an edge preserving filter has been applied. A sigma filter
was best suited for that purpose (Lee, 1983). The way
how the filter works should be explained very briefly.
Within the filter window the mean is calculated of all that
pixels whose difference to the central pixel is less than
the standard deviation of the image noise. The noise
variance must be known in advance and may estimated
by experience or it can approximately be determined by
analysing the derivatives of the image. As a result of the
filler process we obtain an image with smooth homo-
geneous areas but still significant edges and textures.
Figure 2 demonstrates the efficiency of the sigma filter.
Now we begin the image analysis with the forest
segmentation as described in the previous sections by
selecting suitable training areas for deriving the statistics
for thresholding. The next step is dedicated to texture
analysis by applying the Forstner operator. Figure 3
shows on the left hand side the intermediate result
immediately after the feature extraction (black patches
are detected point features, grey areas are the linear
feature elements). The right hand side picture shows the
original IRS image with the extracted settlement class
superimposed to it.
The final step is the rule based classification in a system
that is built up of some 30 rules where for each of the
maximum seven (see chapter 2) - in our test area five -
basic classes uncertainty categories have been assigned.
5.2 Thematic Mapper + SPOT-pan
As we have seen that entire classificaton process, in
particular the rule section, is rather complex. The
6 RESULTS AND ACCURACY
question arises whether our approach is applicable to
other input data than used for designing the system. The
worst case scenario would be the necessity of a
completely new design, or in a less severe case, the need
to modify the decision rules significantly. If this
happened the system would not be very useful in practice.
In order to test our algorithms whether they can fulfil our
expectations a different data set was subjected to the
classification procedure.
This second example uses a Landsat TM and a SPOT-Pan
image as input. By chance both images were acquired on
the same day in 1997. Since the SPOT resolution with
10 m is notably worse than that of IRS we expected less
reliable results, but more importantly, much more
problems with the analysis of the panchromatic image.
Amazingly it turned out that (at least in our case) the
actual resolution difference was not that essential because
the radiometric quality of the SPOT data was by far better
than that of IRS and therefore, noise filtering was not
necessary at all.
The results confirmed that SPOT-Pan is also best suited
for our classification approach. The forest segmentation
as well as the texture analysis yielded results of a quality
comparable to the IRS analysis. Only a minor
modification to the rule system was necessary proving
that the whole classification scheme was not tailored to
one test data set but can successfully be employed to a
new data set without any problem.
1 2 3 4 5 6
Visual Remote Identitical Identical Identical
Orthophoto- Sensing Identitical Classification | Classification | Classification
Classes Interpretation Analysis Classification with High ina ina3x3
Security 3 x 3 Window | Window with
Resol.: 5 meter Resol.: 5 meter High Security
Pixel % Pixel % Pixel % Pixel % Pixel % Pixel %
Water 20789 | 10.4 20546 | 10.3 20449 | 99.5 20392 | 99.6 20458 | 99.6 20394 | 99.6
Forest 78398 | 39.2 78302] 302 73177 | 93.5 72626 | 95.9 74580 | 95.3 73734 | 97.3
Settlement 19694| 9.8 24662 | 12.3 16488 | 66.9 6148 | 92.8 17976 | 72.9 6327 | 95.0
eds, 81119 | 40.6 76497 | 38.2 69897 | 91.4 56293 | 93.8 71898 | 94.0| 57452 | 97.1
total % 90 96 93 97
Sum 200000 | 100| 200000 | 100} 180011 155459 184912 157907
Table 2: Result of classification
The most important question is certainly: How good does
the classification work? Or in other words: How great is
the percentage of correctly assigned classes? For that
purpose a orthophoto map (original scale 1:10000) was
visually interpreted and declared as reference classi-
fication. Then 20 regularly distributed square windows,
i.e. a total of 296 of the whole classified area, were cut out
of the test area where the difference between the
computer assisted and the reference classification had
been analysed in detail. Table 2 shows the result of the
quality check.
Column / lists all pixels of the visual interpretation.
Column 2 contains the respective pixels of the remote
sensing analysis. Comparing Column / and 2 one can
278 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
see, tl
obviot
Settle
overes
„Field
Colum
if con
identic
“Fores
poor r
uncert:
segme
surrou
interpr
tighter
necess
should
very c
(see fi;
81% «
certain
are ide
the p
unexpe
previo
Additi
inaccui
we mu
data sc
the ac
displac
check :
by pix
of pixe
identic
almost
certain
pixels :
Fig 4: C
the clas:
Figure
left im:
the par
are the
the orti
rough |
rectific.
the sar
slight d
must c«
