If two objects can be merged, their respective attribute
values (sizes and sums) are added and stored in the table
entry of the object with the lowest object number. The
other object is removed from the table. Also the index
table is updated: the higher entry will point to the lower
one. Figure 3 shows the states of the index and object table
before and after processing the quadrant in Figure 4.
Figure 4: Segmentation at intermediate level
After the quadrant is finished, also the (new) values at
the outer boundaries are known. They are stored at the
next higher level of the stack, from where they will be re-
trieved when the next larger quadrant (containing this one)
gets processed.
When the entire quadtree is processed in this way, which
1s when the program reaches the highest level, the index
table is updated: All entries that have an object num-
ber associated with them are moved to the beginning of
the table; the pointers of all other entries are updated so
that they will point to the end of the chains. Then the
input quadtree is read again and the output (segmented)
quadtree is produced. Finally, an attribute table is created
from the object table, by transforming sums and sizes into
means and covariances. The attribute table is stored on
disk and can be used in subsequent analysis.
3.1 Iteration
Due to the recursive z-scan order, the process has a slight
tendency to create segments of regular shapes, according to
the quadrants. This effect could be completely removed by
making the process perform a few iterations, with increas-
ing threshold values. With one threshold value, the process
only merges, and because it works quadrant by quadrant,
it first attempts to merge within quadrants. When start-
ing with a lower threshold value than the final one, the
risk of inadvertedly merging sub-quadrants reduces. Irreg-
ular shapes will already be formed, however, and will be
the basis for further merging later, when higher threshold
values come into effect.
3.2 Small objects
The application of the image segmentation process to satel-
lite images causes a large amount of small segments (say,
254
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
less that five pixels in size) to be created. One reason may
be, of course, that due to the limited resolution of satel-
lite imagery, there are many of such small objects in the
terrain.
More important, however, is the effect of mixed pixels,
especially at the boundaries of objects with quite differ-
ent spectral signatures. In the feature space, those mixed
pixels are too far away from both objects, and therefore
they cannot be merged with one of them. The question
is what to do with them. From a segmentation point of
view, we would like them to be incorporated into larger
(neighboring) segments. 'To achieve this, we can relax the
merging criterion, by increasing the threshold value espe-
cially for small segments. However, the spectral values of
the boundary pixels will “contaminate” those of the entire
segment (unless we don’t update the values of the larger
segment when merger is due to criterion relaxation — this
was not investigated, however) and influence a later classi-
fication. Another possibility is to leave the small segments
(mixed pixels) out of the classification procedure and clas-
sify only the large ones. The above-described map calcu-
lation program can be used to make the selection of large
segments, based on the sizes in the attribute table. Under
the assumption that objects are relatively large, compared
to the pixel size, there is a slight preference for the second
option.
4 Experiment
The segmentation process was applied to a Landsat TM
image of the Flevopolder in the Netherlands. The “advant-
age" of this area 1s that there are large fields, so segment-
ation really makes sense. usually, Landsat TM does not
satisfy the previously stated condition that objects should
consist of a significant number of pixels. The method will
be more useful when higher resolution imagery becomes
available.
5 and 7. Using map
The results are shown in Figures
calculation, combining the segment quadtree with the at-
tribute table, only large segments were selected and a ran-
dom grey value was assigned to them. Small segments were
removed.
The image consists of 1000 x 1000 pixels. With a final
threshold value of 6, 180811 segments were created. Des-
pite the large objects in the terrain, many segments are
very small: 136870 single pixels and respectively 20053,
5252, 4009 and 2539 segments of two, three, four and five
pixels. Figure 5 shows small segments in black and reveals
that they are mostly boundary (mixed) pixels.
On the other side of the scale, there are four segments
with more than ten thousand pixels. They are water bod-
ies (IJsselmeer and Randmeren), with 11149, 33317, 44069
and 111375 pixels, respectively. The distribution of the
sizes of the more moderate objects 1s shown in Figure 6
Figure
played
than fi
180 —
160 fy
140 |
120 F
60 r
40 +
5 |
The p
metho:
Image
possib
attribu
bined
excess
dividir
respor
and G
ence n
