The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
280
3.2 Automatic extraction: Requirements through
characteristics of available sensors and images
To enable a fast reaction after a disaster, special emphasis
should be put on the exploitation of multi-sensorial optical and
SAR satellite images, as the current and in particular the future
availability of these sensors allows to acquiring such images
nearly everywhere and at any time. Optical images show strong
advantages concerning the radiometric image quality if they
were taken during good weather and illumination conditions.
On the other hand, SAR holds some prominent advantages over
optical images, which are in particular helpful in crisis
situations. SAR is an active system that can operate during day
and night. It is also nearly weather-independent and, moreover,
during bad weather conditions, SAR is today’s only operational
system available at all. However, as coherent imaging
technique, SAR images are affected by the well-known speckle
noise as well as image derogations due to radar shadow and
layover. The imaged objects are thus subject to drastic changes
in their appearance depending on the radar illumination
parameters. The challenges of man-made object extraction from
SAR data, in particular roads and buildings, can be viewed in
(Wessel & Hinz 2004, Sorgel et al. 2006). The need for and the
potential of further developments in this area has been
recognized years ago, but it will be even more important in the
future as consequence of the high potential of new airborne and
spaceborne sensor systems such as TerraSAR-X, TanDEM-X,
or Radarsat-2.
Driven by the technological advances, recent work in the field
of automatic object extraction shows the importance of
developing models and strategies that combine evidence from
various sources in a sound statistical framework. The evidence
may stem, e.g., from multiple views, different sensors, or
external data. The approach of (Wessel & Hinz, 2004; Wessel,
2006; Hedman et al., 2006, 2007, 2008) may serve as example
for road extraction from single and multiple SAR images. It
models the sensor- and context-dependent appearance of roads
and is meanwhile adapted to Bayesian decision theory in order
to consistently incorporate multiple SAR images and knowledge
about the appearances and relations of objects. The first step
consists of a line extraction in each image, followed by attribute
extraction. Based on these attributes the uncertainty of each line
segment is estimated statistically, followed by an iterative
fusion of these uncertainties supported by context information
and sensor geometry (see Fig. 1). On the basis of a resulting
uncertainty vector each line obtains an estimation of the
probability that the line really belongs to a road. The final step
includes the generation of a road network by calculating optimal
paths through the weighted graph of line segments and
connections between them. Figure 2 shows the result of
employing context information about forest and urban areas into
the extraction of roads. The borders of these areas serve as seed
points for generating the road network, while the interior of
these areas is excluded. An evaluation with manually digitized
reference has shown that a completeness of approx. 70% can be
reached when accepting a correctness also of approx. 70%.
Figure 1. Bayesian fusion module and its input data
As expected, both correctness and completeness increase when
applying a similar version of this extraction system to optical
images. (The only difference between the two versions is the
procedure for primitive extraction and evaluation.) Figure 3
visualizes the result of road extraction on an optical image of
similar resolution and similar scene complexity. Here, a
completeness of approx. 85% is reached while the correctness is
still over 90%. The example shown later in Fig. 6 illustrates,
though, that even a low percentages of cloud coverage may
influence the quality of the results. We refer the interested
reader to (Mayer et al. 2006), where more details and a
thorough comparison and discussion of different automatic road
extraction approaches applied to optical images can be found.
Figure 2. Result of road extraction from SAR X-/L- pair employing global context (from (Wessel 2006)). White lines: extracted roads; black lines:
missing roads; dotted lines: context area “urban” (seed point for extraction); black areas: context area “forest”.