ing and
that the
üch can
as to be
ground.
lect the
orithms
01%. In
1 to less
images
t image
well but
| a good
e result
that the
tage of
and the
ted new
bjects.
vided in
aluating
ta using
ong the
> matrix
'unarity,
Rob Dekker
a fractal based feature, turned out to be one of the more powerful features. It is computed from the content of the total
number of specified windows within a region. When the size of this window is only one pixel, the lacunarity of a region
is given by 0 /J--1, in which c and u stand for the standard deviation and mean intensity of the region (Dobson et al.
1997). Both the mean intensity and lacunarity were used for the segmentation of Dar es Salaam. To reduce the effects of
speckle as much as possible, it was performed on the sum of the image pair of 3-4 October 1997. This way an image
with a number of looks up to 12, for fully uncorrelated areas, is obtained. The result is given in figure 5. For the reason
of comparison, the result of manual segmentation from vertical aerial photographs of 1992 (1:12,500 and 1:52,000,
Sliuzas et al. 1999) is given in figure 4. Figure 5 is of course less accurate than figure 4, but is useful for larger scale
interpretation. Note the difference with figure 3, with respect to the amount of speckle and the number of looks.
Unfortunately no time was left to study the potential of texture for classification, but considering its capability of
segmentation, it is recommended. The next section on interferometry comes up with another feature that can be used for
segmentation and classification.
Figure 4. Manual segmentation of land use from vertical aerial photographs of 1992 (1:12,500 and 1:52,000), produced
by ITC, The Netherlands, projected on the sum image of the tandem pair of 3-4 October 1997.
Figure 5. Automated segmentation of land use from the sum image of the tandem pair of 3-4 October 1997, projected
on the sum image.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 65
