definitions were considered. The combination allows for
great visualisation and analytic capabilities. The ERDAS
software was used and contains various tools not only to
interactively digitise graphic objects in both feature and
image space but also for cursor linking and object-to-
mask mapping in-between image and feature space. This
study provided a description of the classification proce-
dures and a baseline for more detailed mapping and
monitoring using remote sensing-supported change
detection techniques.
STUDY AREA AND OBJECTIVES
The study area was located in Austria south of the city of
Salzburg. It covers mostly an alpine landscape in the
Northern Alps with elevations up to 2700 m above
sealevel. Only the Salzach valley is highly populated
while the alpine parts are covered by very few
settlements. The satellite data rectangle covers the study
area itself as well as great parts of the National Park
Berchtesgaden (Germany) on the other side of the
border. The idea is to implement a National Park
adjacent to the existing one consisting of the same
ecosystem and divided only by artificial boundaries. For
the Berchtesgaden area, various data in digital format
exist and GIS-techniques have been used for more than
ten years. A landcover map was derived in 1985 at an
original scale of 1:10000 from airphoto interpretation
using a very detailed classification theme. The objective
of this work was to use existing information for super-
vised classification and accuracy assessment. Thus, a
hierarchical classification scheme was developed, but
even with generalisation, not all original classes could be
assigned to new classes. Therefore a hybrid parame-
tric/non-parametric classification procedure was chosen
to 'cover additional parts of the feature space.
METHODOLOGY
The Landsat TM data were chosen for use in this project
over SPOT data because they contain spectral infor-
mation from a broader portion of the electromagnetic
spectrum. This was seen as being more important for
vegetation discrimination than the better spatial resolu-
tion of SPOT data. Aerial photography was used as an
ancillary source of information, especially for the image
processing procedures of geocorrection, data enhance-
ment and supervised classification. ERDAS software was
used for image processing and the Arc/Info Geographic
Information System (GIS) for vector data manipulation.
The six non-thermal bands of a TM-scene from August
1990 were classified. Image normalisation was applied
16
on the basis of a 50 meter Digital Elevation Model
(DEM).
Different unsupervised classification techniques were
tested. The lower parts surrounding the study area were
excluded prior to clustering. Repeated experiments with
different numbers of classes were performed but classes
that fell in forest sites could not be comprehensively
identified. The maximum likelihood decision rule for a
supervised classification approach was chosen. This type
of classifier, called parametric, has been the most
commonly applied classification technique because of its
well developed theoretical base and its successful
application with different data types and different
applications (Bolstad and Lillesand 1991). With the para-
metric technique, the classifier must be trained with class
signatures defined by a statistical summary of the mean
vector and the covariance matrix normally acquired by
the analyst selecting samples. It is known that problems
arise if samples are too homogeneous in any one band
or the sample size is too small. The classifier also
assumes that the distribution of the sample data is
normal in all bands, a condition which is sometimes
violated for certain classes (Ince 1987). Therefore, Kloer
(1994) suggests the complementary use of the para-
metric and non-parametric approaches. He underlines,
that the non-parametric classification method has poten-
tials and limitations as well. Some researchers have used
non-parametric methods mainly for the classification of
natural surfaces (Skidmore and Turner 1988) These
methods make no assumption about the shape of the
spectral distribution of data sets except that they can be
grouped by a discriminant function and are expected to
present advantages in spectrally irregular situations.
Masseli et al. (1992) looked upon them as an attempt to
overcome the well known problem of the low corres-
pondence between cover categories and defined spectral
classes. These authors state, that in practice, non-
parametric methods have been demonstrated to perform
far better than the conventional parametric procedures in
many applications. The feature space object-based
classifier has the problem of overlap. Since no proba-
bilities are computed, the only means of resolution is to
consider the order in which the classes are processed.
The pixel in an overlapping area is assigned to the first or
last class for which it is tested and many pixels of an
image will not be assigned to any class in an output
classified image.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
APPI
As [
chos
spec
elect
impa
appr
the s
on 1
com|
(Bla:
supr
usin
and
redu
mas
coar
The
achi
data
Terr
area
belo
defir
bour
com
inclu
Secc
thes
terra
norn
com
In tt
prov
For
Nati:
of tl
inter
as v
use
digit:
Patc
shru
avoi
inac
used
layet
The
resu
the c
