2 
f 
ra 
   
rd LU CN i ELS = 
Figure 7: Segmentation based on cadastral parcels 
After the initial segmentation, a second segmentation 
was performed within the boundaries of the cadastral 
segments. These resulting smaller segments were then 
classified using the nearest neighbour approach. Samples 
were selected and the following features were used in the 
NN classification: mean values for red, green, blue and 
near infrared, brightness, maximum difference and 
NDVI. 
  
  
Figure 9: Object-based classification using the Nearest 
Neighbour method (initial segments derived from 
thematic information) 
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 
  
  
  
  
  
  
  
  
  
Class name Producers | Users KIA per 
accuracy accuracy class 
Building 0.94 0.89 0.90 
Vegetation 0.91 0.87 0.81 
Water 0.60 1.00 0.57 
Overall accuracy 0.89 
KIA 0.80 
  
  
  
Table 3: Accuracy assessment and kappa statistics for 
object-based classification (segments initially based on 
thematic information) 
For this example only buildings, vegetation and water 
were classified and the areas between the cadastral 
blocks, for example large roads, were masked out of the 
classification (see Figure 9). 
DISCUSSION 
It is necessary to move away from dependency on 
individual pixel values into a way of incorporating shape, 
texture and contextual information for image 
classification (Hurskainen & Pellikka 2004). 
Segmentation is a very important step in object-based 
classification. In order to have a successful classification, 
one must have suitable segments that accurately 
represent features of interest. Segmentation based purely 
on spectral information did not result in suitable 
segments. The inclusion of thematic or vector data for 
the initial segmentation in the object-based classification 
resulted in an improvement in overall accuracy when 
compared with the method that was based only on 
spectral information. 
It should be noted that the selection of evaluation or 
check sites has a large influence on the accuracy results 
reported. All check sites were randomly selected and 
were not part of the classification training sites. Since the 
accuracy assessments are always based on a sample of 
the classified scene, it is difficult to get a ‘true’ accuracy 
assessment. 
CONCLUSIONS 
The object-based classification methods show much 
promise and results in better classification accuracy than 
the pixel-based methods that were tested. The inclusion 
of thematic data in the segmentation stage can be used to 
force suitable boundaries that can be further segmented 
and thus improve classification results. 
The decision regarding whether to classify individual 
buildings or larger built-up areas is an important factor to 
consider. Each option has its own merits and drawbacks. 
Individual buildings may be easier to detect based on 
their shape properties, but may vary greatly in spectral 
characteristics due to roofing types and materials used. 
One may need to have sub classes within the building 
class to adequately represent all building types. On the 
other hand, large built-up areas may be less 
homogeneous due to the inclusion of a variety of 
individual features within the built-up area for example, 
buildings, grass, trees, roads, etc. and therefore may be 
difficult to identify adequately and consistently.