Darvishzadeh Varchehi, Roshanak 
  
2.2.3 Pixel Classification. The pixel classification was performed by a standard maximum-likelihood classification 
using the three bands of the scanned aerial photos. The updated roof map guided the classification. The training data 
which was applied for supervised classification comprises the following 17 classes: roofl, roof2,...., shadow, and trees. 
The mean vector and covariance matrix of the seventeen classes was determined from a total of 150 to 300 samples per 
class. Figure 5 shows the result of the classification. 
     
4 m] 
Figure 4. Study area Figure 5. Classificat on result 
    
After classification, all the roof classes were combined into one class (roof) and all the shadows and trees into another 
class (green). There also exist another class, which consists of unclassified pixels. Since there was no information about 
this class, it is named others (see Figure 6). 
This classified map was subjected to accuracy assessment. The accuracy assessment was done by means of a confusion 
matrix and the overall accuracy (expressed as the percentage of correctly classified pixels) was computed. The overall 
accuracy of the classification result was assessed by a cross tabulation of the rasterized updated roof map (reference) 
and the classified map (see Table 2). 
  
  
  
  
  
  
  
  
  
  
Classification results Acc. 
Reference roof others 
map roof 84646 33973 0.71 
others 58523 82858 0.59 
Overall accuracy = 64.42 % 
  
  
Table 2. Confusion matrix when reference map was crossed with classified roofs 
However, classification of remotely sensed imagery is effected by e.g. isolated and mixed pixels and spectral confusion 
of land cover types (Abkar, 1994). Therefore, in the majority of cases, classification based solely on spectral 
observation is not sufficiently accurate for extraction of the roofs, especially in this area where reflectance of the road 
and bare soils is very similar to the roofs. Consequently, no samples were taken for roads and bare soils. This means 
that a lot of roads and bare soils will be classified as roofs. For the roads this is not a problem, because we know where 
the roads are from the digital database. For the bare soils the situation is different; we have no prior information about 
them. This problem can not be resolved by considering bare soils in the classification. If so, the opposite case will 
happen that means many of roofs will be classified as bare soils. 
Ist-stage of classification improvement: As a result, existing knowledge from the digital database (e.g., roads) was used 
to improve the result of the classification. Roads were extracted from the digital data, rasterised and combined with the 
classified map and the roof map (reference). The overall accuracy of this result is shown in Table 3. 
2nd-stage of classification improvement: The segments in the final classified map (1-stage) that are unclassified with a 
size of less than 5 pixels were detected (see Figure 7) and removed (converted to roof class). In Table 4 the overall 
accuracy of this result after this stage of improvement is shown. 
The selection of the threshold (5 pixels) was based on two criteria: 
l. The hypothesis that there are no roofs smaller than this threshold (0.8 m?). 
2. Choosing a larger threshold will affect opposite in the accuracy result. In fact, first only the segments which are 
unclassified and their number of pixels is one were removed. Then those which their number of pixels is two has 
been removed (added to roof class). The same was done for three and four pixels. When this number changes to 
five the accuracy will suddenly come down instead of going up. 
  
316 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.