the settlement class are classified as other features such 
as forest or agricultural areas, due to the high resolution 
of the imagery and wide variety of building roof spectral 
properties. When tested with lower resolution imagery, 
settlements are recognised as uniform but the accuracy of 
the results deteriorates. 
Classes must be decided on beforehand and adequate 
samples that represent the classes must be collected. For 
most super classes there will be subclasses for a feature 
due to the spectral variation within classes. For example, 
many samples of the class ‘water’ were collected as 
subclasses and merged to make the final water class. A 
further challenge lies in the fact that a class may consist 
of various land cover types which are spectrally diverse, 
but need to be grouped together. Such an example is the 
urban built-up class which may consist of buildings, 
gardens (vegetation), swimming pools and bare ground. 
One may consider classifying buildings separately, but 
this decision is influenced by the purpose of the 
classification, and in this case the built-up area was 
required. Even individual buildings can have a multitude 
of different land cover types and colours; for example, 
roof tiles, thatch, metal sheeting, etc. are all spectrally 
diverse. 
In this example the maximum likelihood classification 
method was tested. Final classes consisted of water, 
vegetation, road, built-up and bare ground or sand. 
  
Figure 1: Subset of aerial image used in all classification 
methods tested 
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 
Road 0.67 0.40 0.34 
Bare ground 1.00 1.00 1.00 
or sand 
Water 0.74 1.00 1.00 
Built-up 0.73 0.66 0.43 
Vegetation 0.50 0.60 0.51 
Overall accuracy 0.70 
KIA 0.58 
  
  
  
Table 1: Accuracy assessment and kappa statistics for 
pixel-based supervised method (Maximum Likelihood 
Classification) 
  
   
  
Figure 2: Results of supervised classification — maximum 
likelihood classification 
The classification results indicate that there is overlap 
between the road and the built-up classes due to their 
spectral similarity. Some buildings were also incorrectly 
classified as bare ground or sand and some vegetation 
was incorrectly classified as roads and built-up areas. 
Unsupervised classification 
In unsupervised classification, pixel values within a 
certain land cover type should be close together in the 
measurement (spectral) space, whereas data in different 
classes should be reasonably well separated. The classes 
that result from unsupervised classification are spectral 
classes (Lillesand et al. 2004). 
The unsupervised ISODATA method is popular in the 
classification of heterogeneous high resolution images as 
it is very successful in finding the spectral clusters that 
are inherent in images (Zhang 2001). Unsupervised 
classification may address some of the shortcomings of 
applying supervised classification for land use or land 
cover classification where classes have a high degree of 
spectral variability. Where there is a high degree of 
spectral variability, suitable training sites for relevant 
land use or land cover classes will always be difficult to 
achieve. 
The unsupervised approach is simple and no training data 
or samples are needed, thus making it much faster to 
implement than the supervised approach. Another 
advantage is that the unsupervised classifier identifies the 
different spectral classes present in an image, which 
might not be obvious to an analyst applying a supervised 
classifier. Similarly, there may be so many spectral 
classes in a scene that it would be difficult to train on all 
of them. Since unsupervised classification is the 
identification of spectrally distinct classes in an image, 
the analyst must still use reference data to associate 
spectral classes with the land cover types of interest. The 
spectral classes identified may not be uniquely associated 
with a land cover type, and one may have several 
spectral classes representing a single feature class 
(Lillesand et al. 2004). 
The unsupervised approach was tested using the 
ISODATA method. The results were not satisfactory and 
classes were not easily separated due to the large 
variability within individual classes. The accuracy of the