International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
Figure 1. Brownfields sites in the southeast of Baltimore City 
near the Harbour. The yellow line shows the parcel boundaries. 
2. CLASSIFICATION SCHEME 
2.1 Image Pre-processing 
Ikonos provides both, multispectral and panchromatic data with 
a spatial resolution of four and one meter respectively. Having 
the four multispectral bands with red, green, blue, and infrared 
at hand, the NDVI could be calculated for the image and 
stacked to the four bands as a synthetic fifth band. There is no 
necessity to perform an image data fusion because the software 
eCognition allows a multiresolution segmentation with the 
chosen bands weighted individually. Segmenting the image 1s 
an important prerequisite as it subdivides the image into a given 
number of separate regions. There is a high number of degrees 
of freedom which could be reduced to two degrees to obtain 
optimal objects for this classification scheme on urban 
Brownfields. These levels are connected in a hierarchical 
manner and represent image information in different levels 
simultaneously (Baatz et al., 2000). 
2.2 Class Definition 
Before classes are defined and the analysis should start it has to 
be clarified what the classification is aiming at. So the question 
has to be answered with which classes a Brownfields sites is 
expected to be described. The challenge of this study is to 
classify a land use type so it is tried to cut down the description 
of its use to its spatial appearance. The physical characteristics 
of the site may consist of different objects such as buildings 
(more or less rotten), impervious surface such as parking lots, 
roads and road access, and also sparse, neglected vegetation. 
When thinking about the information for a Brownfields taken 
from parking lots that are either intensely used, have few cars 
only, or are not used at all it is important to take into 
consideration which day of the week the image was acquired. It 
462 
is assumed that on a Brownfields site the parking lots are either 
empty or have few cars only. So if it was taken on a weekend, 
this information needed to be discarded. In this case the image 
information was gathered on a workday (Tuesday) later in the 
morning so it can be assumed that the amount of cars on an 
industrial or commercial parking lot represent some higher or 
lower intensity of business activities. 
Although this definition is not unambiguous the necessity to 
find and describe parameters that sufficiently outline potential 
Brownfields sites by means of their surficial character is 
irrevocable. 
The frame of the knowledge base for the classification is the 
class hierarchy which contains all.classes of the classification 
scheme. On the higher level of segmentation the image was 
being classified into areas with and without vegetation. As 
Baltimore is located at the Patapsco River estuary, an arm of 
Chesapeake Bay with a large harbour area, wafer is a dominant 
class taken separately and not being subdivided. 
On the lower level of segmentation meaningful classes were 
defined for the land use type of Brownfields and assigned to 
their super-objects area wifhout vegetation and area with 
vegetation. The assignment to these ‘super-objects’ was as 
follows (see example given in Figure 2). 
The class hierarchy of area without vegetation contains: 
e Industrial / commercial building I (white roof) 
9 Industrial / commercial building II (black roof) 
e [Industrial / commercial building III (mixed roof struc- 
ture) 
e High intensity commercial / industrial area 
e Paved parking lots - empty 
e Paved parking lots - with few cars or cargo 
e Paved parking lots - with many cars or cargo 
e Bare soil 
As the Ikonos imagery offers very high resolution even in the 
multispectral bands roof types needed to be differentiated which 
increased the number of classes without increasing the semantic 
contents and made the classification more complex than 
demanded. A gain in information is that number of cars and 
their density is easily seen which is an important issue in terms 
of the presumption for the surficial ^ appearance of a 
Brownfields site. 
The class hierarchy for area with vegetation contains: 
e Residential area / single buildings 
e Intermediate intensity commercial / industrial arca 
e Urban woodland 
e Grass (lawn, park) 
e Neglected green spaces 
2.3 Class Description 
To construct an elaborate knowledge base the best suiting 
features are taken to separate the target classes. The range of 
features goes from object features which include spectral image 
information, shape, texture, and hierarchy, to class-related 
features, i.e. relations to neighbour objects, sub- or super- 
objects, memberships and some more. 
So parking lots can be classified by means of their spectral 
reflectance, and the amount of cars on a parking lot can be 
distinguished by its textural information (see Figure 2). Some 
classes also need to be characterised by their distance to a 
neighboring class, so if, for example, parks and some residential 
areas are difficult to distinguish, then the relative border to 
Internati 
water car 
grass tha 
Name 
{paved lc 
Parent cl 
| area wit 
Figure 2 
Figure 2 
class infi 
higher g 
class is 
parking 
NDVI) i 
bands. E 
the thres 
set-up b 
features 
gray lev 
2.4 Fu: 
Once. al 
defined 
each cl 
containe 
for each 
each ob 
image 
member 
So techr 
is done 
suitable 
The cla 
resulting 
classes. 
targeted 
classes ; 
potentia