anbul 2004
ding to the
algorithmic
teraki and
of AIM is
ed, namely
d in case of
iplex areas.
t approach
pproximate
iltaneously,
isi-epipolar
good rays,
pper levels,
grid points
ibsequently
tour points.
multi-patch
masks of
:eometrical
efining the
and single
r extending
continuities
Baltsavias,
n different
of the final
alysis on a
eters to be
has certain
nd blunder
in terms of
yme of its
based grid
is the AIM
eristics of
s different
st to SS. In
sizes have
images has
ls and the
as input in
al filtering
been used.
cluded and
(irregularly
parameters
IM
jatches/
r points/
lges
/6
4
matching
mages
ineously
itasets were
klines have
n estimated
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
accuracy of 20-40 cm and 40-60 cm for Waldkirch and
Yokohama dataset respectively. The same orientation used later
for the automatic extraction was used also for the manual
measurements. Mass points have been separated in two classes:
ground points close to the perimeter of buildings and ground
points defining the bare earth (BE) surface excluding points
close to buildings. This classification is justifled as errors
arising from matching should be treated separately from
modeling errors of the surface in 2.5D. Breaklines have been
extracted along discontinuities on the ground and on buildings
or man made objects (MMO), even small features on roofs.
Points along the breaklines have been interpolated every 0.20 m,
corresponding to approximately 1 pixel in image space. Forest
areas and trees have been excluded from the measurements as it
was difficult to extract reliable 3D coordinates, due to their
imprecise and unclear shape.
Moreover, due to the radiometric quality of the Yokohama
dataset, many roofs, especially the smaller ones lying partly in
the shadowed areas, could not be accurately extracted and for
several manually measured points the estimated accuracy
degraded (~ 60 cm).
Roadmarks were not always easily visible due to the shadowing.
Less points could be extracted in open surfaces and close to
building outlines and therefore the class used to analyze
separately modeling errors was not used.
3. RESULTS OF IMAGE MACHING AND QUALITY
EVALUATION
DSM extraction was tested in selected regions, excluding large
shadowed areas (especially in the Yokohama dataset). As the
Waldkirch region had more variations in land cover, three
regions have been selected’. The relief in area W1052A (Fig. 4
(a)) was undulating and sparse houses, trees and small salt lakes
existed. The area W1052B (Fig. 4 (a)) was relatively flat with
good texture in agricultural fields. The area W1052C (Fig. 4 (c))
included complex objects, well defined roads, single and groups
of trees, in many cases close to the houses. One representative
region has been selected in the Yokohama dataset (Fig. 5) and it
included high buildings (some over 20 m high) and thus large
discontinuities and occlusions. In combination with the low
radiometric quality, the degree of difficulty in extracting a
reliable DSM of this area increased.
The parameters of each system have been modified according to
the characteristics of each area. As mentioned already in Section
22, in SS the AATE method was used in all tests in
combination with the TIN option. For the Waldkirch test areas
W1052A and W1052C, 1 meter grid interval in object space
was used and for W1052B, as the terrain was less undulating,
the grid interval was set to 1.5 m. For the Yokohama test area a
denser matching was employed with 0.5 m grid spacing due to
the density of buildings and large parallax differences. In AIM,
the strategy of each area varied. Single and multiple template
strategies were used for areas W1052A, W1052B and Y0624C,
W1052C respectively. An approximate surface in the upper
pyramid levels was generated by matching of grid points (5 m
interval), and further refined by matching of edges in the lower
levels. However, for area W1052B edge matching was used
only in the two lower levels as the area was less undulating.
The results of the quantitative analysis, performed for the above
areas, are shown in Table 4. The raw DSM points without any
' The naming scheme and coding of the areas indicates dataset, strip and
region of interest. E.g. W1052A is the region A, imaged in strip 1052 of
Waldkirch (W) dataset. The number of the strip is also used to indicate
that the images used in matching are selected only from this strip in case
the area is imaged in multiple strips.
405
postprocessing (e.g. filtering, modeling of breaklines) and the
interpolated elevations of the reference-measured points were
utilized to compute statistics of elevation differences (RMS,
mean with sign, absolute maximum and standard deviation).
(b) (c)
Figure 4. Test areas in the Waldkirch dataset. The size of
W1052A (a), W1052B (b) areas was approximately
2000 x 2000 pixels and of W1052C (c) area 2500 x
1800 pixels.
pL D d E ; = MT NL
Figure 5. Part of test area Y0624C in the Yokohama dataset.
The total size of the area is 1000 x 3000 pixels.
Errors with respect to points on bare earth (BE) and on man
made objects (MMO) are analyzed separately (Pts, Brklin). The
number of single points on BE was less than the interpolated
breakline points and the first were measured in locally flat areas.
Modeling errors are separately computed as error statistics from
points that lie close to the buildings and on the ground (hPts). In
general, AIM delivered more accurate results than SS in all test
areas. In relative flat areas (W1052B), less differences in the
accuracy could be observed. Blunders in DSMs derived with SS
were significantly higher, in locally flat areas and in the case of
MMO but also at breaklines. AIM shows a better performance
in case of discontinuities, as accuracy increased by
approximately two times and blunders were less than SS for all
areas in the case of MMO. For both systems accuracy degraded
for points close to building outlines, as less points could be
automatically derived in these areas. Here, elevation was
interpolated from the nearest points, which apparently were
either on the ground or on buildings, leading thus often to large
errors. In addition, planimetric errors, even if small, can also
cause at surface discontinuities large height errors. SS has
generated more points than AIM in all areas except W1052B.
The reason for this is the inclusion of edges with AIM, leading
to extraction of many points in highly-textured areas, e.g. the
agricultural fields. In the case of Yokohama, as one could
expect, accuracy degraded more, as the geometric and
radiometric quality of the data was poorer and the acquisition of
