gement.
ing, GIS
anges in
rm scale
ast com-
yortance.
ored for
arlier his-
ed. Only
1 the en-
'esenting
teraction
No parts.
jral envi-
S exten-
ice geol-
| into the
analysis.
sat satel-
vention-
1ent data
ch maps
alysis. In
the cul-
The cul
cal data,
he proc-
o the is-
10del the
ironment
and for the re-analysis of the most recent modern local
development plan which was compiled for the Hvar Com-
mune in 1990.
SOLTA ISLAND THEMATIC MAPPER DATA
CLASSIFICATION
The island of Solta has a relatively poor natural environ-
ment database compared to other islands in the area of
Central Dalmatia. The major goal of the research was to
obtain data on the current land use for the entire region and
compare it with the historical data. However, we wanted to
use a low—cost image processing system, one which would
be readily available for local authorities without major soft-
ware and hardware investments. An important issue in the
choice of the test image processing system was also the
easy integration with the other existing geographic data. It
would be of great importance if powerful analysis within the
system could be performed. Thus it was crucial to find a
system that is easy to use and has a good and friendly user
interface. After an extensive search, IDRISI, which is a grid
based spatial data analysis and image processing system,
was used in this part of the research (Ostir, 1995).
The image selected was a Landsat TM quarter scene (90
by 90 km), acquired on July 31, 1993. Preprocessing of the
image, including georeferencing, was performed with EASI/
PACE (PCI, 1994). After the initial stage the satellite data
for the island of Solta was only transferred to IDRISI on
which the processing was continued. Processing comprised
of two steps. In the first stage, unsupervised classification
of satellite imagery with some general information of the
land use was obtained. The process continued with more
refined supervised classification using different statistical
techniques. Results of both methods will be presented in
the following text.
Unsupervised classification
IDRISI, despite being a relatively powerful system, has cer-
tain limitations in image processing. Unsupervised classifi-
cation could be performed simultaneously on three spec-
al bands only (Eastman, 1995). This might represent a
significant drawback if more bands were needed in the proc-
ess of classification. However, after initial inspection of the
image with seven spectral bands, it become apparent that
quite good results could be achieved if bands 3, 4 and 5
Were used.
Spectral band 3 is useful for our application as it enables
easy distinction between vegetation and nonvegetation,
While bands 4 and 5 are very important in the vegetation
analysis. It was our hope that these bands would contain
énough data for effective unsupervised classification. IDRISI
uires, as a first stage in unsupervised classification, the
Creation of a composite image containing data from all three
bands selected (Eastman, 1995). The frequency distribu-
tion of Classes in this composite image suggested that the
Maximum number of classes in the unsupervised classifi-
cation Could be more than 30. However, after the inspec-
tion of the Composite image it was decided that an initial 15
classes cluster analysis of the image would be performed.
The resulting image produced very detailed information on
565
open areas
arable land
Haquis a 5
—
16 km
Figure 2: Results from the unsupervised classification
With seven classes.
the urbanization of the island; however, distinction between
maquis, wooded areas and arable land was rather difficult.
Therefore some additional tests with a smaller number of
classes were performed. The best results were achieved
with the cluster analysis of the image giving seven classes
only. This image produced very clear built areas and there-
fore open land, including the coast, could be easily distin-
guished from arable land, maquis and two types of woods.
The later two probably represent two different covers of
vegetation which resulted either as a different vegetation
type or vegetation density. For the purpose of presentation
only, the landcover map has been simplified to present four
classes only. The coast and open land have been com-
bined in a single class and the woods as well have been
joint in a unique class. The resulting image is presented in
Figure 2. If the number of cluster classes is reduced to six
or five, much data is lost and image classification gives no
useful information.
Supervised classification
The results of the unsupervised classification were very
promising, however, some rather unexpected results oc-
curred. The signal from the coast was, though geographi-
cally easily defined, very mixed. Thus it was decided to try
a supervised classification which would include following
classes: open areas, arable land, woods, coast and maquis.
For each land use type a training area was selected and
digitized on-screen. Each digitized polygon contained 60
to 70 pixels which had to be very homogeneous. Figure 3
presents the spatial position of the training areas. As men-
tioned five different classes were selected:
e Open areas. The spectral signatures of urban and open
areas were very similar. It was hoped that urban areas
and areas of bare rock would be classified in the same
class. The training area was selected in the centre of
Grohote, the largest village on the island. It contained
houses, paved roads and squares.
e Arable land. The largest arable field on the island was
selected as a training area. However, the sample needed
to contain different vegetational cover, from different
vegetables through to cereals.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996