RFACE MODELS 
AGERY 
rmany 
ny 
igh resolution digital surface 
ned automatically, there are 
: surface models. Especially 
which is a growing problem 
illy manual selection of tie- 
for stereo matching. It also 
| estimate of the depth range 
1 this paper an approach is 
point selection, enabling the 
rater masking and elevation 
tests with a large number of 
1 reliability of the proposed 
drastically reduced, the time 
teps became the critical part 
ore, during the last years, 
been developed to solve this 
| 2008; Wohlfeil, 2010 and 
) a very operational solution 
of high resolution digital 
put. As most of the current 
resolution are line scanners, 
g. WorldView 1/2, GeoEye, 
| on this type of sensors. 
eparation 
image data 
r 
processing 
     
oto generation 
e processing steps 
  
ly slow, the search range 
nge in the images can be 
f the terrain. An automatic 
ed in Section 3.6. 
structions from images, the 
f cameras must be known. 
~ (interior orientation) is 
typically known due to camera calibration. For line cameras 
the extrinsic parameters (exterior orientation) essentially 
consists of 6 degrees of freedom for every captured camera 
line. Dense stereo matching requires that the remaining 
geometric error is less than 1 pixel in image space. However, 
if possible it should be below 0.5 pixels. Since the absolute 
pointing accuracies of satellites are much worse, the exterior 
orientation must be optimized with respect to a precise 
relative orientation, using homologous points (also called tie 
points). Especially for line imagery there was a lack of 
software for performing this task robustly and reliably. With 
an integrated approach, presented in the Sections 3.1 to 3.3 of 
this paper, this problem 1s solved. 
If the requirements in terms of relative orientation cannot be 
met, the resolution of the DSM has to be reduced. This task is 
also being automatized with the help of a suitable accuracy 
measure, introduced in Section 3.4. 
Another problem is water, which cannot be matched, since 
images are taken at different times, causing different textures 
in every image based on the movement of waves. Thus water 
should be identified, ignored while matching and smoothly 
interpolated from the shore, later on. The automatic water 
masking is described in Section 3.5. 
An overview of the processing chain is given in Figure 1. The 
only manual interaction that remains is the selection of the 
suitable stereo imagery and its housekeeping data (initial 
exterior and interior orientation, etc.), described in the 
following Section. 
2. PREPARATION 
For the generation of DSMs one or more groups of images 
are specified. Each group contains two or more images that 
are to be matched via SGM. If possible, the images of each 
group should be captured 
- with a large overlap, in which matching can be performed 
- with different along-track viewing angles (pitch angles) 
- at similar seasons and daytimes in order to avoid large 
differences of shadows and vegetation, which can reduce 
the quality of the result 
If it is not possible or not economic to meet these conditions, 
processing is possible as well, but the quality and resolution 
of the resulting products will be suboptimal, but still very 
useful for many application. 
The images of different groups should partly overlap. This is 
important to enable the generation of one large and continu- 
ous DSM of the entire captured area. The homologous points, 
found in the overlapping areas, allow a global alignment of 
the images, reducing spatial discontinuities of the DSM 
between the areas in a high degree. It is also recommendable 
to use groups with different across-track viewing angles (roll 
angles) for the same area in order to resolve most of the 
occlusion in urban areas. At the current stage of development 
the selection of the images is performed manually because 
this only takes a few minutes. There are very good prospects 
to automatize even this step as well (see Section 5). 
3. PROCESSING 
3.1 Height Range Determination 
One important parameter for SGM process is the maximal 
occurring disparity i.e. the size of the search range for 
matching. This parameter negatively influences the result of 
the algorithm if set too low, which results in the search for 
matches being cancelled too early. If — on the other hand - it 
is set unnecessarily high, it slows down processing, since the 
75 
computation time of the process depends on the size of the 
search range. 
Finding the right value for this parameter is essentially a 
matter of finding the maximum and minimum height in the 
scene being processed. We achieve this by consulting the 
digital elevation data of the Shuttle Radar Topography 
Mission (SRTM) of the year 2000. 
The SRTM data is freely available on the web from the U.S. 
Geological survey, providing an elevation model of the 
world in 3 arc-second resolution. This information is 
downloaded and evaluated when needed and a safety buffer 
of 400 meter is added to allow for buildings and other 
deviations from the model. 
3.2 Automatic and robust tie point selection 
The relative orientation of images is optimized by bundle 
adjustment, which requires a sufficient number of homolo- 
gous points visible in two or more images. Several ap- 
proaches for automatic point selection were developed in the 
last decades. The general approach is to select small salient 
image regions (features) in one image and to find the 
corresponding image regions in the other images. This can be 
performed via cross correlation or the efficient implementa- 
tion (Bouguet, 2000) of the KLT feature tracker (Tomasi and 
Kanade, 1991). In case of unknown scale and rotation differ- 
ences between images, approaches like SIFT (Lowe, 2004) 
and SURF (Bay et al, 2008) are preferable. However, 
satellite images are typically provided with a good initial 
orientation and the KLT feature tracker is by far more 
efficient than the other options. 
Independent of the approach used for feature matching, there 
is always the problem of mismatches that can occur under 
suboptimal conditions. Especially if several difficulties like 
moving objects and shadows, repetitive patterns, changing 
vegetation and illumination, specular reflections, water sur- 
faces, perspective distortion etc. come together, the number 
of mismatches easily exceeds the number of correct matches, 
even by multiples. In such cases almost all approaches for 
automatic point selection fail miserably. 
Two steps are vital to successfully process such difficult 
imagery. First, possible radiometric differences between 
images due to different spectral band characteristics of 
sensors, changed vegetation and sun angle have to be 
compensated as far as possible. This is performed by adaptive 
radiometric balancing (as explained in A.2.4 of Wohlfeil, 
2011). Second, the majority of mismatches have to be 
determined and eliminated already during tie point selection 
by a consistency check of redundant matches. Therefore, 
image features are being matched redundantly from every 
image to any other image in all possible directions. The 
consistency of different matches can then be checked. 
  
  
  
  
  
F Position GC | 112434 1.5 
x 1 | X IX X | X 
o XIX X 
3 
4 X] X XX 
SX XIX 
  
  
  
  
  
  
  
  
Table 1: Exemplary information associated with one feature 
in N= 5 possible images. The feature is apparently not visible 
in image 3 and it couldn't be tracked between images 5 and 2. 
The score associated with this feature is 56% (14/25) 
! http://dds.cr.usgs.gov/srtm/