tested to check whether they match with extracted features from
images. To quantify the checking procedure, a similarity
measurement is defined using robust geometric criteria. In the
second step, a digital elevation model (DEM) is automatically
generated from the DSM. From the difference of DSM and
DEM an above-ground mask is derived. From this mask
buildings are generated and compared with the existing
database in order to detect new buildings. According to the
author, the actual delineation of building outlines is not very
accurate, mainly due to shadows.
Malpica and Alonso (2010) developed an approach for urban
change detection integrating multi-spectral satellite imagery,
LiDAR point clouds and a GIS database. SVM (support vector
machine) was used to classify the image, resulting in a
probability layer for buildings. By intersecting the classification
result with the GIS building layer the authors found an increase
in the built up area of a few percent.
Tian et al (2010) used stereoscopic satellite imagery, to detect
height changes by computing the difference between the DSMs
generated at different epochs. A rectangle was fitted to each
extracted blob assumed to be a building. However, most blobs
are highly curved, so the direction of the rectangle edges cannot
be computed reliably.
2. DESCRIPTION OF INPUT DATA AND PRE-
PROCESSINHG
Two pan-sharpened stereo pairs from IKONOS-2 (epoch 1) and
GeoEye-1 (epoch 2), acquired on May-24, 2008 and Sept.-15,
2009 with ground sampling distances (GSD) of 1m and 50cm
respectively, and depicting a suburb of Riyadh the capital of
Saudi Arabia, are used in our study. The slant angle is 11?
toward West for both and the base-to-height ratio is similar with
1:1.75 for IKONOS and 1:1.51 for GeoEye. Figure 1 shows
parts of the images used for our study.
In addition we have at our disposal a GIS database showing all
the buildings existing in the area in May 2008. Reference data
for the building change detection study were generated
Figure 1. Sample of building changes (a) and (c): GeoEye 2009,
(b) and (d): IKONOS 2008
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
manually by delineating and comparing the buildings visible at
the two epochs.
Image orientation was provided by means of Rational
Polynomial Coefficients. For the subsequent DSM generation
we use semi-global matching (SGM, Hirschmüller, 2008). We
thus need to transform the images into epipolar geometry (or at
least something close to it, since for line images epipolar
geometry in the strict sense does not exist). In our case it was
sufficient to rotate the images around the viewing direction
resulting in the x-axes of both image coordinate systems being
parallel to the base, followed by a shift of 2.5 pixels in y-
direction.
As mentioned image matching was carried out using SGM
(Hirschmüller, 2008) resulting in DSMs for both epochs. The
grid spacing was set to Im for both DSM to have comparable
conditions.
3. BUILDING CHANGE DETECTION USING DSM
SUBTRACTION
This section deals with the detection of building changes by
comparing the DSMs of the two epochs. We quickly found that
simply taking the difference between both datasets did not yield
useful results, mainly since georeferencing on the basis of
rational polynomial coefficients (RPC) was not accurate enough
(see Figure 2, a threshold of 2.5m for the absolute difference
was used to show height changes). We thus applied a shift in all
three coordinates to the second DSM with respect to the first. It
was computed automatically using 3D least squares image
matching similar to (Heipke et al., 2002) and amounted to 7.2m
in X, 1.7m in Y and 1.3m in Z.
TUE
e
= ; E
€ gam TB =
AR es noto T x aps *
mo v v M SUME
s > 2 f
JE iA
Figure 2. (a) IKONOS DSM, (b) GeoEye DSM, (c) and (d)
binary change maps with D-DSMs larger than 2.5m: (c) before
(red), and (d) after (blue) shift elimination
After shift elimination, the differences in height were computed
for each position in object space, see equation 1.
AH (7) 5 H HC) A "UG. (D