Effect of the Scene Size
In the above experiments, 60 m was
empirically chosen as the scene
size, and the accuracies, as well as
the visual results, can be regarded
as satisfactory. However, the scene
size can have a significant effect on
the results of the proposed scene-
based approach. A small scene size
can preserve more details, but it
is insufficient to characterize the
spatial pattern of the tea gardens.
On the other hand, as exhibited
in Figure 12, the map with a large
scene size (80 m) shows a better
accuracy than the one with a small
scene size (40 m), but it loses more
information. In order to investigate
the effect of scene size in terms of
both accuracy and detail preserva-
tion, additional experiments were
conducted with a series of scene
sizes: 40 m, 50 m, 60 m, 70 m, and
80 m. More specifically, the best
Kappa value was used to represent
the detection accuracy, and mutual
information entropy (Susaki
et al.
,
2014; Huang
et al.
, 2017), which
measures the dissimilarity of the
entropy between the two images,
was employed to evaluate the detail
loss. Since the map with the small-
est scene size (40 m) lost the least
details, it was selected as the bench-
mark to calculate the mutual infor-
mation entropy. Figure 13 shows
the results of
BOVW
and the
UCNN
in dataset2. As expected, the larger
scene size results in a higher Kappa
value, but a lower mutual informa-
tion entropy (loss of details). It can
be seen that the Kappa value in-
creases slowly as the scene size in-
creases, but the mutual information
Figure 7. The example patches of the classification maps: (a) Raw image, (b) Ground
truth, (c)
BOVW
(spectral-textural), (d)
sLDA
(spectral-textural), and (e)
UCNN
(spectral).
Figure 8. Comparison between the Kappa values of the
different features.
Figure 9. Comparison between the classification maps of
the different features in dataset 4, (d) is the raw image of the
example marked by a rectangle in (e) and (f).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
729
