International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
IMAGE 1 || IMAGE 2
Pixel 0.67 0.58
Meter 6.6 5.69
Pixel 0.63 0.57
Meter 3.38 3.06
Pixel 0.94 0.81
Meter 7.42 6.46
RMSE,
RMSE,
RMSE,,
Table 1. The accuracy of GCPs for the two images
Figure 1. Distribution of GCPs and TPs in the image, the red
point numbers are GCPs and the blue point numbers are TPs.
In the next step these two images were converted to epipolar
projections. Epipolar projected images are required by the DEM
generation routine, since it reduces the error between the stereo
images in the y-direction, so that the stereo matching can be
performed. Measured parallax differences on a pixel-by-pixel
basis are converted to absolute elevations using trigonometric
functions and the orbital data (orbital position, altitude, attitude
and the scene center). The computation relies on the inherent
parallax between stereo images (Siva Subramanian, 2003). An
automated image correlation algorithm (Toutin, 1995) is used to
derive elevations from the parallax, by a set of well located
GCPs and tie points (TPs). The image matching technique
operates on a reference and a search window. For each position
in the search window, a match value is computed from gray
level values in the reference window. The match value is
computed with the mean normalized cross-correlation
coefficient and the sum of mean normalized absolute difference
(Marra , 2001). Elevation points are extracted at every pixel.
The 3-D intersection is performed using the computed
geometric model to convert the pixel coordinates in both images
determined in the image matching of the stereo pair to the three
dimensional data. The output elevations are not computed for
the pixels where the image matching fails to find the
corresponding pixel in the reference image, resulting into some
failure areas. In case of small and scattered failures the software
does interpolate and compute most probable values for them.
The generated DEM is in raw format and does not contain geo-
referencing information. So, the DEM needs to be
georeferenced by using GCPs. The workflow of these steps for
generating DEM from HRS stereo images has been shown in
Figure 2.
390
HRS stereo images Digital 3D map
GCPs selection
TPs selection
Y
Stereo model generation
(bundle adjustment)
Y
Epipolar images generation
Y
Final DEM generation
No
Manual editing of DEM
Accuracy assessment
Figure 2. The work flow of DEM generation
In this study, a DEM of 10 meter grid size for the whole area
was generated using OrthoEngine module of Geomatica
software. Lots of empty patches were found in the generated
DEM, especially in forest areas, due to matching failures. Some
100 tie points were collected but it could not significantly
improve the quality of generated DEM. Because there are lots
of forest regions in these images, lots of empty patches exist in
the extracted DEM. Then, a 60Km by 25Km subset of the
whole area was selected for quality assessment of the generated
DEM. This subset contained fewer empty patches and covered a
rather hilly terrain with height difference of about 350 meter.
This subset is shown in the Figure 3.
Figure 3. The subset of extracted DEM with less failing areas
Wes T w^ vU NN
on
a ap
