The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1209
The unrealistic threshold of R=0.9 for the correlation
coefficient of the least squares matching leads to a not
negligible loss of accepted points. On the other hand a small
step width of just 1 pixel increases the percentage of accepted
points very slightly, but requires a quite higher processing time.
The study gave optimal results for the sub-matrix size for
matching of 10 x 10 pixels.
Optimal results have been achieved in the area with man-made
objects with the threshold for the correlation coefficient R=0.70
of the least squares matching, the sub-matrix of 10 xlO pixels
and the step width of 3 pixels. A lower threshold for the
correlation coefficient may lead to a lower accuracy, requiring a
compromise between both.
Area with man-made objects and trees:
The area includes complex objects such as small and large
buildings and trees close to roofs. Problems are caused by
shadows, repetitive objects, occlusions and poor texture. Few or
no points were extracted in shadow areas close to the base of
buildings.
The matching in dense urban area may not be sufficient to
model discontinuities such as buildings. Buildings often
occlude each other. Furthermore, man-made objects are usually
made with homogenous materials causing large areas of poor
texture. Figure 5 shows the frequency distribution of the
correlation coefficient in a complex area.
parameters
R= threshold for accepted correlation
matching success
R = 0.90 , Step width = 3
sub-matrix 10 xlO pixels
48%
R = 0.80 , Step width = 3
sub-matrix 10 * 10 pixels
61%
R = 0.70 , Step width = 3
sub-matrix 10 xlO pixels
74%
R = 0.70 , Step width = 1
sub-matrix 10 xlO pixels
76%
R = 0.70 , Step width = 3
sub-matrix 7 *7 pixels
35%
The template must have a sufficient size, but a larger size is
corresponds to a low pass filtering, requiring also a compromise.
The identification of points located on top of building may fail
if too large templates are used and the objects may look like
smooth hills and not like buildings. Beside this, noise is caused
by trees close to buildings.
Some “noise*’ of the DSMs has been eliminated by filtering
with the Hannover program RASCOR (Passini et al 2002).
Figure 4 shows the effect of filtering.
Figure 4: Effect of noise filtering on ground and buildings
(a) Before filtering - (b) after filtering
Table2: Matching completeness in area with man-made objects
depending upon chosen parameters
motel ie«l point
□ RHO > 0.95
¡a 0.90 < RHO <
0.95
□ 0.85 < RHO <
0.90
□ 0.80 < RHO <
0.85
■ 0.75 < RHO <
0.80
□ 0 70 < RHO <
0.75
■ 0.65 < RHO <
0.70
□ 0.60 < RHO «
0.65
■ 0.55 < RHO <
0.60
■ 0.50 < RHO <
0.55
□ 0.45 < RHO <
0.50
□ 0 40 < RHO <
0.45
■ 0.35 < RHO <
0.40
■ 0.30 < RHO <
0.35
■ 0.25 < RHO <
0.30
■ 0.20 < RHO <
0.25
□ 0.15 < RHO <
0.20
□ 0.10 < RHO <
0.15
□ 0.05 < RHO <
0.10
□ 0.00 < RHO <
0.05
■ RHO « 0.00