1 2004
ild-
in
1es
uild-
ld-
int
ill
es,
all
Se,
el-
nd
18-
1d-
1al
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
3.3 Results
Results of the matching method applied to both zo-
nes are shown in table 3 and table 4. The first col-
umn provides the percentage of detection, that is
the proportion of well detected objects over the to-
tal size of these objects in the DB. The second col-
umn shows the percentage of false alarm, that is,
the part of the detected objects for which no match
could be found in the DB. Therefore, the addition
of the first and second column does not necessarily
sum to 100. In table 3, the upper number in a cell
refers to the expert's result, and the lower one, to
the novice's one.
Good False
Detection Alarms
87.7% | 14.09 %
69.6% 18.5%
Road Network 84.3% 7.7%
68.4% 12.5%
Built-up area
Table 3: Results of the matching method in the
Sub-urban area (5m)
Good False
Detection | Alarms
Built-up area 64.8% | 31.2%
Road Network 87.995 5.5%
Table 4: Results of the matching method in the ru-
ral area (3m)
These results are much more optimistic than the
results provided by the confusion matrix. More
than 80% of the DB is seen by the expert in the
sub-urban zone. In this zone, RN remains slightly
more visible than BUA, while its visibility is about
13% better than BUA in rural zone. While look-
ing at the BUA false alarms (or missed), a small
part comes from too many small buildings identi-
fied (or missed) and a large part from parking that
have been interpreted as buildings (or vice versa).
The small buildings are more numerous in the ru-
ral zone, explaining the relatively low BUA detec-
tion rate. As far as the RN is concerned, much of
the false alarms come from private roads or from
interpretation errors. The missed roads are small
secondary roads or roads in woody area.
989
4 CONCLUSION
A visibility test of the road network and the built-up
area has been made on SPOTS5 image. Two meth-
ods have been proposed to estimate the portion of
visible objects, and of false alarms. We showed that
the confusion matrix method, often used in classifi-
cation, is not suitable for getting such an estimate.
We thus proposed an object oriented method, called
the matching method, providing more realistic re-
sults, and a better view of what is missed, and what
is misinterpreted.
The object visibility is highly dependent on the ex-
perience of the operator and on the zone type. In
the sub-urban area, approximately 85 % of the DB
was seen by the expert operator while the novice
makes 20 % less. In both tested zones, the Road
Network was more visible than the Built-up area,
with a larger difference in the rural area. The BUA
visibility drops by about 20 % in the rural area, due
to the large amount of small disseminated buildings
missed by the operator. :
Despite results of both experiments cannot be com-
pared due to their difference in landscape (i.e sub-
urban versus rural), we may conclude that SPOT5
5m resolution data are sufficient to detect most of
the Road Network in open area. As far as the built-
up area is concerned, even the SPOTS 3m resolu-
tion seems insufficient to detect individual build-
ings. However, if only most of the building sets
should be detected, the SPOTS 5m resolution data
would be suitable enough.
5 ACKNOWLEDGMENT
The team wish to thank Stephanie d'Hoop and Oli-
vier Defays for having analyzed the satellite im-
ages.
This study is part of the ETATS project, funded by
the TELSAT program of the Scientific, Technical
and Cultural Affairs ofthe Prime Minister's Service
(Belgian State).
REFERENCES
Puissant, A. and Weber, C., 2002. The utility of
very high spatial resolution images to identify ur-
ban objects. Geocarto International 17(1), pp. 31—
4].
