Model 
were 
were 
; with 
asses 
sively 
for a 
; type 
most 
of its 
essful 
ferent 
para- 
class 
mean 
ed by 
olems 
band 
also 
ata is 
times 
Kloer 
para- 
lines, 
)oten- 
| used 
on of 
These 
of the 
an be 
ed to 
tions. 
npt to 
orres- 
ectral 
non- 
rform 
res in 
jased 
roba- 
is to 
ssed. 
rst or 
of an 
utput 
APPLICATION 
As previously mentioned, the Landsat-TM data were 
chosen for use in this project because they contain 
spectral information from a broader portion of the 
electromagnetic spectrum than SPOT. This is a very 
important fact especially in the subalpine vegetation zone 
approaching the tree line, where vegetation is sparse and 
the spectral information of an area may depend largely 
on the combination of trees and shrub and/or the 
combination of single trees or shrub and alpine meadows 
(Blaschke 1992). Analysis of the Landsat-TM data was 
supplemented with airphoto interpretation and mapping 
using a derived SPOT-PAN/TM merge as a base map 
and visual interpretation. This assessment results in a 
reduction of the hybrid classification output to a 15 
master-class system. The new classes can promote a 
coarse scheme where aerial photo derived classes fit in. 
The new classification system is based on the realities of 
achieving mappable units using satellite remote sensing 
data with varying levels of analysis. 
Terrain data were also used to derive slope and aspect 
areas by class and to determine study area borders ('not 
below 900 meters above sealevel’) which were not clearly 
defined by other restrictions like administrative 
boundaries. Management data for the park were also 
compiled and/or updated from the remote sensing data 
including trails, cultural features, and primary and 
secondary park boundaries. It was also attempted to use 
these DEM data to correct pixel radiances for varying 
terrain slope and aspect. These topographic 
normalisation algorithms are today implemented in 
commercially available software like ERDAS IMAGINE. 
In this case study, topographic normalisation could not 
provide better classification results. 
For the German part of the subscene, belonging to the 
National Park Berchtesgaden, Arc/Info polygon coverage 
of the aforementioned landuse derived from airphoto 
interpretation consisted of very detailed classes as well 
as very small patches (some less than one TM-pixel). To 
use these existing data not just as visual support but 
digitally, they were modified with GIS-techniques. 
Patches less than 2500 m? were deleted and others were 
shrunk (using negative polygon buffer) by 30 meters to 
avoid edge effects in the spectral signatures and spatial 
inaccuracy. Some of the resulting isolated patches were 
used for signature training and other parts of this data 
layer were later used for accuracy assessment. 
The classification herein is not hierarchical but the 
resulting classes can be seen as upper-level classes for 
the detailed aerial photo derived landuse map. They are 
17 
not equally distinct. Some classes were readily 
recognised and quite distinct from most other classes 
(e.g. water, wetland). Others were quite heterogeneous 
(coniferous forest) or tend to overlap with other classes 
(e.g. urban and rock). It has been shown that post image- 
processing and GIS techniques make it possible to 
clearly distinguish between some classes utilising 
additional information like DEM data. It has also been 
shown, that the use of ancillary data during and after 
multispectral image classification is not only useful but 
necessary if a single class of objects is not represented 
by a single spectral class. 
Signature training and editing is first done 'traditionally' 
with well approved parametric tools. Polygons repre- 
senting training sets are either created spatially with 
drawing tools or spectrally using a region growing 
function determined by the spectral parameters of neigh- 
bours relative to chosen seed pixels and a spectral 
distance threshold as described by Kloer and Brown 
(1991) for the ERDAS software. Existing evaluation tools 
allow for presentation of the sample statistics and histo- 
grams, image alarms (quick screen parallelepiped 
classifications), and pairwise divergence measurements. 
Parametric signatures can also be generated with unsu- 
pervised clustering algorithms, which were not used 
within this case study. Varying tools provide a compre- 
hensive parametric classification process and may be 
used independently of the non-parametric tools (Kloer 
1994). 
The non-parametric tools are built on the concept of 
feature space. It is an n-dimensional space which 
encompasses all the data values of the image. Practi- 
cally, it is visualised partly in 2-dimensional scattergrams 
or scatterplots. This feature space is handled in the 
ERDAS software with the same data structure as all 
raster images and a lot of tools can therefore be applied. 
The feature space image is displayed using the standard 
visualisation tools which provide identical interactive 
capability as for the image space. 
Both approaches were integrated. First, training classes 
derived from supervised signature training (‘parametric’) 
were displayed and evaluated in various feature space 
images. Signatures were analysed and manipulated in 
order to look at the homogeneities, spectral distances, 
and overlaps. For the maximum likelihood decision rule, 
Transformed Divergence (TD) was used (Swain and 
Davis 1978) for separability listing of every class pair and 
band combination. Fifteen classes resulted and are listed 
below. Cursor linking and image to feature space masking 
between image and feature space windows supported the 
decision process of merging, deleting, and redefining 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996