Darvishzadeh Varchehi, Roshanak 
  
success of segmentation depends on the availability of: high resolution imagery in such a way that the relevant objects 
are represented by a significant number of pixels, powerful hardware, and an efficient implementation regarding the 
size of the remote sensing images (Gorte, 1998). 
The segmentation program uses a multi band image (in this case 3 bands) as input and gives one segmentation as 
output. Moreover, information, like object locations, sizes and perimeters can be retrieved (readers are referred to 
technical report, experimental quadtree software, Gorte 1995). In this study the threshold was chosen by trial and error. 
The result of this segmentation is illustrated in Figure 9. 
This result contains many small segments (noise) and mixed pixels. To remove all these the segments which their 
number of pixel is less than ten have become zero (see Figure 10), meaning that, all the objects with an area less than 
1.6 m? will be deleted. Although by doing this some information may be lost, but those information is of no interest in 
this study as it is here supposed that a roof has at least an area of 2 m?. Further improvement of this result was achieved 
by the segment-based classification, which is described as follows: 
2.2.5  Segment-Based Classification. The aim of segment-based classification is to determine the class (label) of a 
segment (polygon), of which the geometry is contained in the segmented map using the above segmentation program. 
Therefore, the pixels within the polygon are identified from the classification result and the class of the polygon is 
determined from these pixels. 
The following steps were taken to arrive from a pixel-based to segment-based classification: First the georeferenced 
image is classified using per-pixel classification resulting in a label per-pixel. The segmentation result has been Area 
Numbered to effect distinct area numbering; connected raster elements with the same value belong to the same segment. 
The output map from Area Numbering, was superimposed (crossed) with the final classified map in order to get the 
statistics of pixels within each segment. Subsequently, a frequency Table was established to determine the label of each 
segment. Then the most occurring (predominant) class label for each segment was calculated (using the Aggregation 
function) and assigned to the segment (including the unclassified labels). 
The classification accuracy was assessed by comparing the output of the segment-based classification with the updated 
roof map (reference), by calculating confusion matrices and the overall accuracy. This result is presented in Table 5. 
  
  
  
  
  
  
  
  
  
  
  
  
Classification results Acc. 
Reference roof Road others 
map Roof 87939 0 30680 0.74 
Road 0 30411 0 1.00 
Others 38654 0 72316 0.65 
Overall accuracy = 73.33 % 
  
  
Table 5. Confusion matrix when reference map (updated roofs) including roads were 
crossed with improved segmentation including roads 
Roofs 
roads, soil 
green, shadow 
AN others, 
unclassified 
  
    
    
Figure 10. All the small segments Figure 11. Final result of segmention 
With an area of less than 1.6 m? after using classification for 
are detected improvement 
  
  
318 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.