17
REAL-TIME SAR SIMULATION FOR CHANGE DETECTION
APPLICATIONS BASED ON DATA FUSION
Timo Balz
Institute for Photogrammetry (ifp), Universität Stuttgart, Germany
Geschwister-Scholl-Strasse 24D, D-70174 Stuttgart
timo.balz@ifp.uni-stuttgart.de
Commission II WG 1, Commission VII WG 6 & 7
KEY WORDS: SAR, Real-time, Simulation, Change Detection, Fusion
ABSTRACT:
SAR simulators are important tools for developing new SAR systems as well as for supporting the analysis of acquired SAR data.
Using modem graphics cards for SAR simulation, even complex environments can be simulated in real-time. This is realized by
implementing SAR geometry and radiometry within standard graphics hardware, which nowadays offers 3D hardware acceleration
and programmable graphics processing units (GPU). The geometric differences between optical and SAR images are the biggest
challenge for any multi-sensor data fusion approach. High-resolution data fusion should be based on 3D models, in urban areas pro
vided by city models. Differences between the model and the newly acquired data can be detected by comparing the simulation of the
data with the acquired data. Due to unreliable models and inaccurate simulation results, automatic approaches provide various false
alarms. Semi-automatic approaches are more reliable alternatives.
1. INTRODUCTION
The time needed to calculate a SAR simulation is not crucial for
most scientific applications or for applications in sensor design.
For these applications, the SAR simulation has to be as realistic
as possible. A totally different application scenario is the real
time visualization of SAR effects. Here the realism of the simu
lation is not crucial, but the visualization must be in real-time.
Real-time SAR simulations are useful for interactive applica
tions, like mission planning or applications in training and edu
cation. Furthermore, simulation assisted change detection and
data fusion benefit from simulation results in real-time.
Radar images differ in many ways from images acquired by
passive sensor systems. Any data fusion approach has to consid
er these differences. The combination of SAR simulation and
visualization provides new methods for model based data fu
sion. Based on 3D models, e.g. city models, differences between
the model and the newly acquired data can be detected by com
paring the simulation of the models with the acquired data. Data
fusion between high-resolution SAR and optical images must
consider the terrain and the 3D shape of the objects of interest.
Fusing an agricultural area for vegetation classification in a low
resolution, the topography of the area is negligible. Using high-
resolution images in mountainous or urban areas, the terrain and
object shapes have to be taken into consideration. This requires
3D models of the area of interest. Beside the problem of
availability, the quality and reliability of these models is often
questionable. Any change detection based on such models has
to be aware of this problem and has to accept the incom
pleteness of the models as well as the inaccuracies of the physi
cal models or of the simplified implementation used for the si
mulation. Semi-automatic approaches are reliable alternatives.
In the following section, the development of the graphics hard
ware is briefly discussed. In section 3, the real-time SAR simu
lation tool SARViz will be presented. While in section 4 the
need for 3D shape awareness in data fusion is discussed. Fusion
based change detection in urban areas will be demonstrated in
the final part of the paper.
2. GENERAL-PURPOSE COMPUTATION ON
GRAPHICS PROCESSING UNITS
Rapid developments in computer graphics allow for more and
more realistic visualization of extensive virtual worlds in real
time. Visualization applications are realized by GPUs. A mod
em GPU is a data-parallel streaming processor working in a
single-instruction, multiple data (SIMD) fashion. Because
GPUs are highly parallel and specialized on certain arithmetic
operations, these operations can be calculated astonishingly fast.
As depicted in Figure 1, GPUs of the latest generation have
more computational power as standard CPUs, providing nowa
days almost supercomputing speed.
O 150
-I
E
О
-Intel CPU
■ ATI GPU
• NVIDIA GPU
2002 2003 2004 2005 2006
Figure 1. Comparison between CPU and GPU speed (see Buck,
2004; Owens et al, 2007)
Beside visualization, the programmable graphics hardware can
also be used for a variety of general-purpose computations (Ow
ens et al, 2007). A GPU design differs from a CPU design. As
shown in Figure 2, a big part of the CPU is used for branching,
whereas most transistors on the GPU are used for arithmetic op
erations (Owens, 2005).
