International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
The left side building shows a leaning effect whereas the right 
side building is straightened up and the hidden area is shown as 
hatched area. 
Considering the task to obtain information about the area of 
usable agricultural land, the leaning effect is a serious problem, 
especially at forest borders. 
  
Figure 4. Leaning effect in an orthophoto 
Figure 4 shows an orthophoto and the superimposed boundary 
of cadastral parcels. Without any further information such as 
orthophoto center position, there is no way to decide whether 
the image shows the real forest border or a leaning effect. Thus, 
this image might not be very helpful to determine Net Areas. 
The conclusions related to the leaning effect are the same as the 
position error caused by a height error. It should be noted that 
due the nature of line scanner systems, the leaning effect in 
flight direction is totally avoided. 
4. CLASSIFICATION 
In the context of Land Parcel Identification and Net Area 
determination, the land cover types to be taken into account in 
this study is focused on: 
e Arable Land 
e Forage Land 
° Forest 
e Water bodies 
e Infrastructure 
4.1 Classification with eCognition 
For the study of land cover classification, the software package 
eCognition was used. eCognition is based on the idea to carry 
out supervised classification based on "image objects" using 
concepts of fuzzy logic. The program supports a wide range of 
raster formats and has the possibility to process images of 
different resolution (Definiens 2001). 
4.2 Classification Case Study 
In order to proof, whether classification could be used for 
updating Land Parcel Identification Systems, color images taken 
in September 2002 had been used for image classification. 
Training samples for image classification were selected by the 
operator. However, the operator based identification of 
wasteland has been nearly impossible. 
  
  
  
   
   
  
  
  
  
   
= Arable 
e A-Typel 
A-Type2 
A-Type3 
A-Type4 
Forage 
@ F-Type2 
(S F-Typel 
@ FType3 
CO Maize 
=> @ Trees 
@ Single Trees 
@ Forest 
= @ Roads 
& R-Type2 
@ R-Typel 
-| € Houses 
© Eright-roof 
@ Dark-root 
@ Forest-Shadow 
Figure 5. Image used for classification with eCognition and 
class hierarchy (Oesterle 2003; Image by courtesy of the 
State Authority for Land Consolidation and Land 
Development Baden-Württemberg, Germany) 
Figure 5 shows clearly that arable as well.as forage land varies 
significantly in color and brightness. For this reason several 
subclasses for each land cover, as well as training samples have 
been defined. To introduce scale into segmentation, four levels 
from fine to coarse had been established. 
e Level 1 (fine): Single trees, infrastructure 
e Level 2 (medium): Infrastructure (houses) 
e Level 3 (medium): Arable, forage land 
e. Level4 (coarse): Forest areas 
A first classification step based on object features only was not 
satisfactory, as house roofs could not be detected. In a second 
step, classification has been improved by applying class related 
features. The classification results obtained on different levels 
have been merged afterwards. 
  
Figure 6. Final classification result 
The result was an 89% correct classification, found by 
comparing ground truth land cover acquired by field checks 
with the classification result. This result is a very good basis for 
updating the LPIS. Nevertheless, the remaining uncertainty 
requires checking the results by visual inspection. 
To prove these findings a second image from the same image 
flight was processed. The experiments have shown that 
adoptions of to training, including the class hierarchy have 
become necessary to get the excellent classification results 
shown above also in the second test site. The reason for this is 
the quite large variety of color and brightness values of forage 
and arable land in the image. Furthermore, the color of some 
Intern 
types 
buildir 
spectr: 
stabili 
43 U 
The cl 
detecti 
The cl 
work 
availal 
might 
superii 
/ 
| 
Figi 
The Ir 
shown 
the da 
that 
The de 
respon 
The p: 
Parcel 
recomi 
will be 
Ortho 
These 
The ac 
system 
have t« 
the sea 
taken 
disting 
for upx 
satellit 
proble