The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
450
<
o
100
99
98
97
96
95
94
93
92
| - QDC-Validation Data]
2 4 6 8 10 12 14 16 18 20
Number of bands
Figure 5 OA of experiment 1 for validation data
3.2.1 Accuracy assessment
To test the accuracy of the proposed method, the disjoint test
data set were used. We note that in the PSBS method only 40
percent of the training data set was used to train QDC, but once
the features were defined, we used the whole training data set to
train QDC. The classification results are shown in Figure 6 in
terms of OA for the independent test data set. PSBS was
conducted independently by K-means for each feature size. It
can be observed two consecutive feature sizes (number of bands)
will not necessarily result in similar classification accuracies.
However, the trend of OA curves shows the curse of
dimensionality: lower performances for larger feature sizes.
3.3 Comparison of PSBS with other BS methods
The performance of the PSBS approaches have been
compared with Sequential Floating Forward Selection
(SFFS), Sequential Floating Backward Selection (SFBS)
as feature selection methods. SFFS and SFBS were
performed based on the entire training data. OA, AA, and
the classification accuracies of classes corresponding to
maximum OA of three FS methods are shown in Table II.
Class Name
PSBS-
Validation
Data
SFFS
PSBS
SFBS
Hay-windrowed
100
100
99.59
99.59
Com-min
100
80.81
80.81
80.81
Soybean-clean
100
45.18
45.19
40
Grass/Trees
100
100
100
100
Com-min
100
96.97
95.96
97.98
Soybeans-no
till
100
91.66
96.06
93.294
Woods
100
100
97.66
97.51
Com-no till
83.33
48.15
59. 6
55.55 1
Soybeans-min
97.29
76.26
70.737
64.55
OA
98.38
84.75
84.79
82.38 1
|aa
97.85
82.12
82.84
81.031 j
Table II Classification accuracy of Three BS methods for
disjoint test data.
OA of the supervised PSBS approach with OA of the SFFS and
SFBS methods are shown in Figure 8. As illustrated, supervised
PSBS achieved higher OAs with respect to SFBS and SFFS for
almost all feature sizes, thus suggesting improved class
accuracies. The best result in terms of OA was provided by the
PSBS; however, the maximum overall accuracies given by
SFFS and the PSBS approach are very close together (i.e.,
84.75% for SFFS with 4 bands; 84.79% for supervised PSBS
with ten bands). Moreover, the maximum OA of PSBS for
small training sample size in comparison to SFFS is remarkable.
As expected, in the three BS methods the classification
accuracy tends to increase with increasing feature size until
maximum value is reached, but almost monotonically decreases
for larger feature values due to the curse of dimensionality.
Different classification accuracies can be observed by
comparing class to class in Table II. Particularly, the corn-no
till, soybeans-no till, and soybeans-min classes yielded different
accuracies for the PSBS, SFFS, and SFBS. This result may be
due to the use of different search strategies and PSBS as they
explore the bans subsets and prototype space in different ways.
Furthermore, as we pointed out in the supervised PSBS, the
training data set was split into two parts: one for the prototype
space generation and one for the resulting band subset
validation. However, the experimental results show that limited
training data for generating the prototype space is adequate to
achieve comparable results with SFFS. In contrast to the SFFS
and SFBS methods, which are based on the estimation of
scattering matrices in high dimension, the PSBS methods are
based only on the first statistic parameters. As a result, and as
expected, PSBS shows comparable performance with limited
training data set size.
Figure 6. OA of three BS methods for disjoint test data
4. CONCLUSION
In this paper, an innovative band selection method called PSBS
is proposed with two approaches for dimensionality reduction
of hyperspectral data based only on class spectra.
Compared to the traditional BS methods, in PSBS, search
strategies are substituted by K-means clustering to find relevant
bands in order to determine representative band of each cluster.
Moreover, instead of optimising separability criteria, overall
classification accuracy of a validation data set is used to decide
which disjoint optical regions yield maximum accuracy. From
the pattern recognition viewpoint, compared to conventional BS
which possibly examine all set of bands even neighbour bands,
in PSBS the relevant bands are distinguished as a group of
highly correlated bands and the highly correlated bands are
ignore to contribute in the BS process.
Supervised PSBS is assessed and compared with SFFS and
SFBS in terms of classification accuracy. Compared to the