International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
  
aerial photographs, satellite and radar imagery. For example, 
selective interpretation keys were developed for tropical 
woodland of Tanzania (Howard, 1959), and for coniferous and 
hardwoods of Middle-European trees (Grundman, 1984; 
Anthony, 1986; modified by Hildebrandt, 1996). On a SPOT 
false colour composite natural forest, bamboos, scattered trees, 
brush, plantation, reed forest could be identified in northeastern 
Bangladesh (Arquero, 1997), On 1:50,000 scale aerial 
photographs covering moist tropical forest of Kerala and 
Tamilnadu, India, the following forest cover and land use types 
were delineated: tropical evergreen, tropical semi-evergreen, 
moist deciduous, dry deciduous, teak plantation, eucalyptus 
plantation, reeds, bamboos, rubber plantations and tea estates 
(Tomar, 1968). 
2. MATERIALS AND METHODS 
2.1 The Study Area 
The study area is located in the southeastern Bangladesh 
(Figure 1). The forests of the study area can be classified as 
tropical wet evergreen forests and tropical semi-evergreen forest 
(Champion et al., 1965). This forest type is characterized by the 
presence of a considerable amount of evergreen trees in the 
upper canopy. The top canopy reaches a height of 40-60 meters. 
A few semi-evergreen or deciduous species may occur, but 
usually they do not change the evergreen character of the forest. 
The forest is rich in epiphytes, orchids, woody climbers, ferns, 
mosses and palms particularly in shady moist places (Das, 
1990). 
XT r zr" 
E eue = es =e 
N ; ^ bL Le À Ms 
e or adeo etus China 
^ | Bangladesh 
Nepal 
A \ 
A = x. t % 
India +8 
   
  
- Üü ( Myanmar 
Study area 
HL 
f SIN 
Figure 1 Location of the study area 
2.2 Remote Sensing Data 
The study selected Landsat ETM+ data, which is relatively 
cheap and available. It could be easily used by developing 
countries where most of the tropical forests lie. The study has 
decided to develop a selective interpretation key for the Landsat 
ETM+ sensor for the following reason. Because of the relatively 
poor spatial resolution of satellite imagery (in compare with 
aerial photographs, where mostly dichotomous keys are widely 
used), image color, texture and location are important image 
characteristics that must be used to interpret vegetation types; 
these characteristics can easily be presented in selective keys. 
2.3 Image Pre-processing 
The Landsat ETM+ image was received with geometric 
correction form USGS. However, a linear shipment was 
discovered during the field mission and therefore the image was 
re 
158 
shifted accordingly. The current study area consists of small 
hills. But topographic normalization could not be applied due to 
lack of appropriate digital elevation model. The accuracy of 
geo-coding was checked using a portable GPS from the known 
location of the geodetic points of Survey General of 
Bangladesh. The study used a modified dark object subtraction 
method namely COST method (Chavez, 1989; 1996). One 
percent minimum reflectance was fixed for atmospheric haze 
correction (Chavez, 1996; Moran et al., 1992). Using the model 
maker of ERDAS Imagine software the digital count was 
converted to surface reflectance. 
2.4 Visual Interpretation and Field Verification 
A computer-aided unsupervised classification scheme was 
applied to the Landsat ETM+ image. Simultaneously a variety 
of image channels were displayed on computer-screen and tried 
to find an optimal combination where a variety of vegetation 
classes were distinguishable. Additionally a correlation matrix 
of different spectral bands for the sample pixels was calculated 
to ease the band selection process. Interpretation could 
delineate eight different vegetation classes. All variety of 
vegetation classes was reached during 2002-2003. The location 
was identified by a portable GPS (GPS 12, Garmin). A detail 
description of vegetation was recorded and a panoramic camera 
photo was taken for the field-plots. A total number of seventy 
sample plots were collected. 
3. RESULTS AND DISCUSSION 
3.1 Selection of Optimal Band Combination 
The correlation matrix of the spectral bands contains useful 
information about the redundancy of information and selection 
of optimal band combination for interpretation purpose. If the 
bands show strong correlation (value near to 1.000) this indicate 
the bands usually contain similar information to each other. 
When those bands are visualized the minimum separibilty 
among different feature would be noticed. The following table 1 
represents the correlation matrix of the selected sample plots 
where field sampling was carried out. 
  
  
Table 1. Correlation matrix of the representative pixels of 
Landsat ETM- image for the study area 
Bands 1 2 3 4 5 7 
1 1.000 
2 0.861 | 1.000 
3 0.856 | 0.942 | 1.000 
4 0.070 | 0.175 | 0.012 | 1.000 
S 0.609 0.809 | 0.847 | 0.210 | 1.000 
7 0.660 | 0.832 | 0.898 | 0.096 | 0.978 | 1.000 
  
  
  
  
  
  
  
  
  
From table 1 it is observed that among visible (band 1-3) and 
mid-infrared bands (band 5-7) high correlation exist. It means 
that there is a high redundancy of information within those 
bands in the vegetated areas. So it makes more sense to select 
the bands, which contain minimum redundancy. Therefore, one 
In 
ee RAM RR ged 
Cad