ul 2004
—
m
o -5m
«5m
15 m
3D INFORMATION EXTRACTION BY STRUCTURAL MATCHING OF SAR AND
OPTICAL FEATURES
Florence TUPIN and Michel ROUX
GET - Télécom-Paris - UMR 5141 LTCI - Département TSI
46 rue Barrault, 75013 Paris - France
florence.tupin@enst.fr
KEY WORDS: SAR imagery, aerial imagery, optical imagery
ABSTRACT
The subject of this paper is the extraction of 3D information using an optical image and a SAR image. This aim is made difficult by the
very different appearances of the landscape and the man-made features in both images. The proposed method is based on the extraction
of SAR primitives (bright linear and point features). They are then matched to the optical image using the directional and modulus
gradient map computed on the optical image. À set of matches and associated heights is selected for each primitive. In a first step, the
digital terrain model is computed in an iterative way around the predominant height of the SAR primitives. In a second step, using the
external knowledge of a building map, the mean height of each building is computed by means of the previous DTM and the remaining
matching primitives. The results obtained on an industrial test area are evaluated.
1 INTRODUCTION
Three-dimensional reconstruction in urban areas is a very impor-
tant subject. Indeed, 3D models are very useful for many ap-
plications like urban extension monitoring, disaster (flood, earth-
quake) monitoring, mobile phone cells planning, etc. There are
now many ways to obtain digital elevation models (DEM): on
one hand, classical stereovision approaches with optical data are
already operational; on the other hand, radargrammetric or inter-
ferometric approaches with SAR (Synthetic Aperture Radar) data
are still in a research stage (Gamba et al., 2000) (Simonetto et al.,
2003) (Quartulli and Datcu, 2003) (Tison et al., 2004).
We are interested here in the use of heterogeneous sources, spe-
cially the case of optical data and SAR data. Three-dimensional
reconstruction using an optical and a SAR images is theoretically
possible, although the accurate knowledge of the sensor param-
eters is necessary. Besides, since the images of both sensors are
radiometrically and geometrically different (speckle presence and
distance sampling in SAR images, leading to geometrical distor-
tions like overlay and shadow), point matching algorithms with
correlation based methods are un-usable. In this paper we pro-
pose a new method for the extraction of 3D information dedicated
to these two sensors over semi-urban areas. Preliminary results
using the external knowledge of a building map are presented.
There have been many works dealing with the fusion of opti-
cal and SAR data. Two main approaches can be distinguished.
The pixel-based approaches, which often rely on the use of the
joint probability density functions (Schistad et al., 1994) (Lom-
bardo et al., 2003) for segmentation or classification purposes,
and the primitive-based approaches which match features of the
optical and SAR data for registration purposes (Dare and Dow-
man, 2000) or region of interest delimitation (Hellwich et al.,
2000) (Tupin and Roux, 2003).
In the context of 3D information extraction, only figural approaches
are possible. The proposed method is based on the extraction of
SAR strong backscatterers. Indeed, due to the surface smooth-
ness in urban areas (relatively to the wavelength), backscattering
phenomena strongly rely on the surface orientation compared to
the incidence direction. Therefore, specific parts of the buildings
489
usually appear as very bright features (linear and/or punctual),
whereas the rest of the building is more or less of the same grey
level as the background. The radiometrically high features corre-
spond to the corners between the walls of the buildings and the
ground, some parts of the roofs or of the walls; more generally,
everywhere a dihedral or trihedral configuration appears with a
specific orientation towards the sensor. These features are first
extracted in the SAR image, providing a set of salient features.
This step 1s presented in section 2.
These primitives are then projected on the vector gradient map
of the optical image using a range of possible heights (section
3). A set of best matches is then selected for each primitive. In
a first stage, the digital terrain model (DTM) is extracted in an
iterative way around the predominant height of the SAR primi-
tives (most of them lie on the ground since they correspond to
the wall/ground corner of the buildings). This part is described
in section 4. In a second step, using the external knowledge of a
building map, the mean height of each building is computed by
means of the previous DTM and the matched primitives (section
5). Eventually, the results obtained on a test area are evaluated
(section 6). The whole synopsis of the method is shown figure 1.
2 SAR PRIMITIVE EXTRACTION
As said in the introduction, the appearance of the objects in the
SAR image are strongly related to their geometrical properties
compared to the along track direction and the incidence angle,
and to their roughness compared to the wavelength (Hardaway
et al., 1982) (Tupin et al., 2002). Due to multiple bounce scat-
terings, many very bright features corresponding to dihedral or
trihedral configurations are present in the data (Franceschetti et
al., 2002). They correspond to wall / ground corners, balconies,
chimneys, posts, street lamps, etc. Except these features, rela-
tively few information is available on the building surfaces due to
the speckle. Indeed, since rooftop surfaces are quite smooth com-
pared to the wavelength (9cm in S band), their mean radiometry
is often rather close to the one of the ground.
The extracted primitives are thus the very bright features of the
SAR image, either punctual or linear, which correspond to dihe-
