1e list. Features showing a 
> removed as well. 
descending rating and only 
he consistency check and 
| be a relevant amount of 
aining features. But even in 
er is low enough to detect 
dle adjustment. 
rojection can be determined 
SS for every captured line, 
ermination — the rotational 
1 the real camera rotation - 
with the 3D-positions of 
on the platform and the 
hese rotational offsets drift 
> modeled by an orientation 
L = 1..N sets of rotational 
nts in time and interpolated 
rves (Wohlfeil, 2010). The 
orrection parameter sets is 
acteristics of the orientation 
ne images, the appropriate 
ts are determined with the 
joints. While the correction 
s typically expressed by a 
ion correction function of a 
parameters. In both cases 
found that meet best the 
ologous points. 
> bundle adjustment from 
used for performing this 
automatically determined 
tment must be performed 
eliminate incorrect points. 
eir residuals, which are 
of all residuals. In practice, 
worked well in all cases. 
ite orientation 
ologous points, the relative 
es is optimal. But if no 
absolute orientation of the 
stment, the accuracy of this 
'ecially the case for remote 
ld of view, such as most 
narrow swath (ie. few 
ld of view and the large 
e earth, the rays of light 
of the image are almost 
tion is very flat in terms of 
It, very small errors in the 
oints cause the absolute 
ing ground control points 
ct space) in the bundle 
of their positions is very 
mall subset of homologous 
Jstment as pseudo ground 
s determined by spatial 
of sight, using the directly 
one additional bundle 
adjustment is performed including both, homologous points 
and pseudo ground control points. In this way, the accuracy 
of absolute orientation can be retained at the level of direct 
geo-referencing capabilities of the remote sensing system 
while the relative orientation is improved without any ground 
control points. 
3.5 Accuracy Determination 
After bundle adjustment, the accuracy of the relative 
orientation is optimized. However, due to small errors in the 
determination of points, but also due to errors in the interior 
orientation, insufficiently uncorrected atmospheric effects, 
etc. the accuracy is limited and varies from scene to scene. 
SGM requires a relative orientation that allows the prediction 
of epipolar lines with less than one pixel of divergence in 
image space. In order to assure that this requirement is met, 
the resolution of matched images, and so the resulting DSM, 
must be reduced, if necessary. 
In order to check and reduce the DSM resolution 
automatically a special measure is introduced. It measures 
how accurate the epipolar curves (they are actually curves in 
line images, instead of lines) can be predicted. The epipolar 
curve is predicting possible locations of a point of image 7 in 
image j according to the given (interior and) exterior 
orientation of both images. As the images of one scene can be 
captured with different sensors, their resolution and geometry 
can differ. Therefore, the epipolar curve is projected onto a 
plane at mean terrain height, resulting in the curve c;;. 
d(c;;,l;) is defined to be the minimum distance between c;;, 
and the line of sight /; corresponding to the point in image i. 
This distance can be calculated for all homologous points for 
any tuple of images i and j being matched. Finally, the RMS 
of all distances calculable for the given set of homologous 
points is defined to be the epipolar error e,. It represents the 
overall-accuracy of a set of images in terms of SGM. 
The epipolar error should not exceed half the GSD, in order 
to meet the constraints of SGM Matching. If it exceeds half 
the GSD of any of the involved images, the images have to be 
down-sampled to a GSD of 2-:e,. At the same time, this 
GSD is the resolution of the resulting DSM (GSDpgy). This 
means that, after calculating the epipolar error from 
homologous points used for bundle adjustment, the GSDpgy 
can automatically be chosen to achieve both, the best possible 
resolution and quality of the result. 
3.6 Water Masking 
There have been numerous research efforts into water body 
extraction from satellite imagery. (Zhaohui, 2003) relies on 
water bodies having a smooth untextured surface and can 
only detect large bodies. (Bovolin et al, 2006) suggest using 
the IR spectral band. Unfortunately, there exists no general 
solution to segment all sorts of water bodies in a limited 
number of spectral bands or even a panchromatic image. 
Fortunately, the U.S. Geological survey provides data on 
water bodies and coastlines in the resolution of the SRTM 
elevation data. The contours are downloaded for the given 
area. As both, the accuracy of the water mask and absolute 
pointing accuracy of the satellite is limited, water areas are 
dilated a suitable distance. The resulting contours are then 
simplified to facilitate processing (Figure 2). 
The SRTM data has already been manually processed and 
contains all types of water bodies, but it has the obvious 
disadvantage of not always being up to date, due to the 
volatile nature of coastlines. As a fall back mechanism, the 
user is provided the option to manually correct the mask. 
77 
  
Figure 2: Dilated water mask generated from SRTM. The 
simplified water polygons are drawn white with blue edges. 
3.7 SGM Image Matching 
As soon as the mentioned requirements are met, stereo 
matching can be applied to overlapping images for 
computing a dense reconstruction of the scene. For a long 
time, local, correlation based methods dominated 
photogrammetry. Due to the size of the correlation window, 
resulting surface models have a lower spatial resolution than 
the input images. Fine structures are lost and object 
boundaries are smoothed. In contrast, global methods permit 
pixel-wise matching, which is guided by a global smoothness 
constraint. The resulting surface models have the same 
resolution as the input images, but global methods typically 
have a high processing time, which prevent their use for real 
world problems. The Semi-Global Matching (SGM) method 
provides a very good trade-off between accuracy and 
processing speed (Hirschmüller, 2005 and 2008). 
SGM performs matching of corresponding pixels along 
epipolar lines. Since, in general, line images cannot be 
rectified such that epipolar lines are exactly aligned with 
image rows, line images are projected onto a horizontal plane 
and matching is performed along calculated epipolar lines 
(Hirschmüller et al., 2005). As matching cost, Census is used, 
for radiometric robustness (Hirschmüller and Scharstein, 
2009), which is very important in satellite imagery, due to 
typically large time differences, which causes different 
shadows. 
  
  
  
  
Figure 3: Reconstruction of Mt. Everest in 50 cm/pixel from 
World View satellite images, provided by DigitalGlobe