co e GIS TECHNIQUES AND HYBRID PARAMETRIC/NON-PARAMETRIC IMAGE CLASSIFICATION: A CASE 
9D STUDY SHOWING THE POTENTIAL FOR SIGNATURE TRAINING AND ACCURACY ASSESSMENT 
1630 Thomas Blaschke 
cial Salzburg University 
In: Austria 
DE 
ulo. 
Commission Il, Working Group 2 
s € 
iade Keywords: Remote Sensing, GIS, Classification, Signature Training 
5p. 
cas ABSTRACT 
etor 
wh A landuse/landcover map has been developed for a 900 km? area of a planned National Park in the Salzburg 
à da 
Ih Calcareous Alps (Austria). The resulting map should serve as a supplement to the adjacent National Park 
es Berchtesgaden on the German side of the border. The remote sensing approach employed the analysis of 
Landsat TM-data in conjunction with a partly available forest map and geological and terrain data to derive a 
llite thorough classification procedure. The goal was the highest possible compatibility to an existing classification 
etin developed through airphoto interpretation. Existing data from the Berchtesgaden National Park cover a third of 
127- the used subscene and were integrated in order to design a comprehensive protected area system to support 
environmental assessment, monitoring, and restoration activities. Because the airphoto classification scheme 
was too detailed for the satellite data, a remote sensing classification scheme was developed not analogous to 
ES, the airphoto interpretation but comparable through hierarchical organisation. A hybrid parametric/non-parametric 
| de classification strategy (followed by post-classification processing techniques) was used to map 15 main classes. 
The study provided a description of the classification procedures and a baseline for more detailed mapping and 
0 da 
ista 
-24. 
onal 
siro. 
r de 
tado 
| de 
cola 
monitoring using remote sensing-supported change detection techniques. 
INTRODUCTION 
The availability of landuse/landcover data for nature 
conservation is a growing issue in most parts of the 
world. Until recently, in highly structured and densely 
populated areas as Central Europe, there was a strong 
preference for air photo interpretation and a kind of 
,mental reservation" concerning the use of satellite data 
except for small scale climatological studies. These 
reservations are not only a question of spatial resolution, 
but also the level of information that can be extracted 
from the satellite imagery and the relatively poor 
accuracy of classification of these machine-classified 
data compared to the hand-classified cartographic data 
(Mason et al. 1988). Natural and  near-natural 
ecosystems are rare and the need for protecting such 
environments is more and more recognised. Among 
other projects, a National Park is planned in the Salzburg 
Calcareous Alps (Austria) next to the existing National 
Park Berchtesgaden on the German side of the border. 
The border divides a great and unique ecosystem 
administratively into two countries. As a first step 
15 
towards a more comprehensive protection of this alpine 
ecosystem and the implementation of a National Park 
administration, a digital data base of the whole area was 
developed. In this paper, the issue of a landcover classifi- 
cation is elucidated because it is generally the most 
common theme extracted from remote sensing data. 
Because the existing aerial photo classification scheme is 
too detailed for the satellite data, extensive field work and 
statistical analysis of the environmental, biological and 
remote sensing data were performed to develop a remote 
sensing classification not analogous to the vegetation but 
comparable through hierarchical organisation. A hybrid 
parametric/non-parametric classification strategy (follo- 
wed by post-classification processing techniques) was 
used to map 15 main classes. While both classifiers 
have advantages and shortcomings, newer software 
products provide a high level of integration between 
image space and feature space for a hybrid approach to 
classification. A mixed set of signatures derived from 
both image and feature space required decision rules 
where both the parametric and non-parametric class 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996