which had an accuracy of 77.9%. About 5% of 
the grass test fields were confused with the 
lowland forest and 18% with the shifting 
cultivation. 
The supervised classification of Landsat TM 1990 
data resulted in 12 spectral classes. Some were 
merged to obtain the final nine classes because the 
low frequency of pixels in some classes. The nine 
classes were: lowland forest (primary forest); 
shifting cultivation with secondary forest; beach 
with bare soil; wetland rice field; bush with grass; 
settlement with homestead garden; lake; sea and 
cloud. 
It was difficult to distinguish between different 
forest types such as swamp forest, tidal forest and 
rubber plantation. As in the Landsat MSS 
classification, the TM classification result also did 
not show the difference between wetland rice and 
river. The overall classification performance of 
the TM data was 91%, which is even better than 
the results from MSS data. 
A multisensor spectral classification of TM with 
ERS-1 and TM with JERS-1 data was attempted. 
Because of the high speckle noise of the radar 
data, the classification result was not as good, and 
resulted in a small number of classes. These 
classes were lake, sea, vegetation, settlement and 
the wetland rice. The settlements were 
significantly clearer on the radar image compared 
with the Landsat TM image because of the radar 
comer reflection phenomenon. The wetland rice 
fields, lake and sea appeared very dark on the 
radar image because of the specular reflectance 
from the water. 
The following classes were obtained from the 
visual interpretation of the ERS-1 image of 1993: 
lowland forest; shifting cultivation and secondary 
forest; clearcut with bush; wetland rice; rubber 
with bushes; swamp forest; beach with coconut 
trees; settlement; lake; river and sea. The visual 
interpretation of the JERS-1 image resulted in the 
same number of classes as with ERS-1, but tidal 
forest was clear in the JERS-1 image. A 
comparison of these two interpretations shows that 
it was easier to delineate the boundaries of 
wetland rice fields and water bodies (such as river, 
lake and sea) on the JERS-1 image than the ERS-1 
image because JERS-1 has a longer wavelength. 
However, the settlements in the ERS-1 data were 
very clearly delineated because the original spatial 
resolution of the ERS-1 image is much higher than 
the JERS-1 image and also the effect of VV 
polarization. The ability to recognize lowland 
forest, swamp forest, rubber plantations, and 
coastal coconut plantations using the JERS-1 
image were better than ERS-1 because the L-band 
energy penetrates through vegetation canopies 
better than C-band. Both maps showed 
318 
differences in the location of shifting cultivation 
with secondary forest and bushes classes. 
In general, comparing the ERS-1 and JERS.| 
interpretations and TM classification, the radar 
images provided four more classes than the optical 
image. These classes were swamp forest, tidal 
forest, rubber plantations and coastal coconut 
plantations. 
Figure 1 shows the changes for different classes 
between 1973 and 1990. To analyze land use 
changes, the classification result of 1990 was 
overlaid on the result of 1973. The major change 
took place in the lowland forest area. The 
reduction in lowland forest was replaced mainly 
by shifting cultivation with secondary forest 
(10595.5 ha or 13.1%), followed by bush and 
clearcut (3657 ha or 4.5%), settlement (2603.4 ha 
or 3.2%), wetland rice (1007.8 ha or 1.2%) and 
bare land (0.27%). Shifting cultivation increased 
from 19.1 % in 1973 to 38% in 1990; 13.1 % of 
this increase came from lowland forest while 
12.9% came from bush and 11.1% remained as it 
was in 1973. The main factors effecting shifting 
cultivation expansion are the population pressure 
and socio-economic aspects. High population 
growth, both natural and immigration, in a limited 
area puts pressure on the environment, which will 
finally reduce the sustaining capacity of the land. 
So we can say that the increase of population 
results in the increase of shifting cultivation. 
Figure 2 shows the relationship between the 
population, shifting cultivation and lowland forest 
in Bengkulu province. The high positive 
relationship shows that the increase of shifting 
cultivation area is related mainly to the population 
increase. Conversely, the relationship between the 
population and lowland forest area is negative, 
which indicates that the increase of population 
results in a decrease in lowland forest. 
The results from data analysis of the second data 
set were consistent with the results from the first 
test site. The classification result of Spot data 
shows that 11 classes were able to be recognized, 
while Landsat TM recognized 9 classes. Multi- 
temporal ERS-1 image recognized 8 classes, while 
single JERS-1 image shows only 5 classes. JERS- 
] image was able to recognize the forest cover 
type classes better than ERS-1 images because the 
first used longer wavelength. The classification 
accuracy of all four optical and radar data clearly 
shown on Figures 3, 4, 5, and 6. 
4. CONCLUSIONS 
The following conclusions can be drawn from the 
results of the first test site: 
- The classification of radar data (ERS-1 and 
JERS-1) did not give decent results because of the 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996