ı Graph-
r Aided
(1), pp.
> Planar
ramme-
'eatures
esearch
d Data
62.
nterna-
thods,
XIX
e Mit-
)JAGM
197 —
Fitting
Aided
Olaf Hellwich
FUSION OF OPTICAL IMAGERY AND SAR/INSAR DATA FOR OBJECT EXTRACTION
Olaf HELLWICH, Manfred GÜNZL and Christian WIEDEMANN
Chair for Photogrammetry and Remote Sensing
Technische Universität München, D-80290 Munich, Germany
Phone: +49/89/289 22677, Fax: +49/89/280 9573
E-mail: Olaf.Hellwich@photo.verm.tu-muenchen.de
URL: http://www.photo.verm.tu-muenchen.de
Working Group III/6
KEY WORDS: Data fusion, Multi-spectral data, Radar, Multi-temporal, Object recognition
ABSTRACT
Optical imagery such as high-resolution panchromatic or multispectral data, and SAR/INSAR data show different infor-
mation about the imaged objects, and have different advantages and disadvantages when used for object extraction or
landuse classification. Multispectral optical image data is largely determined by the type of the material an object consists
of. Panchromatic data which is often available with a higher resolution than multispectral data emphasizes geometric
detail of the objects, e.g. the complex structure of anthropogenic objects such as road networks. In contrary to this, SAR
data contain information about small-scale surface roughness and - to a lower degree - soil moisture. Height informa-
tion derived by interferometric processing of SAR data contains large-scale surface roughness. These different types of
information are referring to completely different object qualities and are, therefore, largely uncorrelated which helps to
reduce ambiguities in the results of object extraction. The main advantage of passive optical imagery with respect to SAR
data is the lack of the speckle effect leading to images with a far better extractability of linear as well as areal objects. A
major advantage of SAR is its all-weather capability which allows the acquisition of time series of imagery with exact
acquisition dates under any climatic conditions. In this paper, these complementary properties of SAR and optical image
data are demonstrated and used to improve object extraction and landuse classification results.
1 INTRODUCTION
A concept for the fusion of optical image data with SAR/INSAR data for scene interpretation is presented. In Section
2 it is formulated as a semantic network which is implemented in form of a Bayesian network incorporating uncertainty
information. The uncertainty information is provided by different algorithmic components of the developed approach
using approximate probabilistic reasoning. In the Bayesian framework also contextual information is considered. Such
information is e.g. that agricultural fields commonly have a rectangular shape. A shape parameter for the segmentation
of grid data allowing to judge computationally very efficiently the deviation of the segment's shape from a rectangle
is introduced in Section 3. In Section 4 multisensor fusion is demonstrated using two examples. In the first example,
boundaries of areal objects and roads are extracted from optical data, and most of the landuse classes are found with the
help of multispectral information. In the second example, a multitemporal SAR data set is combined with a multispectral
optical image for the purpose of landuse classification.
2 BAYESIAN NETWORK FOR MULTISENSOR MULTITEMPORAL CONTEXTUAL CLASSIFICATION
In this section a Bayesian network for classifying multisensor multitemporal image data considering contextual criteria
is introduced. Only the application of Bayesian networks is explained; for the theoretical background we refer to the
literature. On the conceptual level the Bayesian network is based on a semantic network as previously described in
(Hellwich and Wiedemann, 2000). The relations between real world objects and image data are modeled on three levels.
The real world level contains the topographic objects such as forest, the sensor level the image data. The geometry and
material level plays a mediating and connecting role between sensor and real world level. Its task is to take into account
that the objects are often not directly the cause of the data contained in the images, but that certain material or geometric
properties of the objects are more directly linked to the measured data.
Figure 1 shows a simple example of a Bayesian network for classification of image data. In this case the nodes or
primitives of the network correspond with the pixels of the image. The network is used for landuse classification which is
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 389