The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
1260
3. SEGMENTATION EXPERIMENT
3.1 Experiment data
To evaluate the performance of the proposed segmentation
approach, a multi-spectral QuickBird imagery at 2.44-m
resolution and a panchromatic QuickBird imagery at 0.61-m
resolution, which were acquired in May 2005 in HeFei city of
China were used. The area is about 1023 X 822 pixels and
represents a complex urban environment. The selected part of
the city is characterized by classes of road, highway, grass, and
building. Initially, the QuickBrid imagery were geometrically
corrected to the universal transverse Mercator (UTM)
projection, and re-sampled to 0.61-m to match the image pixel
size, and then fused by the Smoothing Filter-based Intensity
Modulation(SFIM) method using the CASM Imagelnfo®
remote sensing imagery processing software (CASM Imageinfo,
2007) developed by Chinese Academy of Surveying and
Mapping. The SFIM is a superior fusion technique for
improving spatial detail of multispectral images with their
spectral properties reliably preserved (Liu, 2000), and fusion
strategy is helpful for improving classification accuracy (Xu,
2004; Cao,2006). Figure 3(a) shows the panchromatic
QuickBird image, and figure 3(b) shows the multi-spectral
QuickBird composite image of band 4(infra-red), band 2(green),
and band l(blue). Figure 3(c) shows the fused image
compositing from the same bands as figure 3(b).
3.2 Segmentation Experiment
Initially, we segmented the fusion image by the improved SRM
segmentation algorithm to get the initial segmentation result on
the basis of the SRM software package. The key issue is trying
to adjust parameter until getting better initial segmentation
result, then these segmented imagery was vectorized in
coverage format using ERDAS imagine 8.7® (ERDAS imagine,
2003). Figure 3(d) shows the vector image overlaid on the
initial segmentation result.
Then, the objects from the initial segmentation were merged by
the MHR method, meanwhile the objects information such as
mean, boundary length, deviation is recomputed for latter
merging operation. Figure 3(e) and figure 3(f) show the course
scale imageries when the scale threshold is 40 and 60
respectively. Table 1 shows the detail parameters.
Segmentation
level
Shape
Scale Color J
W i m,oth ^CCMP! numbers
Level 1
Level 2
Level 3
8 0 0 0 702
40 0.6 0.3 0.7 374
60 0.5 0.2 0.8 185
Table 1. The parameters of the new method
In order to evaluate the performance of the proposed method,
we also segmented the fused QuickBird imagery using FNEA
that the eCognition soft adopted. The multi-scale segmentation
results are shown in figure 3(g), figure 3(h), figure 3(i), and
detail parameters are shown in table 2. Moreover, we carried on
multi-scale SRM segmentation, and the results are shown in
figure 3(j), figure 3 (k), figure 3 (1), and detail parameters are
shown in table 3.
Segmentation
level
Shape
Seal. Colo, — 0bje , C '
W >m>o,h w ccmp> numbers
Level 1
Level 2
Level 3
100 0.6 0.3 0.7 528
150 0.6 0.3 0.7 270
180 0.5 0.3 0.7 186
Table 2. The parameters of eCognition
Segmentation
level
Scale 0bi "*
numbers
Level 1
Level 2
Level 3
8 702
16 401
22 294
Table 3. The parameters of SRM
3.3 Accuracy Assessment
As shown in figure 3(g), we notice that the FNEA method
always divides the big homogeneity region into lots of small
regions with the similar size, especially highway. The limitation
may be resolved by merging the same classes, but it is based on
initial segmentation objects. When the scale becomes bigger,
there still has the phenomena shown in figure 3(f). This is
caused by the assumption that the objects with same scale have
similar size which is in consistent with nature phenomena. As
we know, building, road, grass and woodland belong to the
same level of land cover class, however, their sizes are different
greatly.
We also notice that there have small redundant objects shown in
figure 3(j). The redundant objects may be wiped off when the
scale becomes large, but there remains the limitation all the
same shown in figure 3(1).
However, the new proposed method overcomes these
limitations by SRM acting as initial segmentation. As shown in
figure 3(d), we notice that the highway is integrally detached
avoiding of smash objects. When the scale becomes bigger,
some classes such as highway, building, grass are easy to
extract, and there is less small objects shown in figure 3(f). The
bigger the scale is, the fewer the object numbers are, the bigger
the object region is, and the boundary of region may disappear
or remain.
Obviously, the new multi-scale segmentation method shows
advantages over traditional multi-scale SRM and FNEA
algorithm in the following aspects:
(1) It makes full use of shape, spectral, scale, local, global
information and the parameters could adjust for various
classes.
(2) It integrates the superiorities of SRM segmentation
method and MHR merging method, and avoids the
disadvantages of these methods.
(3) The segmentation result described both in vector and
raster format integrates RS and GIS expediently, and
establishes better foundation for higher level image
processing such as pattern recognition and automatic
image interpretation.