1091
GENERATION OF COARSE 3D MODELS OF URBAN AREAS
FROM HIGH RESOLUTION STEREO SATELLITE IMAGES
Thomas Krauß, Manfred Lehner, Peter Reinartz
German Aerospace Center (DLR), Remote Sensing Technology Institute
PO Box 1116, 82230 Wessling, Germany, thomas.krauss@dlr.de
Commission I, WG 1/5, ThS-3
KEY WORDS: digital terrain model, surface modelling, stereo image, high resolution, optical satellite sensors, image matching,
classification, contouring, visualization, information extraction
ABSTRACT:
With the emergence of more and more satellites delivering very high resolution (VHR) imagery with ground sampling distances in
the range of one meter or below the generation of three dimensional urban models directly from space may become possible. Such
models are required for many applications in areas where no up-to-date detailed urban mapping exists like in developing countries.
Besides the creation and updating of maps from sprawling urban settlements, such three dimensional models are also very useful for
simulation and planning. For example simulations of catastrophic events like flooding, tsunamis or earth quakes rely on digital
terrain models (DTM) populated with three-dimensional man made and natural objects. Using VHR satellite imagery is often the
faster and cheaper alternative to acquiring aerial photos or even laser DEMs or in some times even the sole source of information for
remote areas. In this paper a method for an automatic processing chain for urban modeling based on stereo images from VHR
satellites is proposed. After import and preprocessing of the images a digital surface model (DSM) is derived from the stereo data.
Subsequently a digital terrain model (DTM) and true ortho images are generated based on the DSM. Using a high objects mask and a
vegetation mask based on the normalized difference vegetation index (NDVI) a coarse classification will be derived and suitable
models for the detected objects (e.g. houses, trees, ...) are selected and finally the 3D models get exported in a suitable format like
VRML.
1. INTRODUCTION
The launch of WorldView-I on the 18 th of September 2007
added a new member to the fleet of very high resolution (VHR)
earth survey satellites. VHR satellite images with ground
sampling distances (GSD) in the range of 0.5 to one meter are
now available from Ikonos-2 (launch 24.09.1999, GSD pan
1 m, multispectral (MS) 4 m), QuickBird (launch 18.10.2001,
GSD pan 0.6 m, MS 2.4 m), OrbView-3 (GSD pan 1 m, MS
4 m 06/2003-03/2007), and WorldView-I (GSD pan 0.5 m).
Also some more systems are planned for the near future like
GeoEye-1 (planned launch mid 2008, GSD pan 0.4 m, MS
1.65 m, (DigitalGlobe, 2008)), WorldView II (2009, GSD pan
0.5 m, MS 1.8 m), or the Pleiades-HR (2010, GSD 0.7 m, MS
2.8 m).
Imagery with such high resolutions allow for the first time the
generation of high resolution urban models directly from space
for nearly every point on earth. Throughout the following
investigations Ikonos VHR stereo image pairs were used.
Already in 2001 C. S. Fraser showed that Ikonos stereo
imagery has the potential for high resolution building
reconstruction (Fraser et. al., 2001). A semi automatic
approach can be built upon such VHR data with rather good
results. However only few papers exists on fully automatic
processes for extraction of urban objects from high resolution
satellite data.
In contrast many approaches exist for city modelling from
high resolution aerial imagery, laser scanner data, or semi
automatic modelling from high resolution imagery. For a short
overview or further information please refer to (Brenner, 2003)
or the Ascona proceedings (referenced also in Brenner, 2003).
These methods are mostly based on cadastral data, aerial
images, aerial and terrestrial laser scanner data, terrestrial
photographs and more additional information since the aim of
these methods are often near photorealistic city models in
industrial countries integrating data from several sources in
often intense manual work for the urban models (CyberCity,
2008, 3D Geo, 2008). For the future Pleiades satellites
(Flamanc, 2005) proposed a framework for the generation of
3D city models using a digital surface model (DSM) and a true
ortho image generated from the high resolution satellite data
together with 2D footprints of the buildings.
In contrast we want to propose a (mostly) automatic method
for deriving 3D city models from only one single stereo scene
of any area in the world. So no additional data besides the two
VHR stereo images will be available since the results should
also be usable for rapid mapping purposes in catastrophic
events.
In this case only coarse city models with a level of detail
between LoD-1 and LoD-2 will actually be needed. Following
the definition in (CityGML, 2008) LoD-1 describes a block
model with buildings as polygons with only a flat roof and
LoD-2 describes buildings with a more detailed outside cover
including more detailed roof structures and simple textures. In
our case a coarse distinction between buildings with flat roofs
and gabled roofs but no more roof details may be possible
depending on the quality of the generated DSM.