split into several short line segments what makes the middle 
point move along the line. The last argument is also true for 
the distance a which is, therefore, of quite low reliability. 
Basic strategy for matching: 
Two lists of line segments (here: one for the image, one for 
the cadastre) are built and sorted in such a way that the most 
reliable parameters (i.e. firstly 9, secondly M) are sorting cri- 
teria. According to 0, line segments of equal orientation are 
grouped while M represents spatial location. If the values of 
Ÿ of two line segments are similar the normal distance of the 
middle point of one line segment to the other line is com- 
puted and vice versa. Line segments with small distances are 
lying approximately on the same straight line and are tested 
for spatial overlap, i.e. if both line segments are projected to 
a common straight line, they must have an overlap. Figure 4 
shows that all lines k;, à — 0,---,3 match with / but not ks. 
To be more strict, one could require that the midpoint of at 
least one line segment must be within the region of the other 
segment or fix a specific amount of overlap in percentage of 
the shorter line segment. 
The output of this matching process is a m-to-n mapping, 
i.e. each line segment in one geometry may correspond to 
multiple line segments in the other one. This yields robust- 
ness against additional /missing line segments due to reasons 
discussed above. 
  
  
  
  
Figure 4: Types of matches for a line segment | 
4 RESULTS 
As discussed in section 2, our real world data set consists of 
a Landsat TM image portion and the accompanying cadas- 
tral borders (see figure 2(a)) which are roughly registered to 
the image. Line segments extracted from the image (see fig- 
ure 2(b)) are to be matched to the cadastral data by the 
algorithm presented in the previous section. 
Figure 5 shows the result of performing the matching algo- 
rithm. The cadastral line segments are shown in white and 
the corresponding image line segments in black. By compar- 
ing to the input data, it can be seen that numerous additional 
lines remain untreated and do not influence the quality of the 
overall result. Matches are found wherever proper lines of im- 
age and cadastre were available. In the upper centre region of 
the image no fields could be identified because of the very low 
image contrast. But at the lower centre region, for example, 
all the meaningful image lines are matched whilst others are 
correctly ignored, e.g., the one in the bright field and the one 
parallel to the field border but lower. Thus, the performance 
of the algorithm appears to be satisfactory. Nevertheless, 
problems can occur, if the quality of the preregistration is 
not sufficient. 
120 
  
Figure 5: Matching result: matched cadastral line segments 
are shown in white and corresponding image line segments in 
black 
5 CONCLUSION AND OUTLOOK 
In this work a method for matching line segments has been 
presented which is used for the fusion of cadastral boundaries 
and lines identified in a satellite image which has been en- 
hanced by spatial subpixel analysis. Lines are represented in 
a special form with the most reliable parameters used as the 
prior matching criteria. The method is robust to a large num- 
ber of extra line segments which do not have corresponding 
partners but is limited by the need for a preregistration. 
Based on the achieved matching result, in future work a local 
transformation for two corresponding line segments shall be 
determined which maps the cadastral information onto the 
image. Additionally, cadastral borders which do not corre- 
spond to line segments in the image can be projected into 
the geometry of an image according to the position of other, 
already transformed cadastral borders. Thus, complete rela- 
tions between cadastral information and land use can even- 
tually be established. 
The image-map-fusion can also be used to improve or fa- 
cilitate segmentation of Landsat TM images [7, 15], as the 
cadastral information indicates possible region borders. This 
task itself is an essential part of a recent project on remote 
sensing image understanding based on physical model inver- 
sion, discussed in [12][Schneider,1996]. 
REFERENCES 
[1] R. Bartl, M. Petrou, W.J. Christmas, and P.L. Palmer. 
On the automatic registration of cadastral maps and 
Landsat TM images. 1996. submitted to European Sym- 
posium on Satellite Remote Sensing lll, Taormina, Italy. 
[2] R. Bartl, A. Pinz, and W. Schneider. A framework for 
information fusion and an application to remote sens- 
ing. In S.J. Póppl and H. Handels, editors, Proc. 15th 
DAGM Symposium, Lübeck, Germany, Informatik ak- 
tuell, pages 313-320. Springer, 1993. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
  
I 
[1 
[1