The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
The segmentation results in Figure 6 represent the real cropland 
segments in the GIS object much better than those in Figure 5b. 
Most of the small disturbances could be eliminated. In the 
uppermost segment there are still a lot of disturbances caused 
by rows of trees that indicate a land use as a small orchard 
(class special cultures in ATKIS). Despite being small, these 
objects were not merged with their surrounding segments 
because they were bordered by more than one segment. In an 
ATKIS cropland object, special cultures objects are permitted 
as long as they not exceed a size threshold. For this reason it is 
important for the application not to merge such non-cropland 
objects with their surrounding cropland areas, because 
otherwise the next analysis step might reject the larger segment 
as a cropland object even though it is consistent with ATKIS 
specifications for that class. 
a) 
Figure 5. a) Watershed segmentation of the image in Figure 1 
using a smoothing scale of . = 1. b) Segmentation 
results after region merging. 
Figure 6. Segmentation results after removing segments having 
only one neighbour and being smaller than 1000 nT. 
Other disturbances are located near the object border due to a 
changing tilling direction. These structures caused by turning 
agricultural machines would disturb the verification algorithm 
and are thus excluded from the analysis of the predominant 
edge direction (Helmholz et al., 2007). For that reason, the 
segmentation can also be restricted not to consider areas close 
to the object boundaries. An example for the influence of this 
restriction on the segmentation results is shown in Figure 7. 
Most of the disturbances close to the object boundary could be 
eliminated. The few remaining ones have no influence on the 
next analysis step due to their small size. Of course, the small 
segments corresponding to the orchard in the uppermost part of 
the GIS object remain. The two management units at the lower 
end of the GIS object are both split into two parts. This is 
caused by slightly different reflectance properties of these areas 
due to different soil characteristics. However, all of these 
segments are large enough for the verification step to detect a 
sufficient number of parallel lines for success. 
The label image in Figure 7 is the basis for the analysis of 
parallel lines that is used for the verification of the original 
cropland object. The verification algorithm is applied to each of 
the segments in Figure 7 exceeding a certain size rather than to 
the whole area corresponding to the ATKIS cropland object. 
After that the individual results of verification are merged in a 
final analysis step taking into account the definitions of ATKIS 
for the representation of cropland objects. Compared to the 
original algorithm (Helmholz et al., 2007) this is expected to 
lead to better results because the smaller segments should be 
more uniform in their main tilling direction. 
Figure 7. Results of segmentation if regions near to the object 
border are not considered. 
Figure 8 and Figure 9 show two more examples for ATKIS 
cropland objects that are taken from the same IKONOS scene 
as Figure 1. Note that the ATKIS cropland object Figure 8 
consists of only one management unit, whereas both in Figure 1 
and in Figure 9 there are multiple units. It is obvious that the 
segmentation algorithm detects homogeneous image regions 
that do not necessarily coincide with management units, 
because the algorithm is affected by characteristics of the soil 
such as humidity or soil material. Typical examples are the field 
at the bottom of Figure 7, the area in the left of the field in 
Figure 8, and the field in the middle of Figure 9. The different 
reflectance properties of the soil are the main reason for 
remaining small disturbances. The number of these disturbances 
is of course higher when the smoothness parameter of the 
Watershed algorithm is lower (compare Figure 8b and c). If a 
field is thus split into segments that are large enough for the 
following analysis to succeed (bottom management unit in 
Figure 7), the appearance of different soil characteristic has no 
influence on the application. This is also true if the major part 
of a segment is correctly extracted and the remaining 
disturbances are small enough to be disregarded in the 
following analysis (the management units in the middle of 
Figure 7). However, the management unit in the middle of 
Figure 9 is split into too many small segments, which would 
prevent the verification algorithm from correctly classifying 
that region. Also the large segment in the left part of Figure 8c 
would not be verified correctly despite being too large for being 
discarded: no parallel straight lines are detected (cf. Figure 8d). 
However, in the case Figure 8c, the largest part of the ATKIS 
object would be verified correctly. The distance metric between 
the large segment and the disturbing object suggests that they 
could be merged, but the measure Ty for the strength of the 
boundary between them prevents the algorithm from merging 
them. It might be possible to consider this when the 
classification results of the individual are merged: if an object 
that could not be verified is surrounded entirely by a verified 
389