In: Wagner W„ Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
587 
different approaches have been adopted to highlight the real 
horizontal and vertical changes. For the horizontal changes, 
we apply an edge detection approach followed by a box 
fitting method in order to extract the real changes relative to 
the constructed/destructed building borders and remove the 
virtual ones coming from the different nature of the used 
DSMs. In the case of the vertical changes, we extract the 
height values of the changed objects based on a statistical 
method. Finally, the detection results are compared with the 
manual extraction records. 
2. APPROACH TO THE 3D CHANGE DETECTION 
2.1 Workflow of the proposed 3D change detection 
We aim at extracting height information from the stereo 
imagery, then generating a change-detection map that 
represents 3D changes between the two datasets. We focus 
our research on noise reduction and change detection areas 
extraction. 
The overall workflow of the proposed 3D change detection 
method is shown in Figure 1. The first step consists in 
generating DSMs from two pairs of registered optical stereo 
imagery acquired over the same area (here the city centre of 
Munich) at two different epochs and t 2 . In this paper, the 
DSMs are computed using the Semi-Global Matching (SGM) 
method (Hirschmiiller 2008, d’Angelo 2008). A further co 
registration between the two resulting DSMs has been 
necessary to remove any shift in three dimensions that might 
exist between the two DSMs. After that, the “difference 
image” is produced where real changes are highlighted and 
the influence of the noise (or virtual changes) is reduced 
(sub-section 2.2). Then, this difference image is analyzed by 
means of building edge detection in order to retrieve the 
borders of the different constructed/destructed buildings. 
Both of the positive change (new constructed buildings) and 
negative changes (destructed places) in vertical and 
horizontal direction are extracted. 
Figure 1. 3D Change detection process proposed in this 
paper. 
2.2 Noise reduction 
A major problem to cope with during change detection is the 
reduction of different kinds of noise. In our research, the 
noise is defined caused from: 1) Mis-coregistration. In 
practice, when two images are co-registered at sub-pixel 
accuracy, the true location of a pixel’s central point may be 
anywhere within the pixels surrounding the point (Goodchild 
et al., 1994, Guillermo et ah, 2009). It is unlikely that the 
footprints of two coincident pixels of the DSMs correspond 
to the same area (Bruzzone et ah, 2003). 2) Quality of the 
DSM. Due to the erratic variations of the stereo images 
acquisition conditions, the DSM that generated from stereo 
imagery has some missing information (called holes in the 
DSM) caused by the unsuccessful stereo image matching of 
corresponding pixels. If we analyse the DEM from the pixel 
level in the change detection procedure, the “holes” will be 
detected and displayed as noise in the difference image. 
Many noise reduction methods have been developed in the 
literature (e.g. Gong et ah, 1992; Bruzzone et ah, 2003; Im et 
ah, 2005 and Guillermo et ah, 2009), In this research, we 
assume that each pixel in the first DSM shows the least 
difference with its true corresponding pixel in the second 
DSM. Therefore, we have chosen the “robust image 
differencing" method proposed in the work of Guillermo 
(2009). The robust difference between the initial DSM jc and 
the final DSM x 2 for the pixel (i,j) , is defined as the 
minimum of differences computed between the pixel jc 2 (/, j) 
in the final DSM and a certain neighbourhood (with size 
2*w+l) of the pixel x^(i,j) in the second DSM jc . In 
mathematical words, the robust positive and negative 
differences X pdif (/', j) and X (i, j) relative to the 
pixel (i, j) are defined as written in equations (1) and (2), 
respectively: 
, r . .min. . 11 ,fc 2 ('>y)-^,(A?))>o} 0) 
(pe[i-w,i+w],qe[j-w,j+w]]) 
{(x 2 0,/>-*,(/>><7))<°} (2) 
Figure 2. Noise reducing procedure: (a) Original image 
difference result; (b) Robust image difference result. 
Figure 2 shows a comparison between a simple pixel-to-pixel 
difference between the DEMs and a robust DEM difference. 
In Figure 2 (a), although the changed area are highlighted to 
some extend, the background is very noisy, which will 
because a problem in the change information extraction 
procedure. In contrast, after executing the robust image 
difference in Figure 2 (b), the noise in the background is 
successfully reduced, while the white areas, which are more 
likely to be real changed areas, are not influenced 
significantly.