International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
grey value difference, additionally considering a minimum size 
(cf. Figure 5, each grey value is one segmented field area). 
3.2.2 Deriving Missing Field Boundaries by Line 
Extraction and Grouping: The ficld arcas obtained from the 
segmentation step are only intermediate results. The reason is, 
that the case of identical vegetation in neighbouring fields — and 
therefore a missing boundary — is not taken into account. 
Accordingly, a line extraction (Steger 1998) is carried out 
within each field area to derive missing boundaries. The 
extracted short pieces of lines arc grouped to straight long lines 
in consideration of a minimum length due to the characteristics 
of the field boundaries. In addition, intersection points of the 
lines are calculated, if the end points of the corresponding lines 
have a minimum distance in between. Furthermore, the lines are 
extended to the boundaries of the field areas, if the distance lies 
again below a threshold. Results of the extracted lines arc 
depicted in Figure 5 in white. 
Further GIS knowledge referring to fixed field boundaries 
within the regions of interest (e.g. dead-end streets or tree rows) 
is introduced to support the extraction of field boundaries (cf. 
Figure 2, in Figure 5 the lines are depicted in black). The field 
areas are split by the extracted or additionally introduced lines 
yielding the preliminary field boundaries (cf. Figure 6, 
boundaries of the fields are depicted in black). 
3.2.3 Using Snakes to Improve the Geometric Quality of 
the Results: The field boundaries are in some part 
geometrically inaccurate, which is why a classical snake 
algorithm is used to perform the precise delineation. To 
initialize the processing, the preliminary field boundaries are 
taken. Additionally, most fields are four cornered polygons and 
this knowledge can be exploited by using the information about 
the four corners as well as the straight boundaries at the sides. 
Snakes were originally introduced in (Kass et al. 1988) as a 
mid-level algorithm which combines geometric and/or 
topologic constraints with the extraction of low-level features 
v 
  
Figure 3. Example for the measurement of the outline with a 
snake: Initialization is depicted in white, different 
optimization steps in black 
from images. The principal idea is to define a contour with the 
help of mechanical properties such as elasticity and rigidity 
(internal energy) to initialize this contour close to the boundary 
of the objects. In Figure3 an example is shown: The 
preliminary result of the field area is used to initialize the snake 
(depicted in white) and furthermore the processed different 
iteration steps are depicted to show the movement of the snake 
(black lines). The contour can be looked upon as a virtual 
rubber cord which can be used to detect valleys in a hilly 
landscape with the help of gravity. If the snake is initialized 
close enough to the valleys of the landscape, the gravity drags it 
into the valleys. The "landscape" may be a surface model, an 
image, or the edges of an image. The movement originates in a 
field of gradients, which can be computed on the base of an 
edge detector's result. 
The whole energy of the snake E,,4,. t0 be minimized, is the 
sum of the internal energy E,,, ,, and the external energy Eo 25 
defined in (Kass et al. 1988). The internal energy is described in 
the following in detail due to the speciality of the here 
presented work, given in equation (1): 
E Aer opes) o 
The application field boundary leads to a possibility to select 
special weight functions a(s) and p(s), which are used to control 
the elasticity and rigidly of the contour v(s.0); s is the arc length 
and 7 the iteration number. The weight functions have to lead to 
stiff edges along the expected straight lines of the field as well 
as to allow the snake to form corners. 
As external force the absolute values of the gradients of the 
given red channel of the imagery are used. Additionally, the 
boundaries of the region of interest and the further introduced 
knowledge within the region are manipulated to form “deep 
valleys” to steer the snake to this fixed boundaries (cf. 
section 2). 
3.3 Extraction of Wind Erosion Obstacles 
The strategy extracting wind erosion obstacles is divided into 
two parts, as described in section 3.1: The definition of buffers 
alongside GIS-objects as roads, rivers or railways allows a 
focussed view to narrow search areas. If high NDVI-values can 
be extracted within the search area, evidence of dense 
vegetation is given. The additional extraction of higher DSM- 
values than the surrounding area verifies the potential wind 
erosion obstacles. 
Objects not located alongside GIS-objects, have to be extracted 
without prior information about their location: In a course scale 
of the NDVI-image, a line extraction of high values is carried 
out as well as a line extraction of higher DSM-values than the 
surrounding area. In addition, the geometric model of the wind 
erosion obstacle has to be introduced processing the extracted 
short pieces to final objects in consideration of a minimum 
length, width and height. 
Another possibility to extract wind erosion obstacles is the use 
of texture. In addition, the extraction of single trees in a finer 
scale is an alternative, as for example presented in (Straub 
2003), and a subsequent linking of the resulting objects to lines, 
if they are side by side. 
1068 
Inte. 
— 
[n 1 
pres 
sele« 
regi 
preli 
cach 
supe 
furtl 
of tl 
reco, 
satis 
of tl 
ident 
initia 
these 
iterat 
Figur 
  
Figure