Darvishzadeh Varchehi, Roshanak 
  
cts 3 CHANGE ANALYSIS 
he 
The segment-based classification result was used to analyze the remaining differences between the old roof map (1992) 
and the roofs as determined with the segment-based classification (as it serves as the best result). The implementation of 
as the segmented-based classification method resulted in a final map, which is presented in Figure 11. This result was used 
to to derive the changes. For this reason, the final result of image segmentation (see Figure 11) was subtracted from the 
Or. old (1992 roofs) data (see Figure 12). In the ideal case, the remaining parts of the subtraction should be only the 
changes. The result of changes found in the on-screen digitizing was overlaid with that result (see Figure 12). Visual 
analysis of this result shows the following types of error: 
   
eir 
an 1. Existing error in the boundaries of roofs. As seen in Figure 12, there are a large number of small areas (Slivers) 
in indicating errors along the boundaries of roofs. This incorrect classification can be largely explained by spectral 
ed confusion and incorrect referencing of roof segment geometry. 
2. Existing thematic error in the old roof map (1992 digital data) which served as ground truth. Good examples of this 
type of error are the detected changes shown in Figure 12. 
fa 3. Error due to spectral similarity of roofs with e.g. roads, soils and the surrounding environment. The remaining 
m. unclassified and misclassified segments can be explained by this type of error. This type of error, can be seen from 
is the visual interpretation of Figure 12 (large areas). This is an inherent problem of spectral classification and 
segmentation. 
ed Derived changes 
ea 
nt [1 old roofs (1992) 
he [1 (on-screen) derived 
ch changes 
on 
ed 
Figure 12. Derived changes from on-screen digitizing overlaid with 
changes from segmentation 
4 CONCLUSION 
The aim of this study was to compare the changes derived from on-screen digitizing with extracted changes by means of 
classification and segmentation. Although the result of on-screen digitizing had a better accuracy and completeness, it 
should be kept in mind that the knowledge and intelligence of an expert has been applied. In this case study, the time 
oil used for on-screen digitizing was less than for classification and segmentation but this may not be the case in a 
hadow production environment. In the case of dealing with a bigger database it is obvious that on-screen digitizing may need 
fied more time. For classification and segmentation, once the method is developed, is much faster. However, a compromise 
may be made between the cost of achieving high levels of accuracy and the need for rapid change detection. 
Classification and segmentation may give a better result where the buildings are larger than those found in this photo 
area. The result of segmentation and classification could be improved by adding extra knowledge, for example adding 
some information about the shapes of the roofs or their textures (usually new buildings because of the new materials 
have different reflectance). Also this result can be improved by the accurate spatial registration of the image, which is 
possible through making an orthophoto. 
Therefore, Scanned high resolution aerial photographs could play an important role in the various fields for spatial 
information gathering and map updating. They could provide the necessary data to create information for urban 
planners and decision-makers. The integration of remote sensing in a GIS environment makes it also possible to use 
knowledge on the history and the dynamics of urban land use to improve the image classification results. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 319