Spatial Model of Deforestation in Sumatra Islands Using Typological Approach

High rate of deforestation occurred in Sumatra Islands had been allegedly triggered by various factors. This study examined how the deforestation pattern was related to the typology of the area, as well as how the deforestation is being affected by many factors such as physical, biological, and socio-economic of the local community. The objective of this study was to formulate a spatial model of deforestation based on triggering factors within each typology in Sumatra Islands. The typology classes were developed on the basis of socio-economic factors using the standardized-euclidean distance measure and the memberships of each cluster was determined using the furthest neighbor method. The logistic regression method was used for modeling and estimating the spatial distribution of deforestation. Two deforestation typologies were distinguished in this study, namely typology 1 (regencies/cities with low deforestation rate) and typology 2 (regencies/cities with high deforestation rate). The study found that growth rate of farm households could be used to assign each regencies or cities in Sumatra Islands into their corresponding typology. The resulted spatial model of deforestation from logistic regression analysis were logit (deforestation) = 1.355 + (0.012*total of farm households) – (0.08*elevation) – (0.019*distance from road) for typology 1 and logit (deforestation) = 1.714 + (0.007*total of farm households) – (0.021*slope) – (0.051*elevation) – (0.038* distance from road) + (0.039* distance from river) for typology 2, respectively. The accuracy test of deforestation model in 2000–2006 showed overall accuracy of 68.52% (typology 1) and 74.49% (typology 2), while model of deforestation in 2006–2012 showed overall accuracy of 65.37% (typology 1) and 72.24% (typology 2), respectively.


Introduction
Tropical forests in Sumatra s have enormous Island benefits in social, economic, cultural, political, and ecological aspects.To describe the enormous benefits of forests in the Islands of Sumatra, Van Beukering et al. (2003) reported that the total economic value of Gunung Leuser -1 National Park (GLNP) is USD316 million year , which represents total value of economic benefits of water supply, fisheries, flood control, agriculture, energy, ecotourism, biodiversity, carbon storage, forest fire prevention, nontimber forest products, and timber forest products.Forests of Sumatra are also the habitat of more than 10,000 species of plants, 201 species of mammals, and 580 species of avifauna (Whitten et al. 2000).
The forests in the sland continue to experience pressure i as the forest cover continually declines every year.Data issued by the Ministry of Forestry mentioned that forest cover in 2011 was 14.84 million ha (M 2012) then oF declined in 2012 to 13.97 million ha (M 2014).According oF to data released by Forest Watch Indonesia (FWI), Sumatra as one of the major islands in Indonesia was experiencing high rates of deforestation at 412,421.9 ha year -1 (Sumargo et al. 2011).Large-scale deforestation in the tropics for the provision of agricultural lands was one example of land use changes that has a major impact on biodiversity, land degradation, and the earth's ability to support human needs (Lambin et al. 2003).The impacts of deforestation and degradation of tropical forests in the form of forest destruction have become an important global issue due to the magnitude of cumulative impacts on biodiversity, soil productivity, and global climate (Geist & Lambin 2001;Laurance et al. 2014).
Geist & Lambin (2002) have conducted a study on deforestation in tropical forests, including in Indonesia.The study found that the main factors that drive deforestation in tropical forests were land expansion for agriculture, logging, and infrastructure development.Meanwhile, the underlying factors were economic, institutions and policies, technological, social, cultural, population growth, as well as other factors such as land characteristics, biophysical properties of crop/land and social unrest.Indonesia as an , archipelago country with a variety of forest biophysical ecosystem characteristics and different socio-economic conditions of its people, have a diverse driving force of deforestation in each region.
The studies conducted by Allen Barnes (1985) and , Meyer Turner (1992) lso Pahari Murai (1999) and , a and stated that land use changes, including deforestation on a global scale were driven by population growth.Quantitative research that mention common factors such as wealth and population growth as the main cause of deforestation considered has ignoring the specific factors such as biophysical factors, proximity, and other socio-economic factors (Rudel 2007).In regards to related policies, this simplification in which the driving forces of deforestation considered as the same for any region was considered inappropriate.
Administratively, tra is divided into 10 provinces Suma and 93 regencies/cities with a total area of 47,322,331.3 ha.The types of forest landscapes in each administrative region in the sland are specific in characteristics and diverse in i problems.Consequently, the driving forces that triggers the occurrence of deforestation in each administrative region in Sumatra were also varies.Hence, it was necessary to investigate further on the driving forces for different regencies, i.e. based on their typology.Typology was defined as a classification or grouping of objects based on similarity of the basic traits into certain types.This definition is modified from the Oxford English Dictionary (Soares & Stevenson 2009).
Several researches on the spatial model of deforestation have been conducted in some locations in the sland, i including Melati (2012) in Riau Province, Mulyanto and Jaya (2004) in the orest oncessionaires of PT Duta Maju f c Timber West Sumatra Province, as well as Linkie . et al (2004) in the Tapan Hill, Jambi Province.However, studies on spatial model of deforestation for a large extent using typological approach in Sumatra have never been conducted.Considering its potential information to be generated, this study may provide input in the preparation of deforestation mitigation policies and strategy of baseline formulation in REDD+ mechanism in Indonesia in general, as well as in the Islands of Sumatra in particular.The aim of this study was to develop a spatial model of deforestation based on the driving force for each typology in Sumatra .Islands

Data
Terra MODIS MOD13Q1 satellite image data used in this research was acquired in 2000, 2006, and 2012 with 250 m spatial resolution.The band combination used were middle infrared (MIR) band on the red layer, near infrared (NIR) band on the green layer, and blue band on the blue layer (Wedastra et al. 2013), while selection of good pixel data were based on the band information that determines the image quality of the Pixel Reliability and VI Quality bands (Solano et al. 2010;Wedastra et al. 2013).The MIR, NIR, and blue band selection was based on the spectral characteristic of the objects, particulary related to forest and non-forest.The MIR band was superior in determining soil, NIR band has its advantage in vegetation determination, while blue band as a strong chlorophyll absorber may enhance the occurance of vegetation (Richards & Richards 1999).
The other supporting data were administrative boundary map, road network map, river network map, digital elevation model from Shuttle Radar Topographic Mission, and statistical data of social economic for each regency/city from Bureau of Statistic Service in Sumatra Islands for the period 2000−2012.
Deforestation identification Deforestation in this study was defined as forest cover changes into permanent non-forest.In this study, the forest land cover changes into timber estates were not included in the calculation of deforestation, because logging activities in timber estates is a regular activity that is not permanently occurred.The analysis of land cover classification on Terra MODIS satellite images was performed using supervised classification approach (Otukeia &

Determination on typology of deforestation
In this study, regency/city was considered as the smallest unit to define the typology of deforestation.The typologies were built based on clustering approach.In theory, cluster analysis is a multivariate technique that has the main purpose to categorize objects based on their own characteristics.Cluster analysis classifies objects, thus each object that has the closest similarity to other objects are classified in the same cluster.The formed clusters should have high internal homogeneity and high external heterogeneity (Ediyanto et al. 2013).
In details, tandardized uclidean istance (SdED) was s e d used to perform cluster analysis.The use of SdED was based on the difference in the range of values and the difference in variable units used in the typology determination.The distance between data was calculated using equation as 2 described in quation : note: Si = the diversity of i-th variable Xij = the value of i-th variable from cluster j Xik = the value of i-the variable from cluster k Regional Domestic Product (GDP) (Romijn . 2013) were et al predicted to be the sosio economic driving forces of deforestation.Therefore those factors were used to characterize a regency/city and determine its typology.In addition, the rate of deforestation was used to validate the typology categorization for each regency.

Accuracy assessment of typology of deforestation
The accuracy assessment on typology of deforestation was performed by calculating the overall accuracy between clustering result and deforestation rate of each regency/city in Sumatra.The deforestation rate was resulted from the analysis of land cover classification using satellite image of Terra MODIS MOD13Q1.The formula of deforestation rate was based on the formula developed by Fearnside (1993) as shown in Equation [2].
[2] note: r = deforestation rate t 2 = final year t 1 = first year A 2 = forest area at final year A 1 = forest area at first year Determination of deforestation spatial model Binary logistic regression analysis (Linkie . 2004;Mulyanto & et al Jaya 2004;Prasetyo . 2009;Arekhi 2011;Arsanjani . et al et al 2013) was used to obtain the deforestation model.The driving forces such as socio-economic data, as well as biophysical and cultural factors were used as independent variables in the model.The selection process of driving forces in deforestation was conducted using the stepwise method by reducing or removing the driving factors that have no effect on the model.
Pixel samples were randomly selected to build a deforestation model in each typology.The dependent variables in logistic regression were binary, where the value of 1 meant there is change in forest land cover into non-forest and the value of 0 means there is no change in forest land cover during 2000 2006.The resulted regression equation was used to predict the probability of deforestation in the future (2006 2012).
-Processing of binary logistic regression models was conducted using SPSS software.Socio economic factors ( ) were vector data with regencies/cities as a spatial X −X [3] note: x 1 = population density (inhabitants km ) x 2 = the number of farming households (household) x 3 = number of poor (people) x IDR 4 = GDP (billion ) x 5 = slope (%) x 6 = altitude (m) x 7 = distance from the road (m) x 8 = distance from the river (m) x 9 = distance from the settlements (m) In the logistic regression model, the model is considered to be fit if the value of Hosmer-Lemeshow test shows a value bigger than p-value (0.05), while the contribution of predictor variables can be seen from the Exp β (odds ratio) value which is an exponential function of regression coefficient.The farther from 1, the contribution of a variable will be greater, in contrary, the closer to 1, the contribution of a variable in a model will be smaller (Hosmer & Lemeshow 2000).
The independent variables ( variables) have different x unit and scale, thus it is necessary to standardize the scores of each variable.The standardization process of variables x was conducted by standardizing the values of each variable into the range of 0−255 using the equation formulated by Jaya (2006)  In order to observe the association of each independent variables, multi collinearity test was performed.Logistic regression model should avoid the use of multi collinear variables.The correlation coefficient between the predictors were calculated using the Pearson correlation coefficient.In this study, correlation coefficient of 0.65 was used as the limit for selecting variables as referred in (de Almeida . et al 2002;Aguayo . 2007).

et al
The accuracy of spatial model of deforestation was assessed by comparing the deforestation prediction map with the actual deforestation map derived from the land cover classification of Terra MODIS image (image to image comparison).The process of transforming the results from deforestation model into raster data formats was performed using Arc GIS 9.3 software.In brief, the flowchart of model establishment showed in . is Figure 1 Results and Discussion    Margono et al. (2012) which stated that land cover area in Sumatra was continued to decline every year.

Typology of deforestation
Clustering results using number of farm households as a determining variable can be seen in Figure 2. Typology 1 was characterized by a relatively low rate of number of farm households as to compared with typology 2. In order to validate the clustering result (typologies), the rate of deforestation was used.This study found that the limiting threshold of low and high deforestation rate was at the magnitude of 2,562.7 ha year -1 .In other words, a regency having deforestation rate of < -1 2,562.7 ha year was categorized as low deforestation rate regency.Conversely, a regency having deforestation rate of ≥ -1 2,562.7 ha year was categorized as high deforestation regency.In overall, the membership of clustered regencies and its corresponding rate of deforestation resulting accuracy of 73.1%.
Based on the information in Figure 2, there were 63 regencies/cities that belonged to the typology 1 (low deforestation rates) and 30 regencies/cities that were members of the typology 2 (high deforestation rates).The average value of deforestation rate for typology 1 was 1,156.53ha year -1 -1 and for typology 2 was 3,968.81ha year , respectively.Typology 1 was dominated by urban areas that relatively more advanced and regencies/cities that have relatively limited forest area.Meanwhile, typology 2 was dominated by regencies/city with relatively larger extend of forest area than those belonged to typology 1.
The increase in the expansion of agriculture sector, particularly oil palm plantations in umatra has believed of S as a major factor leading to high deforestation rate.Increasing farming/plantation areas occasionally accompanied by increase in the number of farm households.Thus, indirectly the increase of number of farm households was one of the driving forces of deforestation.This results were in line with Gatto .(2015) and Setiawan .(2015 et al et al which stated that the high rate of forest conversion into agricultural land/plantation in Sumatra has led to high deforestation rate in the area. 1 shows the results of binary logistic regression analysis using stepwise method to obtain the model of deforestation probability for each typology.Overall test results in typology 1 and typology 2 showed the significance (sig) value of 0.000, which means that there was at least one predictor variable that affects variable Y (H /the null hypothesis rejected) so that the model 0 could be analyzed further.The test results on each of the influencing factors (partial test) also showed a significant value under 0.05 (H rejected), thus it could be interpreted 0 that the driving forces of deforestation both in typology 1 and typology 2 showed a significant influence on the model (Table 1).Model fit test using Hosmer-Lemeshow significance test in typology 1 deforestation model was 0.959 (> 0.05) and in typology 2 was 0.468 (> 0.05), thus both models were said to be fit or worthy (H accepted).In other 0 words, there were no significant differences between the models with their observation values, so that the deforestation model built could be accepted or feasible to be used.

Spatial models of deforestation Table
There were differences in the number and types of socioeconomic and biophysical factors that trigger the occurrence of deforestation in typology 1 and typology 2 (see Table 1).-Deforestation occurred in typology 1 was influenced by factors such as number of farm households, elevation and , distance from the road, while in typology 2 was influenced by the number of farm households, slope, elevation, distance from the road and the distance from the river.The value of -, R squared in typology 1 was 0.605, which means that 60.5% of the deforestation can be influenced by the number of farm households, elevation and distance from the road, while the , remaining percentage was influenced by other factors.-R squared value in typology 2 was higher than -squared value R in typology 1 at the magnitude of 0 641, which means that the .deforestation in typology 2 can be influenced by the number of farm households, slope, elevation, distance from the road , and distance from the river, while the remaining percentage was influenced by other factors.
In typology 1, elevation and distance from the road factors showed a negative correlation with the deforestation probability, while the number of farm households showed the opposite.Elevation factor has an odds ratio of 0.923, which means that one unit increase of the elevation factor value will lower the deforestation probability by 7.7% (1− 0.923); this value was the highest among other driving forces.Distance from the road factor has an odds ratio of 0.981, which means that one unit increase of the distance from the road will lower the deforestation probability by 1.9% (1−0.981), while the number of farm households has an odds ratio of 1.012, which means that one unit increase of the number of farm households value will raise the deforestation probability by 1.2% (1.012−1).Areas with low elevation values, close to the road and have a high number of farm households would have a high deforestation probability.The correlations between the driving forces of deforestation to the extent of deforestation probability in typology 1 can be seen in Figure 3.
In typology 2, in which the regencies/cities having high deforestation rate, the driving forces of deforestation were more complex in comparison to typology 1.It could be seen from many driving forces that played a role in typology 2. Similar to typology 1, elevation, distance from the road and slope showed a negative correlation with deforestation probability, while distance from the river and number of farm households showed the opposite.In typology 2, elevation was the most contributed driving force to the deforestation probability model, followed by other factors such as distance from the road, slope, distance from the river and the number of farm households, respectively.The relationship between the driving forces with the intensity of deforestation probability in typology 2 shown in . is Figure 4 Similar to typology 1, accessibility factors such as elevation and distance from the road have significant influence to deforestation in typology 2. Deforestation probability in areas having lower elevation, shorter distance from the road, flatter slope, away from the river and have , high number of farm households were predicted to have high probability of deforestation.On the contrary, areas having lower elevation, shorter distance from the road, near from the river but have high number of farm households and steeper slope were predicted by model as areas with low probability of deforestation.
Distance from the river in typology 2 was directly proportional to the deforestation probability, which means that deforestation in typology 2 was more commonly occurred in the forest area that far from the river as to compared to the forest area that close to the river.This happens due to the decrease in timber potential or forest areas in near the river.The phenomenon could be connected to many cases of deforestation occurred in the areas close to the river back in the 1970s 1980s, where rivers were used to − transport the timbers.The results are consistent with research Etter .( 2006) which states that deforestation in the et al Amazon region in Colombia has a positive relationship with the distance from the river.
In typology 2, one of the driving forces of deforestation was the distance to the road, in which contributed to higher deforestation probability in comparison to in typology 1. Areas belonged to typology 2 such as Riau Province has relatively more forest concessionaires (either active or inactive), where its road network was often used as means of illegal logging.Geist and Lambin (2002) suggested that the road construction could be a driving force of deforestation.
The study found that elevation was the most influencing driving forces towards the deforestation probability both in typology 1 and typology 2. The contribution of elevation in typology 1 gave higher effect in comparison to the elevation contribution in typology 2. Low topographic forest areas in typology 1 were more potentially undergone deforestation in comparison to low topographic forest areas in typology 2. Low topographic forest area in typology 1 such as in the east coast of North Sumatra Province was undergone massive forest conversion into oil palm plantations.Forests conversion to plantations in typology 2 occurred in areas with The deforestation models in typology 1 and typology 2 in this study were in line with the study results of deforestation models which mentioned that elevation factor and the closeness to the road were factors that affecting deforestation as built by Linkie . (2004) in the lowland forests of et al Sumatra Islands, Arekhi (2011) in the Northern Forest, Ilam Province.Iran and Kumar . (2014) in Kanker istrict, et al D Chhattisgarh Province, India also mentioned that deforestation was inversely proportional to the distance from the road, slope and elevation.
The number of farm households in typology 1 provided a higher contribution to the deforestation probability when compared to typology 2. This factor was directly proportional to the deforestation probability.This result showed that areas with a high number of farm households in typology 1 were more likely to undergo deforestation in comparison to those in typology 2. This result was different from Prasetyo . (2009) which stated that the percentage et al of farm households has a negative correlation with the deforestation probability in Java Island.
Generally, models of deforestation probability in each typology was formulated as shown in Equation Figure 4 Relations of driving factor with probability of deforestation in typology 2. typology 2. The result of overall accuracy assessment in the deforestation probability models were smaller than similar study in Java Island conducted by Prasetyo .( 2009), et al which obtained an overall accuracy of 88.70%.However, spatial modeling on deforestation was part of a complex land-use change model, thus the accuracy of more than 85% was difficult to achieved (Huang . 2007).et al Distribution value of deforestation probability during 2006−2012 in typology 1 was ranged between close-to-0 to 0.9533 (Figure 5a), while the distribution value of deforestation probability during 2006−2012 in typology 2 was ranged between close-to-0 to 0.9859 (Figure 6a).If the probability of deforestation less than 0.5, then it was classified as a non-deforestation, whereas if the probability of deforestation is higher than or equal to 0.5 then it was classified as a deforestation.The predictive deforestation maps of both typologies were depicted in Figure 5b and Figure 6b.Those areas that were predicted to have high deforestation probability in each typology should attract more serious attention from the government.Policy implications that could be taken including minimizing the development of road infrastructure that passes through wooded areas, both on the typology 1 and 2. This is due to the proximity of the road is one of the factors that have contributed greatly to the chances of deforestation in the typology 1 and typology 2.
Spatial modeling of deforestation on a large scale (Sumatra Islands) by using a typology of deforestation as done in this study, has generated different driving forces of deforestation on each typology.The results of this study confirmed that the triggering factors of deforestation in each locations were different and could not be generalized.The studies which simplify the population growth factor (Allen & Barnes 1985;Meyer & Turner 1992;Pahari & Murai 1999) as the cause of deforestation on a global scale becomes inappropriate.The information that oversimplified the factors in deforestation can lead to policies that are less precise in handling the issue.

Conclusion
Typological deforestation in Sumatra Islands could be classified into typologies, namely typology 1 with low 2 deforestation rate and typology 2 with high deforestation rate.The number of farm households could be used as the determining factor to assign an area into appropriate typology.The spatial model of deforestation probability in typology 1 was influenced by driving factors such as elevation, distance from the road, number of farm households, while in typology 2 the deforestation probability was influenced by elevation, distance from the road, slope, distance from the river, and number of farm households.The spatial model of deforestation in 2000 2006 was sufficient − to be used to predict the deforestation probability that occurred during 2006 2012 with an overall accuracy of − 65.37% for typology 1 and 72.24% for typology 2 , respectively.
into raster format sized 250 × 250 m, and the data of biophysical driving factors (x −x ) were raster 5 9 data having 250 × 250 m pixel size.In general, the logistic regression equation is formulated as shown in Equation [3].
as shown in Equation [4].[4] note: R out score = score resulted from rescaling E input score = estimated score of input E min score = minimum value of estimated score E max score = maximum value of estimated score R max score = highest score resulted from rescaling (255) R min score = lowest score resulted from rescaling (0)

Figure 1 Figure 2
Figure 1 Flow charts stages of development of deforestation spatial models using binary logistic regression.

Figure 3
Figure 3 Relations of driving factor with probability of deforestation in typology 1.Total of household farmer ( ), elevation (), distance to road ( ).

Figure 5
Figure 5 The distribution of deforestation probability of typology 1 in 2006−2012 based on a logistic regression model (a), Deforestation predicticve model of typology 1 in 2006−2012 based on a logistic regression model (b).

Figure 6
Figure 6 The distribution of deforestation probability of typology 2 in 2006−2012 based on a logistic regression model (a), Deforestation predicticve model of typology 2 in 2006−2012 based on a logistic regression model (b).

Table 1
Statistical test of deforestation models in typology 1 and typology 2