however, their average
JM
distances are among the lowest
(around 1.573) and thus all have a poor separability. This
implies residential classes are potentially more difficult to be
spectrally separated from others. Without any extra improve-
ment effort, the overall accuracy for the spectral classification
is very poor: 52.50%. The producer’s accuracy, user’s accu-
racy, and Khat for each
LULC
categories are reported in Table 3
Overall, water, and forest are well classified with their Khats
all higher than 0.85, indicating these are spectrally different
classes. The classification results of grass and commercial
and industrial are acceptable (Khat = 0.67). Despite its good
separability, cropland and pasture is mostly mixed with low
density residential, forest, and grass, and thus poorly classi-
fied like the three residential groups since their Khats are all
between 0.03 – 0.26. This suggests their characterized spectral
features alone are not enough for an acceptable image classifi-
cation and other information is thus needed to improve their
classification results.
The results from texture image evaluation show that the
optimal window determined by one criterion for
ML
image
classification needs not be optimal for another. Table 4 lists
the window sizes that best or better satisfy each of the nine
measures. Evidently, most listed window sizes can satisfy at
least two criterions of considerations except the 21 window.
Besides, some windows tend to meet more criterions than
others, for example, the 59 window satisfied seven criteria,
followed by the 47 and 55 windows that each met six crite-
ria. Consequently, only sixteen texture bands derived using
the following window sizes were used for the texture clas-
sification along with the three spectral bands: 15×15, 19×19,
21×21, 23×23, 27×27, 31×31, 33×33, 35×35, 37×37, 39×39,
41×41, 47×47, 49×49, 55×55, 59×59, and 61×61. Table 2
also lists the
JM
distances for all textural classifications. The
highest and the lowest overall JMs are reported by the 27
(1.864) and 39 (1.773) windows, respectively, which high-
light two major transitions in the overall separability in all
textural classifications. For windows of 15 – 27 and 41 – 59,
their overall
JM
distances are high (OAve >1.812) with most
of
LULC
types indicating the excellent/good separability and
also better than the spectral classification (1.806) mostly due
to the improved separability for commercial and industrial.
It hence suggests these may be ideal windows to produce
textural bands with good separation between different classes
of interest, which are likely to improve the classification
result. For windows of 31 – 39 and 61, however, with the
majority of classes only representing the acceptable/poor
separability, their overall
JM
distances are low (OAve <1.804)
and poorer than the spectral classification (1.806) mainly
because of the degraded separability for water and cropland
and pasture. It thus implies these windows are less useful to
create textural bands that would benefit image classification.
Among the eight
LULC
groups, only the three residential types
consistently report poor separability with low
JM
distances
across all window sizes (RAve = 1.598 – 1.673), with the low
density residential has the highest average
JM
value (1.657),
followed by the high density (1.637) and medium density
(1.609) residential. Nevertheless, they are also the only classes
produce higher
JM
distances by all windows than those from
the spectral classification (RAve = 1.573). This suggests the
fractal-based textural information may be the most valuable to
assist the classification of the residential classes, although its
contribution is still limited.
Overall, the resultant texturally classified images tend to be
very noisy especially those from the band combinations con-
sidering smaller window sizes. The overall accuracy and Khat
for individual textural image classification were summarized
in Table 5. Obviously, both the overall classification accuracy
and Khat change with band combinations based on different
texture bands, demonstrating a clear window size effect of
texture band creation on
ML
image classification. Generally
speaking, the textural classifications based on the fractal
information are inclined to result in very poor classification
accuracy. The best classification result is given by the 31 win-
dow with an overall accuracy of only 25.75% and an overall
Khat of 15.00%. The 27 window, while having the highest
overall
JM
distance, reports one of the lowest Khat (1.26%),
although its overall accuracy is the third highest among all
window sizes (23.50%). Figure 9 graphs the producer’s and
user’s accuracy and Khat for individual
LULC
categories for
all textural classifications. Generally speaking, the producer’s
accuracy tend to be higher than the user’s accuracy for
LULC
types that have a smooth texture, such as cropland and pas-
ture and water, throughout window sizes. Nevertheless, for
those roughly characterized
LULC
classes like commercial and
industrial, residential, forest, and grass, the user’s accuracy
is conversely higher than the producer’s. With an average
Khat of 20%, the fractal texture seems to be most useful in
classifying low density residential. Its contribution becomes
limited for classes like cropland and pasture, commercial and
industrial, medium density residential, and forest that report
an average Khats of around 11%. This texture feature is use-
less for water, high density residential, and grass that result in
near zero Khats. For certain
LULC
classes, their highest Khats
appear to be denoted by a specific window size. For example,
cropland and pasture and low density residential with the 31
window, water with 39, commercial and industrial with 19,
and medium density residential with 27. Yet except the one
for the medium residential class, none of these windows are
those where these classes’ highest
JM
distances are marked
(Table 2). Therefore, a good or excellent separability does not
always guarantee a good classification result. With varying
window sizes, both of cropland and pasture and low density
residential experience two sharp drops at the 31 –33 and
the 59 – 61 windows, with the former greater than the latter.
Besides, their Khats fluctuate more and are higher before the
31 window than those after. The medium density residen-
tial, however, behaves a bit more complicated, as its Khats
indicate three peaks with very similar values at 27, 47, and 59
windows where major changes are found. The Khats for the
rest five classes change little across window sizes, suggestion
a limited window effect on them.
Table 4. A list of window sizes that best or better meet all
criteria for the fractal-based texture bands (the superscript
indicates the number of criteria that are met by a given
window).
Best Texture
Bands
Better Texture
Bands
#
Correlation 15
4
, 21
1
, 27
2
31
3
, 47
6
, 55
6
, 59
7
, 61
3
1
Mutual
Information
59
37
4
, 39
5
, 41
6
, 43
4
, 45
4
,
47, 49
4
, 51
4
, 53
3
, 55, 57
4
2
Coefficient of
Variance
33
3
, 35
3
, 37, 39 15, 19
2
, 23
2
, 41, 43, 45, 47, 61
3
Skewness
33, 61
15,19, 23, 27, 35, 37,
39, 41, 43, 45, 47
4
Kurtosis
59
31, 39, 41, 43, 45, 47, 49, 55
5 Local Variance
33, 35, 37, 39,
41, 47
49, 51, 53, 55, 57, 59
6
Isarithmic
59
55
7 Triangular Prism 57
59
8
Variogram 13, 15
41, 51
9
Moron’s I
31
49, 51, 53, 55, 57, 59
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
703
