The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
The change detection is clearly limited by the resolution of 
input satellite data in relation to the size of changes to be 
detected. The completeness and correctness values are (0.85, 
0.50) and (0.80, 0.55) for the two methods, respectively. Such 
values clearly reflect that detecting 2D building changes in a 
satellite context is too hard a challenge for the current state-of- 
the-art. This observation is corroborated by the fact that the 
ground classification performed in (Matikainen et al., 2007) 
does not give acceptable results for Toulouse when carried out 
in a fully automatic way. The results may appear to be 
disappointing. However, the completeness of the two systems 
both turn out to be very close to the values found in (Mayer et 
al., 2006) and expected for a system to be operational. 
Regarding the correctness, the low rate is mostly related to a 
high number of FPs during the detection of new buildings, 
especially for (Rottensteiner, 2007). As mentioned at the 
beginning of section 4, it is less problematic as the 
corresponding FP objects would inescapably be checked by a 
human without the support of automatic techniques. The results 
that are achieved here clearly demonstrate that it is worth while 
to carry out research in the satellite context, especially towards 
the reduction of FPs. 
Figure 3. Change Detection Evaluation in Toulouse Test Area 
(with the same colour code as in Fig. 1). 
4.3 Factors affecting the accuracy 
In this section, we will try to sum up some preliminary results 
based on the experiences of this EuroSDR project. We will 
focus the analysis on the impact of input data on the change 
detection performance on the one hand, on the features of the 
method on the other hand, more specifically on the type 
(geometric/radiometric) of primitives to use. 
4.3.1 Impact of the DSM 
The limiting factor of change detection appears to be the quality 
of the DSM. The erroneous height values present in the initial 
DSM between some buildings (i.e. nearby step edges) and the 
quantization effect observed in both areas and that prevents to 
exploit surface roughness in the change detection process 
clearly affect the quality of output change maps. It should be 
possible to overcome such drawbacks by using LIDAR data, as 
indicated by the completeness and correctness computed in 
(Matikainen et al., 2004) and (Rottensteiner, 2007). The results 
achieved for the EuroSDR test dataset based on LIDAR data 
have not been evaluated yet but should confirm those results. 
Improving the performance of an image-based change detection 
system implies a higher robustness of stereo-matching 
techniques with respect to shadow areas and a higher 
preservation of object details, especially step edges. 
(a) Rottensteiner: FN case (b) Rottensteiner: FP case 
(c) Champion: FN cases (d) Champion: FP cases 
Figure 4. Evaluation details in Toulouse. 
4.3.2 Impact of the DTM 
As previously highlighted, the extraction of buildings and 
consequently the performance of the change detection process 
are the better the more accurate the DTM is. In this study, the 
morphology-based method used in (Rottensteiner, 2007) and 
the surface-based method (Champion and Boldo, 2006) both 
fail in the presence of topographic discontinuities (cliffs). 
Refined approaches should be considered in the future to better 
model such terrain features. It must also be noted that in order 
to keep the process fully automatic, manual corrections were 
not employed for this study. Manual corrections of difficult