ated. One reason may
ed resolution of satel-
h small objects in the
effect of mixed pixels,
ects with quite differ-
we space, those mixed
objects, and therefore
f them. The question
segmentation point of
corporated into larger
this, we can relax the
threshold value espe-
the spectral values of
we” those of the entire
1e values of the larger
rion relaxation — this
influence a later classi-
ive the small segments
on procedure and clas-
-described map calcu-
e the selection of large
attribute table. Under
tively large, compared
~ference for the second
led to a Landsat TM
erlands. The “advant-
rge fields, so segment-
Landsat TM does not
on that objects should
xels. The method will
tion imagery becomes
| 5 and 7. Using map
quadtree with the at-
ere selected and a ran-
|. Small segments were
)0 pixels. With a final
its were created. Des-
n, many segments are
nd respectively 20053,
ro, three, four and five
ts in black and reveals
ed) pixels.
iere are four segments
They are water bod-
th 11149, 33317, 44069
he distribution of the
is shown in Figure 6
na 1996
Figure 5: Detail of segmented image. Objects are dis-
€ o eo J
played with random grey values, those that are smaller
than five pixels are black
180 T T T T
number of objects as function of object size ——
140
120 + | zl
il
|
60 Fr
Î
40 F W, d
I p
20 Fr V q
Wis
0 i L 1 AAA AN]
50 100 150 200 250 300
Figure 6: Distribution of object sizes
5 Conclusions
The paper gives a description of an image segmentation
method. which is embedded in a quadtree based GIS and
Image Processing system. Generally, the system gives the
possibility to integrate remote sensing data, map data and
attribute data. It offers raster processing capabilities, com-
bined with high resolutions of maps and images, without
excessive storage and processing requirements. By sub-
dividing an image into segments, assuming that these cor-
respond to objects in the terrain, the integration of RS.
and GIS can be strengthened. However, the correspond-
ence must probably be further established using classifica-
tion procedures.
A new aspect in the segmentation process is that it
is based on multi-spectral image data. The difference
between the proposed method and the existing grey-scale
segmentation methods is that in the second case different
results are obtained from the individual bands, which must
be later combined by overlaying. This comparable to the
difference between parallelepiped (box) classification and
minimum (Euclidean or Mahalanobis) distance classifier -
the latter are usually superior.
For the time being, the merging criterion is a simple one,
based on mean spectral values and covariance matrices. In
the introduction of this paper, we said that the human vis-
ion system tends to segment the image first and to classify
the segments later. Probably, for trained image interpret-
ers, it is more realistic to assume that they do both at the
same time. In the near future, we will therefore incorpor-
ate training statistics in the criterion: subsegments will be
merged if the resulting spectral signature still fits to one of
the sample distributions.
References
[1] Y.-L. Chang and X. Li, "Fast image region growing",
Image and Vision Computing, Vol. 13, pp.559-571,
1995.
[2] B.G.H. Gorte, Experimental quadtree software, Tech-
nical report, ITC, Enschede, 1995.
[3] S.L. Horovwitz and T. Pavlidis "Picture segmentation
by a tree traversal algorithm”, J.ACM, Vol.23, pp.368-
388, 1976.
[4] P.M.Mather, Computer processing of remotely-sensed
images, an introduction, Wiley, 1987.
[5] O.J. Morris, M. deJ. Lee, A.G. Constantinides, ” Graph
theory for image analysis: an approach based on the
shortest spanning tree”, IEE Proc.-Part F. Vol. 133,
pp.146-152, 1986.
[6] H. Samet, The design and analysis of spatial data
structures, Addison- Wesley, 1990.
[7] J.A. Richards, Remote Sensing Digital Image Pro-
cessing, an introduction, Springer- Verlag, 1993.
255
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
