Detection of Lines in Synthetic Aperture Radar (SAR) Scenes
Olaf Hellwich®, Helmut Mayer* and Gerhard Winkler"
* Chair for Photogrammetry and Remote Sensing, Technical University Munich
Arcisstr. 21, D-80290 München, Germany
T: +49.89.2892.2677, F: +49.89.280.9573, EMail: olaf@photo.verm.tu-muenchen.de
"Mathematical Institute, Ludwig-Maximilians-Universität
Theresienstr. 39, D-80333 München, Germany
EMail: winkler@rz.mathematik.uni-muenchen.de
KEY WORDS: line extraction, Markov random fields, SAR, interferometry, GIS, feature extraction, fusion.
ABSTRACT:
Due to the speckle effect of coherent imaging the detection of lines in SAR scenes is considerably more difficult than in optical
images. In spite of this, users of SAR data strongly demand their reliable and accurate detection. Therefore, a new approach to
detect lines in noisy images using a Markov random field (MRF) model and Bayesian classification is proposed. The unobservable
object classes of single pixels are assumed to fulfill the Markov condition, i.e. to depend on the object classes of neighboring
pixels only. The influence of neighboring line pixels is formulated based on potentials derived from a random walk model. Locally,
the image data is evaluated with a rotating template. As SAR intensity data is deteriorated by multiplicative noise, the response of
the line detector is a normalized intensity ratio which results in a constant false alarm rate. The maximum a posteriori (MAP)
estimate of the object parameters is approximated using simulated annealing. To obtain results with less computational effort the
iterated conditional modes (ICM) estimator is applied to the maximum likelihood estimate. The approach integrates intensity,
coherence from interferometric processing of a SAR scene pair, and given Geographic Information System (GIS) data.
1. INTRODUCTION
Lines in SAR scenes can be used for precision geocoding of
SAR scenes or precision registration of SAR scenes with
images acquired by other sensors (Leberl, 1990). Extracted
lines are a basis to verify as well as update linear objects in
GIS or in maps (Caves, 1993). Geologists use SAR scenes to
detect lineaments, as the SAR sensor is very sensitive to geo-
logic structures. The problem with lines in SAR scenes is that
they are not only difficult to detect (Adair & Guindon, 1990;
Hellwich & Streck, 1996), but that they are also partly in-
visible depending on the azimuth of the incident radiation
(Hendry et al., 1988). In this paper as lines we regard narrow,
elongated areas with approximately constant image intensity
which are bounded by bright or dark regions. Note that this
includes also lines which are bounded by a bright region on
one side and a dark region on the other side.
In the past several approaches to the detection of linear struc-
tures in SAR scenes have been taken which can be differen-
tiated into three groups depending on whether they mainly rely
on a local evaluation of the intensity function, whether they
consider more global criteria, or whether combine both. Local
detectors either compute geometric properties like the first or
the second derivative of the intensity function (e.g. Burns et
* This research was partially funded by Deutsche Agentur für
Raumfahrtangelegenheiten (DARA) GmbH under contract SOEE9423.
312
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
al., 1981; Kwok, 1989), or they conduct a statistical evaluation
of regions often defined by a rotating template (e.g. Caves et
al., 1992; Lopes et al., 1993). Several investigations had the
result that a local gradient computation is not suited for the
detection of edges as the speckle effect of coherent imaging
causes a noisy response of the edge operators (Geiss, 1984;
Bellavia & Elgy, 1986; Adair & Guindon, 1990). Among the
operators using statistical parameters those computing the
intensity ratio of neighboring regions have been shown to give
the most reasonable responses (Adair & Guindon, 1990; Caves,
1993). They have a constant false alarm rate, as the standard
deviation of SAR intensity is equal to the intensity itself.
Approaches using more global criteria are the methods based
on the Hough transform (Wood, 1985; Quegan et al., 1986;
Skingley & Rye, 1987; Green et al., 1993), minimum-cost
search (Bellavia & Elgy, 1986) or dynamic programming
(Wood, 1985) to extract thin lines from SAR images. In spite
of promising results the use of the Hough transform is limited,
as in this context it can only be applied to the detection of
straight lines.
Methods combining local operators with a more global evalua-
tion are those developed by Samadani & Vesecky (1990) and
Arduini et al. (1992). They use Bayes’ theorem and a MAP
estimation to combine a conditional probability to observe
certain image data given a linear structure with a prior prob-
ability derived from the generic knowledge that lines are con-
tinuous and neighboring pixels depend on each other.
Our ne
related |
and forr
an MRF
structur
lines, g
knowlec
the kno
narrow,
intensity
ate the
fore inc
(PDF)
second
tinuous
line car
physica
evidenc
model :
betweer
tion to
structur
linear «
approac
close tc
ability
directio
As SA
coherer
SAR sc
the full
In sect
framew
structui
of cont
model
local e
dedicat
SAR s
section
in secti
2. BA
The ex
approa
tion of
& Sch
density
accord
