open areas 
   
  
   
  
  
coast 
arable land 
Maquis 
maquis 2 
a 5 18kn 
D nnm) 
d 
Figure 3: Position of training areas selected. 
e Woods. Spatially well defined and clearly seen as darker 
areas on infrared bands of the TM image. 
e Coast. Also spatially well defined, but has very mixed 
signal — probably due to different coast types resulting 
in the geology, proximity of vegetation etc. It was hoped 
that inclusion of coast in the classification would facili- 
tate classification of open areas. 
e Maquis. Probably the most problematic sample. Most of 
the maquis has developed on abandoned pastures and 
can be very dense in some cases. Sometimes even re- 
cently abandoned grasslands and pastures could be in- 
cluded in it. As most of the problems were predicted in 
this class, two samples were taken. 
Following the initial inspection of the training area samples 
it was decided to use only bands 1, 2, 3, 4, 5 and 7 in the 
analysis. Band number 6 was excluded due to its large pixel 
size which resulted in the significant loss of data. During 
the more detailed inspection of sample signatures it became 
apparent that some of them overlap and therefore they had 
to be modified. In Figure 4 the mean values of corrected 
signature samples are plotted. It is evident that in general 
the bands do not overlap. 
    
open areas 
arable land 
  
Haquis o 5 18 kn 
Woods d 
Figure 5: Results from the supervised classification. 
  
566 
  
140 
120 
1005 
| 77$ open area 
80 5 —8— arable lad 
—à— woods 
60 —— Maquis 
> coast 
  
  
04 i ! } } | 
IM TM ™3 TM TVG TW 
Figure 4: Mean reflection values of signatures of 
corrected samples. 
Several methods of classification were used in the classif. 
cation procedure. Since the parallelepiped method is Very 
simple and robust, the results obtained were inappropriate 
(Mather, 1989; Sabins, 1986). Thus minimum distance and 
maximum likelihood methods were tested. They are both 
provided as an integral part of the IDRISI software pack- 
age. As expected, the best results were produced when 
maximum likelihood method was applied (see Figure 5). 
The results of the classification were compared with the 
1:25,000 map produced in 1981. Though recent air photo- 
graphs should have been used for quality control, the only 
available comparable sources were maps. There was evi 
dence of the growth of maquis in all areas where agricu- 
ture has not been so intensive. Abandonment of arable land 
followed between the 1981 till 1993. It seems that all the 
abandoned arable land has been covered by maquis. The 
wooded areas seem to be constant. 
CONCLUSIONS 
The study has two important implications. First, the land 
use map of the island of Solta was produced. The island of 
Solta probably has the worst spatial data on the natural 
environment in the whole region of Dalmatian islands. Al 
other islands, especially the island of Hvar and the island of 
Braé, have soil and geology maps of some relevance. In 
the study an accurate land use map was produced, despite 
serious problems with the satellite imagery applied in the 
Mediterranean environment. These problems are (Gaffney 
et. al., 1995): 
e Extreme polyculture of the Central Dalmatian agriculture 
on a single land parcel a wide variety of crops may be 
grown together, along with subsidiary tree crops. 
e Arable areas on the island are typified by the use of Very 
small fields. Sometimes up to 45% of the fields can be 
smaller than 30 meters. 
e Last but not least, field boundaries are composed of 
masses of cleared stones and stone terraces which may 
be several meters wide and up to 3 meters in height 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996