Spatial Model of Deforestation in Kalimantan from 2000 to 2013

Forestry sector is the biggest carbon emission contributor in Indonesia which is mainly caused by deforestation. A significant area of forest cover still can be found in Kalimantan Island (one of the largest island in Indonesia) although an alarming rates deforestation is also exist. This study was purposed to established spatial model of deforestation in Kalimantan islands. This information is expected to provide options to develop sustainable forest management in Kalimantan trought optimizing environment and socio-economic purposes. This study used timeseries land cover data from the Ministry of Environment and Forestry (2000–2013) and is validated by SPOT 5/6 images in 2013. The spatial model of deforestation were developed using binary logistic. The results of logistic regression analysis obtained spatial model of deforestation in Kalimantan = 1.1480714 – (0.033262*slope) – (0.002242*elevation) – (0.000413*distance from forest edge) + (0.000045*Gross Regional Domestic Product). Validation test showed overall accuracy about 79.64% and 77.01% for models of deforestation in 2000–2006 and 2006–2013 respectively.


Introduction
Forestry sector is the biggest carbon emission contributor in Indonesia Yet, forest absorbs carbon (Boer et al. 2010).emissions photosynthesis processes and stores it in the by forest biomass.These plays are the main factors why forestry an important o tomitigate climate change.
r le Emission in forestry sector is mainly caused by deforestation for various purposes.The Ministry of Forestry (2014) stated that Indonesia's forest cover in Indonesia in 2013 was around 96.49 million ha out of its 187.9 million 2 ha land area ata shows that total .Furthermore, the d from the the deforestation rate in Indonesia during 2012−2013 about -1 0. 6 million about 0. 7 9 ha year , while the reforestation rate 2 -1 million .Data from other states that deforestation ha year Indonesia in the period 2000−2012 an average of around -1 0.84 million year and the reforestation of about 0.48 ha -1 million year (Margono 2014).ha The Kalimantan is part of Indonesia where the islands forest coverage is large with the high deforestation rate (IPSDH 2014).Kalimantan has experienced heavy deforestation and forest degradation during the past 2 decades (Langler et al. 2007).Generally, the trigger of deforestation is ocio-biophysical and s economic.Biophysical in example are the elevation and slope drives (Prasetyo et al. 2009;Kumar et al.2014).Socio-economic driver in example are the demographic and income rate (Romijn et al. 2013).Deforestation in Kalimantan are caused by elevation and high demand of the farming or plantation area, in which also occurred in protected areas (Scriven et al. 2015).Burn et al. (2015) stated that economic force will impose strong pressure on Kalimantan forests.This condition requires good management to preserve the forests of Kalimantan and to avoid damage to forests such as Java (Prasetyo et al. 2009) and Sumatra (Margono et al. 2012) where forest cover approximately 30% of the total land area.
Spatial model can be u as a tool to find the factors that sed significantly contribute deforestation.Mas .(2004) to et al described that deforestation model potentially gave more benefits : providing a better understanding of that consist of how driving factors govern deforestation, generating scenario of deforestation rate in the future, predicting location of deforestation and support the design of policy ing responses to deforestation.Spatial model of deforestation is expected to provide a more detail information on deforestation in Kalimantan in order to develop plans drivers of sustainable forest management with optimal environment and economics functions.The spatial model of sociodeforestation can be constructed with a variety of techniques or methods.Logistic regression approach is proven to be able to be u ed in analyzing deforestation (Arkehi 2013).s Park (2013) compared the development of critical model by using frequency ratio (FR), analytic hierarchy process (AHP), logistic regression (LR), and artificial neural network (ANN).Those 4 methods give values of accuracy which are not too differently and logistic regression is considered as the most optimal method ogistic regression be used  2011), high resolution image can be used as a reference to validate medium or low resolution image.Validation of locations were systematically arranged according to the national forest inventory (NFI) plots (DirjenPlan 2014).Those combinations resulted in 307 checking points (Figure 2).Validation method using higher resolution image can save up time and cost.The next procedure was the accuracy assesment.The widely used accuracy calculation, i.e error matrix (overall accuracy, producer accuracy and user accuracy) was used in this study (Foody 2002).The drivers of deforestation Several spatial explanatory variables (Table 2) describing potential proximate causes of deforestation are generated as follow: 1 Slope and elevation were constructed from SRTM 30 m resolution with raster surface analysis.2 Peatland were constructed from the MoA's peatland data

Developing eforestation patial
n According to Menard (2002) when VIF value is bigger than 10 indicates that there is a problem in the multicolinearity.Treshhold VIF widely used in study is 10 (O'berin 2007).When variable is indicated having multicolinearity then it has to be eliminated.When there is no problem with multicolinearity in the variable, then it can be proceed to logistic regression analysis.

Calibration of the model
The feasibility of the logistic regression model was showed by result of the value -2 Log Likelihood and Hosmer and Lemeshow test.A model is considered as feasible when there is reduced value of -2 Log Likelihood and the significance value of Hosmer and Lemeshow test bigger than 0.05 (Hosmer and Lemeshow 2004).Discriminative test was conducted to obtain information of how valid the model to differentiate the probability of deforestation.Discriminative test was performed by calculating the reciever operating characteristic (ROC) value (Dahlan 2012).

Spatial model of deforestation
The developed model of deforestation of 2000−2006 later was integrated into the chosen variable map to generate a deforestation probability map of 2000−2006.The generated probability map was later classified into a changing category and unchanging category by using cut value.Cut value was chosen based on the value of the highest kappa accuracy of the generated probability map (Fielding & Bell 1997)

Results and Discussion
Land cover data accuracy Land cover validation of MoEF data of 2013 using available SPOT 5/6 image of 2013 (Figure 2) resulted 96.09%, 95.00% and 96.38%.for forest and nonforest classes of overall accuracy, producer accuracy and , user accuracy.These results were not widely different from the calculation of overall accuracy on land cover data from all over Indonesia (wall to wall) in 2011 that resulted the accuracy 98% for forest and nonforest (IPSDH 2012).Margono (2014) . 2000).In Kalimantan the steep slope lies in the et al northern part of island.
The analysis results showed that the distance from the forest edge effect is negative, it means that the long distance of forest edge causes the deforestation occur is getting lower.Arkehi (2013) states that the distance of the forest edge are factors that influence the occurrence of deforestation, forest closer to the chances of deforestation is increasing.
The results showed that GDRP gave a positive effect, meaning that the greater GDRP the higher the incidence of deforestation.Romijn . (2013) and Ewers (2006) states et al that GDRP is a factor that affects the occurrence of deforestation, the higher the GDRP of a region, the higher the chances of deforestation.The correlation between GDRB and deforestation must be assessed with caution because they are less clear about whether deforestation later declined if countries become richer (Kaimowitz Angelsen & 1998).

Fitting of the model
The results of feasibility model test by using Hosmer and Lemeshow test showed that model was fit since it held statistical significance 0.424 (> 0.05).Nagelkerke R value described that 46.6% variation was 2 explainable by model, meanwhile the rest was explained by factors other than model.The generated Nagelkerke R value 2 was only an approached value, because in logistic regression, determination coefficient can not be calculated as in linier regression.Discrimination test is conducted by calculating ( ) r o c value.The ROC ( eceiver perating haracteristic) generated ROC value from this model is 84.4%.This value is considered as strong category (80-90%) (Dahlan 2012).

Model implementation
The developed model implementation by using variable map resulted in deforestation probability map 2000−2006.Later, the deforestation probability map was developed into deforestation model map 2000−2006 based on cut value 0.81 (Figure 4a and Table 4).Thus, for probability value (0 < P < 0.81) was non-deforested areas, while probability value (0.81 < P < 1) was deforested areas.5 This result considered satisfactory, because the complexity of landuse change (deforestation) make it difficult to make model with an accuracy more than 85% (Huang 2006).Other spatial model which used logistic regression at different location with different variable resulted in overall accuracy 65.51%, an unchanged user accuracy 65.55% and a changed user accuracy 61.10% (Park et al. 2013).Huang (2006) unchanged overall accuracy result 71% and changed user accuracy 73%.Prasetyo .( 2009) et al result overall accuracy 88.70%, producer accuracy, and user accuracy at non-deforested areas 95.76% and 92.44%.While the producer accuracy and user accuracy for deforested areas 2.97% and 13.64%.

Conclusion
This study developed spatial model of deforestation in the Kalimantan using GIS and logistic regression.
developed by using logistic regression based on the data of 2000−2006 and then the generated model was applied for forcasting the 2006−2013 period and was validated by using actual land cover data within 2006−2013.Deforestation was identified by analyzing natural forest conversion into non-natural forest.This data of deforestation was later used as dependent variable.
. Then the deforestation model of 2000−2006 using the best selection cut value was use to generated deforestation prediction map of 2006−2013 by adjusting the variables.The deforestation prediction map 2006−2013 later was validated with deforestation actual data of 2006−2013 from MoEF in order to obtain overall accuracy, producer accuracy and user accuracy values.

Figure 3
Figure 3 Distribution of samples used in the logistic regression.
Indonesia use methods autologistic and von Thunen spatial-autoregressive models.Validation results indicated that both models had less different.The choice of the model will depend on data availability and purpose.The von Thunen model can be useful if the spatial data are scarce or available only at a single time of point.The autologistic model can provide higher accuracy, when the aim is to obtain more spatially accurate predictions than a mechanistic understanding of drivers of deforestation.Validation was conducted by comparing MoEF's landcover map 2013 and the available high resolution image SPOT 5/6 of 2013.According to Doris and Cardille (

Table 1
Data used and iteration technique Figure 2 Cheking point validation of MoEF land cover map 2013 year.The test results multicollinearity between independent variables n Table2 slopes and elevation are the factors that influence deforestation, when the slopes is going steeper the occurrence of deforestation is getting lower.The steep of slope is usually avoided in logging activity or forest conversion because it is relatively more difficult in practice and requires higher costs (Burn

Table 3
Result of logistic regression

Table 4
Best selecting cut off of probability map of deforestation

Table 5
The test result validation of deforestation model