2004 
ment 
anan 
intah 
30% 
are. 
on a 
n the 
Is an 
ty is 
from 
ty of 
ween 
ment. 
1eavy 
under 
n and 
arried 
inting 
panan 
<arya 
nting 
oupes 
years 
ogged 
sively 
have 
ing is 
ibanan 
search 
ticular 
single 
tors in 
of this 
to the 
he test 
le tree 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
felling in the Labanan concession is reduced to the area of 
RKLI for the following reasons: 
2.3.1 Evidence of illegal logging 
Previous research by Bhandari (2003) showed evidence of 
single tree felling in RKLI. In addition early investigation with 
the local people, company officers and my own field 
observations revealed that RKL has suffered a lot from single 
tree felling. 
2.3.2 Accessibility 
RKL1 is located at the official entrance to the concession area 
through the existing road network which provides access to the 
villages and the market in Tanjung Redeb. Further more, a 
major part of RKLI is located alongside this road (Figure 2). 
Consequently the accessibility of RKL1 is very good which 
makes it attractive for illegal loggers. 
2.3.3 Oldest logged RKL 
RKLI was the first logged RKL in 1976 and will be re-entered 
in 2011 for the second harvest cycle according to the long-term 
management plan of the company. The forest has regenerated 
itself and the chance of finding good quality timber with a 
diameter larger than 50cm is high compared to the other areas 
in the concession. This situation also creates opportunities for 
illegal loggers. 
2.3.4 Terrain condition 
The terrain condition in RKL1 has no very steep slopes as 
compared to the protected area and is thus favorable for felling 
activities 
  
  
  
  
  
Figure 2. Location of RKL1 within Labanan concession 
2.4 Collection of Ground Truth 
The purpose of the fieldwork was to collect training data for the 
image classification and testing data for the accuracy 
assessment of the classification output. The fieldwork was 
conducted in September 2003 in the Labanan concession, East 
Kalimantan, Indonesia. Before going into the field for data 
collection, a consultation session was arranged with the head of 
planning & inventory of Pt. Inhutani 1. This was to identify 
areas in the concession where illegal felling of single trees was 
taking place. Four areas were identified ie. RKLI, RKLA, 
RKLS and the protection area. However, RKLI was found to 
have the highest number of single tree felling which was the 
reason to reducing the area for analysis to RKL I. 
Un 
3. REMOTE SENSING DATA ANALYSIS 
This method section describes the activities that were carried 
out to detect single tree felling using remotely sensed data. The 
first step was pre-processing of the Landsat-7 ETM+ image. 
The image was shifted using the main road map and 
georeferenced using the ground control points collected in the 
field. The second step was the image classification using the 
Maximum Likelihood ML (Figure 3) and the Sub-pixel SP 
classifier (Figure 4). 
3.1 Image classification 
The ML image classification was performed on two data sets 
i.e. 30 m resolution and 15 m resolution. The signature for each 
class was selected by displaying the ground truth shape file 
(training data set) over the image and selecting the pixels with 
the respective ground truth one by one. The scatter plot space 
was used to evaluate the selected pixels in each category. New 
logged points (NLP) that were located close to the road were 
excluded, since these were likely to be misclassified as road. 
3.1.1 Sub-pixel Classification 
Although the SP classifier requires raw data, a georeferenced 
image was used because selecting aoi for the training set was 
not possible using a raw image. Prior to the signature 
derivation, pre-processing and environmental correction were 
performed. During the environmental correction cloud pixels 
were selected and removed. 
Signature derivation: 
Signature derivation and evaluation is an important part in the 
subpixel classification. There are two ways to derive a 
signature, manual and automatic. Manual derivation is used 
when whole pixel MOI can be used as training set. In this case, 
the material of interest (MOI) is the opening or disturbance 
caused by single tree felling was in subpixel fraction. For that 
reason the automatic signature derivation was used. The 
training set aoi was selected using the same NLP as in the ML 
classification. There are two other aoi files that can be used as 
input for the signature derivation (i.e. valid and false aoi). NLP 
outside RKL1 were used for valid aoi. 
Signature evaluation: 
The automatically generated signatures were evaluated upon the 
material fraction detection and the SEP value. It was also 
compared with the gap fraction found in the field. 
Image Classification: 
The classification was run using the selected signature. The 
default tolerance value was set at one and the number of output 
classes at eight which will result in eight different MOI fraction 
classes ranging from 0.20 to 1.0 with increments of 0.1. The 
eighth class with MOI fraction 0.90-1.0 was considered as the 
NLP class and the final result was a map with two classes, NLP 
and other. : 
Accuracy assessment and comparison 
Fifty percent of the collected ground truth data (test data set) 
was used for the accuracy assessment. The test points were 
carefully chosen making sure that the test and the training data 
set were equally spread geographically. Each classified image 
was then crossed with the test data to generate a confusion 
matrix. The respective confusion matrices were then used to 
calculate the different accuracy measures 1.e., producer's, user's 
accuracy, class mapping accuracy for each class and the overall