3. IMAGE CORRELATION PERFORMANCE
Automatic matching procedures are used to find
corresponding points in the images and to determine the
respective 3D ground coordinates therefrom. In the
present experiment a straightforward grey value based
correlation was applied, including backward correlation as
a first quality control mechanism.
For the JERS-1 stereo models a more or less coarse grid
of points was correlated to determine representative
accuracy parameters for 3D data extraction. The
elevations of the extracted data were compared to the
corresponding reference elevations extracted from the
reference DEM. Statistics on the resulting height
differences are summarised in Table 4 together with the
percentage of points having been matched under
consideration of the specified correlation criteria.
Concerning the optical data we first of all note a good
stereoscopic performance of the pure JERS-1 stereo
model. The achieved height accuracy of 51 meters RMS
error is even slightly better than proposed within the
geometric modelling. In the experiment related to SPOT
panchromatic stereo data, on the other hand, we have
achieved an RMS height accuracy of around 26 meters.
Hence, the resulting height accuracy achieved from the
JERS data is fairly reasonable if stereo disposition and
pixel resolution are considered, both of them being
significantly worse in comparison to SPOT.
Further a very high correlation success rate of 82.6% was
achieved for JERS-1, whereas the comparative rate for
SPOT was 539^ only. With regard to the image matching
performance it can be seen therefrom, that the
simultaneous in-orbit stereo data acquisition of the JERS
sensor has a distinct advantage versus the multitemporal
data acquisition of the SPOT sensor.
The SAR images were processed in an adequate manner
and, in particular, without any preprocessing (e.g. speckle
filtering) of the image data. Due to the SAR specific
speckle noise and the radiometric peculiarities like
particularly layover the correlation step is significantly
more problematic for these data. This is immediately
documented by the low percental rate of successful
correlation of 17% only. For ERS-1 stereo data, for
comparison, similar correlation rates of up to some 20%
have been achieved.
We can note, however, a fairly high height accuracy of
some 80 meters for the JERS-1 SAR stereo pair, whereas
the respective accuracy achieved from stereoscopic
modelling (Table 2) amounts only to about 140 meters.
Hence we can conclude that the stereoscopic correlation
together with the criteria used for this process is more
accurate than the interactive measurement of (control)
points, which has been done monoscopically in the
individual images.
It can be assumed, that the height accuracy resulting
from JERS-1 SAR stereo data in general may be
674
improved through a proper preprocessing like speckle
filtering of the image data. Then similar accuracies like
those achieved in experiments related to ERS-1 SAR
stereo data, i.e. some 50 meters RMS height errors,
should be feasible, as the geometric prerequisites like
pixel resolution or stereo intersection angle are around
the same.
hk eesesasnsancnscsnsssnanncunnnansnannsans
| Sid.Dov niin. ps nf Bats... anor Sd
| Minimum ( 1966 |... 258.
Maximum 128.6 183.1
Table 4: Statistical values of stereo mapping accuracy
analysis (given in meters).
4. RELIEF MAPPING
4.1. DEM Generation
The correlation results of the optical JERS-1 stereo data
were used in a next step in order to generate a
stereoscopically derived digital elevation model (DEM).
First, 3D ground coordinates were determined for the
matching points through intersection of the respective
projection lines. Then a triangulation procedure was
applied to the 3D point data resulting from the correlation
and intersection procedure. Erroneous correlation data
manifest themselves usually as irregular peaks or holes,
i.e. up-/down-pyramids, in the triangulated network. An
automatic filtering mechanism was applied to identify and
eliminate such outliers, which considers the slope of
triangle edges coinciding at a triangle point. Finally, a
regular raster of elevations was interpolated from the
triangle net with a specified frame and mesh size.
Smoothing procedures were used to eliminate the
triangular terrain shapes in the raster DEM.
This resulting stereo-derived DEM is shown in Figure 3 in
a grey level coded presentation. For topographic
reference a DEM with a cell size of 12.5 is available,
which has been generated from topographic maps in a
scale of 1 : 25000. This DEM is shown in Figure 4. For
ease of comparison selected height levels are indicated
by contour lines. For a detailed quality assessment of the
stereo-derived DEM the height differences to the
reference DEM have been calculated and presented in
Figure 5. In this Figure respective contour lines indicate
height errors larger than +/- 50 meters (dark and bright
areas). The statistical values calculated for this difference
DEM compare well to the values given in Table 4. For
instance, basically the same standard deviation of 47.8
meters or mean value of -18.8 meters have been
determined.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
