ul 2004
onse of
ir asso-
e a first
ach, we
‘ground
ght his-
rincipal
close to
mitives
und hy
+0n])
> match
whole
search
1ed ap-
| a sub-
can be
ulation
on the
uctures
ation.
which
) select
it. This
ght but
itching
uilding
nts ly-
:ularly,
during
y have
associ-
os with
s been
s. The
for the
ng and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
P
Figure 7: Result on a building: on the left the matched SAR prim-
itives superimposed on the optical image (in blue corresponding
to ground matches and in pink corresponding to the building); on
the right the building footprint given by the map and the matches.
For instance, for the building shown on figure 7, a mean altitude
of 20m is obtained which is the true elevation (i.e an height of
11.5m). This good result is due to the small parapet at the end
of the roof which is correctly matched with the corresponding
feature in the optical image.
This is quite difficult to give general conclusions about the poten-
tial of the method since the SAR image appearance strongly influ-
ence the result. But some general remarks can be made, both on
the intrinsic limitations of SAR / optical 3D reconstruction, and
about the method and the improvements which should be made
and are subject of further work.
Concerning the intrinsic potential of SAR / optical 3D informa-
tion:
e one of the main limitation is of course the assumption that a
strong reflector in the SAR image correspond to an edge in
the optical data; in fact, it has often been verified on our test
set, even for short edge;
e the second main limitation is the nature of the SAR data; in-
deed for many buildings with rather smooth roof, the backscat-
tered signal is uniform (although noisy); in this case there is
no 3D information.
Concerning the improvements to be made:
e some good candidates to match are the parapets ending the
roof; a problem arises from the shadow presence in the opti-
cal image since a match with higher response can be caused
by this shadow (an example is shown figure 8);
e the linear features are supposed to be horizontal (for a seg-
ment, both extremities should be at the same height); in
practice for part of the roof with a “V” shape, this constraint
can provide some problem;
e an important limitation is that only the best match of a SAR
primitive is used; furthermore, the method does not intro-
duce any contextual relationship between the primitives (this
is done in a coarse way in the estimation of the building
height by imposing to all the features to have the same height);
in practice the inverse strategy should be tested: an height
hypothesis is made for the building and the corresponding
score is computed (allowing a small tolerance around the
tested height);
e as said before, the suppression of the features corresponding
to ground/wall corners is very important (if they are badly
matched a wrong height would be given); but on the other
hand, the method proposed is too severe and many features
are suppressed due to accidental matching at the ground
level; both ground estimation and building reconstruction
should be performed simultaneously with a complementary
constraint: ground features cannot appear inside a building.
Figure 8 presents a building which perfectly illustrate these prob-
lems. The predominant height is the good one (26m correspond-
ing to the beginning of the roof) thanks to 6 well matched points
and 2 well matched linear features on the roof. But some lin-
ear features are matched against the shadow, thus giving a wrong
height. Besides, many other linear features of the roof have been
suppressed during the ground estimation because of the period-
ical structure of the roof: they have been matched by accident
with the optical image to an height within the ground interval
[hg — dn; hg + bn).
75
Figure 8: Building with a periodical structure generating some
mis-matchings
7 CONCLUSION
This article has presented a preliminary study for height informa-
tion extraction using a SAR data and an optical data. The results
are encouraging if bright features appear in the SAR image since
they often correspond to a corner also visible in the optical data.
Nevertheless, such method gives very sparse results (only a few
features are matched) and the 3D reconstruction step should be
made in the same time as the scene interpretation.
Acknowledgment The authors would like to thank Emmanuel
Meier and Michel Auger for some algorithmic developments.
REFERENCES
Bloch, I., 1996. Information combination operators for data fu-
sion: A comparative review with classification. IEEE Transac-
tions on Systems, Man and Cybernetics SMC-26(1), pp. 52-67.
Canny, J., 1986. A computational approach to edge detection.
IEEE Transactions on Pattern Analysis and Machine Intelligence
PAMI-8(6), pp. 679—698.
