(a) original image 
  
(b) image in subpixel resolution 
Figure 1: Band 4 (near infrared) of Landsat TM 
problem and to obtain an improved image for line extraction. 
In this paper, a Landsat TM image of an area in the north- 
east of Vienna is used. For further processing the near in- 
frared channel is selected due to its high potential of crop 
discrimination and can be seen in figure 1(a). 
2.1 SPATIAL SUBPIXEL ANALYSIS 
In order to extract edges accurately, we apply spatial subpixel 
analysis as proposed by Schneider [11]. This approach has 
the capability to reduce the severe mixed pixel problem if the 
average size of regions of homogeneous spectral signatures is 
not much larger than the pixel size. 
The signature of a mixed pixel is composed of different sig- 
natures of two or possibly more adjacent regions. The border 
between these regions passes through a mixed pixel which is 
analysed within the context of its 8 neighbouring pixels. To 
split one mixed pixel, the parameters of the borderline are 
estimated. Possible parameters can be the orientation of the 
borderline, the normal distance of the line and the pixel to be 
analysed, and the spectral signatures. Subsequently, images 
of reduced pixel size are produced according to the derived 
parameters. 
The result of applying this method to our original Landsat im- 
age, to be seen in figure 1(a), is shown in figure 1(b). Wher- 
ever the spatial subpixel analysis was successful, each original 
mixed pixel is replaced by 9 smaller pixels with "cleaner" spec- 
tral signature. The improvement is obvious as long borders 
appear much more smooth and straight. Problems still occur 
at the corners of objects, e.g. at the right end of the bright 
field in the centre. 
2.2 EDGE DETECTION AND PERCEPTUAL LINES 
The field boundaries are extracted from the subpixel pro- 
cessed version of the near infrared channel (figure 1(b)) of our 
Landsat image. Some optimal edge detector filters [9] and 
the optimized Hough transform [6] provide the boundaries. 
The field edges extracted consist locally of several parallel 
lines, with the lines orthogonal to them missing due to the 
fact that they are made up from many small edgel strings, 
too noisy to be connected into straight lines by the Hough 
module. 
These missing boundaries are perceptual boundaries, per- 
ceived only due to the staggered nature of the long well de- 
fined field boundaries along the orthogonal direction. As it 
is unrealistic to expect an ordinary edge detector to identify 
them, a special algorithm for the identification of percep- 
tual lines has been developed. Details of this method are 
described in [1]. It is based on the idea that the more promi- 
nent edgels which have already been used for line extraction 
are suppressed so that the Hough transform can be applied 
once again to the reduced data set. Finally, the perceptual 
lines are found, too. In figure 2(b) all the identified lines, 
including the perceptual boundaries, are shown. 
Finally, a high number of image line segments are identified. 
A visual comparison shows that not all of them have corre- 
sponding cadastral borders. Some of them are artifacts, but 
most of them refer to "real" structures in the terrain, i.e. 
to several crop types within one parcel of land. Some other 
line segments are missing in the image since there is no grey 
value difference in the image due to the same crop type on 
neighbouring fields. 
118 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996