CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
164
The reminder of this paper is organized as follows: Section 2
gives a brief review of state-of-the-art methods for delineating
the 3D geometry of buildings from SAR images in general,
eventually leading to a discussion of the boundary conditions of
this study. The theoretical background for height estimation
from across-track interferometry as well as error sources are
compiled in Section 3, before Section 4 analyses the accuracy
potential of deriving building heights under various given
prerequisites. Finally, Section 5 draws conclusions in the light
of the TanDEM-X mission and the results of this study.
2. 3D BUILDING GEOMETRY FROM SAR IMAGES
2.1. Overview
Over the past decades, a large variety of approaches for deriving
3D building information from SAR images has been developed.
According to underlying methods and used data the different
methods can be roughly grouped into following categories:
(a) height-from-shadow using mono- or multi-aspect data
(b) fitting prismatic models based on statistical optimization
(c) model-driven segmentation of pre-computed height data
(d) height estimation supported by feature detection / matching
(e) Exploiting layover areas in single or multiple InSAR pairs
To keep the overview focused, we only refer to the original
work of each of these groups. We are aware that numerous
approaches have been developed meanwhile, which could be
assigned to one or more of these groups.
Ad a) Due to the oblique imaging geometry of SAR systems,
buildings cause the well-known RADAR shadow, which
basically corresponds to the occluded area at ground. As for
conventional optical shape-from-shadow approaches, it only
needs simple trigonometry to calculate the object height from
the shadow boundary when knowing the sensor imaging
geometry and assuming horizontal ground (similar for the
layover area (Tupin, 2003)). A compilation of the
corresponding formulae can be found, for instance, in (Sorgel et
al. 2006). It is usually assumed that a shadow edge corresponds
to a certain object edge, whose height is to be estimated. As
only a few number of building edges can be matched to shadow
edges for a specific viewing direction of the SAR, (Bolter &
Leberl, 2000; Leberl & Bolter, 2001) generalize this approach
to multi-aspect SAR and embed it into an iterative height
estimation framework supported by InSAR cues. By this,
building footprint and height are estimated simultaneously,
yielding an accuracy of 1.5m - 2m for airborne SAR.
Ad b) The concept described in (Quartulli & Datcu, 2001;
2003) models the geometry of buildings and geometric relations
between adjacent buildings by a number of parameters
(position, length, width, height, roof slope, distance etc.). After
initialization of model instances in image space, the parameters
are statistically optimized using amplitude, coherence and
interferometric phase information from the images. While this
kind of thorough object-oriented modeling helps to cope with
heavy noise and image derogations, it limits the approach to a
small number of building shapes, not speak about the
computational complexity mandatory for parameter
optimization. This might one of the reasons why the results
cannot prove the general feasibility of the approach and no
accuracy analysis has been carried out; whereas, the
mathematical formulation is very elegant.
Ad c) A purely data-driven strategy that complements the
aforementioned approach is presented in (Gamba &
Houshmand, 1999; Gamba et al., 2000). The procedure starts
with the computation of the interferogram and derives level
lines by segmenting it into height intervals. Level lines fulfilling
certain shape constraints are selected as seed points to start a
regiongrowing algorithm. This algorithm continues as long as
segments can be added without exceeding a predefined
threshold for co-planarity. The achieved accuracy using
airborne C-band data is reported to be 2.5m for large industrial
buildings. This method is in principle independent of the data
source and can be applied to any kind of height models, as so
for LIDAR-based height models (Gamba & Houshmand, 2000).
Ad d) While the former extraction strategy infers the semantics
of buildings purely based on the roof geometry, approaches
following the spirit of (Sorgel et al., 2003; Tison et al., 2007)
include hypotheses of buildings, building parts, and/or adjacent
context objects (roads, vegetation, etc.) from the very beginning
of processing. To this end, a supervised classification and/or
feature detection is carried out before building reconstruction.
This may contain areal objects but also linear features and spots
indicating double bounces at building walls, which become
especially prominent in high resolution SAR (Stilla, 2007). The
cues provided by these hypotheses are then iteratively grouped
and optimized together with the heights derived from InSAR
data until reasonably shaped buildings are extracted or
hypotheses are rejected. Due to generic processing of multiple
cues, this concept is easily extended to multi-aspect SAR data.
The reported accuracy yields again 2 - 3m for the airborne case.
Ad e) The final group of approaches does not only include
image features derived from SAR or InSAR data but models the
complete interferometric phase profile for building walls and
roofs (Thiele et al., 2007). Since vertical walls form layover
areas as consequence of the oblique RADAR distance
measurement, this kind of modelling implicitly contains the
assumption that the main contribution of scattering in such
layover areas is induced by building walls and not by clutter in
front of the building or by the overlayed part of the roof. This
approach can be generalized to SAR tomography (Reigber &
Moreira, 2000; Fornaro et al., 2003) if more than one
interferometric pair of the same viewing direction is available.
(Zhu et al., 2008; 2009) show that deriving 3D information via
tomographic analysis and statistical model selection can be
adapted to pixelwise calculation of dense height maps of urban
areas, thus linking the concepts of SAR tomography with
Persistent Scatterer Interferometry (Ferretti et al., 2001;
Kampes, 2006). These approaches are however in a preliminary
stage so that a thorough accuracy analysis is not yet available.
2.2 Discussion
While each of the approaches is characterized by individual
advantages and limitations, the latter category seems to be a
good compromise between a data-driven strategy and object-
oriented modeling. It is flexible in the sense that it is not
restricted a-priori to specific building shapes. On the other
hand, there are still object-oriented aspects included since
typical building regularities are to identify in the InSAR data.
Concerning the utilization of shadow and layover effects one
has to keep in mind that, especially in urban areas, layover
appears very often and may also cover shadow from
neighboring buildings. Hence, shadow areas are usually hard to
