bul 2004 
tituto de 
use laser 
ithin the 
from the 
quickbird 
SM) was 
© 
1ation of 
rain, the 
sraphy, a 
, already 
yrmalized 
from the 
Figure 2 
ixels are 
  
ig used to 
onsidered 
geometric 
o account 
id that the 
terms of 
tem. The 
tive view 
adir. The 
effect of 
ht of the 
nsate the 
projective 
g, derived 
e the roof 
> image in 
terest, for 
| growing 
ions. This 
r spectral 
process as 
ew image, 
:gion. The 
Baatz and 
was used. 
3, using à 
] spectral 
| allows to 
elements 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
within an urban scene is, in many cases, regular. According to 
Antunes (2003), it is necessary to take into account the scale of 
the problem to be solved and the type of image data in order to 
choose de parameters of the segmentation. 
2.4 Classification 
In the classification step, the degree of association of a region to 
each chosen soil cover class is described by a fuzzy 
membership function fA(x), which can assume values in the 
interval between 0 the 1. The membership functions, obtained 
from different features, as shape or spectral properties, can be 
also combined, in order to model an object. Fuzzy logic 
operators are available to combine the membership functions of 
different features and to draw a conclusion about the most 
suitable class for each object. Therefore, training regions are 
chosen in order to compute parameters that describe each class. 
For the classification, a large set of variables is available since, 
after the segmentation, the image is composed by segments, that 
can be described using spectral and spatial features, as well as 
topological relations within segments. The main problem 
consists in selecting the most appropriate features and combine 
them using the fuzzy logic approach. This task requires 
experience and good knowledge of the data set. The 
classification can be performed using a hierarchical tree 
approach, which builds up a hierarchical network of image 
objects. For the purpose of the hierarchical analysis, each 
segment is considered as an object that has relation to other 
objects within the same level or in other levels. At the bottom 
of the hierarchical tree, coarser objects can be found, results of 
a more generalized segmentation. At the other end, small 
objects, results of a fine segmentation are located. Smaller 
objects can inherit properties of objects on a lower level, and 
are considered specializations. The relations between objects 
stored in the network allows to use local context in the 
classification. 
3. RESULTS 
The data set was processed using the described approach and a 
thematic image was produced. For the classification of the 
image, a hierarchical tree was proposed after analyzing 
different possible networks. The used hierarchical networks is 
mainly based on the elevation of the objects, derived from the 
normalized DSM, and the spectral information derived from the 
satellite image. Spatial parameters, like form or texture, were 
not considered, because their performance was considered low 
compared to the spectral and altitude information. 
After deriving a satisfactory land-cover classification, the 
regions were grouped in 6 categories: “trees”, “grass”, “roads”, 
“yards”, “roofs” and “bare soil”. The main problem was 
associated to “bare soil”, which is easily confused with “roofs”. 
Because the main objective of the study is to detect buildings, 
the thematic image was simplified, producing a binary image of 
the buildings. Figure 3 shows the segments classified as 
buildings. 
In the result, the blocks can be identified and the buildings are 
separated from the other objects. Nevertheless, erros are still 
present. For example, small regions remained and the contours 
of the objects were not exactly located are at the borders of the 
buildings in the data set. The first problem was solved using a 
size filter that discarded small areas. The second problem is 
difficult to solve since it occurs during the segmentation step. In 
591 
some cases, the error is caused by errors in image registration 
and in some cases it is caused because of the poor spectral 
resolution of the image, which causes objects to have similar 
spectral response and difficults the delineation of the borders, 
even during a visual analysis. 
     
  
    
327 . m a 
Figure 3 — Result of the classification of buildings 
Figures 4 and 5 show a perspective view of the data set before 
and after the process. The view was obtained using the 
geometry of the laser data and the colours of the green red and 
near infrared channels of Quickbird. In the first image, the 
elevation (nDSM) of every pixel is displayed. In the second 
one, only the pixels of the normalized DSM that belong to 
buildings have elevation information. It ca be seen that the 
vegetation was successfully eliminated and the buildings were 
recognized. 
  
    
Flgure 5 — Coloured nDSM of the buildings.