Comparison of modelling ann and elm to estimate solar radiation over turkey using noaa satellite data

This article was downloaded by: [Siirt Universitesi]
On: 02 September 2013, At: 01:34
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of Remote
Sensing
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tres20
Comparison of modelling ANN and ELM
to estimate solar radiation over Turkey
using NOAA satellite data
Mehmet Şahin
a
a
Engineering Faculty, Siirt University , Siirt , 56100 , Turkey
Published online: 19 Aug 2013.
To cite this article: Mehmet ahin (2013) Comparison of modelling ANN and ELM to estimate solar
radiation over Turkey using NOAA satellite data, International Journal of Remote Sensing, 34:21,
7508-7533, DOI: 10.1080/01431161.2013.822597
To link to this article: http://dx.doi.org/10.1080/01431161.2013.822597
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions

International Journal of Remote Sensing, 2013
Vol. 34, No. 21, 7508–7533, http://dx.doi.org/10.1080/01431161.2013.822597
Comparison of modelling ANN and ELM to estimate solar radiation
over Turkey using NOAA satellite data
Mehmet ¸Sahin*
Engineering Faculty, Siirt University, Siirt 56100, Turkey
(Received 29 May 2012; accepted 22 June 2013)
In this study, solar radiation (SR) is estimated at 61 locations with varying climatic
conditions using the artificial neural network (ANN) and extreme learning machine
(ELM). While the ANN and ELM methods are trained with data for the years 2002 and
2003, the accuracy of these methods was tested with data for 2004. The values for
month, altitude, latitude, longitude, and land-surface temperature (LST) obtained from
the data of the National Oceanic and Atmospheric Administration Advanced Very High
Resolution Radiometer (NOAA-AVHRR) satellite are chosen as input in developing the
ANN and ELM models. SR is found to be the output in modelling of the methods.
Results are then compared with meteorological values by statistical methods. Using
ANN, the determination coefficient (R2
), mean bias error (MBE), root mean square error
(RMSE), and Willmott’s index (WI) values were calculated as 0.943, −0.148 MJ m−2
,
1.604 MJ m−2
, and 0.996, respectively. While R2
was 0.961, MBE, RMSE, and WI
were found to be in the order 0.045 MJ m−2
, 0.672 MJ m−2
, and 0.997 by ELM. As can
be understood from the statistics, ELM is clearly more successful than ANN in SR
estimation.
1. Introduction
Solar radiation (SR) is a general expression of electromagnetic radiation emitted by the
Sun. Energy can be captured and converted into a useful form of energy, especially heat
and electrical energy. In recent years, many studies for various purposes in the field of solar
radiation have been used. These can be listed under agronomy, hydrology and ecology,
photovoltaic cells and thermal solar systems, solar furnaces, concentrating collectors, and
interior illumination of buildings, etc. (Benghanem, Mellit, and Alamri 2009; Ulgen and
Hepbasli 2009).
Although solar radiation is very important, values of SR cannot be easily obtained like
other meteorological parameters such as air temperature, land-surface temperature (LST),
and relative humidity. SR measurement is limited for various practical reasons such as the
purchase of vehicles engaged in measuring, maintenance and repair costs, and calibration
of instruments (Bakirci 2009). In fact, even in developed countries, SR measurement instru-
ments are not found in all meteorological stations. However, SR values from all stations are
needed to enable the validity of research. In order to overcome this problem, researchers
have tried to acquire SR values by using artificial neural network (ANN) methods that
may be applied to parameters such as latitude, longitude, altitude, sunshine duration, LST,
*Email: msahin@siirt.edu.tr
© 2013 Taylor & Francis
Downloadedby[SiirtUniversitesi]at01:3402September2013

International Journal of Remote Sensing 7509
air temperature, relative humidity, pressure, and rainfall, which are easily obtained by
meteorological stations unable to measure SR.
The efficiency of ANN models in SR estimation is proved through the comparison of
obtained values with real values, and researchers agree that ANN models are suitable for
and applicable to SR estimation (Bechrakis and Sparis 2004). In developing this process,
some researchers have attempted to obtain SR values for measured locations (Koca et al.
2011; Ozgoren, Bilgili, and Sahin 2012). Then, researchers have trained ANN benefiting
from the points of the measurement of SR and applying to the same or different locations
in used ANN models. Although ANN methods are accepted to be successful, especially in
estimation of SR in the early days, lack of the ANN methods has been understood with time
at locations that have not got meteorological stations. So it is clear that there is no facility
to get basic meteorological parameters such as LST, air temperature, relative humidity,
pressure, and rainfall to estimate SR on these places. To overcome this problem, researchers
have begun to use remote-sensing methods to estimate the SR (Cracknell and Varotsos
2007). The satellites are used as effective instruments in remote-sensing methods, and the
data obtained from satellite channels are converted to a suitable form so that SR can be
estimated without the use of ANN models by using various algorithms (Janjai et al. 2011;
Polo et al. 2011). However, ANN methods dependent on satellite data are now being used
to estimate SR (Qin et al. 2011; Lu et al. 2011; Rahimikhoob, Behbahani, and Banihabib
2013).
Nowadays, researchers have developed ANN and various intelligent methods to pre-
dict target properties, one of these being extreme learning machines (ELMs). The classical
learning algorithm in neural networks such as ANN requires the setting of several user-
defined parameters. However, ELM only requires the setting of the number of hidden
neurons and the activation function. It does not require adjustment of input weights and
hidden layer biases during implementation of the algorithm, and it produces only one opti-
mal solution (Cheng, Cai, and Pan 2009). Therefore, it has been determined by various
studies that the training of large data sets and developed network of testing time by the
ELM method requires only a short time according to ANN methods (Huang, Zhu, and
Siew 2006; Yeu et al. 2006; Feng et al. 2009; Huang, Wang, and Lan 2011). This is a dif-
ferent innovation that ELM has contributed to the literature. ELM is used in various fields
depending on these features. Fields that can be expressed include remote sensing (Pal 2009;
Chang et al. 2010), health (Kwak and Kwon 2008; Bharathi and Natarajan 2011; Qu et al.
2011), recognition of handwriting characters (Chacko et al. 2012), image deblurring (Wang
et al. 2011), the effects of the electrical storm transmission (Yang et al. 2011), electricity
price forecasting (Chen et al. 2012), reservoir permeability prediction (Cheng, Cai, and Pan
2009), classification of electronic nose data (Prakash and Rajesh 2011), sales forecasting
(Sun et al. 2008), metagenomic taxonomic classification (Rasheed and Rangwala 2012),
particle swarm optimization (Han, Yao, and Ling 2012), abnormal paediatric gait classifi-
cation (Rani and Arumugam 2010), etc. However, no study has been reported on estimation
of SR by ELM, either with satellite or meteorological data, and this study is the first to use
ELM for SR estimation.
In this study, SR prediction was achieved by using both ANN and ELM for satellite data
pertaining to the same training and testing locations, with the aim of acquiring missing SR
data. A further aim was to determine the success of the ELM method in comparison with
ANN, which is commonly utilized in modelling SR over Turkey. Because the data from
2002 and 2003 are employed to train the network, those for 2004 were used to test the

7510 M. ¸Sahin
accuracy of both methods in 61 locations. Month, altitude, latitude, longitude, and LST
were considered as input data during the training of the network. The 603 LST maps were
obtained using the normalized difference vegetation index (NDVI) and emissivity maps for
2002–2004. Then, 42 monthly mean LST maps were created from related 603 LST maps.
LST values were created using data obtained from the National Oceanic and Atmospheric
Administration Advanced Very High Resolution Radiometer (NOAA-AVHRR) sensor in
the Becker–Li (1990) algorithm.
2. Study area and data sources
Turkey is divided into seven geographical regions depending on the climatic conditions.
These are the Mediterranean Region, Aegean Region, Marmara Region, Black Sea Region,
Central Anatolia Region, Eastern Anatolian Region, and Southeastern Anatolia Region,
each region having its own unique climate characteristics. The sixty-one locations which
are selected as the control points in the study are provided based on the distribution of
property over seven geographical regions (see Figure 1).
The altitudes, latitudes, and longitudes used as input parameters in ANN and ELM to
estimate SR and geographical regions are shown in Table 1. The satellite data used for
the purpose of both training and testing for the period 2002–2004 were provided by the
Scientific and Technological Research Council of Turkey-Bilten. The meteorological values
for related time periods were obtained from the Republic of Turkey Ministry of Forestry
and Water Affairs (Turkish State Meteorological Service).
3. Methodology
3.1. Estimation of NDVI
NDVI is a simple graphical indicator that can be used to analyse remote-sensing mea-
surements and assess whether the target being observed contains live green vegetation
or not. Data from the red and near-infrared channels are taken from satellite sensors in
remote-sensing studies. When received data are analysed, marked differences in reflections
of the red and near-infrared channels of plants are observed depending on spatial resolution.
Accordingly, the value of NDVI in NOAA-AVHRR is formulated as follows:
NDVI =
NIR − RED
NIR + RED
, (1)
where RED and NIR are spectral reflection in near-infrared and visible regions, respec-
tively. If Equation (1) is rewritten relative to NOAA-AVHRR, Equation (2) can be
obtained:
NDVI =
CH2 − CH1
CH2 + CH1
, (2)
where CH1 and CH2 are the reflectance values of the first and second channels on board the
NOAA-AVHRR, respectively. According to Equation (2), NDVI can take values between
−1 and +1, directly dependent on the energy absorption and photosynthetic capacity of the
vegetation (Sellers 1985; Myneni et al. 1995).

Figure1.MapofTurkeyandstudylocations.

7512 M. ¸Sahin
Table 1. Locations used in the study.
Location Altitude (m) Latitude (◦
N) Longitude (◦
E) Geographical region*
Adana 27 37.03 35.21 1
Adıyaman 672 37.45 38.17 2
A˘grı 1632 39.43 43.03 3
Aksaray 960.77 38.23 34.03 4
Amasya 411.19 40.39 35.51 5
Ankara 890.52 39.57 32.53 4
Antakya 100 36.12 36.10 1
Antalya 63.57 36.42 30.44 1
Artvin 628.30 41.11 41.49 5
Aydın 56.30 37.51 27.51 6
Balıkesir-Gönen 37 40.06 27.39 7
Batman 310 37.35 41.07 2
Bilecik 539.19 40.09 29.59 7
Bingöl 1177 38.52 40.30 3
Bitlis 1573 38.22 42.06 3
Burdur 957 37.43 30.18 1
Bursa 100.32 40.13 29 7
Çanakkale 5.5 40.08 26.24 7
Çorum 775.91 40.33 34.58 5
Diyarbakır 674 37.54 40.12 2
Denizli 425.29 37.47 29.05 6
Edirne 85 41.41 26.33 7
Elâzı˘g 989.75 38.39 39.15 3
Erzincan 1218.22 39.45 39.30 3
Erzurum 1758.18 39.57 41.40 3
Gaziantep 854 37.03 37.21 2
Gümü¸shane 1219 40.28 39.28 5
Hakkâri 1727.74 37.34 43.44 3
I˘gdır 858 39.55 44.03 3
Isparta 996.88 37.45 30.33 1
˙Istanbul-Göztepe 32.98 40.58 29.05 7
˙Izmir 28.55 38.23 27.04 6
Kahramanmara¸s 572.13 37.36 36.56 1
Karaman 1023.05 37.12 33.13 4
Kars 1775 40.37 43.06 3
Kastamonu 800 41.22 33.47 5
Kayseri 1092 38.43 35.29 4
Kır¸sehir 1007.17 39.09 34.10 4
Kilis 650 36.42 37.06 1
Kocaeli-˙Izmit 76 40.46 29.56 7
Konya 1030 37.52 32.28 4
Kütahya 969.25 39.25 29.58 6
Malatya 947.87 38.21 38.13 3
Mersin 3.40 36.48 34.38 1
Mu˘gla 646 37.13 28.22 6
Mu¸s 1322.76 38.41 41.29 3
Ni˘gde 1210.50 37.58 34.41 4
Ordu 4.10 40.59 37.54 5
Rize 8 41.02 40.30 5
Samsun 4 41.21 36.15 5
Siirt 895.54 37.55 41.57 2
Sinop 32 42.02 35.50 5
Sivas 1285 39.45 37.01 4
(Continued)

Table 1. (Continued).
Location Altitude (m) Latitude (◦
N) Longitude (◦
E) Geographical region*
¸Sanlıurfa 547.18 37.09 38.47 2
Tokat 607.90 40.18 36.34 5
Trabzon 30 40.59 39.45 5
Tunceli 980 39.07 39.33 3
Van 1670.58 38.28 43.21 3
Yalova 3.81 40.40 29.17 7
Yozgat 1298.33 39.49 34.48 4
Zonguldak 135.35 41.27 31.38 5
Note: *Mediterranean Region (1), Southeastern Anatolia Region (2), Eastern Anatolian Region (3), Central
Anatolia Region (4), Black Sea Region (5), Aegean Region (6), and Marmara Region (7).
3.2. Estimation of surface emissivity
Surface emissivity is defined as the ability of the heat energy of land surfaces to be trans-
formed into light energy as black body modelling. According to this principle, NDVI maps
were used to obtain the following emissivity formulae:
ε4 = 0.9897 + 0.029 ln (NDVI), (3)
ε4 − ε5 = 0.01019 + 0.01344 ln (NDVI), (4)
where ε4 and ε5 are emissivity values related to the fourth and fifth channels of the
NOAA-AVHRR sensor, respectively (Cihlar et al. 1997). Also, ε4 and ε5 are used in the
Equations (5) and (6) to obtain the formula of difference of emissivity ( ε) and average of
emissivity (ε), respectively:
ε = ε4 − ε5, (5)
ε =
ε4 + ε5
2
. (6)
3.3. Estimation of LST by NOAA-AVHRR
Land surface is a key parameter in many applications, such as the Earth’s energy and
water cycles, water–heat balance, energy balance, drought monitoring, agriculture mete-
orology, forest fires, disaster monitoring, etc. (Vazquez, Reyes, and Arboledas 1997). LST
is estimated using satellites that can scan land surfaces at different spectral channels. One
satellite, NOAA-AVHRR, has two thermal channels (4 and 5) operating at 10.5–11.3 µm
and 11.5–12.5 µm, respectively, for land-surface monitoring (Prabhakara, Dalu, and Kunde
1974; McMillin 1975). Various split-window algorithms have been developed based on the
two adjacent thermal channels, one of which is that by Becker and Li (1990), who derived
a local split-window for viewing angles of up to 46◦
from nadir, given as follows:
TBecker−Li−1990 = 1.274 + P
T4 + T5
2
+ M
T4 − T5
2
, (7)
P = 1 + 0.15616
1 − ε
ε
− 0.482
ε
ε2
, (8)

7514 M. ¸Sahin
M = 6.26 + 3.98
1 − ε
ε
+ 38.33
ε
ε2
, (9)
where T4 and T5 are brightness temperatures of channels 4 and 5 of NOAA-AVHRR,
respectively. P and M are coefficients dependent on atmospheric effects and regional
surface emissivity. The coefficients of P and M used in Equation (7) were found by
LOWTRAN 6 simulation program (US Air Force Research Laboratory, Wright-Patterson
AFB, OH, USA).
3.4. Artificial neural network
ANN creates modelling based on a biological neural system. This method is learned from
given examples by constructing input–output mapping in order to perform predictions
(Kalogirou 2000). ANN modelling is composed of an input layer, one or more hidden
layers, and an output layer. Neurons in each of the layers and weights interconnect. One
of most important issues in ANN is the bindings that provide data transmission between
neurons. A binding that transmits data from one neuron to another also has a weight value.
G(x) is a summation function that calculates the exact input reaching a neuron. The input,
by multiplying with variables and weight coefficients, builds up input for G(x) summation
function. The mathematical expression of an artificial neuron can be written as
yi = F [G (x)] = F
n
i=1
wijxj − Qi ; xi = (x1, x2, . . . , xn), (10)
where x = {x1, x2, x3, . . . xn} is an input variable to be processed. On the other hand,
w = {w00, w01, . . . ,wij} is weights and shows the importance of data reaching a neuron and
their impact on it (Karem et al. 2008). The values of weights can change in the process of
training. Qi represents threshold value; F (.) is an activation function. G (.), that comes to
F(.), is the function that produces the output by processing the inputs.
3.5. Extreme learning machine
ELM is a feed-forward neural network model that has a single hidden layer, and calculates
input weights randomly and output weights analytically. The nondifferentiable or discon-
tinuous activation functions can also be used with activation functions such as sigmodial,
sine, Guassian, and hard-limiting in the hidden layer of ELM (Suresh, Saraswathi, and
Sundararajan 2010).
Traditional feed-forward neural networks depend on parameters such as momentum
and learning rate. In this type of network, parameters such as weights and threshold values
should be updated with gradient-based learning algorithms. However, the learning process
takes time and is affected by local point errors to ensure optimum performance. Changing
the momentum value may prevent point of the error locally, but will not affect the long-
term impact of the learning process. ELM also generates input weights and threshold values
randomly, but output weights are calculated mathematically (Huang, Zhu, and Siew 2006).
The ELM network is the customized state of an ANN model comprising a single hidden
layer and feed-forward.
If X = (X1, X2, X3, . . . , XN ) and Y determine input and output features, respectively, the
mathematical expression of the network with M neurons in the hidden layer is indicated as
follows (Suresh, Saraswathi, and Sundararajan 2010):

M
i=1
βig(WiXk + bi) = Ok, k = 1, 2, . . . . . . . . . N, (11)
where Wi = (Wi1, Wi2 . . . . . . . . . Win) and βi = (βi1, βi2 . . . . . . . . . βim) express weights
in the input and output layers, respectively. While bi determines threshold values in the
hidden layer, Ok represents output values. g(.) is the activation function (Rong et al. 2008).
The purpose of N input features in a network is achieving the error as
N
k=1
(Ok − Yk) = 0
or min
N
k=1
(Ok − Yk)2
. Therefore, Equation (11) can be rewritten as follows (Huang, Zhu,
and Siew 2006):
M
i=1
βig(WiXk + bi) = Yk, k = 1, 2, . . . . . . . . . , N. (12)
In addition, the Hβ = Y equation can be used in Equation (12) (Huang, Zhu, and Siew
2006). H, β, and Y are indicated as follows (Suresh, Saraswathi, and Sundararajan
2010):
H =
⎡
⎣
g(W1X1 + b1) · · · g(WM X1 + b1)
...
...
...
g(W1XN + b1) · · · g(WM XN + bM )
⎤
⎦
N×M
, (13)
β =
⎡
⎢
⎣
βT
1
.
.
βT
M
⎤
⎥
⎦
Mxm
and Y =
⎡
⎢
⎣
YT
1
.
.
YT
M
⎤
⎥
⎦
Nxm
, (14)
where H is the input matrix in the hidden layer. Training of the network in a feed-forward
ANN corresponds to searching for the solution of linear least squares in the equation Hβ =
Y by the ELM method. The ELM algorithm can be summarized in three steps, as follows
(Huang, Zhu, and Siew 2004; Liang et al. 2006):
(1) Wi = (Wi1, Wi2 . . . . . . . . . , Win) input weights and threshold values of bi of the
hidden layer are generated randomly;
(2) H hidden layer output is calculated;
(3) β output weights are calculated according to β = H†
Y. Y is the target feature.
3.6. Performance criteria
In statistics, the coefﬁcient of determination (R2
) is used in the context of statistical models
whose main purpose is the prediction of future outcomes on the basis of other related infor-
mation. It is the proportion of variability in a data set that is accounted for by the statistical
model. This provides a measure of how well future outcomes are likely to be predicted by
the model (Steel and Torrie 1960). Mean bias error (MBE) testing provides information
on the long-term performance, with a low MBE being desirable. Ideally, a zero value for
MBE should be obtained. A positive value gives the average amount of overestimation, and
a negative value underestimation. The root mean square error (RMSE) is always positive

7516 M. ¸Sahin
and a zero value is ideal. This test provides information on the short-term performance of
the models by allowing a term-by-term comparison of actual deviation between the cal-
culated and measured values (Katiyar et al. 2010). Recently, Willmott’s index (WI) has
been widely used to analyse comparison studies, and is intended as a descriptive measure.
It is both a relative and bounded measure that may be applied in many different fields
in order to make cross-comparisons between models (Willmott 1982). WI takes values of
0 ≤ WI ≤ 1.
In this study, R2
, MBE, RMSE, and WI are used statistically to establish criteria for
the estimation of LST and SR, and also for comparison of ANN with ELM. These criteria
indicate how input features explain SR, and the criteria are calculated using the following
formulae:
R2
=
n
i=1
(Yi − Yi)2
−
n
i=1
(Yi − ˆYi)2
n
i=1
(Yi − Yi)2
, (15)
MBE =
1
n
n
i=1
ˆYi − Yi , (16)
RMSE =
1
n
n
i=1
ˆYi − Yi
2
, (17)
WI = 1 −
n
i=1
ˆYi − Yi
2
/
n
i=1
ˆYi + Yi
2
, (18)
where n is total sample size, Y is actual SR values, and Y and ˆY define average actual SR
values and estimated SR values, respectively (Erdinç 2005; Sousa et al. 2007). Additionally,
ˆY and Y can be expressed as ˆY = ˆY − Y and Y = Y − Y, respectively.
4. Results and discussion
4.1. Land-surface temperature
First, images of NOAA 12-14-15-16/AVHRR were converted to the format of Level-1B,
which can recognize the format by image processing programs, through Quorum software.
Then, Envi 4.3 (ITT Exelis Company, Colorado Springs, CO, USA) and Idrisi Andes (Clark
Labs Company, Jamestown, NY, USA) image processing programs were used to make
radiometric and geometric corrections of the images. The channels of the first and sec-
ond obtained images were used in Equation (2) to create NDVI images. One of the images,
shown in Figure 2(a), was generated on 20 May 2002, at 06:44 local time. When the NDVI
image is examined, it is clear that the image takes values varying between −0.68 and
+0.75 (see Figure 2(a)). This form of NDVI image is not appropriate to use statistically in
Equations (3)–(4) because the ln(NDVI) function is undefined in the range −1 ≤ NDVI ≤ 0.
Therefore, the values between −1 ≤ NDVI ≤ 0 are removed from the NDVI images (see
Figure 2(b)). When the figure is examined, it will be seen that NDVI values in western
Turkey are between 0.14 and 0.38. While the effective NDVI range in the northwest of the
country is between 0.24 and 0.42, it is occasionally possible to find NDVI values between
0.52 and 0.57 in individual locations, and in northern Turkey the range is 0.28–0.71 where

(a)(b)(c)
(d)(e)(f)
–0.68
–0.59
–0.50
–0.41
–0.32
–0.23
–0.14
–0.05
0.04
0.13
0.22
0.31
0.40
0.49
0.58
0.67
0.750.75
0.71
0.66
0.61
0.57
0.52
0.47
0.42
0.38
0.33
0.28
0.19
0.14
0.09
0.05
0.000.00
0.06
0.12
0.18
0.25
0.31
0.43
0.49
0.61
0.80
0.98
0.92
0.86
0.74
0.67
0.55
0.37
0.00
0.06
0.12
0.18
0.25
0.31
0.43
0.49
0.61
0.80
0.98
0.92
0.86
0.74
0.67
0.55
0.37
0.00
–0.01
–0.05
–0.11
–0.17
–0.24
–0.30
–0.36
–0.42
–0.48
–0.54
–0.61
–0.73
–0.91
≤0.98
–0.85
–0.79
–0.67
0.06
0.12
0.18
0.25
0.31
0.43
0.49
0.61
0.80
0.98
0.92
0.86
0.74
0.67
0.55
0.37
0.24
Figure2.NDVI,regulatedNDVI,ε4,ε5,ε,andεimagesfrom(a)to(f),respectively.

7518 M. ¸Sahin
high rainfall leads to marked plant diversity. NDVI values are in the range 0.28–0.47 in the
eastern part of the country, but may reach 0.61 in individual locations; these values were
recorded for plateaux on high mountains, and the region is rich in vegetation. The southern
part of the country has an NDVI range of 0.14–0.57, while in regions with irrigated farm-
ing the range is 0.61–0.71. NDVI values in some interior regions are between 0.9 and 0.42.
It will be seen from the NDVI map that Turkey’s neighbour, Syria, has poor plant cover.
The emissivity maps of the fourth and fifth channels of the NOAA-AVHRR sensor were
obtained by using the final form of the NDVI image in Equations (3) and (4), respectively
(see Figures 2(c) and (d)). When Figures 2(c) and (d) are examined, it will be seen that ε4 is
between 0.83 and 0.97 while ε5 is between 0.93 and 0.97. The emissivity values in thermal
channels of the same image from different wavelengths have different values. It will be
seen that the channel 5 emissivity value of AVHRR is higher than that of channel 4.
The emissivity images for the fourth and fifth channels of NOAA-AVHRR were used
in Equations (5) and (6) to obtain emissivity difference ( ε) and average of emissivity (ε)
(see Figures 2(e) and (f )). When Figures 2(e) and (f ) are examined, it will be understood
that ε is mostly between −0.17 and −0.04 while ε is between 0.86 and 0.97.
In addition, brightness temperatures of the fourth and fifth channels were created by
Idrisi Andes and Envi 4.3 image processing software. Thereafter, brightness temperature,
ε, and ε images were employed in Equations (7)–(9) to get LST maps according to the
Becker–Li (1990) algorithm (see Figure 3).
When the map of Turkey is examined, it will be understood that the vast majority of LST
values vary between 289 K and 296 K. LST values in the northern part are between 286 K
and 296 K; in the eastern and northeastern parts are between 282 K and 287 K; and in the
western part are between 291 K and 296 K. Although effective LST values vary between
291 K and 298 K in the southern part, it will be observed that some values are between
282 K and 287 K. The air temperature change range is 298–305 K in neighbouring Iraq
and Syria.
LST is not achieved in points where there are seas, lakes, and rivers because these points
have water. This is an expected result because the emissivity values used in the algorithm
are obtained from NDVI values. It is understood that there are not any plants sufficiently in
the points, especially in seas, lakes, and rivers. This problem may seem like a lack in the
method; as operating points are based on pixels that are not selected over lakes, rivers, and
seas, it does not constitute an obstacle.
<273.41–275.68
275.69–277.96
277.97–280.23
289.36–291.63
300.75–303.02
307.59–309.86
305.31–307.58
303.03–305.30
298.47–300.74
296.19–298.46
293.92–296.18
291.64–293.91
287.08–289.35
284.8–287.07
282.52–284.79
280.24–282.51
Figure 3. Land-surface temperature map (in kelvin) for 20 May 2002, at 06:44 local time.

A total of 603 LST images were employed in the study (see Table 2), and the images are
real-time data. Making use of these images, 42 monthly mean LST images were exposed
in 2002–2004 using the same method of calculation. Furthermore, the 2196 LST values
from 61 locations were achieved via satellite data over a period of three years. These values
were compared statistically with meteorological values, using Equations (15)–(18); with
the R2
having a value of 0.970, RMSE, MBE, and WI were 1.790 K, 0.08 K, and 0.991,
respectively (see Figure 4). Recent studies have employed detrended fluctuation analysis
(DFA) in statistical comparisons. One of these studies is ‘new features of land and sea
surface temperature anomalies’, in which Efstathiou et al. (2011) statistically analysed
global mean land and sea surface temperature (LSST) anomalies with DFA, for the period
January 1850 to August 2008, for both hemispheres. These workers carried out a correla-
tion between LSST statistics, proposing that the results of DFA in LSST time series can
enhance the reliability of climate dynamics modelling. In addition, scientists have esti-
mated LST with satellite data derived from various regions of the world. The RMSE values
of all studies researched appear to vary within an error range of 1–3 K (Vidal 1991; Coll,
Sobrino, and Valor 1994; Ouaidrari et al. 2002; Katsiabani, Adaktilou, and Cartalis 2009;
¸Sahin and Kandirmaz 2010). The results of the present study are in accord with the above,
in that RMSE was found to be 1.790 K.
The R2
, RMSE, MBE, and WI values were calculated for selected locations as control
points (see Table 3).
When Table 3 is analysed, the lowest RMSE value is found for the province of A˘grı
(1.356 K) and the maximum for Çanakkale (2.187 K). RMSE in other locations ranges
from 1.356 K to 2.187 K. The highest R2
was obtained for Balıkesir-Gönen (0.989), with
Karaman the lowest (0.942). It is irrelevant whether MBE is positive or negative, providing
it is close to zero. According to this rule, the best and worst MBE values were found to be
−0.014 K and −1.156 K for the provinces of Burdur and Balıkesir-Gönen, respectively. The
lowest WI was recorded for Çanakkale and Karaman (0.983) and the highest for Antalya
(0.996).
4.2. Solar radiation
Although month, latitude, longitude, and LST are very important as input parameters in
acquiring SR values, altitude is also very important. Furthermore, it has been verified using
ANN methods that the altitude of any point in the sky has an influence on SR values.
Alexandris et al. (1999) studied measurements of solar biological effective ultraviolet (UV)
radiation over the period 7–14 June 1997 using an aircraft-based radiometer, at several
different altitudes from sea level up to 13 km. The results showed that an increase in bio-
logical effective UV radiation of about 7% per kilometre occurs throughout the troposphere.
This increase has been compared with the burden ozone content at each height level as it
is derived from concurrent ozone measurements obtained from ozonesonde ascents. This
comparison showed a strong anti-correlation between biological effective UV radiation and
total ozone content above the UV measurement height level. Moreover, it was reported that
global total ozone dynamic’s surface solar ultraviolet radiation has an impact on variabil-
ity and ecosystems. Kondratyev and Varotsos (1996) studied global total ozone changes
and biologically active surface solar ultraviolet radiation variation on the basis of satel-
lite and conventional surface observations. In that study, relevant impacts on terrestrial
and aquatic ecosystems, and biochemical cycles, were discussed. There is a possibility of
remote-sensing techniques being used to obtain atmospheric concentrations of various trace
gases.

7520 M. ¸Sahin
Table 2. Dates of images used in the study.
2002
2 January 2002 21 March 2002 12 June 2002 15 September 2002 15 December 2002
11 January 2002 1 April 2002 22 June 2002 25 September 2002 23 December 2002
24 January 2002 5 April 2002 28 June 2002 1 October 2002
25 January 2002 7 April 2002 30 June 2002 2 October 2002
26 January 2002 8 April 2002 1 July 2002 3 October 2002
1 February 2002 15 April 2002 9 July 2002 10 October 2002
12 February 2002 1 May 2002 28 July 2002 31 October 2002
14 February 2002 2 May 2002 29 July 2002 2 November 2002
15 February 2002 5 May 2002 2 August 2002 3 November 2002
1 March 2002 20 May 2002 22 August 2002 24 November 2002
7 March 2002 28 May 2002 1 September 2002 2 December 2002
10 March 2002 3 June 2002 3 September 2002 6 December 2002
2003
1 January 2003 7 April 2003 1 July 2003 24 September 2003 28 December 2003
4 January 2003 9 April 2003 3 July 2003 26 September 2003 31 December 2003
5 January 2003 10 April 2003 4 July 2003 27 September 2003
(Continued)

30 January 2003 2 May 2003 27 July 2003 19 October 2003
2 February 2003 3 May 2003 28 July 2003 21 October 2003
4 February 2003 7 May 2003 1 August 2003 22 October 2003
3 March 2003 3 June 2003 21 August 2003 24 November 2003
11 March 2003 8 June 2003 29 August 2003 3 December 2003
1 April 2003 27 June 2003 20 September 2003 23 December 2003
2004
7 January 2004 27 March 2004 2 July 2004 29 September 2004 30 December 2004
8 January 2004 28 March 2004 3 July 2004 30 September 2004
9 January 2004 30 March 2004 5 July 2004 3 October 2004
(Continued)

7522 M. ¸Sahin
5 February 2004 26 April 2004 1 August 2004 30 October 2004
6 February 2004 29 April 2004 4 August 2004 3 November 2004
7 February 2004 30 April 2004 5 August 2004 5 November 2004
1 March 2004 28 May 2004 29 August 2004 1 December 2004
Katsambas et al. (1997) showed that daily total ozone observations made by satellite
between 1985 and 1993 have been used to investigate ﬂuctuations in daily broadband and
spectral solar ultraviolet radiation reaching the ground. That study, carried out in summer
over Athens (Greece), showed increases in ultraviolet irradiance reaching the ground of
0.54%, 0.98%, 2.60%, and 0.79% per decade for the month of July at 300 nm, 312 nm,
320 nm, and UVB (280–320 nm), respectively. Similar results were also obtained by Feretis
et al. (2002).
In the present study, ANN and ELM were used to acquire solar radiation values. For
this purpose, the data belonging to 61 centres of localization are chosen as control points
in the period 2002–2004 in Turkey. While the data for the years 2002 and 2003 are used
for training ANN and ELM models, the constructed models have been tested for accuracy
with the data of 2004. The ANN model used in this study consists of the input layer, hidden
layer, and output layer. While month, altitude, latitude, longitude, and LST derived from
satellite data are used as input layer, solar radiation values are obtained from the output
layer (see Figure 5).

270
260
250
250 260
R2
= 0.970
y = 0.9998x
270 280 290 300 310 320 330
Meteorological values (K)
Satellitevalues(K)
280
290
300
310
320
330
Figure 4. Comparison of satellite and meteorological LST values for coefﬁcient of determination
(R2
= 0.970).
There is no mathematical formula to determine optimum nerve cell (neuron) number
in the hidden layer of the ANN model, the number being decided during training of the
network. Neuron numbers increased from 2 to 50 according to the rule of two-by-two in
the hidden layer to achieve the most appropriate ANN model.
However, the creation of ANN initial weights was random and the appropriate ANN
model was determined as a result of trial and error. In addition, different training algorithms
were tested during the training of the network. The best models developed according to
training algorithms and transfer functions, and number of the neurons in the hidden layer
are shown in Table 4.
According to Table 4, the lowest and highest values of R2
will be seen to be 0.846 and
0.943, respectively. The ANN model with the highest value of R2
was that with the train-
ing algorithm trainlm. Its transfer functions in the hidden and output layers were recorded
as logsig and linear, respectively. This network is being developed using the 20 neurons
in the hidden layer. This model has the lowest R2
, showing that the training algorithm,
transfer function in the hidden layer, and transfer function in the output layer are trainscg,
logsig, and linear, respectively. There are 48 neurons in the hidden layer of the network,
and the highest and lowest values for RMSE were found to be 2.458 and 1.604 MJ m−2
,
respectively. The model which is effective to try development of R2
is identical to RMSE
statistics. MBE values were also calculated in the developed models. The best and worst
MBE values were calculated as 0.013 and −0.310 MJ m−2
. The model with the best MBE
had a training algorithm and transfer functions in the hidden and ouput layers as trainoss,
logsig, and linear, respectively. There are 16 neurons in the hidden layer. The training algo-
rithm with the worst MBE was trainlm. The transfer function in the hidden layer is logsig,
while the transfer function in the output layer is linear. There are 36 neurons in the hidden
layer, with the highest WI 0.985 and the lowest 0.961. The ANN model that achieved the
highest WI had the training algorithm and hidden and output functions as trainlm, logsig,

7524 M. ¸Sahin
Table3.MBE,RMSE,R2
,andWIvaluesbystudylocation.
ProvinceMBE(K)RMSE(K)R2
WIProvinceMBE(K)RMSE(K)R2
WI
Adana0.0231.5480.9730.993˙Izmir−0.4381.4580.9780.994
Adıyaman0.7891.8780.9740.992Kahramanmara¸s0.0311.9910.9590.989
A˘grı0.0501.3560.9760.994Karaman0.3852.0620.9420.983
Aksaray0.0541.4930.9710.992Kars0.3311.7200.9690.992
Amasya0.1641.4980.9730.993Kastamonu0.3231.6530.9620.989
Ankara−0.8411.5780.9880.996Kayseri−0.0941.7990.9730.993
Antakya−0.2671.9020.9640.990Kır¸sehir−0.4291.9790.9510.986
Antalya0.3711.6580.9850.996Kilis0.8561.9120.9630.988
Artvin0.1421.5510.9680.991Kocaeli-˙Izmit−0.8231.9040.9570.985
Aydın0.1061.5170.9710.995Konya−0.0161.4810.9760.994
Balıkesir-Gönen−1.1561.6310.9890.995Kütahya0.1941.8270.9430.985
Batman0.1861.7590.9650.990Malatya−0.4891.7300.9710.992
Bilecik−0.7871.8750.9590.987Mersin0.9362.0950.9620.987
Bingöl1.1282.0660.9730.990Mu˘gla0.5642.0100.9570.987
Bitlis−0.4662.0180.9570.988Mu¸s−0.4441.7440.9750.993
Burdur−0.0141.7710.9520.987Ni˘gde0.0221.8200.9670.991
Bursa0.8001.9790.9580.987Ordu−0.4641.8170.9670.990
Çanakkale0.1472.1870.9440.983Rize−0.4861.7120.9490.986
Çorum1.1361.7640.9690.987Samsun−0.2411.6260.9680.991
Diyarbakır0.0902.0970.9580.989Siirt0.6211.8210.9740.992
Denizli0.0961.9410.9670.991Sinop0.2681.5940.9660.991
Edirne0.3001.8330.9670.991Sivas0.3891.4340.9760.993
Elâzı˘g−0.5671.9820.9570.988¸Sanlıurfa−0.1231.4410.9820.995
Erzincan−0.0781.8550.9620.989Tokat0.1921.6580.9720.992
Erzurum0.6911.7990.9740.992Trabzon−0.3831.7260.9570.988
Gaziantep0.4342.0370.9650.990Tunceli0.4611.8600.9640.990
Gümü¸shane0.5091.8790.9620.989Van−0.4481.5650.9810.995
Hakkâri0.2422.0630.9650.991Yalova0.2081.5260.9700.992
I˘gdır0.6531.9070.9720.992Yozgat−0.0791.9570.9500.985
Isparta−0.1971.8560.9620.989Zonguldak0.1361.5190.9660.991
˙Istanbul-Göztepe0.1751.9530.9600.988

Altitude
Latitude
Longitude
Month
LST
Solar radiation
Figure 5. ANN and ELM architecture used in this study.
and linear, respectively. The hidden layer had 20 neurons. The training algorithm of ANN
with the lowest WI value is trainscg, while the transfer function of the hidden and output
layers is logsig and linear, respectively. Moreover, it will be seen from Table 4 that there are
48 neurons in the hidden layer of the model.
In this study, although the ANN network was trained with the data for 2002 and 2003,
the success of the network was tested with the data for 2004. The success of a one-year
study is assessed according to RMSE values. The lowest RMSE value, as mentioned previ-
ously, was 1.604 MJ m−2
and this was developed as the most successful model (Table 4).
The result of the tests was that the ANN (5:20:1) structure model gave the most accurate
values in ANN models. The model recorded ﬁve neurons (month, altitude, latitude, lon-
gitude, LST) in the input layer, 20 in the hidden layer, and one in the output layer (solar
radiation). The network training algorithm and transfer function in the hidden layer are
trainlm and logsig, respectively. The linear function is used as transfer function in the out-
put layer. The values of MBE, RMSE, R2
, and WI were calculated depending on location,
and are shown in Table 5.
In this study, whereas the highest RMSE was 3.005 MJ m−2
(the province of Batman),
the lowest was 0.879 MJ m−2
(the province of Kahramanmara¸s).
Moreover, the lowest R2
is found as 0.892, which belonged to province of Isparta.
Malatya has the highest value of R2
as 0.984. And also, the best and worst MBE values
are obtained in order to take 0.009 MJ m−2
and 2.468 MJ m−2
for provinces of Artvin
and Batman, respectively. Other locations take the R2
values between 0.892 and 0.984, and
RMSE values between 0.879 MJ m−2
and 3.005 MJ m−2
. The approach to zero of MBE
values vary between 0.009–2.468 MJ m−2
. In addition, the lowest WI value is obtained for
Batman province as 0938. The highest value of WI is 0.996. This value is calculated for the
province of Kahramanmara¸s.
It is clear from the statistical results of this study that SR as estimated by the ANN
method gave the optimum value for the province of Kahramanmara¸s and the worst for
Batman. The estimated and the actual SR data for these two provinces are given as total
daily SR per month (see Figures 6(a)–(b)).

7526 M. ¸Sahin
Table 4. ANN models developed.
Training
algorithm
Hidden
transfer
function
Output
transfer
function
Number of
neurons in
hidden layer
MBE
(MJ m−2
)
RMSE
(MJ m−2
) R2
WI
trainlm tansig linear 30 −0.241 1.765 0.929 0.981
trainlm logsig linear 16 −0.197 1.690 0.936 0.983
trainlm logsig tansig 14 −0.263 1.634 0.942 0.984
trainlm logsig tansig 50 −0.257 1.744 0.935 0.982
trainlm tansig tansig 26 −0.126 1.761 0.927 0.981
trainlm tansig logsig 18 −0.173 1.609 0.941 0.984
trainlm tansig logsig 44 −0.212 1.624 0.942 0.984
trainlm logsig logsig 20 −0.236 1.769 0.931 0.981
trainscg logsig linear 44 −0.224 2.149 0.873 0.969
trainscg logsig linear 48 −0.099 2.458 0.846 0.961
trainscg tansig linear 28 −0.127 1.988 0.904 0.975
trainscg tansig linear 24 −0.141 2.000 0.903 0.975
trainscg tansig tansig 4 −0.202 2.063 0.898 0.974
trainscg tansig logsig 26 −0.187 1.930 0.910 0.977
trainscg logsig tansig 25 −0.147 2.043 0.897 0.974
trainoss logsig linear 16 0.013 2.172 0.883 0.970
trainoss tansig tansig 16 −0.179 2.108 0.892 0.972
trainoss logsig logsig 46 −0.138 2.021 0.902 0.975
trainbfg logsig linear 26 −0.176 1.959 0.908 0.976
trainbfg logsig linear 44 −0.240 1.935 0.912 0.977
trainbfg tansig linear 32 −0.229 1.978 0.909 0.976
trainbfg tansig tansig 16 −0.092 1.956 0.906 0.976
trainbfg tansig tansig 46 −0.054 1.925 0.910 0.977
trainbfg tansig logsig 36 −0.071 1.972 0.902 0.975
trainbfg logsig logsig 42 −0.227 1.962 0.907 0.976
trainbfg logsig tansig 36 −0.039 1.974 0.903 0.975
Differences between estimated and actual values according to the ANN method
ranged between 0.010 and 1.525 MJ m−2
and between 0.012 and 5.490 MJ m−2
for
Kahramanmara¸s and Batman, respectively. The ELM method was applied to the same data
set to evaluate SR.
Month, altitude, latitude, longitude, and LST derived from satellite data were used as
input in the input layer by the ELM method, and SR was obtained as output from the output
layer (see Figure 5). There are five neurons in the input layer and one in the output layer.
The best model was generated to establish the most appropriate model by increasing the
neurons five by five from 10 to 150 in the hidden layer. The tansig, sinus, sigmoid, radial
basis, and probate transfer functions were used in the hidden layer, while the linear transfer
function was selected in the output layer (see Table 6).

Table 5. MBE, RMSE, R2
, and WI values by province.
Province MBE (MJ m−2
) RMSE (MJ m−2
) R2
WI
Adana −1.133 1.591 0.972 0.981
Adıyaman −0.513 1.541 0.966 0.982
A˘grı −0.439 1.472 0.949 0.983
Aksaray 0.328 1.294 0.961 0.989
Amasya 0.174 1.438 0.952 0.987
Ankara −0.607 1.229 0.965 0.990
Antakya 0.337 1.037 0.969 0.991
Antalya −1.731 2.512 0.953 0.966
Artvin 0.009 1.561 0.939 0.983
Aydın −2.189 2.683 0.964 0.962
Balıkesir-Gönen −0.340 1.765 0.936 0.979
Batman 2.468 3.005 0.945 0.938
Bilecik −0.715 1.636 0.953 0.979
Bingöl 1.655 2.335 0.946 0.971
Bitlis −1.579 1.858 0.974 0.980
Burdur 0.919 1.710 0.952 0.982
Bursa 0.101 2.008 0.923 0.975
Çanakkale −0.967 1.578 0.957 0.985
Çorum −0.103 1.175 0.961 0.990
Diyarbakır −0.347 1.224 0.964 0.991
Denizli −0.478 1.271 0.948 0.986
Edirne −1.447 1.939 0.967 0.977
Elâzı˘g −0.523 1.048 0.981 0.994
Erzincan 0.277 1.248 0.962 0.989
Erzurum −1.577 1.948 0.959 0.967
Gaziantep −0.146 1.221 0.963 0.989
Gümü¸shane −1.399 1.928 0.931 0.980
Hakkâri 0.293 1.042 0.969 0.991
I˘gdır −0.421 1.320 0.959 0.988
Isparta 0.873 2.226 0.892 0.966
˙Istanbul-Göztepe 1.172 1.711 0.959 0.982
˙Izmir −0.689 1.473 0.963 0.990
Kahramanmara¸s 0.198 0.879 0.981 0.996
Karaman −0.211 1.055 0.974 0.994
Kars −0.238 1.347 0.954 0.983
Kastamonu −0.097 1.810 0.908 0.969
Kayseri −0.343 1.104 0.977 0.992
Kır¸sehir 0.773 1.519 0.953 0.984
Kilis 0.443 1.185 0.970 0.991
Kocaeli-˙Izmit −0.875 1.767 0.955 0.976
Konya 0.094 1.391 0.957 0.988
Kütahya −0.523 1.588 0.945 0.985
Malatya 0.600 0.994 0.984 0.994
Mersin −0.054 0.991 0.978 0.994
Mu˘gla −0.383 1.370 0.955 0.986
Mu¸s −0.812 1.567 0.957 0.987
Ni˘gde −0.765 1.206 0.981 0.992
Ordu −0.024 1.363 0.939 0.985
Rize −0.149 1.358 0.913 0.977
Samsun 0.587 1.804 0.928 0.979
Siirt −1.261 2.301 0.951 0.968
Sinop 2.196 2.567 0.960 0.966
Sivas −0.289 1.225 0.971 0.990
(Continued)

7528 M. ¸Sahin
Province MBE (MJ m−2
) RMSE (MJ m−2
) R2
WI
¸Sanlıurfa 0.541 1.090 0.979 0.993
Tokat −0.288 1.613 0.946 0.984
Trabzon −0.132 1.263 0.921 0.981
Tunceli 0.045 1.298 0.967 0.991
Van 0.159 1.183 0.971 0.993
Yalova −0.416 1.193 0.968 0.992
Yozgat 0.298 1.541 0.943 0.984
Zonguldak 0.620 1.351 0.971 0.989
The most successful ELM model had the structure (5:150:1), with 150 neurons in
the hidden layer. The transfer function model is tansig in the hidden layer, and the
transfer function in the output layer is linear. If the results are evaluated statistically
according to the criteria, R2
, RMSE, MBE, and WI values are calculated as 0.961 and
0.672 MJ m−2
, 0.045 MJ m−2
, and 0.997, respectively. At the same time, these values were
obtained depending on location by taking into consideration the ELM (5:150:1) model
(see Table 7).
It is clear from Table 7 that the lowest R2
was recorded for Aksaray (0.940), with the
highest 0.993 for the province of Isparta. Tunceli recorded the lowest MBE and RMSE
(0.012 and 0.347 MJ m−2
, respectively), while the highest (0.158 and 1.257 MJ m−2
,
(c)
(d)
(b)(a)
30
25
20
15
Month
Month
Month
Month
Solarradiation(MJm–2
day–1
)Solarradiation(MJm–2
day–1
)
day–1
)
day–1
)
10
5
0
January
February
M
arch
April
M
ay
June
July
August
Septem
ber
O
ctober
Novem
ber
Decem
ber
January
February
M
arch
April
M
ay
June
July
August
Septem
ber
O
ctober
Novem
ber
Decem
ber
January
February
M
arch
April
M
ay
June
July
August
Septem
ber
O
ctober
Novem
ber
Decem
ber
January
February
M
arch
April
M
ay
June
July
August
Septem
ber
O
ctober
Novem
ber
Decem
ber
30
25
20
15
10
5
0
30
25
20
15
10
5
0
30 Actual value
ANN value
Actual value
ANN value
Actual value
ELM value Actual value
ELM value
25
20
15
10
5
0
Figure 6. Estimated and actual SR values for Kahramanmara¸s (a), Batman (b), Tunceli (c), and
Aksaray (d).

Table 6. ELM training and testing parameters.
Number of layers 3
Number of neurons in layers Input: 5
Hidden: 10 . . . 150
Output: 1
Activation functions Tangent sigmoid; sinus; sigmoid; radial basis; probit; purelin
Learning rule The ELM for SLFNs
Sum-squared error 0.0001
Note: ELM, extreme learning machine; SLFN, single-hidden layer feedforward neural network.
respectively) were recorded at Aksaray. R2
varied between 0.940 and 0.993 for most
locations, with MBE between 0.012 and 0.158 MJ m−2
. RMSE was in the range
0.347–1.257 MJ m−2
. The lowest WI was for Gaziantep (0.989) and the highest (for
more than one city) was 0.999 (Table 7). The names of these provinces were Bingöl,
Çanakkale, Diyarbakır, ˙Istanbul-Göztepe, ˙Izmir, Karaman, Kilis, Koceli-˙Izmit Kütahya,
Malatya, Ni˘gde, Samsun, Tokat, and Tunceli. The best estimation was for Tunceli, and
the worst for Aksaray. Using the ELM method, actual and estimated daily total SR monthly
data for these two locations are shown in Figures 6(c)–(d).
The actual and estimated data obtained by the ELM method are compatible with
each other for Tunceli, but not for Aksaray. According to monthly data, while differences
between actual and estimated data ranged between 0.046 and 0.831 MJ m−2
in Tunceli, in
Aksaray these ranged from 0.102 to 1.608 MJ m−2
.
5. Conclusion
In this study, SR was estimated using both ELM and ANN in 61 locations with varying
climatic conditions. Both methods were trained with data from 2002 and 2003, while
model accuracy was tested with data from 2004. Solar radiation values obtained from
the use of ANN and ELM models were compared statistically with the values of SR as
measured by meteorological stations. The (5:20:1) model proved to be the most success-
ful ANN model, calculating SR with statistical values of R2
, MBE, RMSE, and WI as
0.943, −0.148 MJ m−2
, 1.604 MJ m−2
, and 0.996, respectively. In addition, this model
has a training algorithm that is trainlm, with transfer functions in the hidden and output
layers being logsig and linear, respectively. There were 20 neurons in the hidden layer.
The (5:150:1) model proved to be the best ELM model, with 150 neurons in the hid-
den layer. The transfer functions of the ELM (5:150:1) model in the hidden and output
layers are tansig and linear, respectively. R2
, MBE, RMSE, and WI were calculated as
0.961, 0.045 MJ m−2
, 0.672 MJ m−2
, and 0.997, respectively. Use of RMSE is a general
precept rather than that of MBE, especially in short-term (e.g. 1 year) comparison. Since
the RMSE of the ELM method (0.672 MJ m−2
) was lower than that of the ANN method
(1.604 MJ m−2
), these results show that ELM was more successful than ANN. The EML
method to obtain result signiﬁcant statistically in SR calculation is more successful than
ANN method, is an innovation in terms of literature.
Estimation of SR cannot be achieved with only a very small error using the ELM
method and depending on satellite data. Construction of a suitable network of meteoro-
logical stations throughout any country and the recording of permanent measurements are
very difﬁcult and burdensome economically. Moreover, even if these were to be established,

7530 M. ¸Sahin
Table7.Errorvaluesbylocation.
Province
MBE
(MJm−2
)
RMSE
(MJm−2
)R2
WIProvince
MBE
(MJm−2
)
RMSE
(MJm−2
)R2
WI
Adana−0.0250.4970.9710.998˙Izmir−0.0180.4260.9850.999
Adıyaman−0.0610.7830.9890.994Kahramanmara¸s0.0390.6210.9640.998
A˘grı−0.0360.6030.9670.997Karaman−0.0340.5840.9750.999
Aksaray0.1581.2570.9400.994Kars−0.0350.5880.9800.996
Amasya−0.0280.5290.9820.998Kastamonu−0.0460.6810.9530.995
Ankara−0.0950.9770.9670.994Kayseri−0.0310.5600.9840.998
Antakya−0.0450.6680.9650.996Kır¸sehir0.1221.1030.9600.994
Antalya−0.1191.0920.9520.994Kilis0.0140.3770.9820.999
Artvin−0.0250.5000.9710.998Kocaeli-˙Izmit−0.0170.4100.9900.999
Aydın−0.0290.5410.9790.998Konya0.0500.7100.9860.998
Balıkesir-Gönen−0.0610.7830.9650.996Kütahya−0.0260.5070.9800.999
Batman0.0300.5480.9640.998Malatya0.0200.4530.9860.999
Bilecik−0.0560.7500.9640.995Mersin−0.0350.5880.9820.998
Bingöl0.0240.4860.9760.999Mu˘gla−0.0650.8060.9470.995
Bitlis−0.0310.5600.9690.998Mu¸s−0.0760.8720.9690.996
Burdur0.0350.5940.9850.998Ni˘gde−0.0200.4520.9880.999
Bursa0.0450.6720.9740.997Ordu−0.0310.5530.9670.998
Çanakkale−0.0230.4810.9790.999Rize−0.0420.6480.9640.996
Çorum−0.0340.5840.9870.998Samsun0.0200.4470.9800.999
Diyarbakır−0.0180.4250.9920.999Siirt−0.0350.5950.9770.997
Denizli−0.0550.7380.9620.996Sinop0.0830.9110.9620.996
Edirne−0.0390.6220.9730.998Sivas−0.0660.8100.9760.996
Elâzı˘g0.0530.7270.9650.997¸Sanlıurfa0.0390.6270.9820.998
Erzincan0.0260.5100.9540.998Tokat−0.0170.4090.9810.999
Erzurum−0.0370.6100.9680.996Trabzon0.0500.7090.9660.995
Gaziantep−0.1481.2160.9510.989Tunceli0.0120.3470.9880.999
Gümü¸shane−0.0630.7910.9740.997Van−0.0550.7390.9600.998
Hakkâri0.0470.6840.9800.997Yalova−0.0350.5880.9610.998
I˘gdır−0.0350.5940.9740.997Yozgat−0.0370.6110.9720.998
Isparta−0.0440.6650.9930.998Zonguldak0.0510.7170.9870.997
˙Istanbul-
Göztepe
0.0150.3930.9810.999

the distribution of the stations might not adequate. Rather, it would be more appropriate to
utilize meteorological satellites which are capable of scanning all regions. For this reason,
SR obtained from satellite data using the ELM method is recommended for researchers
studying SR.
Acknowledgements
I would like to express my gratitude to the Republic of Turkey’s Ministry of Forestry and Water
Affairs (Turkish State Meteorological Service) personnel, who provided a wide range of facilities for
acquiring meteorological data; and to the Scientific and Technological Research Council of Turkey-
Bilten personnel, who provided a wide range of facilities for acquiring satellite data.
References
Alexandris, D., C. Varotsos, K. Y. Kondratyev, and G. Chronopoulos. 1999. “On the Altitude
Dependence of Solar Effective UV.” Physics and Chemistry of the Earth Part C Solar Terrestrial
& Planetary Science 24: 515–517.
Bakirci, K. 2009. “Correlations for Estimation of Daily Global Solar Radiation with Hours of Bright
Sunshine in Turkey.” Energy 34: 485–501.
Bechrakis, D. A., and P. D. Sparis. 2004. “Correlation of Wind Speed Between Neighboring
Measuring Stations.” IEEE Transactions on Energy Conversion 19: 400–406.
Becker, F., and Z. L. Li. 1990. “Toward a Local Split Window Method over Land Surface.”
International Journal of Remote Sensing 11: 369–393.
Benghanem, M., A. Mellit, and S. N. Alamri. 2009. “ANN-Based Modelling and Estimation of
Daily Global Solar Radiation Data: A Case Study.” Energy Conversion and Management 50:
1644–1655.
Bharathi, A., and A. M. Natarajan. 2011. “Cancer Classification Using Modified Extreme Learning
Machine Based on ANOVA Features.” European Journal of Scientific Research 58: 156–165.
Chacko, B. P., V. R. Vimal Krishnan, G. Raju, and P. Babu Anto. 2012. “Handwritten Character
Recognition Using Wavelet Energy and Extreme Learning Machine.” International Journal of
Machine Learning and Cybernetics 3: 149–161.
Chang, N. B., M. Han, W. Yao, L.-C. Chen, and S. Xu. 2010. “Change Detection of Land Use and
Land Cover in a Fast Growing Urban Region with SPOT-5 Images and Partial Lanczos Extreme
Learning Machine.” Journal of Applied Remote Sensing 4: 043551.
Chen, X., Z. Y. Dong, K. Meng, Y. Xu, K. P. Wong, and H. W. Ngan. 2012. “Electricity Price
Forecasting with Extreme Learning Machine and Bootstrapping.” IEEE Transactions on Power
Systems 27: 2055–2062.
Cheng, G. J., L. Cai, and H. X. Pan. 2009. “Comparison of Extreme Learning Machine with
Support Vector Regression for Reservoir Permeability Prediction.” Computational Intelligence
and Security 2: 173–176.
Cihlar, J., H. Ly, Z. Li, J. Chen, H. Pokrant, and F. Hung. 1997. “Multi-Temporal, Multichannel
AVHRR Data Sets for Land Biosphere Studies – Artifacts and Corrections.” Remote Sensing of
Environment 60: 35–57.
Coll, C., J. A. Sobrino, and E. Valor. 1994. “On the Atmospheric Dependence of the Split-Window
Equation for Land Surface Temperature.” International Journal of Remote Sensing 15: 105–122.
Cracknell, A. P., and C. A. Varotsos. 2007. “Fifty Years after the First Artificial Satellite: From
Sputnik 1 to ENVISAT.” International Journal of Remote Sensing 28: 2071–2072.
Efstathiou, M. N., C. Tzanis, A. Cracknell, and C. A. Varotsos. 2011. “New Features of the Land and
Sea Surface Temperature Anomalies.” International Journal of Remote Sensing 32: 3231–3238.
Erdinç, A. 2005. “Stock Market Forecasting: Artificial Neural Network and Linear Regression
Comparison in an Emerging Market.” Journal of Financial Management and Analysis 18: 18–33.
Feng, G., G.-B. Huang, Q. Lin, and R. Gay. 2009. “Error Minimized Extreme Learning Machine with
Growth of Hidden Nodes and Incremental Learning.” IEEE Transactions on Neural Networks 20:
1352–1357.
Feretis, E., P. Theodorakopoulos, C. Varotsos, M. Efstathiou, C. Tzanis, T. Xirou, N. Alexandridou,
and M. Aggelou. 2002. “On the Plausible Association between Environmental Conditions and
Human Eye Damage.” Environmental Science and Pollution Research 9: 163–165.

7532 M. ¸Sahin
Han, F., H. F. Yao, and Q. H. Ling. 2012. “An Improved Extreme Learning Machine Based on Particle
Swarm Optimization.” Bio-Inspired Computing and Applications 6840: 699–704.
Huang, G.-B., D. H. Wang, and Y. Lan. 2011. “Extreme Learning Machines: A Survey.” International
Journal of Machine Learning and Cybernetics 2: 107–122.
Huang, G.-B., Q.-Y. Zhu, and C.-K. Siew. 2004. “Extreme Learning Machine: A New Learning
Scheme of Feedforward Neural Networks.” IEEE International Joint Conference on Neural
Networks 2: 985–990.
Huang, G.-B., Q. Y. Zhu, and C. K. Siew. 2006. “Extreme Learning Machine: Theory and
Applications.” Neurocomputing 70: 489–501.
Janjai, S., P. Pankaewa, J. Laksanaboonsong, and P. Kitichantaropas. 2011. “Estimation of Solar
Radiation over Cambodia from Long-Term Satellite Data.” Renewable Energy 36: 1214–1220.
Kalogirou, S. A. 2000. “Applications of Artificial Neural-Networks for Energy Systems.” Applied
Energy 67: 17–35.
Karem, C., B. M. J. Q. Taha, H. Stuart, G. M. Hosni, and G. Hugo. 2008. “Comparison of Ice-Affected
Stream Flow Estimates Computed Using Artificial Neural Networks and Multiple Regression
Techniques.” Journal of Hydrology 349: 383–396.
Katiyar, K., A. Kumar, C. K. Pandey, and B. Das. 2010. “A Comparative Study of Monthly Mean
Daily Clear Sky Radiation Over India.” International Journal of Energy and Environment 1:
177–182.
Katsambas, A., C. A. Varotsos, G. Veziryianni, and C. Antoniou. 1997. “Surface Solar Ultraviolet
Radiation: A Theoretical Approach of the SUVR Reaching the Ground in Athens, Greece.”
Environmental Science & Pollution Research 4: 69–73.
Katsiabani, K., N. Adaktilou, and C. Cartalis. 2009. “A Generalised Methodology for Estimating
Land Surface Temperature for Non-Urban Areas of Greece Through the Combined Use of
NOAA–AVHRR Data and Ancillary Information.” Advances in Space Research 43: 930–940.
Koca, A., H. F. Oztop, Y. Varol, and G. O. Koca. 2011. “Estimation of Solar Radiation Using
Artificial Neural Networks with Different Input Parameters for Mediterranean Region of Anatolia
in Turkey.” Expert Systems with Applications 38: 8756–8762.
Kondratyev, K. Y., and C. A. Varotsos. 1996. “Global Total Ozone Dynamics – Impact on Surface
Solar Ultraviolet Radiation Variability and Ecosystems.” Environmental Science and Pollution
Research 3: 205–209.
Kwak, C., and O.-W. Kwon. 2008. “Cardiac Disorder Classification Based on Extreme Learning
Machine.” World Academy of Science, Engineering and Technology 48: 435–438.
Liang, N. Y., G.-B. Huang, P. Saratchandran, and N. Sundararajan. 2006. “A Fast and Accurate
Online Sequential Learning Algorithm for Feedforward Networks.” IEEE Transactions on Neural
Networks 17: 1411–1423.
Lu, N., J. Qin, K. Yang, and J. Sun. 2011. “A Simple and Efficient Algorithm to Estimate Daily Global
Solar Radiation from Geostationary Satellite Data.” Energy 36: 3179–3188.
Mcmillin, L. M. 1975. “Estimation of Sea Surface Temperatures from Two Infrared Window
Measurements with Different Absorption.” Journal of Geophysical Research 36: 5113–5117.
Myneni, R. B., F. G. Hall, P. J. Sellers, and A. L. Marshak. 1995. “The Interpretation of Spectral
Vegetation Indexes.” IEEE Transactions on Geoscience and Remote Sensing 33: 481–486.
Ouaidrari, H., S. N. Gowarda, K. P. Czajkowskib, J. A. Sobrinoc, and E. Vermotea. 2002. “Land
Surface Temperature Estimation from AVHRR Thermal Infrared Measurements: An Assessment
for the AVHRR Land Pathfinder II Data Set.” Remote Sensing of Environment 81: 114–128.
Ozgoren, M., M. Bilgili, and B. Sahin. 2012. “Estimation of Global Solar Radiation Using ANN over
Turkey.” Expert Systems with Applications 39: 5043–5051.
Pal, M. 2009. “Extreme-Learning-Machine-Based Land Cover Classification.” International Journal
of Remote Sensing-Letter 30: 3835–3841.
Polo, J., L. F. Zarzalejo, M. Cony, A. A. Navarro, R. Marchante, L. Martin, and M. Romero. 2011.
“Solar Radiation Estimations over India Using Meteosat Satellite Images.” Solar Energy 85:
2395–2406.
Prabhakara, C., G. Dalu, and V. G. Kunde. 1974. “Estimation of Sea Temperature from Remote
Sensing in the 11 to 13 µm Window Region.” Journal of Geophysical Research 79: 5039–5044.
Prakash, J. S., and R. Rajesh. 2011. “Random Iterative Extreme Learning Machine for Classification
of Electronic Nose Data.” International Journal of Wisdom Based Computing 1: 24–27.
Qin, J., Z. Chen, K. Yang, S. Liang, and W. Tang. 2011. “Estimation of Monthly-Mean Daily Global
Solar Radiation Based on MODIS and TRMM Products.” Applied Energy 88: 2480–2489.

Qu, Y., C. Shang, W. Wu, and Q. Shen. 2011. “Evolutionary Fuzzy Extreme Learning Machine for
Mammographic Risk.” International Journal of Fuzzy Systems 13: 282–291.
Rahimikhoob, A., S. M. R. Behbahani, and M. E. Banihabib. 2013. “Comparative Study of Statistical
and Artificial Neural Network’s Methodologies for Deriving Global Solar Radiation from NOAA
Satellite Images.” International Journal of Climatology 33: 480–486.
Rani, M. P., and G. Arumugam 2010. “Children Abnormal Gait Classification Using Extreme
Learning Machine.” Global Journal of Computer Science and Technology 10: 66–72.
Rasheed, Z., and H. Rangwala. 2012. “Metagenomic Taxonomic Classification Using Extreme
Learning Machines.” Journal of Bioinformatics and Computational Biology 10: 1250015.
Rong, H.-J., Y.-S. Ong, A.-H. Tan, and Z. Zhu. 2008. “A Fast Pruned-Extreme Learning Machine for
Classification Problem.” Neurocomputing 72: 359–366.
¸Sahin, M., and Kandirmaz, H. M. 2010. “Calculation Land Surface Temperature Depending on
Becker and Li–1990 Algorithm.” Journal of Thermal Science and Technology 30: 35–43.
Sellers, P. J. 1985. “Canopy Reflectance, Photosynthesis and Transpiration.” International Journal of
Remote Sensing 6: 1335–1372.
Sousa, S. I. V., F. G. Martins, M. C. M. Alvim-Ferraz, and M. C. Pereira. 2007. “Multiple Linear
Regression and Artificial Neural Networks Based on Principal Components to Predict Ozone
Concentrations.” Environmental Modelling & Software 22: 97–103.
Steel, R. G. D., and J. H. Torrie. 1960. Principles and Procedures of Statistics. New York: McGraw-
Hill.
Sun, Z. L., T. M. Choi, K. F. Au, and Y. Yu. 2008. “Sales Forecasting Using Extreme Learning
Machine with Applications in Fashion Retailing.” Decision Support Systems 46: 411–419.
Suresh, S., S. Saraswathi, and N. Sundararajan. 2010. “Performance Enhancement of Extreme
Learning Machine for Multi-Category Sparse Data Classification Problems.” Engineering
Applications of Artificial Intelligence 23: 1149–1157.
Ulgen, K., and A. Hepbasli. 2009. “Diffuse Solar Radiation Estimation Models for Turkey’s Big
Cities.” Energy Conversion and Management 50: 149–156.
Vazquez, D. P., F. J. O. Reyes, and L. A. Arboledas. 1997. “A Comparative Study of Algorithms for
Estimating Land Surface Temperature from AVHRR Data.” Remote Sensing of Environment 62:
215–222.
Vidal, A. 1991. “Atmospheric and Emissivity Correction of Land Surface Temperature Measured
from Satellite Using Ground Measurements or Satellite Data.” International Journal of Remote
Sensing 12: 2449–2460.
Wang, L., Y. Huang, X. Luo, Z. Wang, and S. Luo 2011. “Image Deblurring with Filters Learned by
Extreme Learning Machine.” Neurocomputing 74: 2464–2474.
Willmott, C. J. 1982. “Some Comments on the Evaluation of Model Performance.” Bulletin of
American Meteorological Society 63: 1309–1313.
Yang, H., W. Xu, J. Zhao, D. Wang, and Z. Dong. 2011. “Predicting the Probability of Ice
Storm Damages to Electricity Transmission Facilities Based on ELM and Copula Function.”
Neurocomputing 74: 2573–2581.
Yeu, C.-W. T., M.-H. Lim, G.-B. Huang, A. Agarwal, and Y. S. Ong. 2006. “A New Machine Learning
Paradigm for Terrain Reconstruction.” IEEE Geoscience and Remote Sensing Letters 3: 382–386.

Comparison of modelling ann and elm to estimate solar radiation over turkey using noaa satellite data

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (18)

Similar to Comparison of modelling ann and elm to estimate solar radiation over turkey using noaa satellite data

Similar to Comparison of modelling ann and elm to estimate solar radiation over turkey using noaa satellite data (20)

More from mehmet şahin

More from mehmet şahin (11)

Recently uploaded

Recently uploaded (20)

Comparison of modelling ann and elm to estimate solar radiation over turkey using noaa satellite data