Istanbul 2004
ian scan pixel
ed tie points
formed using
angulation is
1, no manual
of the stereo
rtuozo, like in
tion of the tie
nage Station,
ns and offers
asured at each
ve can not be
extraction in
sitions no tie
adjustment in
value can not
the analytical
fact, that only
le adjustment,
sitions where
| o; does not
esulting from
t up the areas
the tie point
ally measured
jint extraction
isurements for
each of the
s of tie points
tion, therefore
act enough for
ults could be
esented results
tive character.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
NT EP M \ 4 Sn
ZN NE co. TT
referri en Det E ag e
T. f : : A x s [^
e 1 : |. A Á.
* Te ; Jj & pai .
E E
Figure 3: Distribution of tie points, red: automatically generated
points, blue: manually measured points
3. DTM GENERATION
3.1 Reference data
In the area of investigation, no control points of superior
accuracy exist. Thus, for the comparison of automatically
generated DTMs, again accurate reference data measured on an
analytical plotter Wild S9 was acquired using the software
XMAP by Aviosoft™. For 5 stereo models from two strips,
parallel profile measurements were performed with an average
point distance of 20m along the profiles and a profile distance
of 20m. Breaklines were not measured due to the flat areas
except for model 116 117. Using our software DTMZ, grids
with 5m mesh size were interpolated from the measured points.
The grids then could be imported to ESRI ArcGIS™ 8.3 and
directly compared with the grids acquired from the DIPS by
calculating the height difference:
AZ = Lpies - ZREFERENCE- (1)
From the resulting height difference grid, mean value, RMS
error and the minimum and maximum offset were computed.
Potential trends in the performance of height errors and their
spatial distribution can be extracted. To visualise the
distribution of the height errors, histograms of the difference
grids and contour lines were produced, which allow for an
examination of the geomorphological accuracy.
The topography covered by the processed stereo models is
predominantly flat with only few discontinuities except for one
model, consisting of the images 16 and 17 in the first strip,
which contain parts of 3 quebradas (figure 1).
For DTM generation, the orientations computed using Image
Station Digital Mensuration were transferred to the analytical
plotter and used for the manual measurements. On Virtuozo, 4
control points per stereo model, transferred from ISDM, were
used for absolute orientation of the images. To minimise the
marginal differences caused by varying spatial attitudes,
Geomagic Studio 4.1 by Raindrop Geomagic Inc. was used to
register the automatically derived DTMs to the reference data
(global registration function).
3.2 Image Station™
The method of DTM generation applied by Image Station
Automatic Elevations (ISAE) is described in detail in the
Nn
wo
manual which is integrated in the graphical user interface (Z/I
Imaging Corporation, 2002). For each level of the image
pyramid, an initial DTM is derived by matching homologous
points, starting with a horizontal plane in the first level. From
this initial DTM, a DTM is modeled with bilinear finite
elements which then serves as the initial DTM for the next
level. For matching, ISAE uses the interest operator and
correlation coefficient while the matching area is geometrically
defined by a parallax bound and the epipolar line.
ISAE offers a lot of different strategies for DTM generation.
Users can choose different terrain types, adaptive or non-
adaptive grid, parallax bound and matching modes, correlation
thresholds or define terrain types by themselves. Different
smoothing filters with user-defined weights and sampling
factors can be applied. Some tests with the default terrain types
showed, that the best results could be achieved using terrain
type "flat" with adaptive mode. The smoothing filter was set to
"high", keeping the default values for smoothing weight (2.0)
and sampling factor as recommended in the manual. After a first
attempt with a correlation threshold of 0.95, a value of 0.75 was
chosen because of the low point density and thus a strong
terrain filtering achieved with 0.95 (figure 4). The result is a
grid with 5m mesh size which the software interpolates from the
measured points.
The height differences obtained for the different models show
clearly, that not only texture but also terrain characteristics,
especially steep slopes like on the border of the valleys in the
images 16 and 17, affect the accuracy of the automatically
generated DTMs. Table | shows the mean height error, its
standard deviation and the minimum-maximum range of the
height differences between the automatically derived DTM and
the manually measured DTM.
Table 1: DTM generated using Image Station compared to the
manually measured DTM on an analytical plotter
Wild S9
Model | AZ | Std. Deviation Min. — Max.
116 117 025m | 3.10m -19.1m — 24.2m
210 211 ON : :
211 M2 022m | 1.99m -4.4m — 36.4m
212 213 0.25m | 1.53m -6.3m — 18.8m
213 214 -0.02m | 1.33m -12.9m — 7.0m
223 224 -0.01m | 0.77m -3.3m — 7.6m
Mean height errors, standard deviations and the minimum-
maximum-ranges of the different stereo models show a
noticeable heterogeneity. In model 210_211, for an unresolved
reason no correct DTM could be calculated.
The differential grid of model 116 117 shows another
phenomenon: The big blunders show a coincidence with areas
of steep slopes and are predominantly positive (figure 4).
