International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
3.2 Object-oriented image analysis 
3.2.1 Image Segmentation The despeckled intensity image 
and the generated texture layer are used for the segmentation 
procedure described in this section. 
In various tests we found that slightly different input images 
resulted in significantly different segmentations of the scenes, 
even though the segmentation settings were identical. As this 
severely restricts the ability to automate the classification 
process a first classification phase is interposed. The goal of 
this phase is to create comparable image objects on a specified 
segmentation level, independently of the input image 
characteristics. 
This is achieved by iteratively performing a classification based 
border optimisation on an initial, relatively coarse image object 
level. For that purpose a set of sub-levels with constantly 
decreasing segment size is generated underneath the initial 
image object level. Then for each segment of the new layers it 
is successively tested whether its intensity value significantly 
diverges from both, its respective super-object located at the 
initial segmentation level and its directly adjacent segments. If 
it does, it is classified as a “significant substructure”. 
Consequently, the appropriate super-object on the initial 
segmentation level is cut according to the shape of the 
identified structure. After the procedure has finished, only the 
initial, henceforth trimmed image object level is retained for 
further analysis. The effect of this adjustment is illustrated by 
Figure 2. 
The optimised level serves as the basis for the actual 
classification. Additionally, a level with small objects is created 
underneath this base level and another one with very large 
segments is generated above. The newly generated fine level is 
best suited for the characterisation of single small-scale 
structures like houses or roads, while the segments of the third, 
coarse level cover large areas. Thus they represent complete 
quarters, agricultural fields or forest stands optimally. 
The described image segmentation finally results in three image 
object levels. 
3.2.2 Image classification The images are then classified 
according to a set of class rules collected in a “rule base”. A 
key issue of the rule base development is the use of robust 
features for the class description. This is achieved by basing 
the classification procedure on textural and contextual features 
primarily. 
For the definition of the rule base the segments of the coarse 
level are analysed with respect to the spatial composition of the 
underlying small-scale structures using textural features, in 
particular Haralick parameters (Haralick, 1979). Moreover the 
shape of these segments is utilised, e.g. to separate built-up 
areas from spectrally and texturally alike agricultural fields. 
The latter are typically by far more symmetric and uniform in 
shape. 
The fine segments of the lowest level are classified to 
characterise small-scale urban structures like houses, other 
significant scatterers or shadows. They are mainly defined on 
the basis of their intensity, their difference in brightness to the 
surrounding objects and the composition of the neighbouring 
area. In addition, the textural and shape-related characteristics 
of the appropriate super ordinate segment situated at the coarse 
level are considered by obtaining the according information 
from the segments of the third level. The information provided 
by the fine and the coarse level are then combined at the initial 
medium level to calculate the final “settlement mask”. 
  
Figure 2. Adjustment of image objects shown for two differing input segmentations (a: initial; b: optimized ) 
480 
Inte 
The 
to 1 
tran 
vali 
data 
disti 
for 
regi 
pho 
The 
an c 
91% 
illus 
righ 
show 
The: 
9094 
thes: 
textu 
their 
and 
deve 
inclu 
Erro 
possi 
non- 
The 
areas 
repre 
only 
the e 
the f 
the « 
Cons 
signi 
from 
lowei 
the 
chara 
agrici 
uncla 
The s 
polari 
Nevel 
along 
intera 
range 
of ob 
ambig 
inforn 
conte? 
well a 
of sol 
appro: 
tools 
hierar