[S. 
IS 
id 
al 
an 
IS 
IS 
IS 
are 
is 
Roeland de Kok 
the landscape project, build around a set of image and thematic landscape data. The (image-) objects are embedded in a 
multi-level landscape matrix which allows both an Eco-topological description of the objects as well as a modeling of 
the chorical dimensions (Leser, 1997) of the landscape. The way certain layers are build into the landscape model is very 
much depending upon the origin of the data-type, such as sensor based mapping, land-surveying point data, 
administrative boundaries etc. The layered multi-scale landscape description is very much in line with standard GIS. 
The huge pre advantage of eCognition above standard GIS is the availability of automatic extracted spatial objects 
available from automatic multi-level segmentation analysis. 
3.2 The ‘project’ in Landscape analysis 
Each analysis session starts with the construction of a ‘Project’. The user has to define in advance the objects of interest 
and the thematic layers needed to construct a landscape that is defined through inter-linked ‘landscape objects’. These 
objects represent land cover surfaces and have at least a minimal size and a unique position linked to layer. This exclude 
point and line objects, these should be defined as surfaces with a minimal dimension. Analyzing a landscape model 
using ‘objects’ is partly using modern concepts from cartography as well as object-oriented approaches in computer 
science, this allows possible confusion ( after Molenaar.....). Although introducing new vocabulary is only useful when 
generally accepted, it fits very well to describe objects from the initial segmentation round in terms of image object 
primitives. After a cycle of classification and segmentation, the resulting output delivers objects of interest and the 
database table output that belongs to it. The introduction of expert knowledge is essential in three different phases of the 
analysis. The most complicated one is the construction of the semantic network that depends very much upon the 
sequence of segmentation and the basic spatial objects desired. The second important phase is the definition of classes 
per object layer. The third step is the construction of the fuzzy logic decision curves for each class. If raw image data 
needs to be classified, nearest neighbor or fuzzy logic decision rules can be applied for object classification. If required, 
fuzzy logic decision curves can be constructed on the basis of training areas. If for particular classification, the fuzzy 
decision curves are known, no training areas are required. To deal with modern developments in image analysis, the 
following combination applied in the eCognition software package, proved to be a successful one: 
e An advanced segmentation algorithm, that is linked to an object oriented database. Through this link, the output of 
the multi level segmentation is embedded in a hierarchical semantic network. 
e One single database for GIS and (satellite) image information to guarantee full synergy. 
* Query of the database through fuzzy logic decision curves, allowing the formalization of the expert knowledge. 
e Full input format facilities through ‘PCI-Geogateway and ASCILBMP and TIF output facilities, allowing the 
export of raster based objects with their table link. 
4 ACASE STUDY; SELECTED EXAMPLES 
To give more insight in the classification procedure with an object hierarchy, several typical examples from a 
classification session in mountainous forest area have been selected. The imagery used in this case study is similar to 
actual satellite data and has a direct link to the forest GIS available. The basic object in the forest GIS is the forest stand. 
This is an area which is defined by homogeneous treatment and therefore an administrative boundary. This condition 
does not mean that the surface within a forest stand has also a homogeneous spectral/textural response. This problem is 
widespread in forest GIS and deviates from the per-field analysis in agricultural domain such as used in the studies of 
Janssen (1994) or van Leeuwen (1996) Therefore, at first, the attributes from the GIS layers are important but not the 
polygon definition. The area descriprion is needed in a further stage of the GIS analysis. (relationship to sub-objects). 
The resulting objects of interest can be easily integrated as vector data into the forest GIS under ArcInfo/ArcView. 
However, the strategy is to reach an advanced level of analysis with image objects using the eCognition software 
package, before continuing further GIS analysis using administrative forest stand boundaries in standard GIS software. 
4.1 A GIS synergy with 5 Meter panchromatic and SPOT multi-spectral data 
Remote sensing using 5 (to 10) meter panchromatic data and multi-spectral imagery around 20 meters are the major 
part of the mainstream satellite imagery from past and for coming decades. Landsat, SPOT, MOMS and IRS can be 
seen as major workhorses and their sensor specifications have been proven to be effective in median mapping scale (1 
:100.000 up to 1:50.000). The quality, quantity, temporal availability and familiarity of this data type among remote 
sensing specialists, assures that new data products are generally compared to standard products from this sensor type. 
Although the 5 meter panchromatic band in this case study is derived from a mosaic of 15 orthophoto’s with 80 cm 
resolution, the way this data is handled shows very well the possibilities of an object oriented analysis applied to 
panchromatic data ranging from 5 to 15 meter and multi spectral data from 4 to 30 meters pixel resolution. 
The first possibility of introducing expert knowledge is the selection of the different aggregation/segmentation levels. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 225