the 
ung 
or a 
and 
itrol 
atic 
n in 
rect 
VIFs 
tage 
1 be 
the 
Vere 
  
,We 
ages 
mid 
tner 
000 
cted 
mes 
n of 
etric 
nd a 
hing 
etric 
first, 
tion 
0 m 
yere 
and 
hout 
the 
is of 
two 
n in 
ions 
ally 
of 1 
Poli, Daniela 
  
pixel on level 4 would result to ca. 370 m height error, so we considered blunders (3 RMS) heights with errors larger than 
ca. 1100 m. The accuracy of the remaining match points is very high at ca. 0.2 pixels, showing the potential of our 
matching method, if large blunders can be excluded. 
  
  
  
  
  
  
  
  
  
  
  
  
  
Transformation Match point dataset used # of match # of reference Mean Max. RMS 
used in matching points points Abs. 
conformal before tests 8891 55 1100 7189 2776 
after 2 tests (after 1st test) 6148 (6643) 42 904 7189 2313 
% rejected points 30% (25%) 
after 2 tests and blunder elim. 6148 36 41 203 74 
rotations before tests 8891 55 1162 6994 2786 
after 2 tests (after 1st test) 6539 (7366) 43 886 6994 2304 
% rejected points 26% (17%) 
after 2 tests and blunder elim. 6539 37 35 154 66 
shifts before tests 8891 55 1168 7006 2798 
after 2 tests (after 1st test) 6636 (7731) 45 986 7006 2427 
% rejected points 25% (13%) 
after 2 tests and blunder elim. 6636 38 38 210 69 
  
  
  
  
  
  
  
Table 1. Statistics of height differences (in m) between semi-automatic measurements in the original images 
and automatic matching in low resolution images (288 m). The matching in the low resolution 
images was provided using different transformations (conformal, rotation, shifts) and radiometric 
adjustment in every iteration. Two automated tests to detect blunders were applied. Blunder 
elimination was manual to show the accuracy potential of matching. 
After comparing the results, we chose as best version the one obtained with rotations. Contour lines with a interval of 200 
m were interpolated, and overlapped on a high resolution image. Two groups of clouds at different height were identified: 
the group in the top of the image, with a top height of 4000-7400 m, and the group in the bottom, which is very low 
(down to 600 m). Fig. 5 shows contour lines (for Z >1000 m) of the region containing the group of high cirrus clouds and 
the same region in the original image. This figure demonstrates nicely some of the matching problems encountered. 
Using area-based matching, land areas between the clouds and close to their borders, are attracted during matching due 
to the bright higher-contract clouds to a totally wrong height, i.e. those of the clouds. These land points can not be 
recovered during matching in lower pyramid levels as they are very far away from the correct points. Furthermore, since 
they were within the acceptable height range and they had stuck to a side minimum, they could not detected by the two 
quality tests applied. These problems are nicely shown in the contours of the lower middle part (Fig. 5), where land 
points are lifted up at the cloud-top heights, whereas in the upper middle part, where large land areas with no clouds 
exist, points are correctly matched. Another disadvantage of area-based matching is that it does not allow a good 
modelling of cloud borders, leading to smoothing. These effects become worse, the higher the pyramid level is, so 
feature-based matching and/or starting from lower pyramid levels will lead to reduction of these problems. Additional 
matching problems were caused for this stereopair, due to the large parallax (height) range. To start of 1-2 pixels 
approximations, a pyramid level higher than the 6th one would be needed, but then the image would become extremely 
small. Thus, we started from level 6th living with the fact that some points could never be correct due to poor 
approximations. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 1165