006), the previous values of the
input (NDVI, NDVI-DEV, VCI, or TCI) and the output
(corresponding drought condition based on SPI values)
were used to train MLP and RBF networks. To this end,
76 percent of the data were used as training data and
12 percent as validation data. The training data were used
to obtain the parameters of the models, and validation data
were used to decide the early stoping point. Since we do not
have many samples, early stopping usually improves the
generalizability of the model [44]. Finally, 12 percent of the
data was considered as test data set, to test the general-
izability of the model on an independent sample. This way,
probable overfitting could be prevented resulting in good
generalizability of the trained network [44]. The data are
resampled to avoid sensitivity to specific choice of training,
validation, and test data sets. Finally, the output of the
network, i.e., the predicted drought condition, was com-
pared with that of the target month and the forecasting
accuracy was determined.
Figs. 4, 5, and 6 show the accuracy of forecast for MLP,
RBF, and SVM, respectively. The networks have one
output that is the drought condition corresponding to the
SPI value. However, we trained the networks four times
each with an input that is NDVI, NDVI-DEV, VCI, or TCI.
As it is seen (Figs. 4, 5, and 6), in RBF and MLP, TCI
resulted in the best performance (i.e., highest accuracy)
than the other three inputs, which indicates that TCI has
the highest capability in monitoring drought conditions.
However, when SVM was used as a classifier, TCI was the
best performing measure only in some months. Appar-
ently, combining all these features enhances our ability in
modeling the correct drought conditions. From these three
models, SVM showed the most improvement when all
features (NDVI, DNVI-DEV, VCI and TCI) were consid-
ered as inputs to the model. The MLP networks could
achieve prediction accuracy of up to 90 percent for some
months (Fig. 4). In 2006, the best accuracy was obtained for
January and May followed by April and August, while
January and October had the best forecasted drought
96 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
Fig. 4. Forecasting accuracy of MLP network with input as NDVI, NDVI-DEV, VCI, or TCI and output as drought condition (extreme dry, moderately
dry, dry, neutral, wet, moderately wet, and extremely wetbased on Table 1). ALL indicates that all these parameters as (NDVI, NDVI-DEV, VCI, or
TCI) are considered as inputs to the model. The results are for months January-December (month No. 1-month No. 12 in the figure) in years 2006,
2007, and 2008.
condition in 2007. The model could forecast the drought
condition of July, August, and December 2008 with an
accuracy of about 90 percent. We found a performance for
April-May 2008 and June 2007. This is due to extreme
drought condition happened in many parts of Iran in 2008.
There was no such a drought condition in 1987-2007 period
for which we had the data. The model learns based on
previous experiences and if it has not received such an
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 97
Fig. 5. Forecasting accuracy of RBF network. Other designations are as in Fig. 4.
Fig. 6. Forecasting accuracy of SVM. Other designations are as Fig. 4.
experience, its predictions will not be accurate. If enough
data are available from different drought conditions, the
model will be able to learn all the patterns.
Fig. 5 shows the forecasting results of RBF network. For
each case, the best RBF parameters were optimized based
on the time series of past times. Its performance is very
similar to that of MLP network with a maximum accuracy
of about 90 percent. Also, similar to MLP, RBF could not
predict well the extreme drought condition happens in
April-May 2008. When the models input is the time series
of TCI or VCI, RBF network resulted in higher accuracy as
compared to the cases with NDVI or NDVI-DEV as input.
In SVM, no single feature was the best performing in all
months (Fig. 6). However, in general, classification accuracy
of TCI and VCI as input was better than that of NDVI and
NDVI-DEV.
The reason for superior performance of TCI and VCI
over NDVI and NDVI-DEV can be explained as follows.
Most of the vegetation indices are efficient indicator of
water stress in relatively homogenous terrain. However,
this is not the case in heterogeneous regions. Both TCI and
VCI are calculated from the long-term record data and also
they are normalized indices. Therefore, they allow for a
comparison of different ecosystems (in heterogeneous
regions such as Iran). They try to separate the short-term
climatic signal from the long-term ecological signal, and as
such, they are better indicators of water stress condition as
compared to NDVI and NDVI-DEV.
To better compare the performance of the classifiers,
Fig. 7 demonstrates the prediction accuracy of these models
when the input is TCI. As it is seen, MLP always resulted in
better performance than RBF and SVM. When all features
were considered as inputs, the performance of the models
was almost the same. It is worth mentioning that the
performance can be improved if the SPI is also fed into the
model as input; however, this was not the aim of this work.
We also compared the performance of the model in
different climate zones (Tables 2 and 3). Fig. 8 shows the
accuracy of MLP in different climate zones (averaged over
all the stations in the zone) in years 2006, 2007, and 2008. As
it is seen, the best performance is obtained for arid and
semiarid regions (Zones 6 and 5, respectively) in which the
majority of the stations are located. RBF also resulted in
similar performance (results not shown here).
One of the key aims of the project was to create a tool to
help decision makers in their policies for minimizing
drought risks. Often, such persons have no or minimal
knowledge about modeling and forecasting tools. Thus, we
created J2EE-based web application for visualizing the
drought maps. We developed a user-friendly tool with web
services, enabling it to be integrated within various portals.
Various open-source technologies such as Hibernate,
Spring, and Vaddin were customized in the implementation
of the back-end services and user interfaces. The tool has a
computational engine that was implemented in C++ and
was integrated into the system. A snapshot of a sample map
is shown in Fig. 9.
The software has a computation engine based on MLP,
SVM and RBF networks. This engine takes the time series of
inputs (NDVI, NDVI-DEV, VCI, or TCI) and output (drought
condition based on SPI), learns the parameters of MLP, SVM
and RBF networks and predicts the drought severity index
for the desired period. To run the program, by pressing the
Run Process button, a windows-based executable file
(.exe) is called by the Java Runtime Environment and a
98 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
Fig. 7. Forecasting accuracy of MLP, RBF, and SVM. The networks get TCI as input and drought condition as output. Other designations are
as in Fig. 4.
background thread is started, which is responsible for
handling the process. After selecting the desired neural
network type and desired satellite-based vegetation index as
inputs, the tool creates a text file containing the predicted
values of SPI for each station. Then, these predicted values
are stored in a database, which has been designed for this
purpose. Users can view the predicted values of SPI on the
graphs and compare different results of runs. We also create
the map showing drought condition (class).
6 CONCLUSION
Droughtas one of the most obstinate climatic extreme
eventsis a destructive climate condition affecting many
parts of Iran and causing considerable damages on both the
environment and human lives. Forecasting drought can
have a significant role in better managing the available
resources and reducing the damages. There are a number of
measures quantifying the severity of drought of which SPI is
the mostly used one. SPIwhich is used in this study-
takes into account the precipitation index normalized in a
period (e.g., 30 years). One of the major challenges for
modeling drought (through SPI) is unavailability of mea-
suring stations in many places, which makes it impossible to
monitor drought conditions for that places. NOAA-AVHRR
images are freely available for all over the world giving
useful information on a number of parameters such as land
cover and vegetation conditions.
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 99
Fig. 9. Snapshot of the tools created for drought prediction based on remote sensing data.
Fig. 8. Accuracy of MLP in different climate zones (Tables 2 and 3) in years 2006, 2007, and 2008 (averaged over all stations in the climate zone)
with various satellite-based vegetation indices as input.
In this work, we used satellite images to forecast the
drought conditions in Iran. To this end, intelligent models
including MLP, RBF, and SVM were considered, and they
were trained based on the time series obtained for the
period 1978-2008. The model took a number of satellite-
based features (NDVI, NDVI-DEV, VCI, or TCI) as an input
and outputted the drought conditioncharacterized by SPI.
Our experimental results showed that this model could
successfully forecast the drought condition with an accu-
racy of up to 90 percent. Also, we found that TCIamong
satellite-based featuresresulted in the best performance
and MLP as the best performing model. We created a web-
based application for visualizing the results producing
forecasted nationwide maps of drought conditions. Our aim
in this work was to find the best performing satellite-based
vegetation index to model drought conditions, and we
found that TCI is such a measure.
Helping public policy making is an important issue in
sustainability applications. Policy makers have a major role
in minimizing the risks of a natural hazard like drought
condition. To help the policy makers to have access to real-
time drought conditions (and future forecast), we created a
user-friendly software tool targeting such audience.
ACKNOWLEDGMENTS
This work was supported by Remote Sensing Center of Iran
under Grant No. 900/139. The authors would like to thank
Dr. B. Saghafian, Mr. H. Tajari, Mr. M. Jannati, Mr. S.
Naeeni, Mr. M. Ansari, and Mr. M. Salehifar for their help
in processing the data and interpreting the results.
REFERENCES
[1] R.R. Heim, A Review of Twentieth-Century Drought Indices
Used in the United States, Bull. Am. Meteorological Soc., vol. 83,
pp. 1149-1165, 2002.
[2] W.C. Palmer, Meteorological Drought, US Dept. of Com-
merce, Weather Bureau, Research Paper No. 45, Washington
DC, 1965.
[3] T.B. McKee, N.J. Doesken, and J. Kleist, The Relationship of
Drought Frequency and Duration to Time Scales, Proc. Eighth
Conf. Applied Climatology, 1993.
[4] A.H. Weerts, J. Schellekens, and F.S. Weiland, Real-Time
Geospatial Data Handling and Forecasting: Examples from
Delft-FEWS Forecasting Platform/System, IEEE J. Selected Topics
in Applied Earth Observations and Remote Sensing, vol. 3, no. 3,
pp. 386-394, Sept. 2010.
[5] A. Diouf and E.F. Lambini, Monitoring Land-Cover Changes in
Semi-Arid Regions: Remote Sensing Data and Field Observations
in the Ferlo, Senegal, J. Arid Environments, vol. 48, pp. 129-148,
2001.
[6] C.J. Tucker and B.J. Choudhury, Satellite Remote Sensing of
Drought Conditions, Remote Sensing of Environment, vol. 23,
pp. 243-251, 1987.
[7] A.J. Peters, E.A. WalterShea, L. Ji, A. Vin a, M. Hayes, and M.D.
Svoboda, Drought Monitoring with NDVI-Based Standardized
Vegetation Index, Photogrammetric Eng. and Remote Sensing,
vol. 68, pp. 71-75, 2002.
[8] C. Gouvia, R.M. Trigo, and C.C. DaCamara, Drought and
Vegetation Stress Monitoring in Portugal Using Satellite Data,
Natural Hazards and Earth System Sciences, vol. 9, pp. 185-195, 2009.
[9] J.D. Bolten, W.T. Crow, X. Zhan, T.J. Jackson, and C.A. Reynolds,
Evaluating the Utility of Remotely Sensed Soil Moisture
Retrievals for Operational Agricultural Drought Monitoring,
IEEE J. Selected Topics in Applied Earth Observations and Remote
Sensing, vol. 3, no. 1, pp. 57-66, Mar. 2010.
[10] C.M. Rulinda, A. Dilo, W. Bijker, and A. Steina, Characterising
and Quantifying Vegetative Drought in East Africa Using Fuzzy
Modelling and NDVI Data, J. Arid Environments, vol. 78, pp. 169-
178, 2012.
[11] C. Bhuiyan, R.P. Singh, and F.N. Kogan, Monitoring Drought
Dynamics in the Aravalli Region (India) Using Different Indices
Based on Ground and Remote Sensing Data, Intl J. Applied Earth
Observation and Geoinformation, vol. 8, pp. 289-302, 2006.
[12] S.K. Jain, R. Keshri, A. Goswami, and A. Sarkar, Application of
Meteorological and Vegetation Indices for Evaluation of Drought
Impact: A Case Study for Rajasthan, India, Natural Hazards,
vol. 54, pp. 643-656, 2010.
[13] C. Hopfner and D. Scherer, Analysis of Vegetation and Land
Cover Dynamics in North-Western Morocco during the Last
Decade Using Modis NDVI Time Series Data, Biogeosciences,
vol. 8, pp. 3359-3373, 2011.
[14] B. Li and S. Tao, Relations between AVHRR NDVI and
Ecoclimatic Parameters in China, Intl J. Remote Sensing, vol. 23,
pp. 989-999, 2002.
[15] A. Lotsch, M.A. Friedl, and B.T. Anderson, Coupled Vegetation-
Precipitation Variability Observed from Satellite and Climate
Records, Geophysical Research Letters, vol. 30, pp. 1774-1777, 2003.
[16] M.J. Kinyanjui, NDVI-Based Vegetation Monitoring in Mau
Forest Complex, Kenya, African J. Ecology, vol. 49, pp. 165-174,
2011.
[17] X. Song, G. Saito, M. Kodama, and H. Sawada, Early Detection
System of Drought in East Asia Using NDVI from NOAA AVHRR
Data, Intl J. Remote Sensing, vol. 25, pp. 3105-3111, 2004.
[18] L. Ji and A.J. Peters, Assessing Vegetation Response to Drought
in the Northern Great Plains Using Vegetation and Drought
Indices, Remote Sensing of Environment, vol. 87, pp. 85-98, 2003.
[19] S.M. Quiring and S. Ganesh, Evaluating the Utility of the
Vegetation Condition Index (VCI) for Monitoring Meteorological
Drought in Texas, Agricultural and Forest Meteorology, vol. 150,
pp. 330-339, 2010.
[20] T. Gebrehiwot, A. van der Veen, and B. Maathuis, Spatial and
Temporal Assessment of Drought in the Northern Highlands of
Ethiopia, Intl J. Applied Earth Observation and Geoinformation,
vol. 13, pp. 309-321, 2011.
[21] A.K. Mishra and V.R. Desai, Drought Forecasting Using Feed-
Forward Recursive Neural Network, Ecological Modelling,
vol. 198, pp. 127-138, 2006.
[22] M.E. Keskin, O

. Terzi, E.D. Taylan, and D. Ku cu kyaman,


Meteorological Drought Analysis Using Artificial Neural Net-
works, Scientific Research and Essays, vol. 6, pp. 4469-4477, 2011.
[23] M.T. Dastorani, H. Afkhami, and B. Borroni, Application of
Artificial Neural Networks on Drought Prediction in Yazd
(Central Iran), Desert, vol. 16, pp. 39-48, 2011.
[24] S. Barua, B.J.C. Perera, A.W.M. Ng, and D. Tran, Drought
Forecasting Using an Aggregated Drought Index and Artificial
Neural Networks, J. Water and Climate Change, vol. 1, pp. 193-206,
2010.
[25] A.R. Sadeghi, A.A. Kamgar-Haghighi, A.R. Sepaskahah, D.
Khalili, and S. Zand-Parsa, Regional Classification for Dryland
Agriculture in Southern Iran, J. Arid Environments, vol. 50,
pp. 333-341, 2002.
[26] A.P. Raziei and B. Saghafian, Annual Rainfall Trend in Arid &
Semi-Arid Regions of Iran, Proc. 21st European Regional Conf.,
pp. 20-28, 2005.
[27] N.D.K. Dayawansa, R.P. De Silva, and M.D. Ratnasiri, A
Comparison of Methods Used in Estimating Missing Rainfall
Data, J. Agricultural Sciences, vol. 3, pp. 101-108, 2007.
[28] T. Buishand, Some Methods for Testing the Homogeneity of
Rainfall Records, J. Hydrology, vol. 58, pp. 11-27, 1982.
[29] J.D.J. Salas, Analysis and Modeling of Hydrological Time Series.
McGraw-Hill, 1993.
[30] J.V. Bradley, Distribution-Free Statistical Tests. Prentice-Hall, 1968.
[31] T.B. McKee, N.J. Doeskin, and J. Kieist, The Relationship of
Drought Frequency and Duration to Time Scales, Proc. Eighth
Conf. Applied Climatology, pp. 179-184, 1993.
[32] D.C. Edwards, Characteristics of 20th Century Drought in the
United States at Multiple Time Scales, Master thesis, Colorado
State Univ., 1997.
[33] T.B. McKee, N.J. Doeskin, and J. Kieist, Drought Monitoring with
Multiple Time Scales, Proc. Ninth Conf. Applied Climatology,
pp. 233-236, 1995.
100 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 1, JANUARY 2014
[34] Y. Moradi, M.A.B.M Said, and I. Bin, Drought Impacts and
Vulnerability in Isfahan Province, World Academy of Science, Eng.
& Technology, vol. 53, p. 740, May 2011.
[35] F.G. Alonso et al., Drought Monitoring in Spain Using Satellite
Remote Sensing, Sensors and Environmental Applications of Remote
Sensing, 1995.
[36] J.W. Rouse, R.H. Haas, J.A. Schell, D.W. Deering, and J.C. Harlan,
Monitoring the Vernal Advancement and Retrogradation (Greenwave
Effect) of Natural Vegetation. Texas A&M Univ., 1974.
[37] C.J. Tucker, Red and Photographic Infrared Linear Combinations
for Monitoring Vegetation, Remote Sensing of Environment, vol. 8,
pp. 127-150, 1979.
[38] A.T. Anayamba and C.J Eastma Jr., NDVI Anomaly Patterns over
Africa during the 1997/98 ENSO Warm Event, Intl J. Remote
Sensing, vol. 22, pp. 1847-1859, 2001.
[39] F.N. Kogan, Operational Space Technology for Global Vegetation
Assessment, Bull. Am. Meteorological Soc., vol. 82, pp. 1949-1964,
1995.
[40] F.N. Kogan, Application of Vegetation Index and Brightness
Temperature for Drought Detection, Advances Space Research,
vol. 15, pp. 91-100, 2001.
[41] L. Giglio, J. Kendall, and C. Tucker, Remote Sensing of Fires with
the TRMM VIRS, Remote Sensing of Environment, vol. 21, pp. 203-
207, 2003.
[42] L.O. Chua and L. Yang, Cellular Neural Networks: Applica-
tions, IEEE Trans. Circuits and Systems, vol. 35, no. 10, pp. 1273-
1290, Oct. 1988.
[43] G.P. Zhang, Neural Networks for Classification: A Survey, IEEE
Trans. Systems, Man, and Cybernetics C: Applications and Rev.,
vol. 30, no. 4, pp. 451-462, Nov. 2000.
[44] S. Haykin, Neural Networks: A Comprehensive Foundation, second
ed. Prentice-Hall, 1999.
[45] D.E. Rumelhart, G.E. Hintont, and R.J. Williams, Learning
Representations by Back-Propagating Errors, Nature, vol. 323,
pp. 533-536, 1986.
[46] S. Chen, C.F.N. Cowan, and P.M. Grant, Orthogonal Least
Squares Learning Algorithm for Radial Basis Function Networks,
IEEE Trans. Neural Networks, vol. 2, no. 2, pp. 302-309, Mar. 1991.
Mahdi Jalili received the BS degree in electrical
engineering from Tehran Polytechnique in 2001,
the MS degree in electrical engineering from the
University of Tehran in 2004, and the PhD
degree from the Swiss Federal Institute of
Technology Lausanne in 2008. In 2009, he
joined the Department of Computer Engineering,
Sharif University of Technology as an assistant
professor. His research interests are in dynami-
cal networks, synchronization in complex net-
works, computational neuroscience, and human brain functional
connectivity analysis. He received the 2009 presidential award of the
15th Razi Research Festival. He is a member of the IEEE.
Joobin Gharibshah received the BSc degree
in computer engineering (Software) from the
Iran University of Science and Technology,
Tehran, Iran, in 2009, and the MSc degree in
artificial intelligence from the Sharif University
of Technology, Tehran, Iran, in 2012. His
current research interests are focused on
complex dynamic networks, social networks
analysis and data mining, neural networks and
operating systems.
Seyed Morsal Ghavami received the BSc
degree in civil engineering (Surveying) from the
University of Zanjan, Iran, in 2009, and the MSc
degree in geomatic engineering from K.N. Toosi
University in 2012. He is currently working
toward the PhD degree in geospatial information
system at K.N. Toosi University. His research
interest includes GIS-based artificial intelli-
gence, agent-based spatial modeling, spatio-
temporal GIS, disaster management, remote
sensing and applications of GIS and RS in land resources.
Mohammadreza Beheshtifar received the MSc degree in remote
sensing of the environment. He is currently with the Remote Sensing
Centre, ISA, Tehran, Iran.
Reza Farshi graduated from the Sharif
University of Technology in 2004 with a
degree in computer engineering. He has been
active in building web applications since then.
For the last five years, he has helped various
programming teams to adopt open-source
technologies in the field of Enterprise Java
such as Spring, Hibernate, Apache, Struts,
and use them effectively.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
JALILI ET AL.: NATIONWIDE PREDICTION OF DROUGHT CONDITIONS IN IRAN BASED ON REMOTE SENSING DATA 101

Você também pode gostar