This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
This study introduces artificial neural networks for the estimation of land surface temperature using meteorological and geographical data in Turkey 26-45 °E and 36- 42 °N. A generalized regression neural network was used in the network. In order to train the neural network, meteorological and geographical data for the period from January 2002 to December 2002 for 10 stations (Adana, Afyon, Ankara, EskiÅŸehir, Istanbul, Izmir, Konya, Malatya, Rize, Sivas) spread over Turkey were used as training (six stations) and testing (four stations) data. Latitude, longitude, elevation, and mean air temperature are used in the input layer of the network. Land surface temperature is the output. However, land surface temperature has been estimated as monthly mean daily sum by using NOAA-AVHRR satellite data in the thermal range over 10 stations in Turkey. After, RMSE between the estimated and ground values for monthly mean daily sum with ANN and Becker and Li method values have been found as 0.0766 °K and 0.0907 °K (training stations), 0.0447 °K and 0.0029 °K (testing stations), respectively.
Land surface temperature (LST) is an important factor controlling most physical, chemical, and biological processes on Earth. Knowledge of land surface temperature is necessary for many environmental studies and management activities of the Earth's resources (Li and Becker 1993). In order to monitor macro-scale spatial changes in surface temperature, scanners designed for sensing in the thermal bands are placed onboard platforms for remote sensing of the Earth's resources from space (Sabins 1997). The extensive application and significant importance of temperature in environmental studies and management is the main force driving the study of LST in remote sensing. With the availability of thermal sensing data, such as channels 4 and 5 of Advanced Very High Resolution Radiometer (AVHRR) data as well as Landsat Thematic Mapper 6 (TM6), the study of LST has become one of the hottest topics in remote sensing during the last two decades (Vogt 1996). Thus, two approaches have been developed to recover land surface temperature from multispectral thermal infrared (TIR) imagery (Schmugge et al. 1998). The first approach utilizes a radiative transfer equation to correct the at-sensor radiance to surface radiance, followed by an emissivity model to separate the surface radiance into temperature and emissivity (Schmugge et al. 1998). The second approach applies the split-window technique for sea surfaces to land surfaces, assuming that the emissivity in the channels used for the split window is similar (Dash et al. 2002). Land surface brightness temperatures are then calculated as a linear combination of the two channels. Today there are so many investigations for using new techniques such as Neural Network for predicting total atmospheric ozone. Artificial Neural Networks (ANN) have been applied to many different environmental sectors, specially in meteorological forecast (Aires et al. 2002; Gardner and Dorling 1998). Govindaraju (2000a, 2000b) reports a number of studies which have used ANNs to forecast rainfall over a short time interval. It is clear from many studies that usage of ANN method is suitable and applicable for estimating global solar radiation especially for regions where very large distances exist between meteorological stations and also having abundant solar energy (Bechrakis et al. 2004; Kalogirou 2001). Nowadays, renewable energy such as wind energy is one of the most attractive sources of energy. Researchers may need to prepare an inventory on the availability of wind energy in an area where there is no measured wind speed data. For this type of situation, it seems useful to predict the wind energy potential using the ANN method (Bilgili et al. 2007). Kalogirou et al. (1999) states that the predicted variations of meteorological parameters such as wind speed, relative humidity, solar radiation, air temperature, etc. are needed in the industry of renewable energy for design purposes, performance analysis and running cost estimation of renewable energy systems. Of special interest to this area is the use of ANNs for forecasting the room(s) air temperature as a function of both forecasted weather parameters (mainly solar radiation and air temperature) and actuator (heating, ventilating, cooling) state or manipulated variables, and the subsequent use of these mid/long-range prediction models for a more efficient temperature control, both in terms of regulation and energy consumption (Ruano et al. 2006).
In this study, the ANN is used to model the LST from near-surface air temperature and geographic information. In addition, the Becker-Li method is used to retrieve the LST from National Oceanic and Atmospheric Administration advanced very high resolution radiometer (NOAA AVHRR) data. The Becker-Li method is presented for the determination of monthly global land surface temperature from the NOAA AVHRR satellite data, which provide wide coverage together with adequate spatial resolution (around 1.1 km at the nadir). The temporal resolution of NOAA-12-14-15 images is at the same location which is scanned by the satellite 5 cloudless day of every month. The satellite images were received in a raw image and zipped data format. They were unzipped and processed with the software `Quorum to Level1B' in order to convert this raw image data to Level1B format, so that remote sensing software can be applied for processing. Radiometric and geometric calibrations were applied first to the images to correct the deficiencies and flaws that could result from the imaging sensor in the platform (satellite). Then, land surface temperature is predicted with Becker and Li Method. Usage of these methods is more suitable for large places. Both methods are applicable in any region for the very large distances between the stations. The objective of the present study is to apply the ANN, Becker and Li methods in collaboration with each other for the prediction of the land surface temperature of the target station using neighboring measuring stations. This will show that these methods can be applied to predict the land surface temperature for any location around sampled measuring stations. The studies which have used ANN, Becker and Li method show a general perspective of land surface temperature in Turkey. For application, ten stations (Adana, Afyon, Ankara, EskiÅŸehir, Istanbul, Izmir, Konya, Malatya, Rize, and Sivas) have been selected from different regions of Turkey. The geographical locations of these land surface temperature stations are shown in Fig. 1. The stations selected can give a general idea about the land surface temperature values in Turkey. The estimation of land surface temperature in Turkey was based on meteorological and geographical data (latitude, longitude, elevation, and mean air temperature). These selected locations in Turkey have different values, as seen in Table 1. Generalized regression neural network (GRNN) is used in the ANN. Meteorological and geographical data are used as input, and land surface temperature is the output.
Fig. 1. Land surface temperature measuring stations in Turkey.
Geographical and meteorological parameters for the stations.
2. Methodology and data sources
2.1. Split-window method
Split-window algorithms for the retrieval of LST from NOAA-AVHRR data have been proposed by different scholars in past decades (Vogt 1996). The algorithm given by Becker and Li (1990b) is worthy of detailed description since it has been cited in many papers (Franca and Cracknell 1994) on the study of LST. The algorithm requires
- brightness temperatures in channels 4 (T4) and 5 (T5)
- the mean emissivity in these channels, .=( . 4+ . 5 )/2=0.975
- the spectral emissivity difference, ..= . 4 -. 5= -0.005 (Caselles et al. 1997; Chrysoulakis et al. 2001).
Becker and Li (1990b) presented a local split-window algorithm for viewing angles of up to 46° from nadir, given as follows:
ker 1990 0 Bec Li
T T T T
T A P M - -
= + + (1)
where A0 , P, and M are coefficients influenced by a number of factors in the process of radiance transmission from the ground to the sensor. For NOAA-AVHRR data, coefficient A0=1.274 and A2=0.15616, A3=-0.482, B1=6.26, B2=3.98, B3=38.33. Later, Li and Becker (1993) modified their algorithm into a general one keeping the form of equation (1). The only difference in the modified algorithm is that the coefficients are determined in terms of water content calculated from a radiance simulation using the LOWTRAN 7 program (Caselles et al. 1997). Based on Becker and Li (1990b), a generalized split-window algorithm has been developed by Wan and Dozier (1996). The form of the algorithm is the same as equation (2), but the coefficients P and M are given as follows:
P A A A
= + + (2)
M B B B
= + + (3)
where A1~A3 and B1~B3 are parameters estimated by the method given by Li and Becker (1993). The difference is that Wan and Dozier (1996) defined A1 in their model as a variable while Becker and Li (1990b) defined it as a constant equal to 1. The mean land surface temperature for 10 stations over Turkey has been determined using the Becker and Li method.
2.2. Artificial neural networks
Artificial neural networks (ANNs) are information processing systems that are non-algorithmic, non digital, and intensely parallel (Dinçer et al. 1996). The use of ANNs for modeling and prediction purposes has become increasingly popular in recent decades (Çam et al. 2005). Researchers have been applying the ANN method successfully in various fields such as mathematics, engineering, medicine, economics, meteorology, psychology, and neurology, as well as in the prediction of mineral exploration sites, in electrical and thermal load predictions, and in adaptive and robotic control. ANNs have been trained to overcome the limitations of conventional approaches to solving complex problems. This method learns from given examples by constructing an input output mapping in order to perform predictions (Kalogirou 2000). . Generalized Regression Neural Network (GRNN) proposed by (Speckt 1991) does not require an iterative training procedure as in back propagation method. It approximates any arbitrary function between input and output vectors, drawing the function estimate directly from the training data. Furthermore, it is consistent; that is, as the training set size becomes large, the estimation error approaches zero, with only mild restrictions on the function. The GRNN is used for estimation of continuous variables, as in standard regression techniques. It is related to the radial basis function network and is based on a standard statistical technique called kernel regression. By de.nition, the regression of a dependent variable y on an independent x estimates the most probable value for y, given x and a training set. The regression method will produce the estimated value of y which minimizes the mean-squared error. GRNN is a method for estimating the joint probability density function (pdf) of x and y, given only a training set. Because the pdf is derived from the data with no preconceptions about its form, the system is perfectly general. If f(x,y) represents the known joint continuous probability density function of a vector random variable, x, and a scalar random variable, y, the conditional mean of y given X (also called the regression of y on X) is given by:
yf y dy
f y dy
E é Cù = ë û
When the density f(x,y) is not known, it must usually be estimated from a sample of observations of x and y. The Probability estimator f(X,Y) is based upon sample values X and Y of the random variables x and y, where n is the number of sample observations and p is the dimension of the vector variable x:
( 1)/( 1)
( ) ( )
, exp exp
2 2 2
i i i n
n s s p s
é ù - C - C C - C é ù - U - U
ê ú C U = ´ ê ú
ê ú ë û ë û
A physical interpretation of the probability estimate f(X, Y) is that it assigns sample probability of width . for each sample Xi and Yi, and the probability estimate is the sum of those sample probabilities (Speckt 1991). De.ning the scalar function Di
( ) ( ) 2 i i
= C - C C - C (6)
and performing the indicated integrations yields the following:
æ ö -
U ç ÷
è ø U C =
æ ö -
å å (7)
The resulting Eq. (7) is directly applicable to problems involving numerical data. When the smoothing parameter is made large, the estimated density is forced to be smooth and in the limit becomes a multivariate Gaussian with covariance .2I. On the other hand, a smaller value of . allows the estimated density to assume non- Gaussian shapes, but with the hazard that wild points may have too great an effect on the estimate (Speckt 1991). The GRNN consists of four layers: input layer, pattern layer, summation layer and output layer. Schematic diagram of a GRNN architecture is presented in Fig. 2. The input units are in the .rst layer. The second layer has the pattern units and and the outputs of this layer are passed on to the summation units in the third layer. The fnal layer covers the output units (Cigizoglu, Alp 2006).
Fig. 2. Schematic diagram of a GRNN architecture.
3. Results and discussions
ANN is used for modeling land surface temperature in Turkey. Mohandes et al. (2004) state that during the training procedure, the weights of the connections between neurons are adjusted in order to achieve the desired input/output relation of the network. This procedure goes on until the difference between the actual output of the network and the desired output is equal with a specified remainder value. Here, the criterion is put forward as the network output which should be closer to the value of desired output. This training procedure has to be repeated for the rest of the input-output pairs existing in the training data. Input variables have been used to validate the ANN. MATLAB software has been used to train the ANN on a personal computer. A GRNN structure GRNN (4, 1.0, 1) corresponds to 4 input nodes, a spread value equal to 1.0 and a single output node. The selected ANN structure is shown in Fig.3. This network consists of input layer, pattern layer, summation layer and output layer. The stations used for training data are at Adana, Afyon, Ankara, EskiÅŸehir, Izmir, Sivas. In test data, stations, Istanbul, Konya, Malatya and Rize are used. In order to train the neural network, meteorological and geographical data measured by the Turkish State Meteorological Service (TSMS) for the period from January 2002 to December 2002 in Turkey from the above 10 stations were used as training and testing data. Inputs for the network are latitude, longitude, elevation, and monthly mean daily sum air temperature; output is land surface temperature. In addition the monthly mean land surface temperature data were measured with land thermometer from TSMS. The ANN structure and weights of neurons could help to understand the relationship between air and surface temperature. Separately, the method of Becker and Li was proposed for the estimation of monthly global land surface temperature values from meteorological satellite data. Land surface temperature (LST) was estimated as a monthly mean by using the Becker and Li method over 10
stations in Turkey. It is suggested in this study that both methods describing the LST throughout a diurnal cycle very well. Data of TSMS are deficient throughout a nocturnal cycle than a diurnal cycle in Turkey.
Fig. 3. Topology of a generalized regression neural network (GRNN) artificial neural networks (ANN).
In the current study, ground data of the ten national stations were used. These stations were selected in such a way as to represent the widely changing climatic conditions of Turkey. The monthly mean daily sum global land surface temperature over Turkey was determined to be a correlation coefficient 94.08% and RMSE 0.0907 °K (Fig. 4), 91.13% and RMSE 0.0029 °K (Fig. 5) for Becker and Li values (training and testing stations). In the case of monthly mean daily sum correlation coefficient, RMSE was found to be 98.06% and 0.0766 °K (Fig. 6), 98.01% and 0.447 °K (Fig. 7) for ANN values (training and testing stations).
Fig. 4. Comparison between measured and estimated monthly mean daily land surface temperature using Becker and Li method in training stations.
Fig. 5. Comparison between measured and estimated monthly mean daily land surface temperature using Becker and Li method in testing stations.
Fig. 6. Comparison between measured and estimated monthly mean daily land surface temperature using ANN
method in training stations.
Fig. 7. Comparison between measured and estimated monthly mean daily land surface temperature using ANN method in testing stations.
The performance values for all stations, such as root mean square error (RMSE) and mean bias error (MBE) for ANN and Becker and Li method values (training, testing) are given in Table 2. The RMSE values, ranging from 0.0144 °K to 9.3184 °K, differ from the actual value for all stations. The maximum RMSE was found to be 9.3184 °K for the Rize station (testing) in the Becker and Li method values, while the best result was found to be 0.0144 °K for the Afyon station (training) in the ANN values. The maximum MBE was found to be - 2.6900 for Becker and Li method values of Rize station, while the minimum MBE was found as 0.0042 for Afyon and Izmir stations. Moreover, another significant point in this table, the performance values of the training are generally better than the performance values of the testing. Fig. 8 and 9 shows a comparison between measured, ANN and Becker and Li values for the ten stations (training and testing stations). Using the ANN and Becker and Li methods is a cheap and effective way to estimate global land surface temperature and construct a land surface temperature database. The ANN model, which needs no satellite data, was used to estimate the monthly mean daily sum at ground level. On the other hand, since the other method needs no ground data and presents valuable results, it can be applied to any region. The application of these methods is suitable particularly for places where the distances between the stations are very large.
Error values of the ANN and Becker and Li Method method approach
Fig. 8. Comparison for the land surface temperature between the ANN, Becker and Li and measured values
Fig. 9. Comparison for the land surface temperature between the ANN, Becker and Li and measured values
A generalized regression neural network (GRNN) was used to estimate land surface temperature by using meteorological and geographical data. The generation of a typical land surface temperature is significant for the calculations concerning many land surface temperature methods. By using the ANN and Becker and Li methods, a production of land surface temperature was used at over 10 stations in Turkey. The monthly mean daily sum values were found as 0.0766 °K and 0.0907 °K (training stations), 0.0447 °K and 0.0029 °K (testing stations), respectively. Construction of a land surface temperature database is very useful for environmental, agricultural, and other applications. According to the results of these 10 locations, correlation values indicate a relatively good agreement between the observed ANN values and the predicted satellite values. These methods can be used by researchers in Turkey and other countries.