ibul 2004 
ssified as 
| and 2). 
ildings as 
| of new 
n. In the 
rmat, but 
vn in the 
icy of the 
ut of the 
ence map 
tained by 
| pixel by 
buildings 
; the total 
neg 18 the 
sification 
pixel by 
ouse area 
All 
94.2% 
3.2% 
2.6% 
  
In this 
ige of the 
(e.g. Over 
uilding in 
ig in the 
the map). 
d values. 
and small 
sons were 
of the 
er 0.75). It 
ation was 
ne for the 
area). If a 
it became 
 building- 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
  
  
  
| 
  
     
  
  
  
  
  
  
  
  
* Detected 
New, certain 
Buildings of the old map: 
   
Buildings from building detection: 
Figure 1. 
Partly detected B Not detected 
New, uncertain 
  
* 
s ^ v, v 
: e 
* = * 
* v * N 
: + 
2 iM e = 
+ EN - am 
», : 
> er «e ne 
ie : . UAM 1 
i 
» : ; 
x - . * e 
* «t 
"2 
Da - *» 
A t 
Enlarged je 
  
  
Results of automatic building detection (upper part) and change detection (lower part) for the industrial area (left), 
apartment house area (middle) and small-house area (right). The width of each area is 900 m. Buildings of the old map © 
The National Land Survey of Finland, permission number 49/M Y Y/04. 
  
So "Ww 
v 
$e 
  
  
  
  
  
  
  
  
Figure 2. Old building map (left), final segmentation (middle) and change detection (right) results for a subarea of 255 m x 255 m. 
The legend for the change detection result is presented in Figure 1. Buildings of the old map © The National Land Survey 
of Finland, permission number 49/M Y Y/04. 
As shown by the estimates in Table 1, a relatively high accuracy 
was achieved. Interpretation accuracy was over 90% for each 
area. The highest accuracy was obtained for the industrial area 
(96.7%), which is natural due to the large building size. Object 
accuracy was lower than interpretation accuracy, ranging from 
72.4% in the small-house area to 86.1% in the apartment house 
area. As already mentioned in Section 2, the reference map does 
not exactly correspond to the laser scanner and aerial image 
data, and part of errors result from this. It can also be observed 
437 
that buildings are typically slightly smaller in the map than in 
the classification result, especially in the small-house area. 
Several reasons can be found for this behaviour: e.g. large 
roofs, use of first pulse data and formation of the DSM by 
selecting the highest point for each pixel. It is likely that the 
lower object accuracy in the small-house area is partly due to 
roof types (ridge roofs reaching over building walls typical) and 
gencralized representation of small buildings in the map 
(including elimination of small polygons in data preprocessing)