signatures. Parametric signatures from objects in feature 
space may be drawn from those pixels that fall inside the 
object. 
A hybrid classification approach was used to utilise these 
different class definitions. Within the same process, 
maximum likelihood as an parametric rule and feature 
space as an non-parametric rule were performed. An 
option was chosen: if the number of non-parametric class 
definitions in which the pixel lies is O, the pixel will be 
submitted to the parametric rule. Among four possible 
options in the used software (Kloer 1994), this decision 
option provided the best results. 
RESULTS 
As discussed before, the selected classification scheme 
was designed to include all the major land covers 
encountered in this area whereby the valleys outside the 
proposed park boundaries were excluded prior to 
classification. 
After the initial classification the accuracy of the new land 
cover layer was compared based on random sites inter- 
preted on 1:10000 scale orthophotographs. Blocks of 3 x 
3 pixels around randomly located points were used as 
sample sites. These blocks were stratified according to 
class prevalence. The verification was supported partially 
by available biotope maps. Therefore, these maps were 
digitised and converted to Arc/Info vector format. 
Accuracy assessment was performed according to Story 
and Congalton (1986). These binomial probabilities are 
based on the percent correct and do not account for 
errors of commission or omission (Lunetta et al. 1991). 
Table 1: Final land cover classes and accuracy 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Class Area (km^) | Row. Marg. | Accuracy 
1 rock, no vegetation 138 67 9196 
2 rock, debis, sparse | 37 24 77% 
vegetation 
3 meadow 75 48 7196 
4 subalpine/alpine pinus | 120 64 8496 
mugo societies 
5 shrub and bush 26 19 7196 
6 water 4.7 7 100% 
7 moor 4.1 7 100% 
8 subalpine shrub and | 83 35 76% 
bush 
9 hardwood/conifer mixed | 175 80 70% 
10 deciduous trees 45 24 77% 
11 pine 37 26 86% 
12 pine, young plantation | 35 24 81% 
13 coniferous trees 59 35 84% 
14 alpine meadow 41 26 88% 
15 residential, streets 19 14 68% 
overall 500 80,3% 
  
  
18 
An overall accuracy level of 80 percent was attained 
(table 1). This level is generally high but it is not fully 
satisfying when the availability of existing landcover data 
for a third of the area, additional data and hybrid 
parametric/non-parametric image classification tech- 
niques are taken into account.The combination of GIS 
and satellite imagery provides a huge potential not only 
within this case study. The GIS allows for the overlaying 
of multiple forms of data which makes signature creation 
and accuracy assessment much easier. 
DISCUSSION 
This small study cannot solve the well known problems 
of using data derived from photo interpretation for 
supervised classification of satellite data. Nevertheless, it 
is considerable, that even with very good data availability 
and the use of advanced hybrid image classification 
techniques, it is obviously difficult to raise the overall 
accuracy over 80 percent. Using graphic tools in feature 
space layers, non-parametric techniques allow for the 
complete" utilisation of the area covered by the 
scattergram. In some cases it is difficult to interpret the 
corresponding pixels in the image using image alarm 
tools and previews. Similar to the integration of super- 
vised and unsupervised techniques, it should be possible 
to gain from the advantages of two different techniques 
rather than to be handicapped by the disadvantages of 
both. Commercially available software provided us with 
sufficient tools but more research and further 
applications are necessary. 
REFERENCES 
Blaschke, T., 1992, Integration von Satelliten- und GIS- 
Daten im geplanten Nationalpark Kalkalpen (Ober- 
österreich). Salzburger Materialien 18, pp. 135-148. 
Bolstad, P. and Lillesand, T., 1991, Automated GIS 
Integration in Landcover Classification. Technical 
Papers, ACSM-ASPRS Annual Convention, pp. 23-32. 
Ince, F., 1987, Maximum Likelihood Classification, 
Optimal or Problematic? A Comparison with the Nearest 
Neighbour Classification. Int. Journal of Remote Sensing 
8(12), pp. 1829-1838. 
Kloer, B., 1994, Hybrid parametric/non-parametric image 
classification. Technical Papers, ACSM-ASPRS Annual 
Convention, pp. 307-316. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996 
Kloer 
Searc 
ASPF 
Lune 
McG 
Geog 
Sour: 
Engir 
Mase 
Inclu: 
rame 
Phot: 
pp. 2 
Masc 
Petrc 
in the 
imag 
pp. 1 
Skidr 
accul 
metri 
Engi 
Story 
user' 
Rem 
Swai 
quan