alpine
etween
tigation
hanges
jnificant
|, in the
) digital
esented
can be
ssue in
ut very
on was
n of the
jression
covered
ocedure
1sidered
ear cuts
rised by
forested
S) to be
nreshold
). Areas
resulting
s (AND-
rs in the
vill result
reas. As
is study,
ritical in
the case
o reduce
n of the
ats which
ow. The
in values
eters are
yrmalised
alibrated
signature
alibration
objective
ihich has
applied to
indmeier,
od takes
s well as
>, ozone,
monoxide
on of 2.5
sen from
! user.
signed in
methods,
iIbsequent
processing steps described in the following chapters can
be based both on relatively or absolutely calibrated
images. By following this approach any future success in
image calibration can be incorporated into the monitoring
system.
4.4 Classification and Signature Analyses
llection of ground truth
For the classification and change detection detailed
information on the parameters to be monitored is
indispensable. Additionally, ground truth information is
necessary to validate the performance of the different
image normalisation and calibration methods. The ground
information to be included was derived from aerial
photographs for two reasons: First because there are
sufficient suitable photographs available of the areas to
be investigated and secondly, because these are very
well suited to bridge the gap between satellite images and
the actual state of the ground.
Classification
As a first step of the classification procedure forest masks
are created by applying threshold values for each
acquisition date. The mask to be used for the
classification of the forest parameters (classification
mask) is derived by overlaying the multitemporal forest
masks. Since also open forests (clear cuts, deforested
areas due to storm, insects, fire, etc.) have to be
considered areas which are covered by forests at least at
one of the image acquisition date (OR-Operator) will be
included into the classification mask. The result of this
procedure will then be a classification mask consisting of
areas which are
e covered by forests at all acquisition dates
e areas which are open at acquisition date t1 and closed
at acquisition data t2
e and areas which are closed at acquisition date t1 and
open at acquisition data t2.
Basically, a separation of closed forest and non-forest
using satellite data is possible. However, using the
signatures only, it is not possible to separate areas with
severe or total deforestation with vital and lush grass
vegetation from some agricultural areas, like grassland.
Therefore, a multi-temporal approach as described above
is necessary in order to support the separation of forest
and non-forest.
As the next step, a supervised maximum likelihood
classification is applied for each image, which provides
pixel-wise classification results for each acquisition date.
For the classification the TM-bands 2,3,4 and 5 or the
SPOT-bands 2,3 and 4 were selected. The classified
categories are
e forest density
* forest types
* age classes
The classification of the calibrated images taken at
different dates are performed by using only the
signature statistics calculated from training areas
selected from one of the multi-temporal data sets. The
use of different signature statistics calculated from multi-
temporal sets of independently selected training areas
should not be used since
e itis hardly possible to delineate exactly the same area
in two data sets in practice
e itcannot be guaranteed that these training areas have
not changed
e ground truth is not in general available for different
acquisition dates.
The results of the classification now are different data
sets providing information on the distribution of the forest
classes at different times. The classification results will
then be used for interpreting the change of spectral
signals as described in the next chapter.
4.5 Calculation Change Vector
In a next step the change vector of the actual and the
historical image is calculated in order to detect the
changed areas. The calculation of the change vector can
be based either on single pixels or on larger areas such
as systematically applied windows or specific forest areas
whereby the latter requires additional GIS information on
the distribution of forest units. In the following both
procedures (single pixels and larger areas) will be
discussed.
Change vector based on single pixels
Several change detection methods have been developed
either based on classified or on calibrated images. À main
restriction for the application of change detection methods
is the position accuracy when overlaying the different
image data sets. To investigate the resulting
displacement effects for forest classifications we
performed a displacement simulation for a actual forest
classification. The simulation was carried out by shifting
the forest type classification (for three classes) 0.5 pixel to
east and 0.5 pixel to north which was assumed to be the
optimal geometrical accuracy level to be reached by
parametric geocoding of satellite images. The
superposition error resulting from the presumed image
distortion of 0.5 pixels can then be estimated by
comparing the shifted classification with the original
classification. The remaining correspondence of both data
sets in dependence on the different forest types is listed
in table 2.
Table 2 has to be interpreted in the following way: If no
displacement errors occur, then the percentage of the
comparisons conif. ---> conif. should amount to 100%. But
in this case only 86.74% of the coniferous pixels still
belong to the class "coniferous" after the displacement
procedure. This means that the displacement leads to an
error of 13.26%. This error splits up into the coniferous
pixels which are incorrectly displaced into the class
"mixed forest" (11.17%) and coniferous pixels which are
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 269
