The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
them perform well in terms of completeness. The differences
almost entirely appear in terms of correctness: classification-
based methods, i.e. (Matikainen et al., 2007) and (Rottensteiner,
2007), seem to be more efficient than (Champion, 2007) and
deliver fewer FPs. The per-building correctness rates obtained
with the three approaches (0.54, 0.58 and 0.45, respectively) are
relatively low, and none of the approaches appears to be a
viable basis for a practical solution. However, this consideration
must be modulated by the good correctness values (of 0.79,
0.83 and 0.75), that are computed on a per-pixel basis. Since the
per-pixel values are directly linked to the area that has been
classified correctly or not, these values clearly highlight that the
change detection is mostly uncertain for small buildings.
(a) Matikainen
(b) Rottensteiner
(c) Champion
Figure 1. Change Detection Evaluation in Marseille Test Area.
In green, TP cases; in red, FN cases; in orange, FP
cases; in blue, 77V cases.
4.2 Toulouse Test Area
4.2.1 (Rottensteiner, 2007)
The evaluation of this method is illustrated in Fig. 3-a. Again,
major changes are well detected. However, changes affecting
small buildings are missed, which results in a high number of
FNs for small buildings (Fig. 4-a). There are also many FPs
that are mostly caused by inaccuracies in the DSM. Shadow
areas are also systematically overestimated in the DSM, which
generates FPs during the detection of new buildings. Two very
large areas of false alarms appear in the eastern part of the
scene (a sports field - Fig. 4-b) and in the north-eastern comer
(a parking lot) and are related to classical problems of stereo
matching algorithms, namely repeating patterns (demarcation
lines in the sports field, rows of cars on the parking lot) and
poor contrast. This entails height variations larger than 4 m in
the surface model in areas that are essentially horizontal.
4.2.2 (Champion, 2007): The evaluation of this method is
illustrated in Fig. 3-b. Overall, the results are similar to those
computed by (Rottensteiner, 2007) and particularly poor for
small buildings (Fig. 4-c). The difficulty to extract pertinent
primitives for small buildings entails 8 FN cases during the
detection of new buildings and many FPs during the
verification of the database (cf. Fig. 4-d for an example).
(a) Matikainen: FP case (b) Matikainen: FP case
(c) Rottensteiner: FN case (d) Rottensteiner: no FP case
(e) Champion: FN cases (f) Champion: FP cases
Figure 2. Evaluation details in Marseille. The same colour
code as Fig. 1 is used.
4.2.3 Remark concerning the satellite context