4. RESULTS
4.1 Primary Results
The time consumption of the program triplet varies
with the array size and the homogeneity of tested
images. The fastest process is the extraction of trai-
ning sites for further computation. Here time needed
to proceed is a function of the array size of the mo-
ving window. Secondly it depends on the number of
control steps for each centerpixel and its neighbours
carried out. Relief textures smaller then the moving
window disables the TA-finding algorithm.
The most processor time consuming step is the class-
building algorithm. The number of operations perfor-
med for the comparision of the computed training are-
as (TA's) and clusters increases by
(n: 2)-1 (2)
where 'n' is the number of TA's and clusters to be
compared. The number of iterations to proceed de-
pends on the homogeneity of the imagery, and there-
fore the necessity to rearrange the class statistics.
The process of attaching the grid data to the spectral
classes is dominated by the number of classes and
AN +,
Me "Eas SRA A 2
fig.4: A sample of a classified image (upper left of fig.3) as
computed from the tested subset ‘A’ and applied on the
wider vicinity. The values are:
- correlation coefficient =0.9
- standard deviation
- clusterdistance
- feature space distance
for pixels
dx I AD a
=
2
2
3
46
the size of the pixel array to be assigned to the clas-
ses. Obviously the time needed would be exaggera-
ted enormously if a large distance from the classes is
allowed and the feature-space of interest around each
class vector intersects with another.
4.2 An Example
If using the descricbed procedure one will quickly
realise that a heterogeneous composed image in
terms of high and low frequent spectral areas is not
easy to handle when the ‘short waved’ part is under
research. In our example smooth sandy areas will
dominate the result. A work-around would be the eli-
mination of terrain influence by a correction model
(under preparation) or alternatively a smoothing ope-
ration which would obviously effect the resolution and
spectral response (fig.3).
The following tables 1 and 2 show the combination
lists which were obtained from the subarea ‘A’ in figu-
re 3. The band combination used is 7-4-1. The subset
has a size of 2562 pixel, 65.536 Pixel per Layer.
The distinction of metamorphic rocks under the condi-
tions of the proposed algorithm is in parts interesting
especially if certain spectral features of potential oc-
curences can be discriminated. This implies an alrea-
dy good knowledge of the geology at least in points,
but nevertheless it allows to see similar fine distinguis-
hed features in the surrounding area, as can be seen
in figure 4.
Size of Correlation Standard Number of
search coefficient deviation TA's
window
3X3 0.95 2.0 1
9x3 0.95 2.5 14
3X3 0.95 3.0 71
3X3 0.90 2.0 19
3X3 0.90 2:5 105
3X3 0.90 3.0 406
3X3 0.85 2.0 65
3X3 0.85 2.5 302
3X3 0.85 3.0 913
Table 1: Results produced by different correlation coeffie-
cients and standard deviations with calcta-routine.
Cluster
distance
Number of
classes
Iterations
1
1
14
69
18
102
359
61
272
md | eh | eh | od | cmd | mh | eh | eh
733
BIWWWNINN ==
Table 2: Results produced by unique cluster-distance for
the computed number of TA's from table 1.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
0 0 (0)z -
Tl — >
