International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
2. METHOD
2.1 Study area, remote sensing data and software
The Pelangkaraya study area is situated in an extensive peat
swamp forest, roughly between the cities Palangkaraya and
Banjarmasin, South Kalimantan, Indonesia (Figure 1). It covers
roughly 115.000 ha (45 by 25 Km). Three main rivers (i.e.,
Barito, Kahayan and Kapuas cut across the flat area. For
centuries, shifting cultivation has been the main practice in this
area. Since the eighties, however, numerous logging firms and
large transmigration projects have been active in the area. As a
result of both logging and transmigration, the forest changed
into a mosaic of logged forest, heavily logged forest, patches of
original forest, agricultural fields (trees and crops). fields
covered mainly with grasses (as a result of fire in the past),
abandoned fields covered mainly with shrubs, and water areas.
Eus pus die "- a a
ha au ad
ut Shi
= pe” x
^ bi gh
E imi rem b Xa ; ot wy
Ww — # i " 1
auras a votes
x épris
Figure 1: Location of study area
In this research, two Landsat TM remote sensing images of
30/08/1990 and 10/05/1996 covering the Pelangkaraya study
area were used. Colour composites (Landsat TM bands 453) of
both images are given in Figure 2. The images were
geometrically corrected using cubic convolution for resampling
to a spatial resolution of 30 meters. Software packages ILWIS
academic 3.0 and eCognition 3.0 were used for LCM
classification. SPSS was used to calculate KHAT and Z
statistics, and Fragstats 3.3 was used to calculate four
Landscape Pattern Indices.
Figure 2: Colour composites of two Landsat TM images of
1990 and 1996 (bands 453, histogram equalization)
2.2 LCM classification method
The LCM classification concept is a multi-scale classification
approach based on two levels (Figure 3). The first level
contains elementary objects consisting of conventional land
cover classes. The second level contains composite objects
consisting of land cover mosaic classes, which are based on
two LCM parameters that is the size and mix of elementary
objects (Obbink, et al., 2002; Schaijk-Obbink et al., 2000).
Remote Sensing
data
Segmentation
Classification
Land cover || Elementary
Level 1 classes objects
Segmentation Parameters’
-spatial size
-mixture
Classification
Land cover | Composite
Level 2 mosaic classes objects
Figure 3: The Land Cover Mosaic (LCM) classification
concept.
In this study, elementary objects were created by segmenting
the remote sensing image into (radiometrically) homogeneous
segments. The eCognition segmentation algorithm was used to
segment the Landsat TM images (band 3,4,5.7) using the
following parameters: scale 10, color 0.9, shape 0.9. Nearest
neighbor was used for a supervised classification of the
elementary object into eight land cover classes (ie, logged
forest, heavily logged forest, shrub, agriculture, grass, water,
river, and clouds). Composite objects were created by merging
all adjacent elementary objects that represent similar land cover
classes. Based on the spatial size (expressed as minimum-area,
MA) and mix (expressed as relative- border-to-neighbor, BN) of
the elementary objects, composite objects were classified into
seven land cover mosaic classes (i.e., mainly logged forest,
mainly heavily logged forest, mainly shrub, mainly agriculture
mainly grass, mainly water, and mainly clouds), and one land
cover class (i.e., river). Mainly refers to predominance of a land
cover class. This means that if the size of an elementary object
with a certain land cover type is smaller than a certain threshold
(MA in ha), and it is surrounded by another elementary object
with a different land cover type as set by a threshold ( BN in %),
then both elementary objects are part of a higher level thematic
class that is a land cover mosaic class.
2.3 Sensitivity Analysis and evaluation methods
A sensitivity analysis was performed to test the impact of the
LCM parameter size (M A) on classification outcome. Fourteen
different MA values were se S from 5 ha to 18800
ha(je,5,5.5. [35. 25, S0, 100. . 200, 250, 300, 350, 400
15000, 1 S ha) These HEN were chosen according to the
size of the image objects present in the classified images: from
