Figure 12. Effect of parameters on the classification performance for the Indian Pines
image: (a) effect of
λ
1
, (b) effect of
λ
2
, (c) effect of the number of training samples, and
(d) effect of the patch size. OA = overall accuracy; AA = average accuracy.
Figure 13. Sparse coefficients of the pixels in an image patch of the Indian Pines image:
(a) , (b) , (c) , and (d) . The pixels are from Class 6 and the corresponding training
samples are from 304 to 377.
was used and the parameters were
assigned as
λ
1
= 0.5 and
λ
2
= 0.05 in
the proposed
CSCSR
method. It can
be seen that for most classes, the
proposed
CSCSR
method outper-
forms the other methods and shows
the best performance.
The effects of the parameters
on the classification results of the
Indian Pines
HSI
data are shown
in Figure 12. As can be seen, the
proper setting of
λ
1
and
λ
2
has a
certain positive effect on the clas-
sification results, which are clearly
better than the results of other
methods for any parameter settings
within an appropriate range. When
the percentage of the training data
is varied as {5%, 10%, …, 50%},
the overall accuracy increases
from 98.15% to 99.78%. More
training data means a larger size
of the dictionary and an increased
number of samples from each class.
While this increase in training-data
size may generally improve clas-
sification accuracy, it may reduce
accuracy if testing samples are well
represented by different classes in
the dictionary. In this article, 10%
of pixels were used as training
samples for all methods.
For patch sizes varied as {1, 3,
5, 7, 9}, the overall accuracies are
respectively {86.82%, 99.30%,
99.14%, 98.68%, 98.11%}. When
the patch size is set as 1×1, the
classification accuracy degrades
clearly because there is no col-
laboration between the neighboring
pixels. As the patch size becomes
larger, the number of smoothed
sparse representation vectors
increases. From Figure 13, it can be
seen that more training samples of
different classes are activated when
the patch size is 9×9, resulting in a
decrease in classification accuracy.
The classification result of the 3×3
patch size is slightly better than the
results based on other patch sizes.
Experimental Results: University of Pavia
The second
HSI
was acquired by the
Reflective Optics System Imag-
ing Spectrometer, which has 115
spectral bands ranging from 0.43 to
0.86 μm. The image has only 103
spectral bands, a size of 610×340
pixels, a spatial resolution of 1.3 m,
and nine classes.
The numbers of training and
testing samples from each class are
shown in Table 3. The training and
testing sets are visually shown in
Figure 14a and 14b, respectively.
The classification accuracy for
each class, the overall accuracy, the
average accuracy, and the kappa
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2019
667
