In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
(a) Champion, 2007
(b) Matikainen et al., 2007
(c) Rottensteiner, 2008
Figure 2. Completeness (diamonds) and correctness (squares) of the detection results as a function of the building size [m 2 ].
of 20 m 2 in width. Note that the buildings from all the test areas
for which results were submitted are combined in order to have
a significant number of changes for each bin. The graphs for
(Champion, 2007) and (Rottensteiner, 2008) also contain the
results from the Toulouse area. The completeness and
correctness rates, computed independently for each bin, are
presented in Figure 2 and demonstrate the close relation
between the quality of change detection and the change size.
This is true for the completeness with (Champion, 2007) and
(Rottensteiner, 2008), but it is even more obvious for the
correctness in all three graphs. Correctness is particularly poor
for buildings smaller than 100 m 2 . Looking at these graphs it
becomes obvious that the two major problems observed in
Section 4.1, namely the potentially critical rate of missed new
buildings, which limits the qualitative effectiveness of change
detection, and the poor correctness for demolished buildings are
caused by the same underlying phenomenon i.e. the fact that
small objects cannot be detected reliably by an automated
procedure. Attentive readers may also notice that a very low
correctness occurs with (Matikainen et al., 2007) with buildings
covering about 235m 2 . It is caused by large ground areas in the
Marseille test area that were mistakenly classified as above
ground objects and then wrongly alerted as new buildings.
4.3 Impact of the Quality of the Input Data
Our experiments show that many FP cases are related to the
quality of the input DSM. The correlation DSMs used in the
imagery context contain a lot of erroneous height values,
especially in shadow areas (where stereo-matching algorithms
are known to have problems) that are almost systematically
alerted as new buildings, as depicted in Figures 3a, 3b, 3c, and
3d. These errors contribute to lower the correctness rate,
especially for new buildings, which drops to 63.5% with
(Champion, 2007) in Marseille. The high rate of 96.3%
obtained here with (Rottensteiner, 2008) may be related to the
use of the initial description of the database as a priori
information for producing and improving the building label
image. In Toulouse, FP new buildings were also related to DSM
errors, caused by repeating patterns. Another problem concerns
the quantisation effects i.e. the fact that the numerical resolution
of height values in the correlation DSM is restricted to the
GSD, which for instance prevents the use of surface roughness
as an input parameter for the Dempster-Shafer fusion process in
(Rottensteiner, 2008) and ultimately contributes to lower the
correctness rate for demolished buildings.
Regarding the Lyngby test area, it was a problem that the
original data were not available. Single points inside water areas
were not eliminated from the data, but used in an interpolation
process based on a triangulation of the LIDAR points,
producing essentially meaningless data in these water areas that
for example caused FP new cases with (Olsen and Knudsen,
2005). The other problem was that first pulse (rather than last
pulse) data were provided, which caused FPs in areas with
dense vegetation, e.g. along rivers with (Rottensteiner, 2008).
Combined with a relatively low resolution (1 m), these
problems contribute to lower the correctness of the systems.
4.4 Impact of Other Topographic Objects in the Scene
In our experiments, some confusion occurs between buildings
and other above-ground objects that are present in the scene and
wrongly alerted as new buildings. Again, this contributes to
lower the correctness achieved for new buildings. The methods
deal with this problem, but currently they only focus on one
class of above-ground objects that is to be separated from
buildings, namely trees. In general, these trees are identified
with indicators based on the NDVI and then eliminated, as
shown in Section 3. Even though this strategy appears to be
efficient, our experiments show that such confusions are not
limited to vegetation but concern other objects that not
considered in the approaches presented in this study. For
instance, bridges or elevated roads are highlighted as FP new
buildings in the Lyngby test area by (Rottensteiner, 2008) and
(Olsen and Knudsen, 2005), as shown in Figures 3g and 3h. To
limit the impact of these problems, two strategies could be
considered in the future. The first one consists in developing
more sophisticated methods that are capable of simultaneously
extracting multiple object classes such as buildings, roads, and
vegetation. Such methods would need to incorporate complex
scene models that also consider the mutual interactions of the
object classes in a scene. They could make use of recent
developments in the field of Computer Vision that are related to
the modelling context in image classification (Kumar and
Hebert, 2006). The second strategy consists in using additional
information on other objects, e.g. by incorporating an existing
road database in the building change detection procedure.
Additional Remark: Beyond the statistical aspects, our
experiments show that the errors generated by the change
detection approaches are often identical. Thus, the FP cases that
occur in the Marseille test area because of the DSM
inaccuracies (Section 4.3) are both present in the outcomes of
(Matikainen et al., 2007) and (Champion, 2007), as illustrated
in Figures 3a and 3b respectively. Some of other errors shared
at least by two approaches are also illustrated in Figure 3.
5. CONCLUSION
Four building change detection approaches have been tested in
three different contexts. If the satellite context appears to be the
most challenging for the current state-of-the-art, the aerial
context and the LIDAR context appear to be a viable basis for
building an operative system in the future. Thus, the high
