| 2004 
g in 
10r- 
re- 
the 
are 
ll of 
the 
d by 
] on 
  
HO HO 
vos 
On vo 
  
e N 
noo OO 
oo ow; 
  
  
o0 00 
Su a 
nn ON 
  
  
1rea, 
n of 
king 
Sev- 
 de- 
01). 
racy 
| in- 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
e There is a large amount of NO pixels in the 
DB. These pixels thus influence much the global 
index, generating an optimistic view. 
e Even though optimistic, this global index pro- 
vides a rather poor classification rate. 
e [n the sub-urban area, less then half of the DB 
built-up area and a little more than half of the 
DB road network is seen by the expert (see 
producer indexes). These values significantly 
drop in the rural area, as they both reach about 
15%. 
e Proportionally, the road network is better de- 
tected than the built-up area. 
e There is a large amount of false alarms: around 
30% in the sub-urban, up to more than 60% in 
the rural zone 
e The results are dramatically worse in the ru- 
ral zone. The reason is that the image was not 
well geo-referenced, so that the detected ob- 
jects often fall next to the DB elements. 
e The expert's performance is approximately 1596 
better in both classes, while keeping a lower 
false alarm rate. 
A look at the confusion classes produced by the su- 
perposition of the detected objects onto the DB can 
give another view of the detection errors. In Fig- 
ure 2, the detected elements that are superimposed 
to the correct DB elements are displayed in green, 
while the missing parts of the DB elements are in 
red, and the detected parts that are not in the DB 
(i.e. false alarms) are displayed in black. The left 
and right images show parts taken from the sub- 
urban and from rural area, respectively. 
Eu o ue emt gan 
  
    
7 Hia ra] dT P i. 
ui M La EE i Pa E 
Figure 2: Detail of superposing the DB onto image 
DB; Green- correct detection; Black- false alarm; 
Red- missing parts. Left: Rural area. Right: Sub- 
urban area. 
987 
This method, while giving some indications con- 
cerning the visibility, has several drawbacks, high- 
lighted by Figures 2. 
e It does not make the difference between a whole 
object not detected and parts of non-detected 
objects. 
e [t is highly sensitive to a slight displacement 
of the DB. 
e Dilated points and lines have fixed sized while 
small buildings have various sizes. 
e Dilation of points and lines provides some tol- 
erance in positioning objects but might also in- 
troduce false error pixels. 
Because of these drawbacks, we have designed a 
matching method. 
3 MATCHING METHOD 
In order to avoid the confusion matrix method draw- 
backs, a more sophisticated method has been im- 
plemented; the confusion matrix is pixel based, while 
the matching method is object-based. The DB is 
divided in the road network and the built-up area, 
then matching rules for each possible pair of de- 
tected elements and DB object are defined and tested. 
e Detected Buildings: 
— Points: point — small building, 
point — polygonal building 
— Lines: line — small buildings, 
line— polygonal building 
— Polygon: polygon — small buildings, 
polygon — polygonal building 
e Detected Roads: 
— Lines: line — linear road 
— Polygon: polygon — linear road 
3.1 Road network 
The roads in the DB are either represented by their 
circulation lanes or by their central axis, and by 
their contour. The latter are stored as polygons and 
the former as polylines. In the matching method, 
we use the polyline representation only.