on
ata
es.
ET
er
ut
on
all
e)
on
al
ad
ns
re
ut
ed
he
|a
re
as
to
Darvishzadeh Varchehi, Roshanak
Classification results Acc.
Reference roof Road others
map Roof 84646 0 33973 0.71
Road 0 30411 0 1.00
Others 43438 0 67532 0.61
Overall accuracy = 70.23 %
Table 3. Confusion matrix when updated roofs (reference map) with roads was crossed with
classified roofs plus roads
2:3 Roofs
[1 roads, soil
others,
unclassified
ME No of pixels < 5
ME unclassified
Forex
Figure 6. Combined roof classes Figure 7. classified result when segments
with green areas and shadows segments with number of pixels less than 5
on roofs were detected
Classification results Acc.
Reference roof Road others
map Roof 85270 0 33349 0.72
Road 0 30411 0 1.00
Others 43736 0 67234 0.61
Overall accuracy = 70.35 %
Table 4. Confusion matrix when updated roofs (reference map) including roads was crossed
with the improved classified result
The explanation for this is that, although there are some segments on top of the roofs, which are unclassified and have
five pixels or even more, there also exist some misclassified roofs that have the same situation. By adding to the
threshold the area of the misclassified roofs also will increase. So the accuracy, instead of going up, will come down.
The final classified map after the second stage classification is displayed in Figure 8.
m
Roofs k
[] roads, soil
green, shadow
Figure 8. Final classification result Figure 9. Segmentation result
2.2.4 Quadtree Segmentation. Human image vision generally tends to divide an image into homogeneous areas
first, and will characterize those areas more carefully later (Gorte, 1995). Applying this approach to digital image
analysis leads to segmentation, which divides the image into segments that correspond to objects in the terrain. The
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 317
