1174
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
4. RESULTS
The edge features are computed based on the first fundamental
form of the multispectral bands. It is shown in Figure 3 that the
edge detection results are uninterrupted, and important edges
have been preserved, but there is much noise in the texture
region. So the texture marking is required to segmentation.
The multiscale texture features of each band are calculated from
the response of log Gabor bank filtering. Each band is filtered
separately using 2 octave bandwidth log Gabor filter bank over
6 orientations and 4 frequencies. The wavelength of the highest
frequency filters is 3 pixels. The scaling between successive
filters is 2. Thus, 4 scale texture features are produced for each
band. Then, the texture features of all bands are fused in scales
based on first fundamental form.
Figure 3. Edge features of multispectral bands
Figure 4. Texture marker in scale 2
For obtaining marker image for watershed segmentation, the
texture features are segmented using the moving threshold
algorithm, in which 50 pixels is used as the minimum region
size. Figure 4 shows the texture marker in scale 2, in which the
texture regions are marked distinctly.
Then, edge features are segmented based on texture-marked
watershed transform, which produce 4 scale segmentation
results. The segmentation result marking with texture features
in scale 2 is plotted in Figure 5. There are 104 regions in the
result. It is shown that the region boundaries are congruent with
most landscape objects. With the texture marking, the over
segmentation problem is solved so that the count of segments is
decreased to a meaningful number. But it is also can be seen
that there are some over-segmentation in severe texture areas
and under-segmentation between different colour areas.
Figure 5. Segmentation result in scale 2
RS (%)
Figure 6. Segmentation accuracy in different scales
A reference map, as shown overlaid in Figure 1, is produced for
quantitatively assess the segmentation accuracy in appropriate
scale. The count of reference polygons is 48. The RC and RS of
the segmentation results marking with different scale texture
features is shown in Figure 6. Scale 1 is corresponding to the
highest-frequency texture features, whereas scale 4 refers to the
lowest ones. It can be seen that the region count parameter RC
is increased with the texture marking from low frequency to
high. That is, the segmentation will become more fragmental if