le at
ional
ation
. We
or at
polar
| was
ction
eing
n y-
SGM
The
rable
s by
| that
yield
s of
ough
ence
in all
st. It
nage
7.2m
d)
fore
uted
(0)
where AH, , Hag; 9nd Ho,
and the DSM values for position (i, j) in the first and the second
epoch, respectively. We call the result a Difference-DSM or D-
DSM and created a binary map by keeping only heights with
absolute values larger than 2.5m.
represent the height change
Moreover, DEMs were generated for each epoch by filtering the
DSMs according to (Niemeyer et al 2010), and normalised
digital surface models (nDSM) were computed as the point-by-
point difference between DSM and DEM. Only heights above a
threshold (we also used 2.5m here) were considered as building
blobs in the nDSM, in this way accumulation of matching errors
in the D-DSM is reduced. For epoch 1 these blobs were
additionally compared to the existing GIS building layer and
where kept only, if they covered at least 75% of a GIS building
object.
Ideally, positive and negative change values in the D-DSM
indicate construction and demolition of buildings. The used
threshold of 2.5m corresponds to the floor height of buildings.
This threshold eliminates non-building changes caused by cars,
low shrubs as well as other small changes not corresponding to
buildings and reduces the information to potential building
changes. However, height changes may also be caused by
dumps, land excavation or different filling heights of petrol
tanks. In addition vegetation and in particular trees normally
pose problems when just subtracting height values, although in
Saudi Arabia, this issue did not prevail.
It becomes clear when comparing Figures 2 (c) and (d), that
although applying the shift results in a major improvement,
even after proper georeferencing areas around the buildings
show apparent height differences above the chosen threshold.
This is mainly due to well-known image matching artefacts
stemming from occluded areas and different shadows due to
differences in view and illumination direction: the DSM is
widened with respect to the actual building in shadow and
occluded areas (e.g. Alobeid et al., 2010; Le Bris and Chehata,
2011, see also Figure 3).
Figure 3. (a) Sample DSM of image matching, (b)
corresponding image with cross section line, (c) schematic cross
section from the side (red: DSM, black: original object)
Therefore, the 3D change map has to be refined appropriately.
Shape and size information can play an important role, as does
the normalised difference vegetation index which is known to
be able to separate vegetated from sealed areas. Additional
challenges due to occlusions of buildings by trees have to be
dealt with separately.
Most of the false change alarms have an elongated shape;
however some of them are small blobs like salt-and-pepper
noise. In order to improve the results we use morphological
filtering. The filter mask is chosen based on the usual minimum
size and width of a building, we chose a size of 50m? and a
width of 4m.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Figure 4 shows a diagram of the method for 3D building change
detection.
DSM
(Epoch 2)
DSM
(Epochl)
jnduj
DSM
Co-registration
yv
D-DSM
Generation
Refinement by
Height, Shape
and Size Info
UOIBULIOJU] AIRI[IXNY 3uis() juoureuro»
Sutissooo1q
3D Building
Change Detection
nding
Figure 4. Diagram of 3D building change detection
4. EXPERIMENTAL RESULTS
4.1 Binary D-DSM
In this section we present first preliminary results of the
described method using the test data from Saudi Arabia. The
area has a size of 550 meters on each side (0.30 Sq. Km).
Figure 5. D-DSM of the whole area as a binary map after
applying a height threshold of 2.5m
