r. It is
immed
89).
lly and
Iso be
all rms
of a
tes that
erflow.
cover
tif the
ay fall
)sed to
y pixel
ared to
ne. To
el must
is pixel
le tests
ranged,
d. The
ers. An
a pixel,
el used
ww. An
le to an
at pixel
‚present
v. The
med by
1e the
only or
1e pixel
> in the
cing up
a pixel
e in the
areas is
stration
B
all six
mber is
; and/or
. Smith,
s, using
alues of
cted the
values of an endmember in 1986 is very similar to 1991.
Differences are due to the fact that the images were not
spectrally calibrated and to seasonal changes. The
endmembers have a high spectral contrast which is
necessary for a successful separation.
| Endmember | Vegetation | Built-up Area Water
TM Channel (86/91) (86/91) (86/91)
1 85/70 147/151 99/96
2 33/32 72/79 43/45
3 27/24 86/99 40/49
4 152/161 77/92 26/27
5 88/106 133/165 9/7
7 25/3] 86/97 6/5
Table 1:DN-Values of Endmembers for 1986 and 1991
Using the selected endmembers the mixture rule (1) was
applied to both satellite images individually, resulting in
three fraction image per image, giving the proportions of
vegetation, built-up areas, and water plus the shadow and
the rms-error images. The fractions are rescaled according
to the rules of table 2 to allow visualization.
Fraction | Fraction Image Value
«-] 0
-1 to 0 0 to 100
0 to 1 100 to 200
1to 1.55 |200 to 255
155 285
Table 2: Rescaling of Fraction Images
The next step is to use to use the fraction images to
determine those areas, where building activities have taken
place.
3. CHANGE DETECTION
The method used here to detect urban growth is closely
related to image differencing. The main use of this
technique so far has been in subtracting bands or principal
components from one another. Both have the disadvantage
that neither individual bands nor principal components
contain information which may be regularly related to a
special land cover type. Fraction images, on the other hand,
offer the advantage of containing a priori defined
qualitative information (certain land-cover type). The
method to detect changes suggested here is to subtract
fraction images from one another, which represent the same
land cover type, calculated from satellite images recorded
at different dates. If the fraction images represent the same
type of information changes should be clearly seen, as the
fraction images values must be higher or lower for pixels,
Where the land cover type has changed compared to a pixel
from an earlier date. Two premises must be satisfied before
à change detection may be attempted. First it is necessary to
make sure that the fraction images actually represent the
same information. To check this, the histograms of both
fraction images are compared. If the information is the
Same, then the general shape of the curve must be
381
approximately the same as well except for minor
differences which are due to land cover and seasonal
changes. Also, one must make sure that the information
shown by the fraction images is as pure as possible, i.e.
only the land cover type in question is represented. If that is
not the case, the change detection will be negatively
influenced, and methods must be found to remove these
influences. To do that the inclusion of one or more other
fraction images in the change detection might be advisable.
Figure 1 shows a comparison of the histograms of the
fraction images for built-up areas for 1986 and 1991.
40000 or
30000 | qr
20000
Number of Pixels
VERS t A -e
80 100 120 140 160 180 200 220 240
DN-Values
Figure 1: Histogram of the Fraction Images for Built-Up
Areas for 1986 and 1991
As the histograms of 1991 has a shift of 5 DN-values to the
left, as compared to the histogram of 1986, the 1991
histogram was corrected by these 5 DN-values for the
comparison. The histograms have a very similar shape,
with the exceptions of a peak at DN-value 100 in the 1986
histogram and a peak at DN-value 105 in the 1991
histogram. These peaks are due to different cloud covers in
different parts of the image and changes in vegetation. As
these differences are in areas where there are no buildings
(a DN-value of 100 is equivalent to a fraction of 0), these
slight differences will not affect the change detection.
The next step is the subtraction of the fraction image for
1986 from the fraction image for 1991. The result is a new
image, and to show the areas of interest, all pixels which
have a positive difference of more than 20 are highlighted.
The threshold value of 20 was found to be most suitable
after examining different values. A problem encountered
here is the differentiation between bare soil and built up
areas. To overcome this problem, the fraction image for
water from 1991 is also included to make the
differentiation more reliable. It was found that very low
values in the fraction image for water and high values in
the fraction image for built-up areas is an indication of bare
soil rather than buildings. Although the change detection
was carried out for the whole city of Vienna, a
development area in the north-east of Vienna was chosen to
examine the results of the method in detail. The result of
this analysis is a map (Figure 2) which shows where
building activities have taken place (black), or might have
taken place, but are more likely to be fields (grey).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
