Fr
RF
a
ARRA
KFPDEL
1
ieieoesia(lefoWiu
h1
Ecological Indicators 45 (2014) 444–455
Contents lists available at ScienceDirect
Ecological Indicators
jo ur nal ho me page: www.elsev ier .com/ locate / ecol ind
orest cover dynamics analysis and prediction modeling using logisticegression model
akesh Kumar, S. Nandy ∗, Reshu Agarwal, S.P.S. Kushwahaorestry and Ecology Department, Indian Institute of Remote Sensing, Indian Space Research Organisation, Dehradun 248001, India
r t i c l e i n f o
rticle history:eceived 10 September 2013eceived in revised form 12 May 2014ccepted 13 May 2014
eywords:orest cover dynamicsredictionependent variablexplanatory variables
a b s t r a c t
Forest cover conversion and depletion are of global concern due to their role in global warming. Thepresent study attempted to study the forest cover dynamics and prediction modeling in Bhanupratap-pur Forest Division of Kanker district in Chhattisgarh province of India. The study aims to examine andanalyze the various explanatory variables associated with forest conversion process and predict forestcover change using logistic regression model (LRM). The forest cover for the periods 1990 and 2000,derived from Landsat TM satellite imagery, was used to predict the forest cover for 2010. The predictiveperformance of the model was assessed by comparing the model-predicted forest cover with the actualforest cover for 2010. To explain the effects of anthropogenic pressure on forest, this study consideredthree distance variables viz., distance from forest edge, roads and settlements, and slope position classes
ogistic regression model as explanatory variables of forest change. The highest regression coefficient (ˇ = −26.892) was noticedin case of distance from forest edge, which signifies the higher probability of forest change in areas thatare closer to the forest edges. The analysis showed that forest cover has undergone continuous changebetween 1990 and 2010, leading to the loss of 107.2 km2 of forest area. The LRM successfully predictedthe forest cover for the period 2010 with reasonably high accuracy (ROC = 87%).
© 2014 Elsevier Ltd. All rights reserved.
. Introduction
The change in forest cover is of global concern as forest is anndispensable natural resource that provides not only a wide vari-ty of ecosystem goods and services but also plays a vital rolen atmospheric carbon balance and thus climate change. Carbonmissions from deforestation and forest degradation are the sec-nd largest source of anthropogenic carbon emission (Le Quérét al., 2009; van der Werf et al., 2009). Although, landscape conver-ion varies significantly throughout the world, its ultimate outcomes mostly the same: extraction of natural resources for immedi-te human needs, often accompanied by biophysical degradationFoley et al., 2005). Hence, assessing the conversion of a forestedandscape may help us to understand the way the natural resourcesxtraction occurs, and consequently the human influences on theorest ecosystem services. For better understanding of the impactf forest cover change, factors affecting it must be fully studied.
ith rapid increases in population and continuing expectations ofmprovement in the standard of living, pressure on natural resourcese has become intense (Eastman, 2001). The consequences of an
∗ Corresponding author. Tel.: +91 135 2524175; fax: +91 135 2741987.E-mail address: [email protected] (S. Nandy).
ttp://dx.doi.org/10.1016/j.ecolind.2014.05.003470-160X/© 2014 Elsevier Ltd. All rights reserved.
ever-increasing pressure of human development have resulted inchanges of vegetation cover or depletion (Becek and Odihi, 2008;Kushwaha et al., 2011), degradation and fragmentation of habitats(Sun and Southworth, 2013), loss of wildlife corridors (Nandy et al.,2007) and an increased human–animal conflict (Gubbi, 2012).
Forest cover dynamics is actually the rate, pattern, spatial dis-tribution, quantity of change in the forest cover to other land coveror land use due to different natural or human induced causes. Theconstant interplay of various human-induced disturbances alongwith topographic and climatic factors can gradually degrade ahealthy forest cover or change it to other land use/land cover cat-egory. Understanding of the causes of land use change has movedfrom simplistic representations of two or three driving forces toa much more profound understanding that involves situation-specific interactions among a large number of factors at differentspatial and temporal scales.
Studies related to forest cover change using satellite-derivedinformation help in understanding the phenomena like carbondynamics, climate change and threat to biodiversity. Estimationof forest cover change and deforestation rate is a major challenge
without the use of satellite imagery, mainly in remote inacces-sible areas. Satellite remote sensing in combination with groundreconnaissance plays a vital role in determining the loss of for-est cover. Many studies have used remote sensing and geographicR. Kumar et al. / Ecological Indicators 45 (2014) 444–455 445
of stu
i2eReedcW2ctiwap
Fig. 1. Location
nformation system (GIS) for effective forest cover monitoring (FSI,011; Kushwaha, 1990; Kushwaha and Hazarika, 2004; Nandyt al., 2007; NRSA, 1983; Singh, 1989; Srivastava et al., 2002).ecent advancement in remote sensing and GIS methods alsonable researchers to model and predict land use/land cover morefficiently than ever-before. Several approaches have also beeneveloped to model and predict the dynamics of land use/landover (Arekhi, 2011; Houet and Hubert-Moy, 2006; Jenerette and
u, 2001; Lett et al., 1999; Ozah et al., 2010; Pontius and Schneider,001; Pontius and Malanson, 2005; Siles, 2009). Temporal forestover analysis linked with geospatial modeling help in genera-ion of future forest cover scenarios. Spatial modeling could be an
mmensely useful activity to understand the future of the forests,hich are undergoing continuous changes such as those broughtbout by deforestation, logging, diversion of forests for non-forestryurposes etc., provided such factors are operative in future too.
dy area in India.
The present study aims to examine and analyze explanatoryvariables associated with forest conversion process and to modelthe forest cover change using logistic regression model (LRM). Loza(2004) studied land cover change and closed forest fragmenta-tion, and established a logistic model to find out the causes offorest conversion in Bolivia. For logistic modeling, the effects of fiveindependent variables viz., distance from roads and settlements,land tenure, soil texture and topography on forest conversion werestudied. The best model of forest conversion (ROC = 71.5%) was pre-dicted by land tenure, distance from roads and settlements. Siles(2009) modeled forest conversion in Bolivia by considering changein forest areas as a categorical dependent variable and distance
from forest edge, roads, and settlements, landscape position andtype of settlement as explanatory variables. Logistic regression wasused for assessing the relative significance of explanatory variableson forest change and for predicting the probability of forest change.446 R. Kumar et al. / Ecological Indicators 45 (2014) 444–455
Landsat TM( 1990, 2000, 2010)
Supervised classification (MLC)
Forest-non forest cover
Ground Truth
Distance fro m roads
Distance fro m settlements
Forest cover change (1990-2000 & 2000-2010)
Slope position classes
Study area extraction
ASTERGDEM
Distance fro m forest edge
Roads(2000, 2010)
Settlements (2000, 2010)
Explanatoryvariables
Depend ent varia ble(Binary: 0 & 1): 1990-2000
Predicte d forest cover (2010)
Forest edge (2010)Settle ments (2010)
Roads (2010)Slope position classes
Probability of forest cover change (2010)
Model calibration (2000)
Prediction (2010)
Depend ent varia ble(Binary: 0 & 1 ): 2000-2010
Forest edge (2000)Settlements (2000)
Roads (2000)Slope position classes
act ua L
ogis
tic R
egre
ssio
n M
odel
igm o
Tirasttsaoiapefcta
2
2
eosa9hafm
Validati on with
Fig. 2. Parad
he study found that the landscape position was the most signif-cant explanatory variable, followed by distance from forest edge,oads and settlements. Logistic regression prediction resulted in anrea under a ROC Curve (AUC) of 85%. Arekhi (2011) modeled thepatial pattern of deforestation using GIS and logistic regression inhe northern forests of Ilam province of Iran. The study analyzedhe effects of six explanatory variables, viz., distance from roads,ettlements, and forest edge, forest fragmentation index, elevationnd slope, on deforestation. The LRM results indicated that mostf the deforestation occurred in the fragmented forest cover andn the proximity to forest edges. Slope and distance from roadsnd settlements had negative relationships with the deforestationattern. However, deforestation rate decreased with increase inlevation. The validation was tested using ROC approach which wasound to be 96%. The present study also considered the forest coverhange as the dependent variable in LRM and the factors like, dis-ance from roads, settlements, forest edge and topography, whichre supposed to be driving these changes, as independent variables.
. Materials and methods
.1. Study area
The present study was carried out in the Bhanupratappur For-st Division (19◦42′02′′–20◦13′32′′ N and 80◦24′11′′–81◦23′44′′ E)f Kanker district in Chhattisgarh province of India (Fig. 1). Thetudy area covers an area of 3157 km2. Almost half of the studyrea is mountainous. The altitude of the area ranges from 362 to90 m. The hillocks are very old and can be categorized into Vind-
yan, Arcian and Dharwar mountain ranges. The important riversre Kotri, Khandi, Medhaki, Bande, Waler, and Doodh which flowrom west to south direction and meet the Indrawati river. The cli-ate of this area is warm and humid. The average daily temperature
l for est cover of 2010
f the study.
is 22.69 ◦C and the average annual rainfall of 1239 mm (Agarwal,2005). The major forest types (Champion and Seth, 1968) existingin the area are: Southern Moist Mixed Deciduous Forest (3B/C2),Southern Dry Mixed Deciduous Forest (5A/C3), Dry Bamboo Brake(5/E9), Moist Peninsular Sal Forest (3C/C2e), Northern Tropical DryDeciduous Sal Forest (5B/C1e), Slightly Moist Teak Forest (3B/C1C),and Dry Teak Forest (5A/C1b).
The forest encroachment and illegal felling is highly prevalentin the area, which is evident from the record of Forest Departmentand was also observed during the field visit. As per the record ofthe Forest Department, 12.4% of the forest area has been illegallyencroached by the people in the past (Agarwal, 2005). This couldbe more now. Local people use wood for constructing and repair-ing their houses, manufacturing agriculture equipment and also asfuel wood. Nearly, every house in surrounding villages has its largewood lot of valuable forest wood, illegally removed from adjoiningforest. Every house in the village possesses large number of cattle,which graze illegally in the adjoining forest area. This disrupts theregeneration of the forest.
2.2. Data and materials
The detailed methodology of the present study is shown inFig. 2. Landsat TM satellite imagery of 17 November 1990, 09December 2000 and 12 December 2010 were used for the presentstudy. The satellite images were downloaded from Global LandCover facility (GLCF) (http://glcf.umd.edu) and U.S. Geological Sur-vey (USGS) Center for Earth Resources Observation and Science(EROS) (http://glovis.usgs.gov). The study area was extracted from
the temporal imagery by overlaying the boundary. The images of allthe three periods were interpreted using maximum likelihood clas-sifier (MLC) in ERDAS IMAGINE. The training sites for classificationwere collected from the field using a GARMIN-12 handheld GPSR. Kumar et al. / Ecological Indic
rcagbtaaat
Fig. 3. Forest cover during (a) 1990, (b) 2000, and (c) 2010.
eceiver. The supervised classification yielded four land use/landover types, viz. forest, agriculture, waterbody and non-forest. Thegriculture, waterbody and non-forest were merged into a sin-le category – non-forest and the remaining was forest. Finally aoolean map of forest and non-forest was generated for all thehree periods (Fig. 3). The accuracy of the classified images was
ssessed on the ground. The overall accuracy, producer’s and user’sccuracy, and Khat statistic of the classified image of 1990, 2000nd 2010 were estimated. The most common way to representhe classification accuracy of remotely sensed data is in the formators 45 (2014) 444–455 447
of an error matrix (Congalton, 1991). The overall accuracy wascomputed by dividing the total correct by the total number of pix-els in the error matrix. In addition, the total number of correctpixels in a category is divided by the total number of pixels ofthat category, which indicates the probability of a reference pixelbeing correctly classified and is the measure of omission error. Thisaccuracy measure is called producer’s accuracy because the pro-ducer of the classification is interested in how well a certain areacan be classified. On the other hand, if the total number of cor-rect pixels in a category is divided by the total number of pixelsthat were classified in that category, then this result is a mea-sure of commission error. This measure, called user’s accuracyor reliability, is indicative of the probability that a pixel classi-fied on the map/image actually represents that category on theground (Story and Congalton, 1986). Another discrete multivariatetechnique of use in accuracy assessment is called Kappa (Cohen,1960). The result of performing a Kappa analysis is a Khat statis-tic (an estimate of Kappa), which is another measure of accuracy(Congalton, 1991). ASTER GDEM (Global Digital Elevation Model),downloaded from Earth Remote Sensing Data Analysis Center(ERSDAC) (http://www.gdem.aster.ersdac.or.jp), was used for gen-erating slope position classes. Road and settlement maps of 2000and 2010 were collected from the Forest Department and digitizedin ArcGIS ver. 9.2 for generating the spatial layers of roads andsettlements for the respective years for use in the study.
2.3. The dependent variable: forest cover change
The forest cover change taken place between 1990–2000 and2000–2010 was considered as dependent variable (Fig. 4). Hence,a boolean image with the categories ‘forest change’ (forest tonon-forest) and ‘no change’ (forest remained unchanged) were gen-erated for the period 1990–2000 and 2000–2010 by subtracting theforest cover of 2000 from 1990 and 2010 from 2000 respectively.
2.4. The explanatory variables of forest cover change
Distances from forest edge, roads, settlements, and slope posi-tion classes were considered as potential explanatory variables offorest cover change (Figs. 5 and 6). Distance measures the Euclideandistance between each cell and the nearest of a set of target fea-tures. Forest edges have a high probability to be deforested (Ludekeet al., 1990) and studies showed that deforestation tends to startfrom the edge of existing forest (Eastman, 2006). Hence, distancefrom forest edge was considered as one of the explanatory vari-ables of forest cover change. The distance from forest edge for 2000(Fig. 5a) and 2010 (Fig. 6a) were generated from the forest bound-ary of the respective years using DISTANCE operator of IDRISI Taiga.Forest cover change is also highly related to proximity to roadsand settlements (Ludeke et al., 1990) and hence these were con-sidered as explanatory variables. To obtain the variable distancesfrom roads and settlements of 2000 (Fig. 5b, c) and 2010 (Fig. 6b,c), the DISTANCE operator of IDRISI Taiga was used.
Topography has certain influence on the forest cover change.Hence, slope position class was considered as one of theexplanatory variables of forest cover change. ASTER GDEMand topographic position index (TPI) (Weiss, 2001) were usedto generate slope position classes (Jennes, 2006) using Topo-graphic Tool 9.2 of ArcGIS 9.2. Thus the study area wasclassified into six categories of TPI, viz., valley (TPI ≤ −1SD),
lower slope (−1SD < TPI ≤ −0.5SD), flat slope (−0.5SD < TPI < 0.5SD,slope ≤ 5◦), middle slope (−0.5SD < TPI < 0.5SD, slope > 5◦), upperslope (0.5SD < TPI ≤ 1SD) and ridge (TPI > 1SD) (SD – Standard devi-ation of TPI values) (Figs. 5d and 6d).448 R. Kumar et al. / Ecological Indicators 45 (2014) 444–455
2000
2e
gwnpcvpatfIwVbcpnaa
Fig. 4. Dependent variable, forest cover change, during (a) 1990–
.5. Statistical test for association between dependent andxplanatory variables: Cramer’s V test
Cramer’s V is a statistic that transforms chi-square (for a contin-ency table larger than two rows by two columns) to a range of 0–1,here unit value indicates complete agreement between the twoominal variables (Liebetrau, 1983). The explanatory variable testrocedure is based on Cramer’s V contingency table analysis whichan test the strength of the association between the dependentariable and both quantitative (distance) and qualitative (slopeosition) variables. The test was performed using explanatory vari-ble test procedure of IDRISI. Before performing the explanatoryest procedure, the qualitative variable needs to be transformedrom nominal to numeric values. The Evidence of Likelihood tool ofDRISI was used to perform this transformation. Each slope category
as tested with the explanatory variable test based on Cramer’s, using dependent variable – the boolean image of forest changeetween 1990 and 2000. The results of the explanatory test pro-edure for each variable were Cramer’s V values and p values. The
value expresses the probability that the Cramer’s V is not sig-ificantly different from 0 (Eastman, 2006). Although Cramer’s Vssessed the relationship between an individual explanatory vari-ble and forest change, a deeper analysis was required to test the
, and (b) 2000–2010, used for calibration and prediction of LRM.
significance of each variable in the forest cover change process.Logistic regression model was used to give a better insight intothis.
2.6. Logistic regression model (LRM)
LRM was used to model and analyze the forest change in IDRISITaiga. The objective of the present study was to assess the impor-tance of the explanatory variables on forest change from 2000 to2010 and predicting the probability of change by 2010. The binarypresence or absence is the dependent variable (1 – forest changeand 0 – no change) for the periods 1990–2000 and 2000–2010. Inlogistic function, the probability of forest change is considered tobe a function of the explanatory variables. It is a monotonic curvi-linear response bounded between 0 and 1 (Pontius and Schneider,2001) and defined by the logistic function:
p = E(Y)eˇ0+ˇ1X1+ˇ2X2+ˇ3X3+ˇ4X4
1 + eˇ0+ˇ1X1+ˇ2X2+ˇ3X3+ˇ4X4(1)
where p is the probability of forest change, E(Y) is the expectedvalue of the dependent variable Y, ˇ0 is a constant to be estimated,ˇi is the coefficient to be estimated for each explanatory variableXi. This logistic function (Eq. (1)) can be transformed (Eq. (2)) into
R. Kumar et al. / Ecological Indicators 45 (2014) 444–455 449
F dge (2s
am
l
l
2
0mcobia
votsmw3totm
ig. 5. Explanatory variables used for calibration of LRM (a) distance from forest elope position classes.
linear function (Eq. (3)) which is called a logit or logistic transfor-ation:
ogit(p) = loge
(p
1 − p
)(2)
ogit(p) = ˇ0 + ˇ1X1 + ˇ2X2 + ˇ3X3 + ˇ4X4 (3)
.7. Model calibration and prediction
All the independent variables were normalized between 0.1 and.9 before introducing them in the model. The natural log transfor-ation was done for the continuous variables (distances). For the
ategorical explanatory variable (slope position class), the evidencef likelihood transformation was applied. The LRM was calibratedefore prediction by including the explanatory variables for 2000
n the IDRISI’s Logistic Regression Module as independent variablesnd the forest change during 1990–2000 as dependent variable.
The stepwise method was used to select the best set of predictorariables among the four predictor sets studied. In the first step,ut of four explanatory variables, one was considered at a timeo make four different one-variable models grouped as predictoret-1. In second step, predictor set-2 was formed with 2-variableodels where combinations of 2 different explanatory variablesere selected at a time. Similarly, predictor set-3 was formed with
-variable models where combinations of three different explana-
ory variables were selected at a time. Finally, predictor set-4 withnly one 4-variable model was formed considering all the explana-ory variables simultaneously. The best-fitted model with theinimum amount of predictors was done by Akaike Information
000), (b) distance from roads (2000), (c) distance from settlements (2000) and (d)
Criterion (AIC) Index, following the methodology used by van Gilsand Loza (2006). The regression equation of the best-fitted predic-tor set and the probability of forest change for 2000 were generated.The probability of forest change for 2000 was used to predict thechange in forest between 2000 and 2010. The dynamic variables,distance from roads, settlements and forest edge, were substitutedwith their respective variable for 2010 for new prediction and thevariable- slope position classes remained the same. The regressionequation of the best-fitted predictor set and the probability of forestchange for 2010 were generated. The threshold value was selectedfrom the residual file of probability of forest change of 2010 andit was used to generate change-no change binary image (0 = nochange (i.e. forest to forest, non forest to non forest) and 1 = change(forest to non-forest)). The no change areas were masked out fromthe binary image, and the rest of the image was intersected withthe forest non forest image of 2000 to incorporate the changes thathad taken place between 2000 and 2010. The resulted image is thepredicted forest cover for the period 2010.
From all the four predictor sets obtained by logistic regression,the model statistics like model chi-square, goodness of fit, pseudoR-square and AUC/ROC were calculated. Model chi-square is the dif-ference between −2 ln L (L = likelihood) for the best fitting modeland −2 ln L0 for the null hypothesis. The model chi-square valueoffers significance test for LRM (Ayalew and Yamagishi, 2005). Forassessing the significance of LRM, the goodness of fit is an alterna-
tive to model chi-square. It is calculated based on the differencesbetween the observed and the predicted values of the dependentvariable. The smaller the difference, the better is the fit (Hosmeret al., 1997). The pseudo R2 (1 − (ln L/ln L0)) value (1 – perfect fit and450 R. Kumar et al. / Ecological Indicators 45 (2014) 444–455
Fig. 6. Explanatory variables used for calibration of LRM (a) distance from forest edge (2010), (b) distance from roads (2010), (c) distance from settlements (2010) and (d)s
0(
2
(Ivaticteob
3
3v
ttae
lope position classes.
– no relationship) indicates how the logit model fits the datasetMenard, 1995).
.8. Validation of prediction
The predicted forest cover of 2010 was validated using ROC/AUCRelative Operating Characteristic/Area Under Curve) module ofDRISI Taiga. The ROC module is an excellent method to assess thealidity of a model that predicts the location of the occurrence of
class by comparing a suitability image depicting the likelihood ofhat class occurring (the input image) and a boolean image show-ng where that class actually exists (the reference image). The ROCurve is the true positive fraction vs. false positive fraction andhe AUC is a measure of overall performance. The predicted for-st cover map of 2010 was compared with actual forest cover mapf 2010 using 100 sampling points. A ROC/AUC curve was generatedetween the false positive (%) and true positive (%).
. Results and discussion
.1. Relationship between dependent variable and explanatoryariables
Fig. 7 shows the relationship between dependent variable and
he explanatory variables. This relationship was useful for selec-ion of the explanatory variables for forest change in the studyrea. Most of the forest change has taken place near the forestdges. Occurrences of forest change within 300 m of forest edgewere noticed to be highly distinct but forest change dropped offto virtually nil beyond that (Fig. 7a), indicating a non-linear rela-tionship between forest edge and forest change. The incidence offorest change near roads was found to be noticeable but droppedoff to zero after 2 km from the roads (Fig. 7b) and after around 3 kmfrom the settlements (Fig. 7c). Hence, these variables were testedwith Cramer’s V and then included in the model. The forest changefor each category of slope position class varied from class to class(Fig. 7d). Maximum forest cover change was found to have occurredon flatland and lower slopes. The hilltops were also observed tobe have undergone forest cover change due to mining activities.Hence, slope position classes were also tested with Cramer’s V toobserve their influence before use in the model.
3.2. Association between dependent variable and explanatoryvariables using Cramer’s V test
The strength of association between forest change and theexplanatory variables is shown in Table 1. For continuous variables,viz. distance from roads, settlements and forest edge, V values werefound to be between 0.5 and 0.6 with a p value of 0.0, indicat-ing a good association. Cramer’s V values for slope position classesshowed a strong association with forest cover change for the classes
flat slope (0.76), lower slope (0.68) and hilltop (0.53). On the otherhand, the classes valley (0.03), upper slope (0.01) and middle slope(0.28) have less association as less forest cover changes occurredunder these slope classes.R. Kumar et al. / Ecological Indicators 45 (2014) 444–455 451
Table 1Association between dependent variable (forest cover change) and explanatory vari-ables using Cramer’s V.
Explanatory variables Cramer’s V p value
Distance from roads 0.623 0.00Distance from settlements 0.534 0.00Distance from forest edge 0.573 0.00Slope position classes:
1. Hilltop/Ridge 0.53 0.002. Upper slope 0.01 0.003. Mid slope 0.28 0.004. Flat 0.76 0.00Valley 0.03 0.00Lower slope 0.68 0.00
Table 2Correlation analysis among the explanatory variables.
Slopepositionclasses
Distancefrom forestedges
Distancefrom roads
Distancefrom settle-ments
Slope positionclasses
1 −0.04 −0.48 −0.44
Distance fromforest edge
1 0.79 0.73
Distance fromroads
1 0.81
3
as0femtr
3
biLtvowdigspCtdbTa
g equ
atio
n
coef
fici
ents
of
4
sets
of
pre
dic
tor
vari
able
s.
ry
vari
able
sR
egre
ssio
n
coef
fici
ents
Pred
icto
r
Set-
1
Pred
icto
r
Set-
2
Pred
icto
r
Set-
3
Pred
icto
r
Set-
4
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
M13
M14
M15
3.00
4
70.9
57
0.89
9
1.60
1
71.5
57
0.27
6
0.64
58
71.5
32
71.5
60
1.64
5
72.9
83
72.0
60
71.5
78
0.97
8
72.1
55it
ion
clas
ses
−0.2
25
−0.0
37
−0.1
12
−0.1
44
−0.0
35
−0.0
34
0.10
7
−0.0
34fr
om
fore
st
edge
s
−26.
854
−27.
032
−26.
737
−26.
776
−26.
872
−26.
929
−26.
711
−26.
892
from
road
s−0
.524
−0.4
69
−0.1
53
−0.4
22
−0.1
48
−0.1
39
−0.3
83
−0.1
36fr
om
sett
lem
ents
−0.5
51
−0.4
68
−0.1
16
−0.1
90
−0.1
14
−0.0
27
−0.1
69
−0.0
26
Distance fromsettlements
1
.3. Correlation analysis among the explanatory variables
The correlation analysis among the different explanatory vari-bles viz., distance from roads, settlements, and forest edge,howed that there exist very high correlation coefficients between.7 and 0.8 (Table 2). High correlation existed between the distancerom forest edge and distance from roads and settlements, as forestdges are directly related to the establishment of roads and settle-ents. The lowest coefficients (less than zero) were found between
he slope position classes and the distance variables indicating noelationship between these variables.
.4. Logistic regression modeling
In the LRM analysis, four predictor sets were compared. Theest fitted predictor set was the combination of all the variables
ncorporated into the model. Tables 3 and 4 show the results ofRM for four predictor sets. The predictor set-4 was found to behe best combination for prediction. In the present study, highestalue (189,896.608) of model chi-square for the predictor set-4 wasbserved (Table 4). A high value indicates that the forest changeas less expected under the null hypothesis (without parametersriving forest change) than the full regression model (with includ-
ng parameters). The predictor set-4 also has the smallest value ofoodness of fit (913,971.504) indicating the better goodness of fittatistic among the predictor sets (Table 4). The pseudo R2 of theredictor set-4 is 0.299 indicating a relatively good fit (Table 4).lark and Hosking (1986) suggested that a pseudo R2 value greaterhan 0.2 indicates that the model is a relatively good fit for theata. Hensher and Johnson (1981) also stated that pseudo R2 valueetween 0.2 and 0.4 can be considered as extremely good fits.he predictor set-4 also attained the highest accuracy (ROC = 87%)mong all the predictor sets (Table 4).
Hence, the regression equation of best-fitted predictor set-4 is
iven below:logit(p) = ˇ0 + ˇ1X1 + ˇ2X2 + ˇ3X3 + ˇ4X4
logit(p) = 72.155 − 0.034(SPC) − 26.892(DFE) − 0.136(DR) − 0.026(DS)(4)
Tab
le
3R
egre
ssio
n
Exp
lan
ato
Inte
rcep
tSl
ope
pos
Dis
tan
ce
Dis
tan
ce
Dis
tan
ce
452 R. Kumar et al. / Ecological Indicators 45 (2014) 444–455
Fig. 7. Relationship between dependent and explanatory variables: (a) Forest cover change vs. distance from forest edge, (b) Forest cover change vs. distance from roads, (c)Forest cover change vs. distance from settlements, (d) Forest cover change vs. slope position classes.
Table 4Statistics of the 4 predictor sets obtained by logistic regression.
Model statistics Predictor set-1 Predictor set-2 Predictor set-3 Predictor set-4
Total number of pixels 6,768,877 6,768,877 6,768,877 6,768,877−2 ln L (L = likelihood) 764,690.125 761,921.100 758,482.982 760,059.907−2 ln L0 949,186.915 948,266.8733 947,177.486 949,956.515Model chi-square 184,496.790 186,345.772 188,694.504 189,896.608Goodness of fit 919,017.000 915,963.189 915,302.439 913,971.504
.196
.781
wSD
a(it(ewpavaaˇcra
The results of calibration for 2000 and prediction of LRM for2010 were presented in Fig. 8. The area under forest and non-forest during 1990, 2000 and 2010 (actual and predicted) is shownin Table 5. It was observed that the study area experienced a
Table 5Areas under forest cover during 1990, 2000 and 2010 (actual and predicted).
Class Area (km2)
Pseudo R-square 0.194 0AUC/ROC 0.758 0
here ˇ0 = intercept; ˇ1, ˇ2, ˇ3 and ˇ4 = regression coefficients;PC = slope position classes; DFE = distance from forest edges;R = distance from roads; DS = distance from settlements.
The relative contribution of the explanatory variables was evalu-ted using the corresponding coefficients in the LRM. The interceptˇ0) can be the value for the dependent variable when eachndependent variable takes zero value (Eastman, 2006). Based onhe coefficients values, all the explanatory variables were rankedTable 3). Among the continuous variables, distance from forestdges was the best single predictor for forest change (1990–2010),ith a ˇ2 value of −26.892 (Table 3, Eq. (4)). This means that therobability of forest change increases near the forest edges thanway from the edges. In other words, the model assigns higheralues of probability of forest cover change to the areas whichre closer to the forest edges. The variables, distance from roadsnd settlements have the regression coefficients of ˇ3 = −0.136 and
4 = −0.026 respectively indicating their significance to the forestonversion process (Table 3, Eq. (4)). Many studies have attributedoad infrastructure to one main cause of forest cover change. Geistnd Lambin (2001) and Krutilla et al. (1995) emphasized that the0.199 0.2990.787 0.872
construction of roads requires clearing of vegetation which leadsto deforestation and therefore greater access to forests can accel-erate the rate of forest cover change. Meanwhile, the categoricalvariable, slope position classes, also has good negative association( ̌ = −0.034) with forest cover change (Table 3, Eq. (4)). It meansthat with the increase in slope, forest change decreases due to lessaccessibility to forest.
1990 2000 2010 (actual) 2010 (predicted)
Forest 2040.7 1937.5 1837.5 1787.6Non-forest 1116.7 1219.9 1319.9 1369.8Total 3157.4 3157.4 3157.4 3157.4
R. Kumar et al. / Ecological Indicators 45 (2014) 444–455 453
Fig. 8. (a) Probability of forest cover change and (b) predicted forest cover map for the period 2010.
Table 6Accuracy estimates of forest cover maps of 1990, 2000 and 2010.
Year Error matrix Accuracy
Class Reference data
Forest Non-forest Total
1990 Forest 35 4 39 • Overall = 90.0%• Producer’s (%)(Forest = 94.59, Non-forest = 82.61)• User’s (%)(Forest = 89.74, Non-forest = 90.48)Khat = 0.78
Non-forest 2 19 21Total 37 21 60
2000 Forest 33 4 37 • Overall = 90.0%• Producer’s (%)(Forest = 94.29, Non-forest = 84.00)• User’s (%)(Forest = 89.19, Non-forest = 91.30)Khat = 0.79
Non-forest 2 21 23Total 35 25 60
2010 Forest 34 1 35 • Overall = 91.6%• Producer’s (%)(Forest = 89.47, Non-forest = 95.45)• User’s (%)(Forest = 97.14, Non-forest = 84.00)Khat = 0.82
Non-forest 4 21 25Total 38 22 60
454 R. Kumar et al. / Ecological Indic
0
20
40
60
80
100
1 21 41 61 81 101
Tru
e po
siti
ve (
%)
True posi tive (% )
False positive (%)
False Pos itive (%)
AUC/ROC of LRM = 87%
cfmh
capdttcbrfwFtf(tw(adtfep
tecicit(eoswM
20 (1), 37–46.
Fig. 9. Validation of LRM prediction (AUC/ROC).
ontinuous loss of forest cover, leading to the loss of 107 km2 oforest during 1990–2010. The accuracy estimate of the forest cover
aps of 1990, 2000 and 2010 is shown in Table 6. All the maps wereaving accuracy of more than 90%.
ROC/AUC graph generated between model predicted forestover and the actual forest cover of 2010 is shown in Fig. 9. Therea under ROC curve is 0.87 which gives an accuracy of 87% for theredicted forest cover of 2010. Loza (2004) studied five indepen-ent variables-distance from roads and settlements, land tenure,opography, and soil texture for forest conversion study and foundhat the distance from road and settlements were highly signifi-ant predictor of forest conversion, the ROC of prediction of changeeing 71.50%. In another study, Siles (2009) modeled distance fromoads, settlements and forest edge, and topography for tropicalorest conversion and found that all these explanatory variablesere highly significant predictor of forest conversion (ROC = 87%).
ive predictors of forest conversion viz., land tenure regime, dis-ance from roads and settlements, topography, and soil suitabilityor farming were tested in a spatial model by van Gils and Loza2006). They found that only three viz., land tenure regime, dis-ance from roads and settlements, out of the five predictors testedere found to be reliable predictors of forest conversion. Arekhi
2011) studied the effects of six factors, viz., distance from roadsnd settlements, forest fragmentation index, elevation, slope andistance from the forest edge, on deforestation. An inverse rela-ionship of forest change with distance from roads, settlements,orest edge and elevation was observed. Fragmentation index andlevation had positive relationship with forest change. The ROC ofrediction of forest cover change was found to be 96%.
The driving factors of forest cover change may vary from placeo place. In the present study, the selected explanatory variablesncompass a substantial share of the factors driving forest coverhanges. Specifically, the accessibility variables seem to be moremportant than the topographical ones. Many studies have indi-ated that most of these factors were also found to be importantn other areas. Proximity to road, town and forest edge were foundo be the important factors of forest change in southern CameroonMertens and Lambin, 1997). Linkie et al. (2004) studied that thelevation and proximity to road were the most influential factorsf forest change in the lowlands of Sumatra, Indonesia. Elevation,
lope, proximity to road, settlement and proximity to forest edgeere found to be the main drivers’ forest cover change in southeastexico (Mas et al., 2004).ators 45 (2014) 444–455
4. Conclusions
The analysis of forest cover change revealed that the forest areahas undergone continuous change leading to the loss of 107.2 km2
forest in the past 1990–2010. The various anthropogenic pressuresin the form of illegal logging and lopping, agriculture expansion,cattle grazing, encroachments etc. have resulted into the loss offorest resources in the area. It was found that the explanatoryvariables, viz., distance from forest edge, roads and settlements,slope position classes were significantly associated with the forestcover change in the study area. The highest regression coefficient( ̌ = −26.892) for distance from forest edge signifies the higherprobability of forest change to areas which are closer to the for-est edges. The frequency of forest cover change gradually decreasedwith increase in the distance from the roads, settlements and forestedges and becomes nil beyond 3 km.
The LRM efficiently modeled all the explanatory variables asso-ciated with forest change and helped in analyzing them for theirrelative significance on forest change process in the area. The pre-dicted set-4 comprising all the four explanatory variables in LRM,was found to be the best model for predicting the forest coverchange. It predicted the forest–non forest cover of the area for 2010with an accuracy of 87%. The present study dealt with four explana-tory variables of forest cover change and prediction, which could beimproved by incorporating more explanatory variables like socio-economic data, forest fragmentation, etc. In the present study onlytwo land cover categories (forest and non-forest) were considered.Further studies involving additional variables could elaborate theeffect of explanatory variables on forest conversion processes. Thisapproach is expected to be helpful in wildlife habitat monitoringand prediction as also for studying the shrinkage/expansion of habi-tat of ecologically and economically important species. The forestprediction maps can be useful to protected area managers for con-servation as well as management purpose and thus can be utilizedfor long-term sustainable management of forests.
Acknowledgments
The authors are thankful to the Director, Indian Instituteof Remote Sensing, Dehradun for encouragement and supportthroughout the study. The authors wish to acknowledge DivisionalForest Officer, Kanker Forest Division, Government of Chhattisgarhand the officers and staff of Chhattisgarh Forest Department forproviding field support.
References
Agarwal, S.K., 2005. Working Plan for East and West Bhanupratappur Forest Divi-sion, Kanker Forest Circle, Kanker (2006-07 to 2015-16), vol. I. Forest DivisionWorking Plan, Durg, Chhattisgarh.
Arekhi, S., 2011. Modeling spatial pattern of deforestation using GIS and logisticregression: a case study of northern Ilam forests, Ilam province, Iran. Afr. J.Biotechnol. 10 (72), 16236–16249.
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regressionfor landslide susceptibility mapping in the Kakuda-Yahiko Mountains, CentralJapan. Geomorphology 65, 15–31.
Becek, K., Odihi, J.O., 2008. Identification and assessment of factors affecting forestdepletion in Brunei Darussalam. In: The International Archives of the Pho-togrammetry, Remote Sensing and Spatial Information Sciences, vol. XXXVII,Part B2, Beijing.
Champion, H.G., Seth, S.K., 1968. A Revised Survey of the Forest Types of India.Manager of Publications, Government of India, New Delhi.
Clark, W.A.V., Hosking, P.L., 1986. Statistical Methods for Geographers. Wiley, NewYork.
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educ. Psychol. Meas.
Congalton, R.G., 1991. Review of assessing the accuracy of classifications of remotelysensed data. Remote Sens. Environ. 37, 35–46.
Eastman, J.R., 2001. IDRISI 32. Guide to GIS and Image Processing. Clark Labs, ClarkUniversity, Worcester, MA.
l Indic
E
F
F
G
G
H
H
H
J
J
K
K
K
K
L
L
L
R. Kumar et al. / Ecologica
astman, J.R., 2006. IDRISI 15 Andes. Guide to GIS and Image Processing. Clark Uni-versity, Worcester, MA.
oley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, S.,Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard, E.A.,Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankutty, N., Snyder, P.K.,2005. Global consequences of land use. Science 309, 570–574.
SI, 2011. India State of Forest Report 2011. Forest Survey of India, Ministry ofEnvironment and Forests, Dehradun.
eist, H.J., Lambin, E.F., (LUCC Report Series No. 4) 2001. What drives tropicaldeforestation? A meta-analysis of proximate and underlying causes based onsub-national case study evidence. CIACO, Louvain-la-Neuve.
ubbi, S., 2012. Patterns and correlates of human–elephant conflict around a southIndian reserve. Biol. Conserv. 148 (1), 88–95.
ensher, D.A., Johnson, L.W., 1981. Applied Discrete Choice Modeling. Croom Helm,London.
osmer, D.W., Hosmer, T., Le Cessie, S., Lemeshow, S., 1997. A comparisonof goodness-of-fit tests for the logistic regression model. Stat. Med. 16,965–980.
ouet, T., Hubert-Moy, L., 2006. Modelling and projecting land use and land coverchanges with a cellular automaton in considering landscape trajectories: animprovement for simulation of plausible future states. EARSeL eProc. 5 (1),63–76.
enerette, G.D., Wu, J., 2001. Analysis and simulation of land use change in the centralArizona-Phoenix region, USA. Landsc. Ecol. 16 (7), 611–626.
ennes, J., 2006. Topographic Position Index. tpi jen.avx, extension for ArcView 3.x;v. 1.3a. Jenness Enterprises, Available at: http://www.jennessent.com/arcview/tpi.htm
rutilla, K., William, F.H., Doug, B., 1995. Periurban deforestation in developingcountries. For. Ecol. Manag. 74 (2), 181–195.
ushwaha, S.P.S., 1990. Forest type mapping and change detection from satelliteimagery. ISPRS J. Photogramm. Remote Sens. 45, 175–181.
ushwaha, S.P.S., Hazarika, R., 2004. Assessment of habitat loss in Kameng andSonitpur Elephant Reserves. Curr. Sci. 87 (10), 1447–1453.
ushwaha, S.P.S., Nandy, S., Ahmad, M., Agarwal, R., 2011. Forest ecosystemdynamics assessment and predictive modelling in eastern Himalaya. In: Interna-tional Society for Photogrammetry and Remote Sensing Archives, vol. XXXVIII,8/W20, Workshop Proceedings: Earth Observation for Terrestrial Ecosystems, 8November, Bhopal, India.
e Quéré, C., Raupach, M.R., Canadell, J.G., Marland, G., Bopp, L., Ciais, P., Conway,T.J., Doney, S.C., Feely, R.A., Foster, P., Friedlingstein, P., Gurney, K., Houghton,R.A., House, J.I., Huntingford, C., Levy, P.E., Lomas, M.R., Majkut, J., Metzl, N.,Ometto, J.P., Peters, G.P., Prentice, I.C., Randerson, J.T., Running, S.W., Sarmiento,J.L., Schuster, U., Sitch, S., Takahashi, T., Viovy, N., van der Werf, G.R., Wood-ward, F.I., 2009. Trends in the sources and sinks of carbon dioxide. Nat. Geosci.2, 831–836.
ett, C., Silber, C., Barret, N., 1999. Comparison of a cellular automata network andan individual-based model for the simulation of forest dynamics. Ecol. Model.121, 277–293.
iebetrau, A.M., 1983. Measures of Association (Quantitative Applications in theSocial Sciences, No. 32). SAGE Publications, Newbury Park, CA.
ators 45 (2014) 444–455 455
Linkie, M., Smith, R.J., Leader-Williams, N., 2004. Mapping and predicting deforesta-tion patterns in the lowlands of Sumatra. Biodivers. Conserv. 13, 1809–1818.
Loza, A.V., (M.Sc. Thesis) 2004. Spatial logistic model for tropical forest conversion:a case study of Carrasco province (1986–2002), Bolivia. International Institutefor Geoinformation Science and Earth Observation, Enschede, The Netherlands.
Ludeke, A.K., Maggio, R.C., Reid, L.M., 1990. An analysis of anthropogenic deforesta-tion using logistic regression and GIS. J. Environ. Manag. 31, 247–259.
Mas, J.F., Puig, H., Palacio, J.L., Sosa-López, A., 2004. Modelling deforestation usingGIS and artificial neural networks. Environ. Model. Softw. 19, 461–471.
Menard, S., (Quantitative Applications in the Social Sciences, No. 106) 1995. AppliedLogistic Regression Analysis. SAGE Publications, Thousand Oaks, CA.
Mertens, B., Lambin, E.F., 1997. Spatial modeling of tropical deforestation in south-ern Cameroon: spatial disaggregation of diverse deforestation processes. Appl.Geogr. 17, 143–162.
Nandy, S., Kushwaha, S.P.S., Mukhopadhyay, S., 2007. Monitoring Chilla-Motichurcorridor using geospatial tools. J. Nat. Conserv. 15 (4), 237–244.
NRSA, (Technical Report) 1983. Forest cover mapping in India from satellite imageryfor the periods, 1972–75 and 1980–82 using Landsat MSS imagery. NationalRemote Sensing Agency, Hyderabad.
Ozah, A.P., Wever, T., Ghys, L., Weissmann, T., 2010. Prospects, challenges andstrategies in the implementation of the Nigerian computerised mining infor-mation system. In: International Society for Photogrammetry and Remotesensing (ISPRS): Core spatial databases-maintenance and services-from theoryto practice, Archive, vol. XXXVIII, part 4-8-2-W9, Haifa, Israel, 2010.
Pontius Jr., R.G., Malanson, J., 2005. Comparison of the structure and accuracy of twoland change models. Int. J. Geogr. Inf. Sci. 19 (2), 243–265.
Pontius Jr., R.G., Schneider, L.C., 2001. Land-cover change model validation by an ROCmethod for the Ipswich watershed, Massachusetts, USA. Agric. Ecosyst. Environ.85 (1–3), 239–248.
Siles, N.J.S., (M.Sc. Thesis) 2009. Spatial Modelling and prediction of tropical for-est conversion in the Isiboro Secure National Park and Indigenous Territory(TIPNIS), Bolivia. International Institute for Geoinformation Science and EarthObservation, Enschede, The Netherlands.
Singh, A., 1989. Digital change detection technique using remote sensing data. Int.J. Remote Sens. 26, 1–6.
Srivastava, S., Singh, T.P., Singh, H., Kushwaha, S.P.S., Roy, P.S., 2002. Mappingof large-scale deforestation in Sonitpur district, Assam. Curr. Sci. 82 (12),1479–1484.
Story, M., Congalton, R., 1986. Accuracy assessment: A user’s perspective. Pho-togramm. Eng. Remote Sens. 52 (3), 397–399.
Sun, J., Southworth, J., 2013. Remote sensing-based fractal analysis and scale depend-ence associated with forest fragmentation in an Amazon tri-national frontier.Remote Sens. 5, 454–472.
van der Werf, G.R., Morton, D.C., DeFries, R.S., Olivier, J.G.J., Kasibhatla, P.S., Jack-son, R.B., Collatz, G.J., Randerson, J.T., 2009. CO2 emissions from forest loss. Nat.
Geosci. 2, 737–738.van Gils, H.A.M.J., Loza, A.V., 2006. What drives conversion of tropical forest inCarrasco province, Bolivia? Ambio 35 (2), 81–85.
Weiss, A., 2001. Topographic Position and Landforms Analysis. Poster presentation.ESRI User Conference, San Diego, CA.