Standard deviations of individual offset differences are under 
0.5 GSD in case of the combined matching, with the exception 
of the height component in Georgian Bay. These numbers re- 
present accumulated errors from human measurements in two 
overlapping data sets on one hand, and SGM and point cloud 
matching on the other hand. An attempt to separate these contri- 
butions is carried out in Table 3, which shows standard devia- 
tions derived from all offsets in each strip overlap, individually 
for each method. Such a computation assumes no offset varia- 
tion along an ADS strip overlap, which is not entirely correct 
but, however, the impact is the same for all methods and doesn’t 
prevent a comparison. 
The numbers in Table 3 confirm the typical accuracy of human 
stereo measurements (note there are two points required to 
derive an offset). They also show that the combined automatic 
approach can achieve better quality, especially for the height 
component. As already found above, the largest standard devia- 
tions occur across flight direction. The largest overall number, 
the X offset standard deviation of geometric matching in the 
New Mexico data set, is impacted by some erroneous results, 
which are discussed in the next section. 
3.3 Performance on Different Terrain Types 
The success rate of the SGM-based info cloud collection and 
subsequent geometric/radiometric point cloud matching is 
strongly dependent on the image content — i.e., on the sensor, 
illumination and viewing geometry and especially terrain pro- 
perties and their variation within a patch. This is influenced by a 
multitude of factors, and the interaction of which is highly com- 
plex. However, image matching requires intensity gradients, and 
geometric point cloud matching is based on height gradients; 
the combined approach utilizes both types. Furthermore, the 
surface approximation by local planes must be valid. Based on 
those theoretical considerations as well as practical tests on real- 
world data, different types of terrain can be characterized regar- 
ding the success rate in providing reliable offsets. So far, this 
performance was investigated on ADS data sets that predomi- 
nantly include mountains, urban/suburban areas and forests. For 
the examples discussed here, the percentages of patches that 
delivered reliable offsets are shown in Table 4. 
  
  
  
  
  
  
  
Data Set Patches C er Ced 
Georgian Bay 113 58.4% 77.8% 
Lansing 808 74.4% 89.4% 
New Mexico 378 66.4% 97.6% 
Subset: Mountains 132 97.0% 100.0% 
  
Table 4. Percentages of reliable strip offsets, based on the total 
number of evenly distributed patches. 
Mountainous terrain inherently features significant height and 
intensity gradients, which leads to a high percentage of results, 
for even up to 100% of the patch locations. The success rate in 
urban areas can range widely, 60-90%. The same holds true for 
trees, where it is generally lower, in the order of 20-50% for 
dense forests, but rapidly increases with the presence of 
clearings and/or different tree species — i.e. intensity and height 
gradients —, which is the case in the Georgian Bay block. The 
main issue, especially in higher vegetation such as trees, is the 
representation by essentially arbitrarily scattered points, which 
impacts or even invalidates the approximation by local planes. 
Patches located entirely in water can be expected to fail. 
106 
  
Figure 2. Patch locations and planimetric offset results for the 
geometric point cloud matching (red) and the combined method 
(black). Examples for the New Mexico strip overlap; offset 
vectors scaled by a factor of 1000. 
The benefit from the combined point cloud matching over the 
geometric approach can be clearly seen in Table 4. Success 
rates are improved for all data sets and all types of terrain. In 
addition, the resulting offsets are more reliable. The benefit is 
minor in mountains (Figure 4, left) but immediately obvious in 
flat areas (Figure 4, right). While offsets from both methods 
agree well in the first case, the solely geometric matching de- 
livers a number of erroneous results in the latter example, which 
were not identified by the offset verification as described in 
section 2.2. This issue could lead to wrong conclusions in the 
Shear Analysis. 
4. SHEAR ANALYSIS / QUALITY CONTROL 
The automated Shear Analysis, based upon the combined geo- 
metric/radiometric matching of dense info clouds, is currently 
becoming a part of North West’s production workflow. For the 
purpose of QC, patch locations are determined along the center 
of each ADS strip overlap in a spacing of 1000-5000 pixels. 
Automatically computed offsets and related statistics for all 
patches are output into a single, comprehensive report. Based on 
that, the user is provided excerpts of different kind, in parti- 
cular: detailed tabular views of all offset parameters that enable 
examination of the matching; comparison of reliable offsets 
against QC thresholds; summaries per image overlap as well as 
for the entire ADS block; graphical outputs of patch footprints 
and offset vectors with attributes assigned, which allows for an 
analysis using commercial mapping software such as Global 
Mapper or ArcGIS. This visualization is the primary tool used 
in the QC process and for further evaluation. 
The 
com 
evali 
not « 
of i 
havi 
the I 
This 
She: 
and 
utili: 
was 
relia 
two 
deln 
resu| 
bloc 
resu 
pron 
Wit 
men 
appr 
cura 
In c 
radk 
vide 
of / 
auto 
tool 
imp: 
The 
the « 
but 
(ster 
appl 
have 
Akc 
tatic 
Che 
trati 
Rob 
Geh 
Roc 
chin 
Vol. 
Geh 
201( 
DSN 
Vol. 
Geh 
Con 
ASP