Ngitude
d in the
[Jambi
ng Hari
orm of
nd with
€ Varies
used the
Part of
nic tuff
Ountain,
it in the
[ Site is
)ber, oil
- Tübber
$ people
I inside
Ise by
outheast
er types
ata were
systems.
used to
nth land
jas 80 x
esenting
t of the
>d on 29
ent land
, green,
R. The
R) data
(5.6 cm
ed... The
th width
During
he study
yes Welt
1ere Was
d in the
Aperture
) and 20
[he dat
length),
of 35.
ie swatl
ata wer
iber 15,
, ERSÀ
94, and
1993.
For the first'test site, our methods included the
classification and interpretation of the microwave
and optical remote sensing data, field checks,
change detection analysis, and comparison of the
results of different types of data.
From the additional information acquired in the
field, a supervised classification was performed on
the available Landsat MSS (all bands) as well as
TM data (bands 2, 3, 4 and 5). Training samples
were first selected from various land cover types:
wetland rice field; shifting cultivation and
secondary forest; bush and scrub; lowland forest
(primary forest); settlement; sea beaches and bare
soil; sea and lake. After selecting the training
samples, a classification was run using a
maximum likelihood algorithm. A quantitative
evaluation of both classification results was done
by testing the accuracy using a confusion matrix
which showed the overall and average accuracies
by class.
An attempt was also made to incorporate radar
data (ERS-1 and JERS-1 data) in multispectral
classification by combining ERS-1 with Landsat
TM3, TM4 and TMS. Similar combinations of
JERS-1, TM3, TM4 and TMS bands were also
classified.
Because ERS and JERS radar data are acquired in
single bands (e.g. wavelength, polarization,
incidence angle), digital image processing
techniques are limited. It was therefore decided to
print the radar images in hard-copies for visual
interpretation. In this case, other information (e.g.
contextual or spatial) can be used as key
interpretation elements to delineate the boundaries
of different land covers. The visual interpretation
maps of the ERS-1 and JERS-1 images were
digitized, polygonized and rasterized for
comparison to see how many classes could be
recognized in each of them. Later, both
Interpretation maps were rasterized to 30 m pixel
Size and registered to the 30 m spatial resolution
of TM, MSS and other materials.
The Landsat MSS classification map of 1973 and
TM classification map of 1990 were compared to
calculate the land use changes (e.g. decreases or
Increases in the areas of different land use classes)
during 17 years. Another comparison of a 1988
Landsat MSS classification map, land use/land
cover map and Landsat TM classification map of
1990 was made to detect the changes in land uses
during that period. Emphasis was put on certain
Classes that were expected to have significant
changes (e.g. forest, agriculture, and settlement
Cover types).
Because urban areas can be detected best on ERS-
1 radar Images, a change detection analysis was
performed by overlaying the settlement area from
317
the ERS-1 image on the TM classification. The
same step was repeated by overlaying the
settlement maps from the 1973 MSS and 1990 TM
classifications to show urban development during
17 years.
Methods used with the data of the second test site
included: image pre-processing (c.g. radiometric
and geometric correction and filtering), object
identification and detection, image classification,
optical and radar satellite image fusion and
comparison to detect forest and deforested areas.
3. RESULTS AND DISCUSSIONS
Using the MSS data it was not possible to have
more than eight classes because of the overlap
between the clusters. It was especially difficult to
separate different forest types (swamp forest, tidal
forest, natural forest and rubber plantation)
because the forest types are not homogeneous in
terms of the tree species. Each forest type consists
of many tree species, resulting in mixed spectral
reflectance characteristics. Thus it was decided to
combine some of these classes.
Shifting cultivation and secondary forest were
placed in one class because the farmers cultivate
several crops (such as coffee, rubber, cereals, etc.)
which makes it spectrally confusing with other
cover types. In addition, the farmers do not clear
the area completely. It was difficult to distinguish
between the river and wetland rice fields because
the river is narrow in some locations and
surrounded by bushes and other vegetation. Also
in some places the wetland rice is located next to
the river. This causes a mixed signature of water
and vegetation. Therefore, both were classified
together.
Different types of soils (such as red soil, brown
soil etc.) did not show distinct spectral signatures.
As a result, they were not separated into different
classes. The same situation occurred with the
settlement and homestead gardens. The homestead
gardens consist of agricultural crops, fruit trees
and bare soil. The settlements and homestead
gardens were also classified as one.
However, it was possible to distinguish between
natural forest (lowland forest), shifting cultivation
and secondary forest, wetland rice, settlement, sea
and lake. The bare soil and beach were classified
together because both classes have almost the
same spectral signatures.
A quantitative evaluation of the supervised
classification results indicated an overall
classification performance of the MSS data of
88.9%, which is relatively good. Most of the
individual classes had classification accuracies
exceeding 80 %, except for the bush and grass
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996