International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
149 
performed between objects from both scene descriptions; also 
intersecting objects inside one scene description are treated. 
The object identification codes and the classes of the original 
objects are stored in the new ones. The resulting new objects 
have now new hybrid classes, like ‘settlement from the DLM’ 
and ‘water from the segmentation’. An object has possibly up 
to 6 subclasses. Altogether 80 hybrid classes are possible. 
These new objects are geometrical disjoint and have an 
unambiguous attribute for their class. 
Therefore, the result after building of the disjoint objects is one 
new scene description, where the semantics of the produced 
disjoint objects are not limited to the topographic object classes 
'settlement', 'forest', 'water' or 'agriculture'. There are new 
hybrid classes consisting of two or more topographic classes. 
The spectral as well as the non-spectral features are extracted 
for the disjoint objects. This information is then used as the 
knowledge base in the next step of semantic classification. 
3.2. Semantic Classification 
Given that the problem of the geometry is treated in the 
previous section, only the landuse semantics of the objects are 
now of interest. The knowledge about the disjoint objects lies 
on a higher symbolic level; thus, a system for knowledge 
representation is used. 
Semantic networks are common schemes concerning 
knowledge representation. So far, semantic networks have been 
used in speech recognition (Kummert, 1992), industrial 
(Niemann et al., 1990a) and medical (Bunke, 1985) 
applications and aerial image analysis (Koch et al., 1997). The 
use of a semantic network for satellite image analysis is new. 
For using and modelling the knowledge about the disjoint 
objects - as well as serving as central control unit, ERNEST 
(Erlanger Semantisches Netzwerksystem) is applied in this 
research (Niemann et al., 1990b; Kummert et al., 1993). Based 
on the experiences with ERNEST in aerial image analysis at the 
IPF (Quint, 1997), a semantic net is designed for the 
classification process (see Fig. 14). 
In the semantic network system Ernest, there are three 
different types of links between two nodes: part-of (bst), 
specialization-of‘ (spez) and concrete-of (kon). The links 
describe the relation between two nodes. The nodes represent 
various objects, events, ideas, or abstract concepts. There are 
three different types of nodes: concept, modified concept and 
instance. At first, only concepts exist. During the analysis, 
modified concepts are distinguished from concepts only by 
more restricted ranges for their attribute values. When a 
modified concept has concrete values, it becomes an instance. 
A semantic network contains two different types of knowledge: 
declarative and procedural knowledge. Declarative knowledge 
consists of concepts and links, while procedural knowledge 
contains methods for determination of attributes of concepts, as 
well as for valuation of concepts and relations. The concept 
primitive is the interface to the database. It fetches unused and 
unclassified objects from the database and stores them in the 
concept object. The semantic net retains in the concept unused 
the object primitives that were not already used. Because 
topographic objects consist of one or more outer and inner 
objects, the contour is built in the concept contour. This 
contour object is classified because of its features to one of the 
semantic meanings. The process is repeated until all disjoint 
objects are classified. 
3.2.1. Data Analysis 
Because the scene consists of the four area-based object classes 
‘settlement’, ‘forest’, ‘water’ and ‘reject’, the learned features 
from Section 2 were stored in these concepts. It is assumed, that 
the ‘reject’ class is the same as the ’agriculture’ class, because 
most of the unspecified area in the DLM200 belongs to this 
class. This assumption simplifies the modelling of the concepts 
on the semantic level. In a future refining, the ‘reject’ concept 
could be split up in a concept for the agriculture class and a 
concept for objects who could not be classified with a minimum 
probability. 
In contrast to the automatically learned statistical class features, 
the valuation and analysis function have to be implemented by 
an operator. 
Fig. 14. Semantic net for the semantic classification. 
The strategy of analysis process is a general, problem 
independent strategy provided by the semantic network 
ERNEST. The analysis starts by creating a modified concept for 
a concept. A modified concept is a preliminary result and it 
reflects constraints for the concept that have been determined 
out of the context of the current analysis state. 
Following the top-down hierarchy in the semantic network, the 
concepts on lower hierarchical levels are each expanded 
stepwise until a concept on the lowest level is reached. Since