anbul 2004
d shapes of
etated areas
0 includes a
- cars) and a
The area is
taken place
ler (E, N) z:
150.6187550).
ages (RGB,
and approx-
strations are
s have been
shed.
nd they are
thorectifica-
gital terrain
e shows the
'osed in yel-
nt has taken
[his is most
ildings have
ing material
>r the classi-
gs, and as it
yn and roads
ering. It can
'etated areas
zes found by
osed in yel-
ted
the test area
hed buildings
cen detected.
ection where
existing reg-
ed. buildings
ht area of the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
Figure 4: Results from the change detection algorithm.
Un
UJ
test area (marked by a red circle). And as can be seen a new build-
ing has been built exactly at the same position as where the two
old buildings were positioned. Due to the comparison method
neither the new nor the demolished buildings are “highlighted”
by the algorithm. Only two new buildings have been detected
by the algorithm. The reason for the poor result is that 9 out of
the twelve new buildings have completely different spectral re-
sponses than the existing buildings in the area, as they are either
blue (6 buildings) or still not finished (3). As one of the two
hypotheses regarding the change detection procedure are not ful-
filled, the algorithm is expected to fail.
45 false alarms (3 times the factual changes found) are generated.
Of those, 26 are located in vegetated areas, 3 are bridges (above
terrain), and 16 caused by existing buildings which apparently
have not be re-detected by the algorithm. If infra-red images are
available, the majority of the false alarms can be eliminated using
the NDVI or by calculation of textural figures using the DSM, as
it can be expected that the texture for forrested areas differs from
buildings. The 3 false alarms caused by bridges can only be elim-
inated by the use of a more "clever" algorithm for nDSM genera-
tion. Looking more into the false alarms caused by buildings not
re-detected, it can be seen that a large proportion of those build-
ings are actually detected (figure 4, examples marked by green
circles), but these detections are eliminated in a later stage of the
change detection algorithm, as part of noise reduction. Refining
the noise reduction method may lead to more existing buildings
being “re-detected”. One of the false alarms (shown by a white
circle in the upper right corner of the area), is a factual difference
but not a change, since it is a roof covering a gasoline station,
and such roofs are not to be registered in the TOPIODK database,
according to the map specification. Such false alarms can only be
verified by a human operator.
6 CONCLUSION
The method presented shows reasonable performance when de-
tecting demolished buildings (12 out of 14 are detected), whereas
the number of new buildings detected is poor (only 2 out of 12).
A reason for this exists, as the new buildings do not share the
spectral response of the existing buildings in the area. One of the
hypothesis for the algorithm is not fulfilled. The algorithm in-
troduces a fairly large number of false alarms (3 times the num-
ber of factual changes detected). Most of these false alarms can
be eliminated by refining some of the processing steps in the al-
gorithm (noise reduction, DSM generation) or by introduction of
additional information e.g. infra-red images or textural measures.
Acknowledgements:
I would like to thank the Danish engineering and mapping com-
pany, COWI, for letting me use their Digital Surface Model. I
would also like to thank my colleague Thomas Knudsen, who
contributed valuable comments and suggestions.
REFERENCES
Asperen, P. V., 1996. Digital updates of the Dutch topographic
service. In: International Archives of Photogrammetry and Re-
mote Sensing 31(B4), ISPRS, pp. 891—900.
AVLBD, 1988. Amtliches topographisch-kartographiches in-
formationssystem (ATKIS). http://www.adv.online.de/
enalish/products/atkis.htm.
