The Use of Artificial Neural Network to Evaluate the Effects of Human and Physiographic Factors on Forest Stock Volume

: Increase in human factors coupled with physiographic factors will impact stock volume in forest ecosystems. The scale of this process and critical information in forestry management provide an incentive for the development of model to predict the forest stock volume. In this paper, we use data derived from Siahrood, Guilan Province, Iran using Field inventory by cluster sampling in a network (1 × 1 km) with 90 clusters and 900 circular plots (1000 m 2 ). To evaluate modeling approaches for stock volume responses to changing condition. The relationship between the standing volume and human factors and each physiographic factor were examined using Pearson and the Artificial Neural Network method. Based on Field observations it was observed that different stock volume exhibit in specific physiographic response to population density, livestock density, distance from village, aspect, slope and elevation. Results show that Multilayer Neural Networks with 12 nodes can predict the forest stock volume with the lowest RMSE (48.76m 3 ). In addition, the artificial neural network designed for the buffer of three populations with 85.5% accuracy was selected as the best model to predict the volume based on the mentioned components. The results suggest ANN is an effective approach to predict exact forest stock volume and human factors in certain topography conditions and provides useful information for the acceptable amount of standing inventory using the present human population in future experiment. © JASEM quantitative influence Multilayer Perceptron network some and and Livestock Vilage) effectiveness role of on

Hyrcanian Forests are surviving from the Tertiary. In terms of geography, they are sub-order of European-Siberian forest and are the continuation of European temperate forests, which start from northern Europe and continue to Caspian forests. Forest areas of Guilan province as parts of the northern forests are generally a Hyrcanian type and a wide range of habitat is covered by them. The rapid growth of human population in recent decades has increasingly threatened forests in the province. Low per capita income, lack of development, basic supplies and the most basic environmental culture has made local communities of forest dwellers to put pressure on natural resources in order to meet their needs (Rastaby et al., 2013). Although a large part of the northern forests including Guilan province is under the management of forestry plans, in many of these managed areas the pressures of human presence disrupts the structure, composition and generally the natural cycle of forests' ecological system. Over sixty thousands forest dweller families are living in forest areas of Guilan province, which are about 370,000 people who temporarily and permanently live there and their livelihood depends on the forest (mohammadi, 2008). Having a proper understanding of the interaction between humans and forests is considered essential in order to reach a successful conservation and management of forests (Bode, 2000), hence the influence of human populations living in forests is inevitable and requires techniques and methods that could measures the disruption caused by the presence due to provide the possibility of a systematic and careful planning to protect the forest (Shadytalab, 1998). Predicting standing volume as one of the most important quantitative features of forest stands' structure has a very considerable role in the planning and assessment of forest resources for managers and planners. Obviously, a lot of natural and unnatural factors are involved in the formation of forest stands; so that, physiographic factors are the most important natural factors (Mitusuda et al., 2007) and human factors are the most influential nonnatural factors. Three variables of slope, aspect and elevation are the most common indicators of physiographic factors that are commonly used in the measurement and mapping of vegetation cover (Liprieur et al., 1988) and can determine the microclimate in many cases (Giger et al., 1966). Based on studies in Hyrcanian Forests of Azerbaijan, the tree density significantly changed due to distance between forest stands and population centers and accordingly the damage severity was defined (Sharnweber et al., 2007). Investigations in the tropical forests of Cambodia about the impact of human population density on the forests' structure and composition have been also conducted (Top et al., 2003(Top et al., , 2004(Top et al., , 2009. In addition, studies on population issues related to forests have been done in other parts of the world including India (Karanth et al., 2006), China (Wang et al., 2001) and Turkey (Sivrikaya et al., 2011). Generally, ecological and socio -economical aspects need special attention in sustainable forest management issues (Wolfslehner et al., 2005). thus, studying how and to what extent the fact that forest resources are affected by demographic pressures is very important to determine the appropriate program for forest management. The aim of this research is to better understand the relationship between humans and forests using quantitative methods of human factors as well as the artificial neural networks to model these components.
This work presents a new approach for predicting the forest stock volume of Siahrood Roodbar hyrcanian Forest by using Multilayer Perceptron Neural Network. We propose a prediction of the forest stock volume by taking into account the measure of the variables (slope, aspect, elevation and human density). The predicted forest stock volume allows us to estimate these volumes by considering only four simple measures, the present method needs only 15% of the sample variables for the training stage making the field work easier and faster.

MATERIALS AND METHODS
The study area is a part of Siahrood Roodbar catchment in the south of Guilan province. The study area lies between 49° 31´ 06´´ and 49° 45´ 56´´E and 36° 56´ 57´´ and 36° 49´ 56´´N with an area of about 15,000 ha. The climate is a humid temperate with average annual rainfall of 903 mm and average annual temperature of 16.2 °C. The dominant tree species in the area are beech with other species such as hornbeam, iron-tree, alder, maple, ash and 5000 people temporary and permanent reside there and are mainly engaged in ranching ( Figure 1). Tree data: In order to measure the tree volume per ha, 90 clusters were systematic randomly determined in a network (1 × 1 km) on the map and nature. Ten circular plots (1000 m 2 ) in 150 × 200 m intervals were placed in each cluster. In each of these plots, trees' data with a diameter more than 12.5 cm were recorded and the standing volume in the clusters were calculated using SPSS ( Figure 2).
Demographic data: to determine population density around each cluster, seven circular layers were first defined at intervals of 1 to 7 km around each cluster using Arc GIS ver.9.3 and their position in each population buffer were determined by recording the coordinates of each village and house using Gps Map 60cx and the population of each layer was set by counting the number of houses and villages and were summarized and calculated in the form of buffer functions by the mentioned software (Top et al., 2006) (Figure 2). Topographic data: Maps of slope, aspect and elevation were prepared by DEM Shuttle Topography Radar Mission (STRM) data with a resolution of 30 × 30 m. Topography properties (slope, aspect and elevation) for each plot were then recorded by GPS with high precision and were extracted peer-to-peer from the raster map.

Tree volume calculation using volumetric equations:
The volumes of all trees were calculated using the Smalian method, according to Eqs. (6) and (7), and the Tarif Table. The tariff table is a volumetric table that estimates tree volume from the DBH 2 . In order to verify the efficiency of the MLP network model proposed. These volumes are compared with volume obtained with 4 parameters (slope, aspect, altitude, population density) (Table 4).
For training the MLP, 20 % of trees were randomly selected from each cluster to collect input and output sample sets. In the testing phase of the MLP, Slope, Aspect, elevation, and population density used to predict the tree volume. This phase is performed fro trees according to the prediction procedure described by Eq. (1). This procedure is described in Table 3.
Data analysis: After preparation and analysis of trees, human and physiographic data in clusters, in order to analyze data and detect the presence or absence of correlation between each of the independent variables (slope, aspect, elevation, population density, livestock density and distanse from village) and the dependent variable (standing volume), the Pearson correlation coefficient (r) was used (Top et al., 2006). This coefficient calculates the correlation between two variables and its value is between +1 and -1. If the obtained value is positive, then changes in two variables occur in the same direction. Conversely, if the r value is negative, then the two variables occur in opposite direction, which means that the value of a variable reduces with the increase in another variable and if the value is zero indicates that there is no relationship between the two variables (soltani et al., 2009).
Prediction methods and artificial neural networks Multilayer Perceptron (MLP) artificial neural network: A Multilayer Perceptron (MLP) neural network is composed of an input layer, one or more hidden layers and an output layer. The input signal propagates feed-forward through the network, layer after layer (Haykin, 1998). Fig.3. shows the MLP used in this work. Several activation functions may be used in the MLP (Haykin, 1998). The activation function used in this work for the hidden layer neurons for the output layer neurons was the hyperbolic tangent, also ;known as the tan-sigmoid or tansig, defined in Eq.(1).

Error calculation of predicted volumes by the MLP:
In order to verify the accuracy of the volumes calculated from the independent variables (slope, aspect, elevation and human density) predicted by MLP, the following percentage errors were calculated: Root Mean Square (RMSE%), Linear correlation (R) Özçelik et al., 2010), according to Eqs. (3),(4),(5), respectively. To calculate the percent of error was used the equation 2.
E%= 1/ N Σ | V pi -V oi / V oi |×100 [2] N: number of clusters V pi : the predicted volume V oi : the actual volume. [3] [4] Where ya and ye are the vectors of actual stock volumes and estimated, respectively, ya and ye are the averages of the actual and estimated stock volumes values, respectively, and n is the number of stock volume measures.

RESULTS AND DISCUSSION
The MLP model proposed in this paper was based on the relationship between four parameters (population, livestock, distanse from village, slope, aspect and elevation) used as input, and the forest stock volume that is predicted as the output of the network. In order to verify the relationship between four parameters and Forest stock volume, the autocorrelation coefficient was calculated through the Pearson correlation coefficient (Eq.5).
Pearson correlation coefficients were calculated for each cluster. Table 2 and table 3 shows the correlation coefficients in 90 clusters. It can observe in Table 2 that there is a significant inverse correlation between standing volume and all of human factors in buffer 3 and this correlation increases from buffer 1 to buffer 3. Also there is a significant positive relationship between elevation and standing volume. Whereas no significant relationship between slope and standing volume.    -To determine the best model for predicting forest standing volume per hectare using human and physiographic components, the Artificial Neural Network Multilayer Perceptron (MLP) was used. The data were first standardized to increase the speed and accuracy of neural network processes based on the equation 1.
X¯= (X i -X min ) / X max -X min [1] X¯: normalized data X i : used data X min and X max the lowest and highest data in each category Standardized data were randomly divided into two parts, i.e. training and testing, for network training. Thus, 20% and 80% of data were used to test the model and for training data, respectively (Hilbert et al., 2001). To design a Neural Network, the The Use of Artificial Neural Network to Evaluate the Effects of Human independent variables of slope, aspect, elevation, population density, livestock density and distance from village were considered as the input layers and the dependent variable of volume per hectare of clusters was considered as output layer.
Forest standing volume prediction: Due to the large number of trees included in this study, Table 4 shows RMSE% and Correlation % (R) value produced by the testing phase of the forest standing volume prediction.  Human population activities in natural ecosystems are considered as a highly influential factor (wang et al., 2001). The results of this study showed that there was a strong and significant inverse relationship between the standing volume and human factors in each of the population buffers and the correlation increased from buffer 1 to buffer 3 because of more villages in the higher buffer. Our results are consistent with the results of tropical forests of Cambodia (Top et al., 2004(Top et al., , 2006(Top et al., ,2009, western forests of India (karanth et al., 2006) and Hyrcanian forests of Azerbaijan (schewnweber et al., 2007). In total, the correlations represented the quantitative influence of natural and unnatural factors. In this study, the performance of Multilayer Perceptron network (MLP) was also examined in order to predict the forest volume per hectares using some topographical and human factors. The results indicated that the aforementioned network with its own training algorithm has the ability to predict the volume per hectare. This technique can predict about 85% of forest standing volume by using topographical (slope, aspect, elevation) and human (population, Livestock and Distance from Vilage) factors. This confirmed the effectiveness of the human population and the effective role of physiographic factors on forest properties that is consistent with (Horsch, 2003;Guisan and Theurillat, 2000;Mitsuda et al., 2007). The use of artificial neural network (MLP) was evaluated in this study to increase the suitable predicting accuracy, which are corresponded to studies of (Hilbert and Ostendorf, 2001;Ingram et al., 2005) however, it is not consistent with Castellanos et al., (2007).
Conclusions: Todays, in most parts of the world, forest management planning is based on habitat and physiographic factors and issues related to human population as effective quantitative parameters receives less attention. Artificial neural network systems allow designers and planners to achieve interactions between them as well as realize parametric relations between them, which are essential for forest planning. It is recommended to use this technique for modeling more effective parameters in the form of function and variable so, the possibility of a greater initiative for forest managers and planners will be provided.