I 
: 
    
unsupervised method was based on a visual inspection. 
Where there are many classes, there is the problem of a 
class being split into more than one class due to the 
spectral differences within a class. Where there are only 
a few classes, there is the problem of unrelated classes 
being classified as the same class. 
  
Figure 3: Results of unsupervised classification — 
ISODATA method — 12 classes 
  
Figure 4: Results of unsupervised classification — 
ISODATA method — 5 classes 
It is thus concluded that the pixel-based approach is 
not acceptable for classifying complex urban 
environments with very high resolution remote 
sensing data. The reasons for this are as follows 
(Hurskainen & Pellikka 2004): 
€ Pixels do not sample the urban environment at the 
spatial scale to be mapped 
€ Building are represented by groups of pixels which 
should be treated as individual objects 
® Buildings produce a wide range of spectral 
signatures 
e Many features in the urban environment appear 
spectrally similar 
OBJECT-BASED CLASSIFICATION 
The limitation of the pixel in tackling issues of location, 
scale and distance has caused a shift towards object- 
based classification (De Dapper et al. 2006). Even 
though traditional pixel-based classifiers are well 
developed and there are sophisticated variations, they do 
not make use of available spatial concepts. The need for 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
context-based algorithms and object-oriented image 
processing is increasing and it is hypothesized that 
object-based image analysis will initiate new 
developments towards integrating GIS and remote 
sensing functions (Blaschke et al. 2000). 
The software used for object-based classification in this 
study is eCognition. A necessary prerequisite for object- 
based image classification is image segmentation. The 
shape of segments derived in eCognition is determined 
by the following parameters (Hofmann 2001): 
e Weight of image channels: specify the weight of 
each spectral band in the segmentation. Channels 
with higher weights have a greater influence on 
object generation. 
e Scale parameter: influences the average object size. 
This parameter determines the maximum allowed 
heterogeneity of the objects. The larger the scale 
parameter, the larger the objects become. 
e  Colour/Shape: the influence of colour vs. shape can 
be adjusted. The higher the shape value, the less 
spectral homogeneity influences the object 
generation. 
e  Smoothness/Compactness: These are attributes of 
the “shape” criterion. If the shape criterion is larger 
than 0, the user can determine whether objects shall 
be more compact or more smooth. 
e Level: determines whether a new generated image 
level will either overwrite a current level or whether 
the generated objects shall contain sub or super 
objects of an existing level. The order of generating 
the levels affects the objects’ shape (top-down vs. 
bottom-up segmentation). 
Using spectral information for image segmentation 
In the first strategy, the image was segmented using the 
multiresolution segmentation algorithm in eCognition. 
All four image layers (red, green, blue and NIR) were 
used with equal weighing in the segmentation process. 
The size of segments was decided on by trial and error. 
Smaller segments were merged to create larger segments 
that consisted of built-up areas as opposed to individual 
buildings. These built-up areas consisted of residential 
buildings, gardens, roads, etc. It was difficult to obtain 
suitable segments using only spectral information. The 
segments were not uniform in shape and size, and some 
contained a mixture of classes that was not ideal. Some 
segments appeared homogenous in nature, but did not 
logically represent features in an image. 
In the Figure 5a, it can be seen that the selected segment 
contains a building, a portion of a road and some trees. 
These segments were created from initially smaller 
segments with scale parameter of 50 (Figure 5b), which 
were then used as the input into a multiresolution 
segmentation to create segments with a scale parameter 
of 100 (as seen in Figure 5a). An initial segmentation of 
100 results in slightly different segments as can be seen 
in Figure 5c.