AUTOMATED PIPELINE EXTRACTION FROM INTERFEROMETRIC SAR DATA
OF THE ERS TANDEM MISSION”
Olaf Hellwich!, Ivan Laptev? and Helmut Mayer!
! Chair for Photogrammetry and Remote Sensing
? Forschungsgruppe Bildverstehen, Informatik IX
Technische Universität München, D-80290 Munich, Germany
E-mail: (olaf|helmut) Q photo.verm.tu-muenchen.de; laptev Q informatik.tu-muenchen.de
URL: http://www.photo.verm.tu-muenchen.de
Commission VII, Working Group VII/6
KEY WORDS: SAR Interferometry, Data Fusion, Line Extraction, Markov Random Field, Snakes
ABSTRACT
A new method for the automated extraction of pipelines and roads from Synthetic Aperture Radar (SAR) scenes is pre-.
sented. It combines intensity data with coherence data from an interferometric evaluation of a SAR scene pair. The fusion
is based on Bayesian statistics and part of a Markov random field (MRF) model for line extraction. Both, intensity and
coherence data are evaluated using rotating templates. The different statistical properties of intensity and coherence are
taken into account by a multiplicative noise model and an additive noise model, respectively. The MRF model introduces
prior knowledge about the continuity and the narrowness of lines. Posterior odds resulting from the MRF method are input
to a ziplock snake-based method for linear object extraction. This processing step is controlled interactively which seems
to be necessary as long as fully automatic processing of noisy data does not provide sufficiently predictable results. The
method is applied to data of the ERS tandem mission.
1 INTRODUCTION
The extraction of linear objects like roads, rivers and
pipelines from SAR data is of great practical interest, be-
cause of the all-weather availability of SAR data. Its au-
tomation is not easy owing to the general speckle effect and
“no-show” effects special to linear features (Leberl, 1990).
The newly developed method for linear object extraction
tries to compensate for these defects as follows: SAR inten-
sity and interferometric coherence data are combined in a
Bayesian data fusion. An MRF model is used to bridge line
gaps caused by speckle effects, thus considering the char-
acteristics of SAR data. In the following processing, object
extraction methods primarily developed for optical data are
applied. Here, interactively initialized ziplock snakes are
used for pipeline extraction.
In the following two sections the basics of MRF- and snake-
based line extraction are explained. Then the method is ap-
plied to an ERS tandem data set from the Siberian lowlands
for pipeline extraction.
2 MRF-BASED LINE EXTRACTION
Previous investigations have shown (Hendry et al., 1988,
Hellwich and Streck, 1996) that SAR intensity and coher-
ence contain complementary information about linear ob-
jects. Over the whole scene both data sources are only
weakly correlated. Visual inspections reveal that linear ob-
jects are often only visible in either intensity or coherence.
To fuse the information contained in both data sources an
automatic method for data fusion was developed based on
a Bayesian approach.
* This research was partially funded by Deutsches Zentrum für Luft- und
Raumfahrt DLR e.V. under contract 50EE9423.
532
According to Bayes' theorem
p(e|y) x p(yle)p(e) (1)
the a posteriori probability density p(e|y) of the object pa-
rameters e given the observations y is proportional to the
product of the probability density p(y|e) of the observations
given the object parameters and the a priori probability den-
sity p(e) of the object parameters. Considering that the ob-
servations are a constant in the estimation of object param-
eters, p(y|e) is also called likelihood function (of the object
parameters).
Bayes' theorem can also be applied recursively by using
the a posteriori probability density p(e|y1) belonging to the
vector of observations y1 as a priori probability density for
the evaluation of a second vector of observations ya (Koch,
1990, p. 8). If y; and y are conditionally independent, the
a posteriori probability is given by
p(elyr,y2) o Pp(vale)p(elu1)
x p(yale)p(y1|e)p(e). (2)
In the newly developed method for line extraction the obser-
vation vectors y1 and y» contain values of the pixels of the
SAR intensity and coherence images. The object parame-
ters e to be estimated for each pixel are the line or no-line
state and - in case of the line state - the line direction. The
a priori probability density p(e) of the object parameters is
formulated as an MRF. In MRF the probability of a state in
a pixel s does not depend on the states of all other pixels
of the image, but only on the states of pixels belonging to a
neighbourhood Os of s.
The goal of processing is the computation of a vector of ob-
ject parameters, i.e. line states and directions as well as no-
line states for the pixels of the image, for which p(e|y1, v2)
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
is |
act
loc
cor
shc
ity :
USi
(Ge
(1)
sint
Inr
funt
whe
prol
In t
tern
can
2.1
Line
sists
et a
bacl
as v
the f
the «
the «
crea
Byr
The
H, (3
sure
rivec
cohe
2.1.1
modi
mali:
two |
It is c
wher
The i
direc
ratios
a dei
giver
wher
no-lir
with
