In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
2.3 Change areas extraction 
In binary change detections, one of the most important final 
steps consists in highlighting real positive and negative 
changes. For spectral images, a simple thresholding of the 
histogram has been widely used to stress real changes and 
remove the virtual ones (Bazi, 2005; Bovolo, 2006; Sen and 
Pal, 2009). While for the 3D change detection, in the DSM 
generation procedure, much information is already missing. 
Simple thresholding on the “difference image” will destroy 
the more original information. Therefore, automatic building 
extraction approach is adopted in our research. It can be 
divided into 3 steps, 
1) Edges extraction: In this step, the Canny edge (Canny, 
1986) extraction method has been adopted. As our focus in 
this work consists in the detection of the urban changed 
man-made structures (building construction/destruction), 
small edges will not be considered in our research. 
2) Mask generation: Since most of the Canny edges are open, 
and could not be filled automatically, we choose to close 
all of the edges with morphological algorithms. We fill 
each closed edge to single mask, which presents the 
changed area. 
3) Box-fitting based building shape refinement: In general, 
according to the quality of the original DSM data and also 
the edge detection and mask filling result, most of the 
edges are highly curved and much information is missed. 
Therefore, the building edges need to be refined so that 
they regain their sharp shapes. In this work, we used the 
box-fitting method proposed by (Sirmacek and Unsalan, 
2008). For this method, we need seed points, which show 
the location of the changed building; the edges which 
control the size of the box for the changed building; and 
also the automatic box growing direction and stop 
condition. In our research, we locate the seed points in the 
centre of each mask that is generated in step 2. Only the 
original edges around each buffered mask area are 
considered to be the edge of this building. 
the original difference image of Figure 2 (b) that is depicted 
in Figure 3 (a). Figure 3 (b) shows the edges extraction 
results where only important edges are kept (small edges are 
removed). The mask generation output is displayed in Figure 
3(c). The result of the box-fitting method is depicted in 
Figure 3 (d) where a more refined version of the changed 
objects is obtained. 
In the change value extraction procedure, the horizontal 
change can be easily calculated according to the pixel 
numbers of the mask area and the DSM resolution. In order 
to get only one vertical change value c for each 
constructed/destructed building defined by Mask or box 
fitting result. We average the pixels values in the “difference 
image” belonging to the same changed object, and define this 
value as the vertical change of each building. In the following, 
we exclude all pixels, which have ‘0’ value (no height in the 
changed area), very low values or very high values and could 
therefore be artifact change, so that these pixels will not be 
involved in the mean value calculation procedure. As 
displayed in Figure 4, only the middle part (gray colour filled) 
of the height difference values are used. 
MinValue Max Value 
-y SH Sr 1 
0 10% 10% 
Figure 4. Vertical change value extraction strategy. 
3. EXPERIMENTAL RESULTS 
3.1 Description of the data 
In order to evaluate the performance of our approach, we 
have chosen the city centre of Munich in Germany as study 
site. Two DSMs from two different epochs have been used to 
detect the potential changes. The first DSM (called in the 
following as IKONOS-DSM) is computed from IKONOS in- 
orbit stereo imagery (level 1A, viewing angles +9.25° and - 
4.45°) with one meter spatial resolution, acquired in July 
2005. It has been generated using the Semi Global Matching 
(SGM) algorithm implemented at DLR (d’Angelo, 2009). 
Due to the lack of another stereo pair, the second one we use 
instead a DSM which is generated from a LiDAR point cloud 
data acquired in February 2003 (called in the following as 
LiDAR-DSM). 
00 1000 ’ " ~ "(b) 
Figure 3. Change areas extraction: (a) Robust difference 
image; (b) Edges extraction; (c) Changed building masks 
generation; (d) Box-fitting-based shape refinement of the 
changed areas. 
Figure 3 summarizes the results of the already described steps 
when applying our changed areas extraction on a subset from 
Figure 5. DSMs of the city centre of Munich: (a) LiDAR- 
DSM (2003); (b) IKONOS-DSM (2005), vegetation cover 
removed. 
Since the LiDAR data used in this work is generated from the 
last return pulses, which represent the bare earth terrain, the 
vegetation cover information is almost fully removed.