nbul 2004 
are either 
weekend, 
the image 
ater in the 
ars on an 
higher or 
'cessity to 
> potential 
aracter is 
tion is the 
ssification 
mage was 
tation. As 
an arm of 
dominant 
isses were 
issigned to 
area with 
s’ was as 
ns: 
of) 
Hot) 
roof struc- 
0 
even in the 
lated which 
1e semantic 
nplex than 
Xf cars and 
ue in terms 
ance of a 
rial area 
best suiting 
he range of 
ectral image 
:lass-related 
- Or super- 
ieir spectral 
| lot can be 
re 2). Some 
stance to a 
e residential 
e border to 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
water can help to assign the ambiguous object rather to the class 
grass than to residential area. 
Name 
[paved lot - empty [ | 
]pave empty | ERE vi 
  
Parent class for display Modifiers 
[ Abstract [7 Inactive 
| area without vegetation vi 
, = -— 
AI je Contained | ‘ve Inherited | 
  
[=e Contained 
- El 
A GLCM Homogeneity, all dir., se_ext_02102001_balt_city_ 
V/* GLCM Homogeneily, all di., se ext 02102001 balt city - 
V/* GLCM Homogeneity, all dir., se ext 02102001. balt city - 
LA Level 
| V*X Mean se ext 02102001 balt city ms&ndvi.img (5) 
| — dox Inherited 
  
| = ® and (min) (area without vegetation) 
| V Level 
U/* Mean se ext. 02102001 balt city ms&ndvi.img (5) 
  
  
  
OK Cancel | 
Figure 2. Class Description for parking lot — empty. 
Figure 2 describes one class in the classification scheme. The 
class information area without vegetation is ‘inherited’ from the 
higher group level. The expert knowledge incorporates how a 
class is best described. It is shown that to assign the class 
parking lot — empty the mean value taken from band 5 (i.e. 
NDVI) is needed as well as textural information from several 
bands. Each feature encloses a membership function in which 
the threshold for the respective band or texture information is 
set-up by the user in a fuzzy range of values. The texture 
features evaluate the texture of the image object based on the 
gray level co-occurrence matrix (GLCM). 
2.4 Fuzzy Logic Classification 
Once all membership functions are set and the classes are 
defined in a best suited manner then each object is compared to 
cach class description in the supervised classification. Its 
contained and inherited expressions produce membership values 
for each object and according to the highest membership value 
each object is then classified. If the membership value of an 
image object is lower than the predefined minimum 
membership value, the image object will remain unclassified. 
So technically speaking, a classification using a fuzzy rule base 
is done by finding out which combination of fuzzy features is 
suitable to distinguish one class from the others. 
The classification does not produce brownfields sites as a 
resulting class, but delineates the above mentioned object 
classes. So the classification is an intermediate step towards the 
targeted land use detection. In the next and final step, certain 
classes are put together on a knowledge based rule to define 
potential brownfields sites as so-called structure groups. 
  
463 
Figure 3. Classified image sample showing large industrial 
buildings (in light and dark grey), various numbers of cars on 
parking lots (in red, purple, and pink), impervious surface 
without distinction (in white) grass (in bright green), neglected 
sparce vegetation (in olive green). 
The classified image sample shown in Figure 3 shows the same 
subset as the Ikonos imagery in Figure 1 in order to visually 
follow-up on the procedure. 
2.5 Structure Groups 
So far the image objects follow the rule on which the image is 
segmented and the output classification underlies the same 
object pattern (see Johnsson, 1994). These segmented and 
classified image objects now need to be structured. Such objects 
of interest are knowledge-based and do not need to meet 
homogeneous standards. They are defined in structure groups 
which contain the criteria for a  classification-based 
segmentation. All classes within one structure group are treated 
as fitting to each other, even if they represent very different 
class descriptions. 
As mentioned in chapter 2.2 the potential brownfields sites are 
presumed to consist of various objects. Therefore not all objects 
of the resulting classification are taken to be structured into 
semantically more meaningful groups but only those needed to 
describe the very target group. So the structure groups for 
potential Brownfields sites follow a certain pattern. They are 
composed of the following objects but vary in their presence or 
absence. Examples of the criteria variations for potential target 
sites are given below: 
Potential Brownfields sites I 
e Industrial / commercial building 
e Paved parking lots — empty 
e Neglected green spaces 
Potential Brownfields sites II