stanbul 2004 
e sign of the 
>gments of a 
tion. of line 
egments are 
classified as 
1e or more of 
yints in small 
egment. The 
1¢ fits 1s used 
nent size and 
e specified to 
iser scanning 
recording of 
n be assigned 
ints into the 
In particular 
be used to 
rials, but also 
fedian values 
within each 
‘he difference 
is known to 
of vegetation 
though large 
he edges of 
differences of 
ild clearly be 
ht differences 
a K-nearest 
ation for each 
segments that 
o form larger 
orrespond to 
S 
) laser scanner 
The data was 
mer with an 
/ was recorded 
n is shown in 
segment-based 
r buildings, a 
; (Sithole and 
vas performed 
‘the segments. 
| ground truth, 
getation points 
ition accuracy 
ind vegetation 
nts that were 
everal kind of 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
  
  
  
  
  
Figure 1: Classification of laser points into classes bare- 
Earth (blue (middle grey)), building (red (dark 
grey)), and vegetation (green (light grey)). 
Classified as 
Ground Bare Building | Vegetation | Total 
truth Earth 
Bare Earth 41956 42 113 42111 
Building 193 30368 5018 35579 
Vegetation 2294 2764 17580 22638 
Total 44443 33174 22711 | 100328 
  
  
  
  
  
  
  
Table 1: Classification results in number of points 
  
  
  
  
  
  
  
Classified as 
Ground Bare Building | Vegetation | Total 
truth Earth 
Bare Earth 99,6 0,1 0,3 100,0 
Building 0,5 85,4 14,1 100,0 
Vegetation 10,1 12.2 77 7 100,0 
  
  
  
  
  
Table 2: Classification results in percentages. 
* One building and larger parts of another building were 
classified as a vegetation segment. These buildings had 
relatively small steep roof faces. With the point distance 
of 1.2 m, the roughness attribute was similar to those of 
vegetation segments. With a higher point density, one 
could probably obtain a better distinction. 
* The classification of the small segments is relatively 
unreliable. The number of points within these segments 
was often too low to generate representative attribute 
values. 
In a post-processing step these small errors could 
easily be repaired. It can be argued that small vegetation 
segments that are rest on building segments should also 
be building segments. Similarly, small building segments 
that are surrounded by bare Earth points and vegetation 
points also need to be reclassified. 
  
2* + 
* “+ 
a X 
~ 5 - 
a . 
* ;t + * 
* 
vt 2 QA 
je A aW a 
KT 
r^ 4 
X n "om 
‘ 
t E 
+ 
“a - * 
t 
-4 w 
^ —- = Li 
A Tu 4: uut m 
" x Wr - fO 
r HX = = a 7m T Ü + 
— » “ : 
M 5 : [. «^ -—. 7 
a 13 = M ns = i ; ET à 
^ d. on d = : M / . 
- oo 
wl i. {a ES 5 4 t* za 
= » 7 cxt s 
Si 3 =“. = "+ wr wor | de 
MATO EE } Tnm 
"Rh . 3 Tu Ae pt ed Les 
oh —— Aa 
* lai Tq e « - -—- f 
eem s AT t : = 
- Pd s ei ” Cc > - ® 
». : = - Tt os = * - 1 x & 
A > e e i * wo - vr o 
a - - T. wf, egw e xx 
it Anat Lo fe rept 11 
T T. = Wt a 1 
iw in ec uw 
x = « 4 
= - E x : 
"gs M i 3 3935 p = ^ 
Ku "tn _— + 
I pt; teu d Wu 
A * * Unt 
ve 
E 1 ru onm 
x ein 
; ^ 
> 4 f - To* 
i. i 
E ud 1 eem 
- f w * 
= 4 57. — =n 
- ar C 
Figure 2: Ground truth for those points of Figure 1 that 
were classified incorrectly. See Figure 1 for the 
legend. 
e Points on walls of buildings were also often incorrectly 
classified as vegetation points. These points stand out 
clearly in Figure 2. Wall points accounted for about 80% 
to the 5018 building points classified as vegetation. A 
correction in a post-processing step could therefore 
significantly improve the classification accuracy. 
However, for the change detection this is of less 
importance, since the sizes of the building segments will 
not increase when the wall points are included. 
e Some patches of vegetation adjacent to buildings were 
grouped in segments with building points and classified 
as building. Such errors may impact the change detection 
as it may be concluded that a building has been extended. 
A larger point density in combination with stricter 
thresholds in the profile segmentation may reduce the 
number of these errors. The stricter thresholds will result 
in smaller segments. The increased point density is then 
required to enable the reliable computation of the 
segment roughness. 
e The large area of vegetation classified as bare Earth near 
the bottom of Figure 2 is an area with very low 
vegetation that was merged with the surrounding bare 
Earth. 
5. CHANGE DETECTION 
Even if the topographic database is up to date and buildings 
are correctly extracted from the laser data, differences will 
exist between the database objects and the extracted building 
segments. These differences need to be taken into account 
during the change detection. Otherwise, many unchanged 
buildings will be presented to the operator for updating. 
Several reasons for differences between database objects and