CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
154
5 CHANGE DETECTION
As mentioned before SAR images are highly affected by noise.
Although the influence of the deterministic speckle effect should
be exactly the same under the same conditions, it is impossible to
assure exactly the same conditions over a longer period of time.
So, if two SAR images are differentiated pixel by pixel the result
is expected to appear very noisy. Alternatively this differentiation
can be calculated in the curvelet coefficient domain. If the input
images are co-registered and same-sized, the images share also
the same combination of curvelet coefficients. Before the differ
ence image is transformed back to the spatial domain, the coef
ficient differences can be either denoised following Section 4 or
weighted quadratically. In the latter case each coefficient is multi
plied by its own magnitude to suppress low and to strengthen high
coefficients. Additionally the influences of the different scales are
equalized by the factor 2 subband (cf. Section 3). As the resulting
image contains positive as well as negative values, the positive
values showing regions that brightened up are coded in green and
the negative values showing regions that darkened are coded in
red. For TerraSAR-X data the geolocation of the detected data
product turned out to be sufficient for the change detection, so
that no further co-registration was necessary.
A disadvantage of this method might be its high demand on mem
ory. The curvelet representation itself is very redundant increas
ing the data volume of an image by the factor 16. Although most
coefficients are nearly zero or set to zero during the image en
hancement process (cf. Section 4), but they have to be processed
during the differentiation as well. If more than three images are
compared the difference matrix including all relative differences
between the input images inflates. But the increasing number of
coefficients goes along with an increasing flexibility in approxi
mating linear features in the input image. Tests with other second
order wavelets proved that critically sub-sampled approaches do
not provide comparable results. To get an impression of the pro
cessing time: The example in Section 5.2 including three input
images of 2091x1113 pixels are processed with a Matlab imple
mentation and require seven minutes on a Solaris workstation.
(a) SAR image 1 (b) SAR image 2
(c) Detected changes (d) Optical image ©GoogleEarth
In the following two examples over the city of Munich are pre
sented. The first one deals with short time changes in the well-
known fairground ’’Theresienwiese”, the second one surveys con
struction activities near the central station over the period of one
year. The processed data sets are acquired by TerraSAR-X in
the High Resolution Spotlight mode and delivered as Multi Look
Ground Range Detected product.
5.1 Short time changes
The two images of the fairground ’Theresienwiese” (Fig. 4(d))
have been acquired in December 2008 and January 2009. Being
processed as spatially enhanced product they have a pixel spac
ing of 0.5 m on ground. Because of the relatively short time lag,
the reflectivity of the surrounding is expected to be the same, so
all changes should be man-made. Comparing visually the two in
put images (Fig. 4(a) and 4(b)) one can remark a brighter area in
the upper middle of Fig. 4(a) that darkened in the second image
(Fig. 4(b)). Especially on the streets inside the fairground many
single pixel changes are obvious. For urban applications single
pixel changes do only disturb the interpretation as one is more
interested in changes happened to structures like streets or build
ings. So, these single pixel changes have to be excluded. Spa
tial averaging would help to find large areas with high changes,
but fine linear structures would be smeared over and probably
get lost. The curvelet approach is able to preserve the structures
while single pixel changes are suppressed. In Fig. 4(c) there
Figure 4: Change detection in the fairground ’’Theresienwiese”
(1: 05.12.2008,2: 18.01.2009)
is one red region in the upper middle of the image, that accords
with the visual interpretation. These changes refer to the ’’Winter-
Tollwood” festival that took place during the first acquisition. The
pavilions caused a much higher reflectivity than the bare soil dur
ing the second acquisition. Additionally there are some small
red and green structures at the bottom left of Fig. 4(c) that were
not visible before. Those refer to buses and cars on a parking
lot. The slightly darkened region in the middle right of Fig. 4(a)
and 4(b) respectively is not marked as change because it does
not contain any structure. In summary, the change image shows
nearly no disturbances as all small scale changes are excluded.
The curvelet approach is very sensitive towards structures (e.g.
buses) and very robust towards slight large scale changes caused
by environmental influences.
5.2 Long time changes
For damage mapping after natural disasters it is only seldom pos
sible to access up-to-date reference data, as most events cannot be
predicted yet. So, seasonal changes in the surrounding of the re
gions of interest have to be taken into account. The three images
of the railway station ’’Donnersberger Brucke” acquired in March
2008 (Fig. 5(b)), September 2008 (Fig. 5(c)), and March 2009
(Fig. 5(d)) are used to map the construction progress inside the
