18 and May,12
th
2001 downloaded from GLCF Global Land Cover Facility http:glcf.umiacs.umd.eduindex.shtml
2. Topographic Map is obtained from SRTM Shuttle Radar Topographic Mission format and downloaded from GLCF Global Land Cover
Facility http:glcf.umiacs.umd.eduindex.shtml with acquisition data 2000
3. Digital Map of Village, Sub-district, was received from PPLH-IPB, in vector or shp format
4. Digital Map of Nature Preserve Cagar Alam Cibanteng and Wildlife Reserve Suaka Margasatwa Citepuh boundary, was received from
Baplan Badan Planologi Forestry Planning Agency Ministry of Forestry.
5. Demographic and socio-economic data, which is collected from BPS Badan Pusat Statistik Bureau of Statistical Center Sukabumi
Province and Head Office Jakarta.
3.3. Supporting ToolsProgram
In this research, supporting tools used, as the terms of software and hardware are as the followings:
1. Software: • ERDAS Imagine 8.7 used for image processing
• ArcView 3.3, used for spatial data processing • See5 C5 Ver. 2, used for classification process and combined
with ERDAS Imagine. • SPSS version 11.5, used for statistical calculation and analysis.
2. Hardware: • PC Dual Processor Xeon™ Intel® Pentium® IV CPU 2.99 GHZ,
1 Gb DDR-SDRAM, Video System NVIDIA GeForce 6600 256 Mb, working at Operating System Window XP Service Pack 2
• Global Positioning System GPS Garmin Type E-Trex Vista, property of PPLH IPB Bogor
19
3.4. Methodology
On the whole of research procedure will be illustrated in Figure 3.2.
3.4.1. Image Preprocessing
The objectives of image preprocessing are to remove some error because of the radiance measured by any given system over a
given object on earth surface is influenced by such factors as changes in scene illumination, atmospheric conditions, viewing geometry, and
instrument response characteristics.
a RadiometricAtmospheric Correction . Histogram Adjustment
is one of method that can be used. Histogram adjustment is used to minimize atmospheric bias. This process is common process
to pre image processing, but in this thesis this process was not done since the ERDAS Imagine can accommodate to identify the
similarity of DN Digital Number value for each area of interest for training sites or signatures area b using mean plot tool. Only
image enhancement method that will be done for improving the visual interpretability in order to increase the apparent distinction
between the features in that Landsat scenes, such as contrast stretching by standard deviation in ERDAS Imagine 8.7
b Geo-Referencing, raw digital image usually contain geometric
distortions so significant that they can not be used directly as a map base without subsequent processing. The sources of these
distortions range from variations in the altitude, attitude, latitude, and velocity of the sensor platform to factors such as panoramic
distortion, earth curvature, atmospheric refraction, relief displacement, and non-linearities in the sweep of a sensor’s
IFOV. The aim of geometric correction is to compensate for the distortions introduced by these factors so that the corrected image
will have the highest practical geometric integrity.
Knowledge Base Classification
Image Stacking
ERDAS img Importing and Reprojection
Image Pre- processing
Classification Processing
Image Analysis
Data Collecting
Satellite Imagery Landsat TM:
- 1990
- 1997
- 2001
DigitalHardcopy Maps -
TopographicContour Map -
Road, River, Coast Line -
Adm Boundary
Secondary Data
Statistical Data
Deforestation 1997 - 2001
1990 1997
Overlaid Cropping using ERDAS Modeler
Maker Visual
Verification Yes
Convert to GRID ArcView
Signature Ares by using Shapefile Polygon
ERDAS - CART Sampling Tool
CART See5 Classification
Converting to Shapefile
ArcView Polygon
Vector Cell 1997-2001
Corrected Image
Logistic Regression
Analysis Vector
Analysis No
Deforestation 1990 -1997
ArcView Polygon
Vector Cell 1990-1997
Model Implementation
Validation Yes
Spatial Modeling
2001 Image Stacking
Converting Continuous to Thematic ERDAS
SEE5 C5 Classifier Construction
ERDAS Recoding
= Data = Process
= Result = Period
1997-2001 = Decision
= Process Direction = Main Process
Boundary = Period 1990-1997
No Defining of
Independent Variable Value for each binary
of Independent
variable
Figure 3.2. Flow chart of research activities and procedures
20
21 Random distortions and residual unknown systematic distortions
are corrected by analyzing well-distributed ground control points GCPs occurring in an image. From another corrected-image,
GCPs are features of known ground location that can be accurately located on the uncorrected image to become a new
corrected-image, also known as rectification process. Actually this was not done in this research, the images
from GLCF just need to be projected to UTM Zone 48 South and Datum WGS 84. In the process also was needed GCP Ground
Control Points, it has about 30 points with first polynomial order. A first order polynomial is normally suitable for a
transformation between two near recti-linear map systems. Landsat images are from year 1995 and 1997 acquisition
are in bsq binary sequential, so it needed to export bsq to img ERDAS format with certain and fix number of rows and
columns. Image acquired in the year 2001, was treated as master, and while the others 1990 and 1997 were referred to as slave
images.
3.4.2. Image Processing
The methodology of image processing in this research is using classification method that obtained from CABS Center for Applied
Biodiversity Science CI Conservation International Washington DC and developed together with Wildlife Conservation Society
WCS – Indonesia Program. In this method, the classification uses CART tool as one of ERDAS Imagine 8.7 plug in and stand alone
program See5 C5, as the additional program to support better classification.
a Image Stacking,
Both Image data of 122065_19901109 and 122065_19970728
have stacked each other, became 12 or 14 with Thermal Band 6 bandslayers into one file
22 122065_19901109_ 19970728
. This stacked-image is called Period 1990-1997. Period 1997 – 2001 that will be used for
validating, was done with the same procedure to stack the images 122065_19970728
with 122065_ 20010512 became 122065_19970728_20010512.
This stacking process always has remaining edge as the result of process, since the wide of both images are not same. This edge
must be cut by using ERDAS Imagine Modeler Maker
b Signature areas, to obtain the interest area was done by cropping
Ciemas and Ciracap Subdistrict based on digital administrative boundary data, with the same projection, and the last
subsetcropping again to obtained both conservation area Nature Reserve Cagar Alam Cibanteng and Wildlife Reserve Suaka
Margasatwa Cikepuh as the study area. Signature areas or training sites are used to define a certain class based on
visualization characteristic. Signature area was made in polygons shapefile format in ERDAS Imagine 8.7. Start to take a signature
area or training site is to zoom the interest area and make one representative polygon that make sure the characteristic spectral
inside that polygon is similar whether first date 1220065_19901109 has one class for instant forest and the
second date 122065_19970728 also has one class for instant non-forest. So the entire signature areas is stand alone layer in
vector format in ERDAS Imagine 8.7 In this study, 5 main land cover categories were distinguished:
- Forest, to identify forest that no change to another uses.
- Non-Forest, assumed as agriculture area, plantation, settlement,
shrub or bush, bare land, and even seasonal flooding -
Water body, to differ between water body such lake and river.
23 -
Cloud, but in the thesis, the land use or land cover will be estimated with the surround condition.
- Shadow, in the thesis, the land use or land cover will be estimated
with the surround condition, since binomial logistic regression can not use cloud and shadow class. So water, cloud and shadow, at
the last process will be included as non-forest category, if the cloud and shadow are among the non-forest area category
ID of each signature area polygon in shapefile format as attributes is one important thing to recode and classify the image.
Two digits can be used for 8 bits of classification result and 4 digits for unsigned 16 bits. Four digits will be more to determine
as many as classes in order to minimize DN Digital Number overlapping for each class.
Note field is just for describing the ID, example for making one particular a forest class to differ with the others by visualization
interpretation. The forest says swamp forest, according to the class that we want to extract only forest and non-forest, and
because of this fact we make particular ID for example start from 1140, 1141, and so on. At the last process, after getting the best
iteration, the image must be converted again to 8 bit data type and recode to common classes forest to forest 11, forest to non
12, non to non 22, water 44, cloud 55, and shadow 66. As mentioned before, that it is better to use 16 bits output data
type, to anticipate overlapping amount training sites, one training site will represent one type of spectral characteristic. In this case,
forest class was collected, and numbered with 4 digits. After obtaining all spectral characteristic of forest, it needs to verify the
similarity of forest spectral characteristic. This task can be done by using Mean Plot of Signature in ERDAS Imagine 8.7. So the
wrong signature area can be deleted and remaining only signature
24 areas which have had similarity of spectral characteristic.
However this technique is only for unchanged class such as forest to forest, and can not be used for changed forest to non,
because non forest class in this case may be depicted many land cover for example seasonal flooding, mixed agricultural or
farming, plantation, and so on. Mean plot of signature is a spectral profile of the mean data file
value of each signature in all bands of the image to be classified was obtained by changing the polygon in shapefile format to
AOI. It uses Copy Selection to AOI, from AOI menu in ERDAS Imagine 8.7 Viewer. After all classes assumed are right, the
polygon shapefile will be converted to grid by using ArcView 3.3.
ArcView program is used for converting shapefile of signature areas polygon into grid format. This grid must be same the pixel
size. Using ERDAS Imagine again is to import the Grid type into img format, with data type unsigned 16 bit in Import
Options. Open one ERDAS Imagine viewer again to see the result of importing process. In this process, the projection
metadata will loose, and it needs to be redefined. From ERDAS Image Information, change the Map Model, the unit into meters
and projection into UTM Universal Transfer Mercator and change Spheroid and Datum WGS 84, UTM Zone 48 South.
The result of above process is still continuous file type, and it must in thematic file type. This process uses Modeler Maker,
and the model will change continuous to thematic file type, with the same data type 16 bit.
c ERDAS Imagine – CART Classification. Classification process
uses CART, from ERDAS tool bar, and choose CART Sampling Tool , Independent Variable File is the original image file,
25 Dependent Variable File is thematic img file after converting
process, uses See5, and the last define the third file output .names, .data, and .test.
After that, process is continued by using stand alone program See5. From See5 main menu, choose File – Locate Data, and find
the path of .data as the result of CART Sampling Tool. Choose File – Construct Classifier, and CC Option window will appear,
check list Boost with 10 default value and the remaining let them as default value, Pruning CF 10 and Minimum 2 Cases in
Global Pruning option. See5 will run the classification process according to the dependent
and independent variable files. If the See5 process is success, the last classification process is to appear the thematic image. Back
to ERDAS Imagine tool bar menu, choose CART again and click Run See5.
Run See5 in submenu CART tool in ERDAS is the last process of classification. The result can be interpreted to the original image,
and if there are many mistaken, it can be iterated and so on. Iteration starts at defining the signature areas polygon again, and
continue the process. This result also can be used to further process, to convert to vector format. Since until this process, data
type 16 bit still inside, the raster must be converted again to 8 bit, in order to be easy to further process. The class ID in the raster
attribute data must be 11, 12, 22, 44, 55, or 66. Recoding of
ERDAS Imagine 8.7 is needed in this case.
Period 1997 – 2001 will be done by the same procedure. Image data 1997 is the same date with image data in period 1990 – 1997,
it is 122065_19901109 and 122065_20010512 for image data 2001.
26 In this case, with the assumption that deforestation in Period 1990
– 1997 will be non-forest in Period 1997 – 2001, since the image data 1997 will be the first date and image 2001 will be the second
date. Both images became one stacked image. By the reason, the deforestation class class number 12 will be clipped by all classes
of raster Period 1990 – 1997. Only class number 12 of Period 1997 – 2001 will be as input theme and Period 1990 – 1997 will
be as clip theme. This process is done by using Modeler Maker ERDAS Imagine 8.7
3.4.3. Vector Processing a Creating Cell Vector
, In this study as the nest for attribute data of each variables is cells. The cell is polygon with square shape,
with size 90 x 90 meters. These vector cells were created by
using ET Vector Grid, a plug in extension of ArcView 3.3. In its process, this tool just need four input extents points, Xmin, Xmax,
Ymin, and Ymax; and XY spacing or grid resolution 90 meters. The result of this process is a squared shape, which consist of cells
or grids.
Each cell has to be filled by either value 0 or 1 for independent
and dependent variables. Joining process to fill each cell, will be
easy by using ArcView Extension, Geoprocessing Assign data by location Spatial Join. The tool will join the ID of cell correlated
by either value 0 or 1, in an attribute table, and was saved by a
name Polygon VectorCell PVC. Name of PVC is based on
ArcView extension to create grid vector polygon, and to differ
with raster grid in this thesis, grid is changed to cell. The results
process are spatial and its attributes, that must be saved into a new
shapefile.
27
b Extracting Variables Data
In this research, dependent variable and independent variable will be list in Table 3.1.
Table 3.1. Binary data and categorization of variables as the factors of
deforestation
Variable Symbol
Value Category Variable
References
Unchanged Forest – Changed Forest
Y 0 Unchanged Forest –
Unchanged Non Forest Saadi, M. and R.
Abolfazi. 2003 1
Changed Forest Deforestation
Saadi, M. and R. Abolfazi. 2003
ElevationAltitude X
1
≥ 250 m Assumed
1 250 m
Assumed Aspect X
2
North and South Saadi, M. and R.
Abolfazi. 2003 1
East and West Saadi, M. and R.
Abolfazi. 2003 Distance from
population centers X
3
≥ 10 km Sitorus, J., E. Rustiadi,
M. Ardiansyah. 2001 1
10 km Sitorus, J., E. Rustiadi,
M. Ardiansyah. 2001 Distance from
Shoreline X
4
≥ 1 km Assumed
1 1 km
Assumed Slope X
5
≥ 25 - 90 degree Assumed
1 0 – 25 degree
Assumed Road X
6
≥ 1 km Assumed
1 1 km
Assumed River X
7
≥ 1 km Assumed
1 1 km
Assumed
Extracting vector data starts from obtaining the polygon that contains the attribute data such as contour data
≥ 250 meters and 250 meters. Sometimes this purpose needs editing manually of
using ArcView extension in order to make easy the process. The binary data must be put into the Polygon Vector Cell as attribute
data by using Assign data by location Spatial Join. It is done the same thing to other variables: road, river, distance from
population center, and distance from shoreline. Independent variable such as elevation was assumed that less than 250 meter
28 tends to occur deforestation. The references and research citation
is presented in Table 3.1. In ArcView 3.3, there is an extension Edit Tool 3.5, can
accommodate the purpose. There are three method to assign the data by location, i.e Inside attribute of polygon source must be
inside of the polygon target, Center Inside attribute of polygon source must be touched the center point of polygon target, and
Intersect attribute of polygon source must intersect to the polygon target. This tool is the same with Geoprocessing extension
Assign Data by Location. In order to anticipate over estimation of deforestation, the Inside
method is applied to this research. The process of extracting slope and aspect are different with
others. Extracting both features are using ArcView 3D Analyst for creating TIN from Features. The next process that uses grid
TIN is using ArcView ModelBuilder. ModelBuilder will identify input automatically of grid theme, in this case the result from
gridding TIN file. Extracting aspect from elevation grid file is provide default by
ArcView ModelBuilder. It will produce classification of aspect in raster or grid format that also as thematic or discrete grid theme;
with defined resolution 90 meters. The grid classes are: • 1 = Flat
• 2 = North • 3 =
Northeast • 4 = East
• 5 = Southeast • 6 = South
• 7 = Southwest • 8 = West
• 9 = Northwest
With the same procedure above, slope can be extracted. Slope is in degrees unit, meter for vertical unit and slope would be
classified into ten classes. • 1 = 0 – 5 degree
• 2 = 5 – 10 degree • 3 = 10 – 15 degree
• 4 = 15 – 20 degree • 5 = 20 – 25 degree
• 6 = 25 – 30 degree • 7 = 30 – 35 degree
• 8 = 35 – 40 degree • 9 = 40 – 45 degree
• 10 = 45 – 90 degree
3.4.4. Logistic Regression Model
Logistic Regression Model, is used to determine the probability of Unchanged Forest Unchanged Non-forest and
Changed Forest Deforestation that will occur by using transformed function of:
π’ = β + β
1
X
1
+ β
2
X
2
+ … + β
p
X
p
where: π’ = probability of dependent variable Changed Forest Y
β = constant of regression
β
1
= coefficient of X
1
independent variable β
2
= coefficient of X
2
independent variable β
p
= coefficient of p
th
of X independent variable According to Saadi and Abolfazl 2003 that they used and
selected a sampling set about 5 of total pixels, but in this research, sampling set will be obtained from all cells that are from the result of
classification result. The entire sampling set that entered into Polygon Vector Cell attribute, will be processed calculation by SPSS
Statistical computer program. Both possibilities value of variables must be input in SPSS editing table. Calculation will be done
automatically by that program, and the calculation formulas have 29
30 been mentioned in Chapter II Literature Review. The result of the
calculation will be recapitulated such Table 3.2. Table 3.2.
Recapitulation table as the result of SPSS calculation of logistic regression
β Standard
Error Wald
df degree of freedom Significant
level exp β
Constant X1
X2 X3
…
Beside calculating and printing the statistic of logistic regression model, SPSS also can calculate the correlation among or inter the
variables, as the result correlation will be display in matrix form, such
Table 3.3.
Table 3.3. Correlation matrix, that defining the correlation among the variables
Constant X
1
X
2
X
3
… Constant
X
1
X
2
X
3
…
Between two variable will have correlation if the correlation value more then or equal 0.5, and less then 0.5, there is no relationship
between them.
3.5. Assumption of Research
1. Supervised Classification by visual interpretation is assumed to be
closed to the real condition and assessed by accuracy assessment.
31 2.
Cloud, shadow, and water, will be included as the closest class around, for example cloud and shadow that are inside the forest area category;
they will be included as forest class. Oppositely, if cloud an shadow are among the non-forest category, they will be non-forest class. This
process is done by recode tool of ERDAS Imagine 8.7 3.
As the consequence of assumption no.2 above, could, shadow, and seasonal flooding are assumed as 0 value. Because of calculation
process, logistic regression can not identify value unless 1 and 0. 4.
Road is class of sub district road that be able to be passed only by car and truck, so only district and province street class are used in this study
5. Logistic regression does not assume a linear relationship between the
dependents and the independents. 6.
The dependent variable doesn’t need not to normally distribution. 7.
Vector data and projection result are assumed correctly. 8.
Center of population is approached by subdistrict office as point feature with buffer 10,000 meters from the point.
IV. RESULTS AND DISCUSSION
4.1. Image Processing 4.1.1. Period 1990 - 1997
Both image data of 122065_19901109 and 122065_19970728
have been stacked each other, became 14 with Thermal Band 6 bandslayers into one file of dataset
122065_19901109_19970728 , and properly overlayed with the same
projection UTM Universal Transverse Mercator zone 48 South, and Datum WGS 84.
This research image processing uses ERDAS Imagine 8.7, from doing subset, geometric correction, until classification process.
The result of stacking process and the origin sources Landsat images can be seen in Figure 4.1, with band combination 4-5-3 or 11-
12-10.
32
Un-scale Image Un-scale Image
a. 1990 b. 1997
Figure 4.1.
Stacking process of Landsat image 1990 and 1997.