Olaf Hellwich
expressed by a landuse node on the real world level. On the sensor level multispectral data and interferometric synthetic
aperture radar (/NSAR) data nodes are shown. The connecting geometry and material level consists of green vegetation,
wood, and soil nodes for the evaluation of multispectral data, and small-scale roughness and large-scale roughness for
INSAR data. The states of the landuse node are landuse classes, e.g. forest, agricultural vegetation, and build-up areas.
The states of the sensor nodes are for instance grey value vectors; those of the material nodes numerical degrees of
presence. The direction of the arrows indicates that the objects to be extracted from the observations are introduced as
root nodes and the observations as leaf nodes (cf. (Pearl, 1988, page 157ff.) versus (Kulschewski, 1999b, Kulschewski,
1999a, page 28f.)). This means that the observations are considered functions of the material properties of the object, and
that the material itself is seen as a function of the object. In this way the direct sensor model is implemented through the
a priori probability density function (pdf) of the root node and the conditional pdf's of the leaf nodes and all remaining
nodes. Thus, the /anduse node is described by the a priori pdf, i.e. by probability values for landuse variable & such as
p(e = forest) = 0.3. The multispectral image data node is given a conditional pdf p(z,,|m,, m,,,m,) where x, is
the grey value vector of the multispectral image data, m, the degree of presence of green vegetation, 7n,, the degree of
presence of wood, and m, the degree of presence of soil.
real world
material
(C wood 7j C soll small rough large rough.
Figure 1: Bayesian network for multisensor landuse classification. (small rough.: small-scale roughness, large rough.:
large-scale roughness)
For the classification of a pixel the leaf nodes are instantiated with the corresponding grey value vectors from the im-
age data sets. Then the Bayesian network is evaluated, i.e. the probabilities are propagated according to the respective
formulas, e.g. given by (Pearl, 1988, Koch, 2000, Kulschewski, 1999b), using the pdf’s.
The model for pixel-based classification is extended to process multitemporal data allowing changes of the landuse node
in time. Figure 2 shows the corresponding Bayesian network, in this case for one multispectral data set and multitemporal
INSAR data sets acquired at later points in time. The arrows between the landuse nodes indicate the dependence of the
state of the node at time #; on the state of the node at time 1; 1. In the Bayesian network this dependence is implemented
with the help of conditional pdf's p(c;|e; ,). In the extreme cases this function either does not allow any changes or
it does not favor any particular state c; with respect to the previous state c; ,. In the first case, for the function values
p(eilei—1) = 0 holds whenever e; # €i—1 which means that the landuse nodes could be united. In the second case,
p(eilei—1) = const., independent of the state of £; ; which means that all classes are equally probable, i.e. that the
landuse nodes could be disconnected. In general the conditional pdf’s of the landuse nodes give transition probabilities of
landuse in time (Bruzzone and Serpico, 1997). For example, when in a certain area a change from forest to built-up land
is to be expected this is expressed by the pdf. The conditional pdf's between landuse nodes developing in time can also
include models describing dynamic developments such as growth and harvest models.
SON.
Figure 2: Bayesian network for multitemporal multisensor landuse classification.
The Bayesian network can be extended to incorporate spatial contextual information. For instance for the determination
390 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.