Rural scene-B is a subset of pansharpened QuickBird image
with 0.6-meter resolution.
This paper takes
MS
and Fractal Net Evolution Approach
(FNEA) algorithm based multi-scale segmentation as two
examples for unsupervised segmentation evaluation. For
MS
based segmentaiton, this experiment only tests the unsu-
pervised segmentation evaluation with change of parameter
M
. For image-A, the spatial bandwidth
h
s
and the spectral
bandwidth
h
r
are respectively fixed as 16 and 5. For image-B,
the spatial bandwidth
h
s
and the spectral bandwidth
h
r
are
respectively fixed as 16 and 7. For
FNEA
based segmentaiton,
this experiment only tests the unsupervised segmentation
evaluation with change of Scale Parameter
(
SP
)
with shape
of 0.1 and compactness = 0.5. Figure 11 displays parts of the
segmentation results using the two multi-scale segmentation
algorithma with different parameters.
Figure 12 demonstrates the unsupervised segmentation
evaluation results by using
NV
,
NVCAR
,
and
NMI
.
There are some researchers that have weighted homogene-
ity and heterogeneity in unsupervised image segmentation
evaluation. Johnson and Xie (2011) gave the same weight (50
percent for both) to homogeneity and homogeneity, and Ming
et al
. (2012) set the weights of homogeneity and homogeneity
respectively of 0.6 and 0.4 because homogeneity plays more
important role in object based image classification. According
to (Ming
et al
., 2015), a good segmentation should synchro-
nously keep high homogeneity and heterogeneity no less
than 0.4, and considering the overall segmentation evaluation
score based on
NV
and
NMI
(Ming
et al
., 2012), the scale selec-
tion results are: for
MS
based segmentation, 40 and 100 are
respectively appropriate
M
parameter value for image-A and
image-B; for FENA based segmentation, 40 is appropriate
SP
parameter value for both Urban scene-A and Rural scene-B.
Figure.11 Segmentation results by
MS
and
FNEA
with differnt scale parameters: (a)~(g) Urban scene-A, (h)~(m) Rural scene-B.
642
October 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
