be true, the corresponding secondary class is assigned to the
center pixel of the window.
Each rule defines the minimum frequency of one or more pri-
mary classes for one secondary class. When compared to the
corresponding elements in the histogram the frequency values
represent thresholds. If all thresholds are exceeded within a
rule, it is recognized as true and the corresponding secondary
class will be assigned. Though experiments with this approach
produced useful results (Steinnocher et al. 1993), only simple
patterns of primary classes could be recognized. Therefore the
design of the rule-set was modified to allow for a combination
of sub-rules within one major rule. Each sub-rule defines a
threshold for one or more primary classes and all sub-rules
have to be true to accept the major rule (Figure 1). Processing
of the rule-set is performed step by step, starting at the top of
the set. As soon as a rule is accepted and therefore applied, the
rest of the rule-set will not be considered any more. If no rule is
found to be true, a rejection class is assigned.
Apart from the design of the rule-set, the size of the analyzed
neighborhood represents a crucial parameter in the postclassifi-
cation process. Choosing a small window size will lead to a
‘noisy’ result since only high frequency structures will be rec-
ognized. If the window is too large the smoothing effect will
become very strong, thus leading to a loss of detail. At this
point it has to be noted that the presented postclassification is a
generalization process and will always suppress some details.
On the other hand, this effect might as well be desired, e. g. for
the generation of thematic maps (Wilkinson, 1993). Since
generalization usually comes with an increase of scale - i.e. an
increase of the pixel size in the raster domain - the algorithm
includes the option of resampling, i.e. the size of the resulting
pixels can be defined as a multiple of the original pixel size.
Since the rule-set and the window sizes are defined by the user,
the right choice of these parameters depends highly on the
user's experience and on the objective of the application.
3. GENERATION OF THE LAND-USE MODEL
3.1 Data description
The data used in this application comprises 12 cloudfree Land-
sat-TM scenes, covering the entire area of Austria. All images
were acquired between August 7 and October 5, 1991, except
one quarter scene, which was taken in August 1992. Due to
stable weather conditions within the period of data acquisition,
this data-set has a homogenous reflectance characteristic and
therefore represents an optimum basis for further processing. In
addition to the image data, a digital elevation model of Austria
with a resolution of 50 m was available.
For training and testing of the classification process reliable
reference information is indispensable. To guarantee a consis-
tent quality of the results, only data available for the entire area
of Austria were used. The Austrian topographic map 1:50.000
(ÓK 50) consisting of 213 map sheets provided information on
major land-cover/use types such as man-made structures, water
bodies, forest, bare rock and glaciers. Though the majority of
the maps were updated in the late 1980's, a visual comparison
with the image data was performed for training- and test-areas
to ensure that no change had occured between the update and
the acquisition of the image data. Since the maps do not distin-
guish between the different uses of open land such as arable
land, pastures, natural grassland etc., a second source of infor-
mation was needed. It was found in a series of analogue satelli-
te photographs, covering about 80% of the Austrian territory.
They were taken by a KFA-1000 camera mounted on the Rus-
sian space-platform MIR in 1991. The images offer two chan-
nels in the red and the near infrared spectrum with a ground
resolution of approximately 7 m. Interpretation of these images
proved to be extremly valuable for generating reliable reference
information.
3.2 Geocoding
To allow for a correct geometrical relationship between re-
motely sensed imagery and other spatial information layers
such as maps, it is necessary to geometrically transform the
images to a map projection system. This transformation is
commonly called rectification or geocoding. In flat terrain it is
sufficient to apply a polynomial transformation based on
ground control points. This approach will not be adequate in
rugged terrain, since pixel displacements resulting from local
differences in elevation are not considered. As most parts of
Austria are extremely mountainous a high level geocoding
method has to be applied to ensure a geometrically correct
result. Based on linear ground control features, the orientation
parameters of each image scan are computed by bundle block
adjustment. Next the image-scans are geocoded with respect to
a Digital Terrain Model. The final result is an Austrian wide
ortho-image mosaic with a ground resolution of 25 m. As this
part of the processing chain was not performed by the author,
no further discussion will be given on this topic. Details on the
theoretical background of high level geocoding and on the
generation of the Austrian image mosaic can be found in Ecker
et al. (1991) and Ecker et al. (1995).
3.3 Spectral classification
As the amount of data to be processed comes up to more than 2
Gigabytes, the ortho images are stratified with respect to the
different Austrian landforms. The average size of the resulting
sub-scenes is about 5000x5000 pixels, including overlap areas
between the scenes.
ELSE IF ...
ELSE rejection class
IF F,, [1F,, ...] ? thr [AND FA[-F,, ...] ? thr...] THEN SC
ELSE IF F,[+F,, ...] > thr [AND F,[+F,, ...] > thr ...] THEN SC
ELSE IF F, [+F,, ...] > thr [AND F, [+F,, ...] > thr ...] THEN SC
with F,,.: relative frequency of primary class; SC: secondary class; thr: threshold
Figure 1: syntax of the rule-set
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
0 JON Ua ER UO VD
=
£.
3.4