'ectral
n this
or the
'ectral
forest
ed on
e area
> Were
1 each
and
owing
iation
| grey
of the
ind 2)
1ction
idates
^
ectral
to |.
The second part of the rule for road identification is based on
the shape of roads. Different directional filters were used to
enhance linear features in the SPOT image. The filters were
constructed to enhance vertical, horizontal and diagonal lines.
Four new raster layers, one for each direction, were created by
filtering. The information in these four layers was combined and
a raster layer showing linear features in four different directions
was created.
CRY
pe )
eed. N
"o^
UAE ix
*
m
y
z^
Figure 4. Membership values for fuzzy statement "line shaped
and thin". Roads have membership values close to 1, and appear
dark in the image.
A second membership function was defined for the class
line shaped. and. thin , based on the pixel values in the layer
with linear features. The membership function was defined in
the same way as for high. spectral reflectance, using median
and standard deviation. However, the minimum and maximum
values for this map layer were set according to:
min = (median + 3 * standard deviation)
max = (median + 6 * standard deviation).
The two resulting layers are fuzzy sets. The logical AND
operator defines an intersection between these two fuzzy sets
(high_spectral_reflectance and line shaped and thin) In
fuzzy logic, the AND operation corresponds to the minimum
value in the fuzzy sets. Thus, in the combination of the layers,
each corresponding pixel in the two layers was compared and
the one with the lowest value was chosen to obtain a map layer
Showing membership in the class road (e.g. Eastman 1993,
Altman 1994). As a final step, a mask based on the topographic
map was used to exclude previously mapped roads.
Example clear-cuts
The rule for identification of clear-cuts was constructed in the
Same way as the rule for roads:
IF high, spectral reflectance
ANDIFNOT road
ANDIFNOT existing clearcut
THEN new. clearcut
537
Here, the combination of map layers is based both on the logical
AND operator and the NOT operator; corresponding to the
complement of the fuzzy set for roads, and the minimum value
of this complement and the membership value for high spectral
reflectance. Finally, previously mapped clear-cuts were masked
out to limit the confusion between old and new clear-cuts.
4. PRELIMINARY RESULTS AND DISCUSSION
The preliminary results are promising. Five known new roads
were clearly detected, with practically no confusion with other
objects of similar spectral characteristics (figure 5). The new
roads are detected with membership values close to 1. Most of
the old roads are also detected, but their membership values are
lower. Since a mask was applied the fact that older roads are
also detected is not a problem, but implies some robustness in
the method. In the final map layer the new roads are clearly
identified and the confusion with other objects is negligible;
few pixels were incorrectly identified as roads.
In this dataset, the statement "line shaped and thin" has more
influence on the result than the statement "high, spectral
reflectance". Almost exactly the same result was obtained from
the single statement as from the two statements combined. The
importance of different parts of the rules would, however,
change from area to area. If, for example, there was a river
present in this area, that too would be "line, shaped, and. thin"
but it would have "low. spectral, reflectance".
\
A
Figure 5. Result from the rule for extraction of new roads. 5
new roads have been detected. The road pixels have
membership values close to 1, and appear dark in the image.
Three new clear-cuts were successfully identified and the
confusion with older clear-cuts was avoided by application of
the mask (figure 6). The membership values differ between the
three clear-cuts The two northern ones mainly have
membership values between 0.8 and 1, while the clear-cut in the
south-east has significantly lower values, mostly between 0.3
and 0.5. However, even without mask they are clearly separated
from older clear-cuts, for which the membership values range
between 0.1 and 0.3, with a few exceptions. There is more
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996