minimum number of samples in any cluster (set to 20), the
minimum distance between clusters (set to 5), the minimum
deviation from mean (set to 1) and the maximum distance
from mean of clusters (set to 100). The initial parameters of
ISODATA
algorithm are obtained by trial and error techniques.
Experimental Results
To demonstrate the efficiency of the proposed method for
dimension reduction of the features set in image clustering, all
features (i.e., 227 features) are generated, and K-means cluster-
ing is applied in full dimension and the reduced dimension
(i.e., 62 features) achieved by the proposed method. The results
for DS1 (i.e., labeled image) is shown in Figure 7. By comparing
Figure 7 and Figure 3, it can be seen that segments, especially
in “bared land” and “crops” classes, achieving in reduced di-
mension by proposed
FS
method is much better than clustering
in full dimension space. It is worthy to note that the focus of
this study is to select the appropriate features that provide bet-
ter discrimination between clusters through clustering algo-
rithms and obtaining the final segments (i.e., semantic objects)
is beyond scope of this research. In other words, the results
with lower over-segmentation (much more segments compared
to manually labeled image) and better separation between clus-
ters are the main goal of this work. As shown in Figure 7a and
7b, over-segmentation caused by the proposed
FS
method is less
than that of result by using all the 227 features (Figure 7b) due
to exclusive redundant features. It must be noted that a differ-
ent legend obtained in both figures is due to random selection
of initial cluster in the K-means clustering algorithm.
The performance of the proposed
FS
method is compared
with the
FS
methods summarized in Table 6, in the same
(a)
(b)
(c)
(d)
Figure 5. Datasets of QuickBird images: (a) DS1: labeled dataset, (b) DS2, (c) DS3, and (d) DS4.
(a)
(b)
(c)
Figure 6. Datasets of GeoEye images: (a) DS5, (b) DS6, and (c) DS7.
(a)
(b)
Figure 7. Results of K-means clustering algorithm on DS1: (a) Clustering with the proposed FS method (i.e., with 62 selected features),
and (b) Clustering without FS (i.e., with 227 features).
218
March 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
