ina 
na 
lization. 
obstacles 
>s can be 
xtraction. 
ng wind 
0 hedges 
Different 
tion, are 
by their 
nshine. 
3). Prior 
le open 
) extract 
potential 
ed work 
ion and 
terest in 
n field 
of wind 
rwards, 
'stacles, 
*b, line 
will be 
d wind 
on 4 to 
Finally, 
d wind 
gery as 
IS-data 
ape of 
Ige and 
e from 
w. The 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
non-vegetation areas in CIR images are removed by NDVI and 
CIE L*a*b. Afterwards, edges of tree rows and hedges are 
extracted with Canny edge extraction algorithm followed by line 
linking, grouping and matching. Lines belong to non-interested 
regions such as urban, forests in the stereo imagery are masked 
out by GIS-data. DSM is also integrated into line grouping 
approach because of usually short distance and low contrast in 
NDVI and CIE L*a*b information of pixels between hedge and 
tree row. Then the matched lines are projected onto the landscape 
with known camera parameters. 3D information provided by 
DSM is used to verify the potential wind erosion obstacles since 
they are always higher than the landscape. Finally, the objects of 
interest, wind erosion obstacles, are described by their 
characteristics and appearance in an overall context with other 
neighboring and influencing objects. 
2.0 Data sources 
~ 
The GIS data with accuracy of about 3 m consists of scene 
description of the German ATKIS DLMBasis (Authoritative 
Topographic Cartographic Information System, basic digital 
landscape model) (Butenuth 2003). Since only wind erosion 
obstacles are of interest, regions where no wind erosion obstacles 
exist (e.g. urban, water, forest) in the imagery can be masked out 
by the available GIS-data. Furthermore, the GIS-objects road, 
river and railway represent the approximate geometric position of 
parts of the field boundaries. They also represent potential search 
areas for wind erosion obstacles, which are usually located 
parallel and near to them. Figure 1 shows the GIS data 
superimposed on the aerial image in the open landscape. Roads 
and field boundaries are depicted in yellow, buildings in white 
and forests in green. A representative region of interest is 
represented in dashed white lines, separately shown in figure 2. 
  
Fig. 1. Open landscape with superimposed GIS data 
CIR images (with ground resolution of 0.5m in this paper) are 
generated in early autumn when the vegetation is in an advanced 
period of growth. The color is almost fully green for wind erosion 
785 
obstacles, while for example light yellow for crops. This 
information is advantageous for automatic vegetation extraction. 
Therefore, the color space RGB, which presents the raw stereo 
CIR images, is transformed into a device type independent color 
space CIE L*a*b since it is powerful in image segmentation. The 
CIR aerial image is classified into vegetation and non-vegetation 
regions. Furthermore, classifying with NDVI is also possible. The 
two approaches are both used to segment the images into 
vegetation and non-vegetation areas. Of course, there are other 
objects than wind erosion obstacles, which will appear in green 
color such as grassland. That means GIS-data and CIR imagery 
are not enough to extract wind erosion obstacles. 
  
Fig.2 Selected region of interest 
  
Fig. 3. DSM superimposed with orthoimage 
Additionally, corresponding DSM with 0.5 m ground resolution 
of the interested open landscape is obtained with VirtuoZo, which 
is developed by Supresoft Inc. The DSM is not precise because