International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
the amount of information that is introduced into the 
classification. 
e We will not only use geometrical, i.e. elevation data, but 
also semantical information as derived from image data. 
This simulates the manual digitising process where 
thematic information about the underlying objects is 
introduced simultaneously and a categorisation of the 
extracted edges is performed. 
e Instead of an edge-based approach we will apply a region 
growing algorithm, thus by-passing the problem of linking 
detected edge pixels to connected lines. 
Figure 4 sketches the outline of the algorithm: A segmentation 
delivers borderlines which can be seen as candidates for 
surface edges. On base of the extracted features of the outline 
and the interior region the follow-up classification performs a 
hypothesis testing on these candidates as well as a 
categorisation into semantical classes (like walls or 
embankments). After the classified edges are converted into 
the vector domain, some post-processing steps (dilation by 
matching with image edges, and smoothing) are performed. At 
both stages, the classification and the post-processing, some 
grouping processes in the sense of perceptual organisation will 
be applied. 
: image 
data (multiple data 
reflections) 
laser scanner 
  
  
segmentation 
y 
feature extraction 
y 
classification 
Y 
vectorisation 
Y 
post-processing 
  
  
  
  
  
  
  
  
  
  
Figure 4. Outline of the proposed surface edge extraction 
approach using multi-sensor data 
In the following sections the whole process will be explained 
in more detail by concentrating on the extraction of only one 
type of surface edges, namely building walls. 
4.4.2 Segmentation: For the case under consideration, the 
detection of walls, both edge detection and region growing are 
aspects of the same processes under the assumption of step 
edges (Pavlidis & Liow, 1990). Hence, instead of an edge- 
based approach we will apply a region growing algorithm, thus 
by-passing the subsequent problem of linking candidate pixels, 
as detected by any edge filter, to connected lines. Furthermore, 
the linkage between segments and attached borderlines is of 
great advantage for the subsequent classification process. 
For the segmentation we use the software system eCognition 
(Baatz & Schüpe, 2000) which uses an extension of a region 
growing method called Fractal Net Evolution Approach 
(FNEA). As we want to extract only building walls at this 
stage of our study, we introduce the lowest elevation values 
from the last laser scanner echo (LE-low) as the heterogeneity 
feature for the segmentation. The LE-low represents only the 
ground surface and buildings (see section 2.2 and figure 2). 
Keep in mind that the LE-low reflection leads to an inside 
“buffering” of the real borderlines into the interior of the 
objects. 
4.4.3 Classification of segments: The classification step 
performs not only a hypothesis testing of the surface edge 
candidates but also their categorisation into semantical classes. 
As already pointed out, in this study we want to concentrate on 
buildings walls only. 
Our two-stage classification procedure starts with an 
elimination approach. Mere: features with rather “weak” 
threshold values are introduced which leads to a set of 
segments that classify nearly all buildings correctly (thus 
minimising the number of omission errors) but still include a 
considerable number of misclassifications (commission errors). 
For the classification of segments that are surrounded by walls 
we are considering the following feature values: 
e Area greater than 50 m! — considering the minimum area 
of a building. 
e Elevation difference to lower neighbours greater 3 m 
considering the minimum height of a building. 
e Normalised Difference Vegetation Index (NDVI) less than 
0.05 — considering the relatively high reflections in the 
red spectrum and the relatively weak reflections in the 
near infrared spectrum due to roof colour and material. 
Consequently, the goal of the second step has to be the 
reduction of the commission errors. In the following only the 
above obtained subset of candidates is taken into account. 
Because the object description using the following features is 
neither geometrically sharp nor standardised we introduce 
partial rather than crisp memberships, i.e. we apply a fuzzy 
logic classification approach. For the classification of 
segments that are surrounded by walls we are considering the 
following feature value ranges that shall make the distinction 
against vegetation segments and that are modelled by a linear 
membership function: 
e NDVI (see above): Based on the hypothesis that with a 
smaller NDVI value the possibility of the existence of a 
building becomes larger, we introduce the value range 
between the minimum value (membership value p=1) and 
the above applied threshold of 0.05 (u=0). 
e Rectangular fit: After creating a rectangle with the same 
area as the considered segment, the area of the object 
outside the rectangle is compared with the area inside the 
rectangle, which is not filled out with the object. For 
buildings a rather high value with a maximum of 1.0 can 
be expected. Thus the fuzzy value range extends from 0 
(p.70) to 1 (n1). 
e Standard deviation of elevation: Due to some very high 
values at the edges (walls) we can expect high standard 
608 
Inter: 
Final! 
throu, 
been 
which 
than ( 
4.4.4 
vecto 
perfor 
The p 
on the 
repre: 
the re 
perfor 
these 
has to 
suitab 
numer 
severe 
distur 
derive 
segme 
Becau 
the ed 
classit 
In ord 
vertex 
surrou 
distan 
necess 
that ar 
corres 
edges 
operat 
bound. 
  
  
Figure