ted and
s crown
lent has
such as
iferous,
well as
yrmation
re pays
ell as of
ps. The
Field of
'S for an
remote
ent for
r Styria,
nountain
chstein",
i| alpine
tion and
nountain
ere the
de and
he sub-
ant. The
more or
soils as
r stress
pogenic
sastrous
tacks as
ced the
rotective
erosion,
risks of
anches).
f alpine
3. DATA REQUIREMENTS
It cannot be assumed that an alpine wide monitoring
approach can rely only on one sensor type due to
different repetition rates of the satellites and cloud cover
problems. Therefore, the studies were concentrating both
on Landsat-TM and SPOT-PAN/XS data. The technical
specifications of the above mentioned satellite systems
are well known and described in several standard remote
sensing manuals.
A critical point is the selection of appropriate multi-
tempor .| imagery where special attention has to be drawn
to the phenological state of the vegetation. This is
particul..rly true for higher alpine regions which are
characterised by a short vegetation period. Therefore,
also meteorological data has to be used to determine the
optimal acquisition date.
Ancill
Apart from ground data which are required as a priori
reference for the definition of training areas and for the
validation of the classification and monitoring results,
digital terrain models (DTM's) are recommended to
improve the classification accuracy. DTM's are to be used
as input to the geocoding and the topographic
normalisation of all image data. It cannot be expected that
for all parts of the alpine regions DTM's will be available.
In that case a lower classification accuracy has to be
accepted which is particularly a problem in areas with a
rough terrain.
4. DATA PROCESSING
The information on environmental changes results on the
complementary information content of the multitemporal
source data layers. However, many restrictions on data
pre-processing have to be taken into consideration for
practical monitoring applications. These restrictions as
well as methods of data fusion for change detection
Studies will be treated.
An overview of the proposed technical methods is given in
table 1. The different image processing steps outlined in
figure 1 are demonstrated and discussed in detail in the
following sections.
Table 1: Workflow of the Change Detection System
ORIGINAL IMAGES / DTM
4.1 PARAMETRIC GEOCODING
Result: Images with mean position accuracy below one
pixel size
—
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
4.2 TOPOGRAPHIC NORMALISATION
Result: Topographic normalised images suited for
classification with only one signature set
4.3 IMAGE CALIBRATION - ATMOSPHERIC
CORRECTION
Result: Comparable image signatures of multitemporal
data sets
4.4 GROUND TRUTH AND CLASSIFICATION
Result: Classification of normalised (4.2) a. calibrated
(4.3) multitemporal images using solely training areas
derived from the actual image
4.5 CALCULATION OF CHANGE VECTOR
Result: Length of change vectors within windows and
determination of length thresholds for the differentiation
of real changes and changes which are due to
insufficiencies in image calibration
4.6 INTERPRETATION OF CHANGES BY MEANS OF
THE CLASSIFICATION RESULTS
Result: Assignment of significantly changed areas into
change categories according to the classified forest
parameters
4.7 VISUAL APPROVAL FOR CRITICAL CHANGE
CATEGORIES
Result: Final classification results for each acquisition
date and for the change categories
4.1 Parametric Geocoding
The investigation has shown that the analysis of
multitemporal/multisensoral remote sensing data sets can
only be efficiently done if the data present itself in a
common geometry. Geocoding of the images therefore
has to meet extremely strict requirements if the data
obtained at different acquisition dates with different
systems are processed multitemporally in one "data
stack".
Geocoding of the individual images of such a data set to
the geometry of a topographic map is the most common
procedure to accomplish comparability. In general,
parametric approaches based on sensor specific mapping
models have been used for geocoding. Achieved
accuracies lie in the order of one pixel size. However, the
experience from monitoring applications has shown the
267
