Forests are living dynamic systems with a biologically diverse and complicated structure. Based on an inexpert point of view forests are only considered as a resource of wood product while from a scientific and expert standpoint the main function of forests and vegetation is that they protect our environment and this outlook is much more important than its economic importance as the source of wood. Forests throughout the world play a crucial role in the persistence of vital processes in the environment, including adjustment of weather, preservation of carbon, and as well as boosting economic growth by serving as wood resources (U.N.E.P, 2007). Therefore, protecting forests and preventing the destruction of these reserves are of high importance. Fire is one of the phenomenas that every year destroys these ecosystems and turns natural environment into arid lands devoid of vegetation. Forest fires are important not only in environmental issues but also in terms of economic, social and security issues. They are also the main issues induced many concern in all over the word (Merino-de-Miguela et al., 2010). Forest fire is one of the main factors in determining plant diversity and spreading according to ecology (Bajocco et al., 2010) and they also produce significant greenhouse gases, smoke and aerosol into atmosphere and cause significant increase in atmospheric CO₂ (Levine, 1999). Modis reflectance and active fire data for burn mapping and assessment are at regional level. Previous statistical data indicated that more than 2 538 hectares of natural and artificial forest of country burned only in 74 and 77 and allotment of Khozestan province was 10.682 hectars (Department of Natural Resources, 1999). The number of forest fire occurred in 1 381 was 63 that they burned 269.6 hectares of forest. The average of each forest fire was 42.8 hectares that indicate 33.3% increase according to the past year (Department of Natural Resources, 2002). The number of forest fire occurred in 2007 was 96 that they burned 2 169.2 hectares of natural and artificial forest (Department of Natural Resources, 2007). Izeh city is one of the most important cities in province of Khozestan that occur many fire forest in that. Fire forest occurs mostly from Farvardin to Aban in this city. Khoradad, Tir and Sharivar has the most forest because they are related to human and they are surface fire and then body fire (Alimahmoodi and feghhi, 2011). One of the first activities for controlling of this phenomena is to determine the most important factor that affect on creation and spreading of fire and it should be used effective and appropriate methods to achieve this important objective. In recent years, various attempts have been made to determine the factors bearing a part in occurrence and expansion of wildfire. Climatic features are among these parameters. Various techniques are used to foresee fire and hazard zonation but these techniques are usually expensive. The absence of required data relating to the third world countries is another defect that makes these techniques questionable. Numerous studies have been carried out on wildfire and we review some of them. Brillinger et al (2003) recommended developing a model for areas with respect to the fire history, elevation from the sea surface, and other data, including days of fire and non-fire days. At present, the United States Forest Services forecasts and tracks fire in forests by means of various instruments, including automatic remote-controlled meteorological stations, but these devices are expensive and cover limited areas. There have been also other studies to improve fire anticipation and detection. Furthermore, collaborative studies to develop response strategies have been carried out (Alonso-Betanzos et al., 2003; Bernabeu et al., 2004). Blackard and Dean (1999) compared artificial neural networks and detected analysis and concluded that artificial neural networks are more accurately in predicting plant cover compared to former statistical techniques. Dimitrakopoulus et al (2011) used the logistical regression to classify climatic characteristics of fire in Greece. Results of this research showed the developed model for forecasting fire in forests is reliable to a high degree. Elmas and Sonmez (2011) carried out a study entitled “A data fusion framework with novel hybrid algorithm for multi-agent”, which made useful conclusions to forecast fire hazard in forests. Fernandes et al (2005) developed a method to locate fires using mechanisms of the artificial neural networks and lidar images. This research employed a monolayer network and results showed this method is reliable by 90%. A monolayer network was employed in this study. Results showed the technique is accurate by 90%. Maeda et al. (2009) used MODIS (moderate resolution imaging spectroradiometer) images in Brazil to forecast fire hazard in Amazon forests. A progressive neural network and data extracted from burn areas in Amazon forests were exploited in the study. The results showed the mean least square of error amounted to 0.07. According to the same results, the neural network is a rapid and relatively precise technique to predict fire in the area under study. Cortez and Morais (2007) used Data Mining Approach and climatic parameters and appraised fire hazard risk in forests of Portugal. Climatic data on temperature, wind, relative humidity etc., as well as data from the Support Vector Machines (SVM) and forest weather indexes, were incorporated to estimate the burnt area. Results of the research showed the best configuration model may be achieved using SVM and most important parameters contributing to the fire hazard in forests, including wind, rainfall, humidity and temperature. Employing neural networks and climatic parameters, Sakr et al (2011) tried to foresee and determine the climatic parameters that contribute to fire occurrence in Lebanon. In this research, two techniques based on artificial intelligence – neural networks and SVM – were used. Results of the research showed that both techniques can anticipate the hazard to a great extent. Since providing required data entails enormous funds, time and high technology, this method impose strict limitations on third world countries developing less expensive with acceptable precision will be very important. Forecasting wildfire in Iran also requires highly accurate data that may be collected within a short time to let responding to fires quickly enough. In our country, we don’t have meteorological stations in many areas and therefore major parts of the country are not covered by the stations and consequently defective data are gathered. Forecasting fire hazard needs for having highly precise data and short response time, and because covering the whole country by meteorological stations to produce sufficient data on the occurrence and expansion of fire requires enormous spending extracting data from climatic factors but with a broader time domain and higher precision and coverage is a preferred method as compared to approaches based on defective data of various factors. However, what is very important is the finding of the most determining factors in occurrence and expansion of wildfire. Khuzestan Province like other parts of Iran enjoys few meteorological stations so that in most cases the data collected are incomplete and lack sufficient accuracy. This is true with Izeh County where only one station works in this area with climatic diversity. The main objective of this research is to find the most important factors in wildfire occurrence and expansion. This enables us to reduce climatic factors and therefore save time and energy and produce data with higher precision that can be gathered through examinations in broader spans. Thus, we have tried to pick out a limited number of feasibly measurable parameters with the purpose of diminishing costs of the system and its maintenance. Meanwhile, the selected parameters should show a great correlation with the wildfires occurred in the past. To meet this end, 12 parameters from the Izeh County meteorological station were chosen.

1 Materials and Methods
1.1 Study area
Izeh County is located in the southeast of Khuzestan province and has an area measuring about 572 219 hectares (31◦ 49’ 34” and 31◦ 37’ 46” N, 48◦ 40’ 32” and 50◦ 08’ 49” S), with a height of 835 m above the sea level. This county is part of the Zagros plant area and 396 973.5 hectares of its lands are covered by forests and ranges (Figure 1). In this study we employed the data gathered by the climatology station of Izeh in a 26-year period in order to investigate the climate of the area. The monthly absolute mean temperature of the area is 0.02◦-23.50◦ and monthly average precipitation varies from zero to 148 mm (Figure 2). 
 

Figure 1 Results of network performed on all parameters

Figure 2 Izeh embrothermic curve

1.2 Research methodology
To carry out this study, data on wildfires and the area of the burnt forests and ranges in Izeh County from 2008 to2010 were provided (through the protection unit of Natural Resources General Department of Khuzestan Province in 2008, 2009 and 2010) (Table 1), as well as climatic data, including temperature (maximum, average and minimum), relative humidity (maximum, average and minimum), wind speed and direction, daylight hours, precipitation, the number of days with and without precipitation on a monthly basis in abovementioned years. These data were received from the Iranian Meteorological Organization and neural networks and multilayer preceptron technique with hyperbolic functions were employed. In this study, artificial neural networks were applied in order to predict the forest fire in Mediterranean region. One of the most important algorithms in this approach is multi-layer perceptron (Figure 3), which is used to predict the phenomena (Fatipour and Najba, 2008; Hosini and Jalilvand, 2006). Preceptrons have the advantage that function based on a simple algorithm (Yuan, 2002). Learning rate is one of the parameters that affect neural networks and speed up the training of the network is significantly faster learning. Choosing the learning rate too small, the network will increase the run time and high values selected for the cause of achieving results are weak. It has been suggested that the learning rate of 0.1 to 0.2 is selected (Sebastian, 2002). In this study, we introduce different values of the learning rate of the network 0.19 is reached. Activator function of the weighted sum of the values of the units in one layer and the next layer of units to communicate. Hyperbolic tangent function is as follows:

 

Table 1 Fire occurrence and burned area in Izeh city in study period

 

Figure 3 MLP algorithm

 

This function takes real values and their values in the interval (-1, 1) is converted. Pararmtr momentum rate is to reduce volatility. This network allows Parameter the slope of the error surface react to changes in the process of training and testing speed can be overlaid (Kia, 2010). The amount of trial and error, a network with a torque of 0.7 has been selected. Acceptable error rate for the network 0.01 were considered. The number of latent neurons are often determined by trial and error; the network with varying latent neurons, for instance 1, 2 or 3 times the input variables are tested. (Sebastian, 2002). In order to determine the preliminary network all climatic parameters that have been omitted firstly were reintroduced and a perceptron network will all possible used parameters was developed. Then, to find important parameters the sensitivity test was carried out and the parameters with highest influence were known. And finally, the suitable topologic structure for our purpose was developed. One of the parameters influencing the neural network is the learning rate. Lower learning rates mean increase in the network response time. Employing high learning rates produces almost unreliable results. Therefore, we suggest adjusting the learning rate between 0.1 and 0.2 (Sebastian, 2002). In our research, with various inputs the learning rate reached 0.19. The momentum rate leads to lower fluctuations. Furthermore, it enhances the overlap between theoretical data and the data received in tests. Using trial and error technique, the momentum rate was set on 0.7. The acceptable error to run the network is 0.001.

1.3 Analyzing Date
After inputting in the spreadsheet, the data were organized. Excel 2007 and NeuroSollutions 5 software programs were applied in this research. A neural perception network with several hidden layers and a number of neurons in each layer, as well as hyperbolic tangent functions, were used to predict the occurrence of fire.

1.4 Model selection
A program was developed in NeuroSolutions 5.07 package for the feed forward and back propagation network. We used the ‘Leave N out’ option, i.e., networks were trained multiple times leaving out different sections of data for each training run. This training procedure is very useful for testing the robustness of a model on small datasets. To objectively evaluate the performance of the network, four different statistical indicators were used. These indicators are - mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R²) and mean absolute percentage error (MAPE) (Zangeneh et al., 2010):

 (1)
 (2)
(3)

Where Y Target and Y Estimated are the actual and estimated benefit to cost ratio values of ith farm, respectively, Y Mean is the mean of actual Benefit to cost ratio values, and N is the number of observations. To objectively evaluate the performance of the network, four different statistical indicators were used.

2 Results

First, in order to examine the correlation between data and the fires occurred in the past the Spirman correlation test was used. Results showed a correlation between the number fires and climatic variables studied in the research (Alimahmoudi et al., 2012). Then, in order to develop a model to incorporate climatic parameters and the frequency of fires neural networks were used. Neural networks employ diverse techniques to predict fire. Among these techniques, the perceptron is the most popular. The acceptable error in the research was set at 0.001. In order to prevent local minimums, we tested the networks with the momentum index. Actually, we tested the network at the momentum index of 0.7. A perceptron network with two hidden layers and 15 neurons in each input and output layer was tested. The result showed the most efficient model with two hidden layers consists of 10 neurons in the input layer and 10 neurons in the output layer, 15 neurons for the hidden input layers and 20 for the output of these layers. For the third layers (output) one input neuron and one output neuron provides the best model that may be applied to forecast fire occurrence in Izeh County. Actually, the neurons of hidden layers may be displayed in this arrangement 10-15-15-1. Firstly, evaluating the results showed the best network is achieved in performance 8 and replication of 680 where the final least square error amounts to 0.0038. Testing the network the correlation coefficient of the network was determined to be 0.00, and the mean square error and NMSE were about 0.73 and 0.018, respectively. In order to find the network ability to fire hazard based on the data used for testing the network the coefficient of determination was estimated to be r² = 0.98. regarding the coefficient of determination obtained, as well as the correlation coefficient, it can be admitted that the network model is satisfactory and may be usable to forecast fires in forests based on climatic parameters.

The result of sensitive analyze showed climatic parameters, such as relative humidity and mean relative humidity with correlation coefficients of 0.94 and 0.93, show highest correlation with fire occurrence and least correlation with rainy days with a correlation coefficient of 0.73. Therefore, the most important parameter in fire occurrence in the area is relative humidity. Furthermore, the results showed the coefficient of determination for minimum, mean and maximum temperature are 0.73, 0.84 and 0.88, respectively. The coefficient of determination for maximum humidity, sunlight hours, rainfall days and dry days were 0.81, 0.87, 0.73 and 0.80, respectively (Table 2). Afterwards, the sensitivity analysis was carried out to know the most important parameters that contribute to fire (Figure 4). Finally, with respect to the coefficient of determinations in table 1, the parameters with higher coefficients (r²>0.85), which actually has a greater correlation with fire occurrence, and were understood to be of highest importance in the sensitivity analysis, were input in the new network. The network was run with various number of layers and with different number of neurons in each layers, and the network with highest coefficient of determination and lowest error was chosen. The final network was developed regarding maximum temperature, minimum relative humidity, mean relative humidity and sunlight hours was designed (Table 3, Figure 5 and Figure 6). Then, in order to develop a model to forecast fire hazard, a network with two hidden layers and 8 run and four neurons in each hidden layer was run 2 000 times. Results showed the 7^th implementation of the network with 680 epoch and with four neurons in hidden layers produces the most desirable network that may be used as the best device to predict fire hazard in Izeh County. In fact, the configuration of neurons of the hidden layers is as 4-4-4-1 (Figure 5, Figure 6 and Figure 7, Table 3).

Table 2 Results of network performed on all parameters

Figure 4 Sensitive analyze curve

Figure 5 Correlation between important influence factors on forest fire

Figure 6 MAPE curve

Figure 7 Correlation curve between observation and prediction data

Table 3 Results of network performed for determine importance factors

3 Discussion
Understanding the factors that determine the occurrence of fire is a crucial element in predicting upcoming changes in the Earth (Zumbrunnen et al., 2010). Every year, we observe the occurrence of fires in Zagros region that destroy biodiversity and stimulate pollution and human hazards and also erosion of soil in the damaged areas. Fire leaves a burn soil with higher hydrophobic property that enhances soil erosion (Pierson et al., 2008). Therefore, as we determine the important parameters that influence the occurrence of fire we will be able to forecast fires. Previous studies show that climatic parameters play acritical role in this process. The ability to predict fire occurrence helps the better management of combating fire and ultimately reduces the burnt area and relieves environmental and economic problems, casualties and damage. On the other hand, there are numerous parameters at work that should be incorporated in a model for forecasting fire. Reducing the number of these parameters to develop an efficient model with high accuracy enables us reducing time and energy and improves our precision in predicting fire. Cutting expenses by employing such model will allow us to broaden the statistical area as much as possible. There has been an interest in using artificial neural networks to predict fire occurrence, and this research tries to take advantage of this method. It was found that the best network for forecasting fire occurrence in Izeh County includes two hidden layers with four neurons in each layer. The arrangement of the neurons in the hidden layes are as follows: 4-4-4-1.

Examining the results showed that effective parameters may be reduced from 10 to 4climatic parameters without damaging the model’s ability to forecast fire occurrence (R²=0.98). Additionally, the results prove one of the very effective parameters is the maximum temperature (R²=0.88) while the correlation between the ambient temperature was lower (R²=0.84) as compared to the maximum temperature. Therefore, maximum temperature may be defined as one of the four parameters with highest effect on fire occurrence. The coefficient of determination relating to maximum relative humidity and its connection with fire occurrence in Izeh County had the highest value (R²=0.94) and this was higher greater for the ambient relative humidity (R²=0.93). The fourth important parameter affecting the fire occurrence was the number of sunlight hours in the area. The coefficient of determination from this parameter amounted to 0.87. Therefore, it can be said when the maximum temperature increases and minimum relative humidity deceases the fire hazard will be at maximum. Biravand et al (2011) with using of GIS and RS IN Iran shown that temperature and Humidity are two factor in forest fire occurrence. Concerning the area under study, regarding its long distance from free waters and low precipitation in summer and long days in this season that intensifies solar radiation on flammable materials all extremely reduce humidity and increase in temperature of the area. Located in a warm and semi-arid climatic region, as well as human factors and other features increase the possibility of fire occurrence in the area. As results of the research shows, climatic elements have a great effect on the occurrence and expansion of wildfires. Cortez and Morais, (2007) showed climatic parameters, including temperature and relative humidity are the effective factors that contribute to fire occurrence. Sakr et al. (2011) also used climatic parameters, such as maximum temperature, solar radiation, relative mean humidity and annual precipitation to predict fire occurrence in Lebanon and made acceptable results.

These results are in agreement with findings of Dimitrakopoulus et al. (2011), Alimahmoudi et al. (2012) and Radpour et al. (1390) that took into account temperature, relative humidity and annual precipitation as the parameters required for forecasting fires. Annual precipitation, the number of rainfall days and the number of days without rainfall also produced average coefficient of determination that showed a moderate connection between these parameters and fire occurrence probability. However, regarding the fact that the final coefficient of determination from the final network has a great importance (R²=0.98) these parameters, as well as the minimum temperature, mean ambient temperature and maximum humidity, can be omitted from the equation so that the model preserve its efficiency for forecasting fire hazard in Izeh County (R²= 0.98).

4 Conclusions
Results of the research showed that in forecasting wildfire hazards in this area only the abovementioned parameters may be used and save time and energy and allocate resources to expand statistical coverage. In our country, one of the major problems concerning the study of climatic factors is the absence of data collection stations in most areas. Since the type and density of plantation and the use type of plants remarkably affect occurrence and expansion of fire in forests including these parameters in future studies is recommended.

Reference
Aleemahmoodi S.S., and Feghhi J., 2011, Review time and cause of fires in forests and pastures in Khuzestan province, Modern Management of Natural Resources, Tehran Azad University of Iran, Science Research

Aleemahmoodi S.S., Feghhi J., Jabarian A.B., Danehkar A., and Attarod P., 2012, Using of regression models to assessment of impotence climate elements on natural fires (forests of western Iran: Izeh), Natural Environment Journal, Iran, under pressing 

Alonso-Betanzos A., Fontenla-Romero O., Bertha Guijarro-Berdi-as B., Hernández-Pereira E., Paz M.I.A., Jiménez E., Soto J.L.L., and Carballa T., 2003, An intelligent system for forest fire risk prediction and fire fighting management in Galicia, Expert Systems with Applications, 25(4): 545-554
http://dx.doi.org/10.1016/S0957-4174(03)00095-2

Bajocco S., Rosati L., and Ricoa C., 2010, Knowing fire incidence through fuel phenology: A remotely sensed approach, Ecological Modelling, 221(1): 59-66
http://dx.doi.org/10.1016/j.ecolmodel.2008.12.024

Bernabeu P., Vergara L., Bosh I. and Igual J., 2004, A prediction/detection scheme for automatic forest fire surveillance, Digital Signal Processing, 14(5): 481-507
http://dx.doi.org/10.1016/j.dsp.2004.06.003

Biravand A., Zare R., and Rahimizade N., 2011, The fire-prone areas brass climatic factors using GIS, National Conference of forests in Iran, Capabilities and bottlenecks

Blackard J.A., and Dean D.J., 1999, Comparative accuracies of artificial neural networks and discriminant analysis in predicting forest cover types from cartographic variables, Computers and Electronics in Agriculture, 24: 131-151
http://dx.doi.org/10.1016/S0168-1699(99)00046-0

Brillinger D.R., Preisler H.K., and Benoit J.W., 2003, Risk assessment: A forest fire example, In: Goldstein, D., and Speed, T. (eds.), Statistics and science: A festschrift for terry speed, IMS Lecture Notes Monograph Series, Institute of Mathematical Statistics, Beachwood, OH, 40: 177-196
http://dx.doi.org/10.1214/lnms/1215091142

Cortez P., and Morais A., 2007, A data mining approach to predict forest fires using meteorological data, In: Maia J., Filipe S.M., and Manuel M.J. (eds.), Progress in Artificial Intelligence, Proceedings of the 13th Portuguese Conference on Artificial Intelligence, 3~7, December, Guimarães, Portugal, pp.512-523
 

Department of Natural Resources Conservation Unit in Khuzestan Province, 1997, 2002, 2007, 2008, 2009, 2010

Dimitrakopoulos A.P., Bemmerzouk A.M., and Mitsopoulos I.D., 2011, Evaluation of the Canadian fire weather index system in an eastern Mediterranean environment, Meteorological Application, 18(1): 83-93
http://dx.doi.org/10.1002/met.214
 

Elmas C., and Sönmez Y., 2011, A data fusion framework with novel hybrid algorithm for multi-agent. decision support system for forest fire cetin, Expert Systems with Applications, 38(8): 9225-9236
http://dx.doi.org/10.1016/j.eswa.2011.01.125

Fatipour A., and Najba M., 1388, Neutral network in SPSS, Kian-rayane-sabz Express, Publish in Iran, pp.151

Fernandes A.M., Utkin A.B., Lavrov A.V., and Vilar R.M. 2005, Design of committee machines for classification of single-wavelength lidar signals applied to early forest fire detection, Pattern Recognition Letters, 26(5): 625-632
http://dx.doi.org/10.1016/j.patrec.2004.09.012

Hosini A., and Jalilvand M., 1385, introduction on neural network, poly copy, publish in Iran

Kia M., 1389, Neutral network in MATLAB, Kian-rayane-sabz Express, Publish in Iran, pp.226

Levine J., 1999, Introduction, In: Levine, J.S. (ed.), Global biomass burning: Atmospheric, climatic and biospheric implications, MIT Press, Cambridge, USA, pp.1-5

Maeda E.E., Formaggio A.R., Shimabukuro Y.E., Arcoverde G.F.B., and Hansen M.C., 2009, Predicting forest fire in the Brazilian Amazon using MODIS imagery and artificial neural networks, International Journal of Applied Earth Observation and Geoinformation, 11(4): 265-272
http://dx.doi.org/10.1016/j.jag.2009.03.003

Merino-de-Miguela S., Margarita H., González-Alonsob F., 2010, Modis reflectance and active fire data for burn mapping and assessment at regional level, Ecological Modelling, 221(1): 67-74
http://dx.doi.org/10.1016/j.ecolmodel.2009.09.015

Pierson F.B., Robichaud P.R., Moffet C.A., Spaeth K.E., Hardegree S.P., Clark P.E., and Williams C. J., 2008, Fire effects on rangeland hydrology and erosion in a steep sagebrush dominated landscape, Hydrological Processes, 22(16): 2916-2929
http://dx.doi.org/10.1002/hyp.6904

Radpour S., Esmaeilzade M., Yazdani V., Darban A., and Jahedpour A., 1390, Application of neural network on estimation of fire occurrence, In: The international conference on forest fire in Golestan province, Iran’s meteorological organization, meteorological data of Izeh city, Tehran
 

Sakr G.E., Elhajj I.H., and Mitri G., 2011, Efficient forest fire occurrence prediction for developing countries using two weather parameters, Engineering Applications of Artificial Intelligence, 24(5): 888-894
http://dx.doi.org/10.1016/j.engappai.2011.02.017

Sebastian S., 2002, Multi perceptreon and back propagation learning, 9.641 Lecture4.September. oi=10.1.1.86.8968

UNEP, 2007, Global Environment Outlook 4, United Nations Environment Programme, Nairobi

Yuan H., 2002, Development and evaluation of advanced data for accurate Land-us/Land-cover mapping, Dissertation for Ph.D., Departement of Forestry, North Carolina StateUniversity, Supervisor: Khorram S., pp.1-235

Zangeneh M., Omid M., and Akram A., 2010, Assessment of machinery energy ratio in potato production by means of artificial neural network, Afr. J. Agric. Res., 5: 993-998 

Zumbrunnen T., Gianni B.P., Menéndez P., Bugmann H., Bürgi M., and Conedera M., 2011 Weather and human impacts on forest fires: 100 years of fire history in two climatic regions of Switzerland, Forest Ecology and Management, 261(12): 2188-2199
http://dx.doi.org/10.1016/j.foreco.2010.10.009