In 1995 the Austrian Ministry of Science and Transport
initiated a research project which was intended to further
applications in remote sensing. The acronym MISSION
(Multi-Image Synergetic Satellite Information for Ob-
servation of Nature) expresses the notion of the project of
extensively utilizing the current possibilities of
spaceborne sensing. The project brought together
organisations, companies and various research institutes
which were interested in solving current or imminent
problems by remote sensing techniques. MISSION has
been subdived into tasks and one of these project tasks
has been carried out by the Austrian Federal Office of
Metrology and Surveying (BEV) together with the
Institute of Photogrammetry and Remote Sensing of the
Vienna University of Technology (IPF) as scientific
partner. The idea behind BEV's project was to check
whether current satellite data may be used for the
completion and the update of its indigenous nationwide
topographic database, the so-called ,Digital Landscape
Model" (DLM). The DLM contains a great variety of
data whose geometric quality is related to the scale of the
Austrian Map 1:50000 or 1:25000. The problem the BEV
was most interested in was the classification of forest
boundaries and detection of changes of settlement areas
carried out by procedures that work automatically to a
high extent. The requirements were defined by setting the
geometric accuracy better than 20 m and the
classification accuracy significantly better than 90%. As a
desirable feature a reliablity code should indicate those
results that were less reliable and needed a further check
by visual interpretation or on site inspection.
Remote sensing imagery that was available for the project
was, firstly, the high resolution panchromatic pictures of
IRS-1C (5 m) and SPOT (10 m). They should guarantec
the requested geometric accuracy of better than 20 m.
Additionally, multispectral datasets are absolutely
necessary for a good quality of classification of landuse
categories. The Landsat TM (30 m) seemed to be most
appropriate, though notably below geometric require-
ment. Already available database contents of the DLM
serve as input too. Some of these data have been derived
from the Austrian topographic map 1:50000 (OeK50) by
simple digitization as provisional input in the DLM. The
geometric accuracy of those data are within the graphical
accuracy of those maps that is about 5 m to 10 m in the
ideal case. Other elements of the DLM were digitized in
the orthophoto maps of the scale 1:10000, where a very
good geometric quality had to be expected, and a third
category originates in stereo compilations of aerial
photographs.
2 MULTISPECTRAL CLASSIFICATION
The multispectral images were highly responsible for a
good classification of landuse. The focus lies on an
acceptable reliability of class discrimination and less on
the geometric accuracy of small details. Although only
wooded areas and settlements were the key classes of the
project, one has to introduce additional classes that are
able to describe the landuse of the area of interest in a not
necessarily complete but at least satifactory way. The
274
classes (and subclasses) defined for the area under invest-
igation were the following:
e Water" (water bodies such as lakes, rivers)
e Forest" (coniferous and deciduous forest)
e Settlement" (urban areas, villages, hamlets and
possibly individual farm houses)
e Field" (agriculturally used ploughed and unploughed
fields)
e Grassland” (meadows, pastures)
e Rock" (high alpline areas without vegetions)
e Glacier
The classification has been performed according to the
decision rules of the well-known maximum liklihood
algorithm. One knows that the quality of a classification
increases if the multispectral signatures of the classes
involved form a homogeneous and normally distributed
cluster. Heterogeneous classes such as urban settlements
that are mixtures of sealed, unsealed areas, vegetation etc.
lead to huge and quite often non-normally distributed
clusters. The discrimination by assigning the class with
the highest probability density is mathematically unique
but not nessarily the only logically correct choice as in
not well-defined classes also the boundaries are rather
uncertain. A typical example is the discrimination of
Settlement" "and. „Fields“ where in many cases
settlement classes are found in agricultural areas.
The multispectral decision rule was therefore slightly
extended by assigning two classes for each image pixel:
the class with. the. greatest probability density (P1) and
that with the second greatest probability density (P2). The
smaller the difference (PI-P2) between this two
discrimination values the less certain and reliable is the
original maximum liklihood decision Pl. A slight
variation of the training samples may have lead to the
opposite result. In order to find a quality or significance
measure for the class difference and the reliability of the
class membership of the final decision the ratio of P1 and
P2 (more exactly the Mahalanobis distances of the P1 and
P2 decision) is calculated and the F-test is applied. We
found out that classes , Water" and ,Forest" could be
classified with high reliability, ,Fields* and
settlements" on the other hand are rather uncertain
classes; both classes may be expected to appear mixed up
in the final result. As already mentioned above the
boundaries between settlements and agricultural fields are
uncertain, settlement pixels may be found in field areas
and vice versa.
3 PANCHROMATIC IMAGES
Since panchromatic images are currently the only ones of
high spatial resolution, they offer the possibility to
achieve the requested geometric accuracy on the one
hand, on the other hand this sort of images can be used
for texture analysis as a supplement to the multispectral
classification. In particular the ,,Settlement“ class, that is
spectrally not well-defined can successfully be processed
and classified through texture analysis due to its high
spatial frequency and very typical textural appearance. In
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
our [
for t
boun
3.1T
Class
alrea
One
been
apply
know
metri
meas
non-c
secon
(W-Ve
w-val
such ;
and |
track:
dense
open
clearl
These
settlei
can bx
The p
featur
expan
areas.
growt
region
is to ri
shrink
of de
textur:
3.2 Th
Forest
image:
seeme
classif
form r
delivei
grey a
bit bri
threshc
begins
find su
deviati
analysi
threshc
threshc
