In: Wagner W„ Sz&ely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
Vertical Change Detection Result Comparison 
| —»— Mask Based ~+~- Box-fitting Based Manual Extraction] 
Figure 7. Vertical Change Detection Result Comparison 
• For the annotation of the horizontal changes (areas of the 
changed objects), the areas of the positive (resp. negative) 
changed objects have been manually extracted using the 
original 1KONOS panchromatic channel (resp. using the 
LIDAR DSM data). 
• For the annotation of the vertical changes (height of the 
changed objects), we extract the height value in the 
changed area from each DSM manually, and calculate the 
difference. 
The final change detection results are summarized in Table 1. 
To compare the two automatic detection result and manual 
extraction result, we show the vertical and horinzontal change 
extraction result separately in Figures 7 and 8(For the 
negative changes, we use the absolute values in Figure 7). As 
the 4 lh and 5 th changes could not be separated in the detected 
mask, we consider them as one changed object in the 
comparison procedure. Also false alarms are omitted in our 
comparison scheme. According to the error bars (get from the 
standard deviation of the height value distribution in the 
mask area), the detected vertical changes fit well with the 
manual extraction result. For the 1 st building, the manual 
extraction result shows relatively larger difference with both 
the mask based and box-fitting based automatic extraction 
result. In order to explain such behaviour, we compare the 
generated DSMs to the IKONOS panchromatic image. As 
showed in Figure 9, the IKONOS-DSM in this area has poor 
quality, resulting in 3 big holes in the middle of the building, 
and the building shape is strongly transformed. This explains 
well the large height difference found between our change 
maps and the manual extraction one. 
Figure 9. 1 st changed building analysis: (a) LiDAR-DSM; (b) 
IKONOS-DSM; (c) IKONOS-PAN-2005; (d) Mask with 
IKONOS-DSM; (e) Masks with the IKONOS-PAN-2005 
Horizontal Change Detection Result Comparison 
Figure 8. Horizontal Change Detection Result Comparison 
In the Figure 9 (d-e), the red mask represents the masked 
shape, while the blue rectangle is the extracted change 
building shape after box-fitting. We can see that the mask fits 
well with the building area in IKONOS-DSM, but has quite 
large difference with IKONOS-PAN-2005, after the box 
fitting, the right parts of the outliers are successfully 
recovered. But the result is limited by the rectangular shape 
assumed in the box-fitting procedure, which displayed by the 
false alarm pixels in the left upper part inside the blue colour 
rectangle. Those situations conduct the relative lower vertical 
changes, while higher horizontal changes in the extraction 
results. 
5. CONCLUSION 
In this paper, a 3D change detection approach based on 
DSMs is proposed and evaluated to detect changes that have 
occurred in the city centre of Munich between 2003 and 
2005. 
The whole procedure is divided into 3 steps. First, we 
generate and co-register DSMs acquired at two different 
epochs. Then, we compute the “robust difference images” in 
order to reduce the noise coming from the different nature of 
the DSMs used. In fact, the random variations of the stereo 
images acquisition conditions as well as the blunders caused 
during the automatic matching and DEM generation process 
makes urban structures look different from one DSM to 
another, especially for building walls and edges. As our focus 
in this paper is the monitoring of urban changes, noise 
reduction is essential. After that, we generate the change map 
with both vertical and horizontal change information. To 
overcome the poor quality of the DSM, we refine the changed 
buildings to rectangular shape. To confirm the validity of our 
approach, we compare our results with manual extracted 
ground truth figures. 
It has been shown that DSMs generated from optical stereo 
imagery could be reliable sources for efficient 3D change 
detection. The extracted change maps demonstrate the ground 
surface changes in most parts of the test areas. But when the 
DSM does not meet the required quality and can not show the 
real situation, the result will be influenced. In addition, only 
rectangular building shapes have been considered in our 
refinement procedure, the change results will certainly 
improve when more building shapes are included. 
ACKNOWLEDGEMENTS 
The authors wish to thank Beril Sirmacek for contributing her 
box-fitting algorithm, and Hossein Arefi for his advice in the 
pre-processing of the DSM.