between the high density residential zoning map and the
reference air photo of the study area indicates this land use is
primarily covered by high-rise buildings, which is potentially
difficult to be located with the
ETM+
image’ 30 m resolution
even when the texture information was utilized. The results of
cropland and pasture from the two types of classification are
also unsatisfying. The different levels of human cultivation
activities (from intensive to extensive) occurred on the land,
such as field crops, grain crops, fallow crops, and orchards,
could induce the possibility of this class sharing similar
spectral and/or spatial properties with others varying from
grass to commercial and industrial. Consequently, this class
is inherently difficult to be isolated from other classes either
spectrally or spatially with the
ML
classifier.
The pairwise test for the significance of Khat was carried
out on the error matrices from the accuracy assessment for
each classification to determine whether the classification
performed with spectral or textural features is significantly
better than the other (Table 6). Overall, when different win-
dow sizes are considered in the textural classification, most
of their classification results are statistically similar with each
other. The only exception is the one relied on the 31 window
as it is prone to be statistically significantly different from
most of other windows (33, 35, 37, 41, 47, 55, and 61). There-
fore, this window size is the most suitable, among all window
sizes that were examined, to create the best fractal-based
texture band that could disclose the major underlying fractal
change of the scene and could facilitate the
ML
logic image
interpretation. However, with all Z scores greater than 1.960,
it is clear that the spectral classification is still considerably
better than any of the textural classification used in the study.
Discussion
By using the triangular prism-based
FD
features, this study
highlights several issues important in the window-based tex-
ture analysis of remotely sensed images. The successful use of
texture features in traditional image classification such as the
ML
classifier, when no additional data are employed, solely
depends on the quality of extracted texture feature that must
consider the nature of landscape, the image’s characteristics
(e.g., spatial, spectral, and radiometric resolutions), the tex-
ture feature extraction (e.g., shape and size of the analytical
unit and the use of specific algorithm), the image analysis and
interpretation, and the interaction between them. To best un-
derstand these factors, one must realize that texture features
must be extracted over a local area that can vary by shape
and size up to the user. When the traditional moving win-
dow with a fixed shape is used, the selection of the optimal
size becomes extremely critical. The results reported in the
current study well illustrate a strong window size effect on
the quality of the derived texture bands and their potential in
the
ML
classification of an urban environment with a total of
twenty-five window sizes.
The Window Effect on Fractal Texture Extraction
The window effect was thoroughly accessed by the use of
a number of measurements at two stages of digital image
processing. At the preprocessing stage, a total of nine assess-
ments were selected to examine the quality of twenty-five
texture features. Among them, some could help to understand
the extracted features’ relationship with spectral bands (cor-
relation analysis, mutual information, variance, skewness,
and kurtosis). Others were useful to quantify the extracted
features’ own properties described as the amount of spatial
content (fractal analysis), whether its surface tends to be
homogeneous or heterogeneous (Moran’s I), or how well it
may approximate the size and condition of the fractal texture
(local variance). At the postprocessing stage, an additional six
measures were also employed to further assess the potential
of the selected sixteen fractal-based texture features on the
ML
classification by considering separability (
JM
distance), accu-
racy assessment (producer’s accuracy, user’s accuracy, overall
accuracy, Khat), and statistical significance (the Z test of Khat)
for each textural classification and amongst classes.
Overall, among all windows of interests, because different
measures evaluated the texture features from various perspec-
tives, the optimal window size that best satisfied one mea-
sure needs not be the one that well met another; the optimal
window for identifying one
LULC
class may not be the one for
another class and for all classes. Nevertheless, based on the
results from all measures being used, certain windows are
highlighted that seem to indicate several major transitions
each corresponding to a critical phase a fractal texture may
evolve through from small to large windows used in feature
extraction, for instance, the windows of 27 – 31, 47 – 49, and
59 – 61. Specifically, these windows marked the beginning of
Table 6. Pairwise comparison of kappa values from the fractal-based texture classifications of all window sizes and the spectral
classification (
GRNIR
) (values more than 1.960 indicating a significant difference at the 95.00% confidence level are highlighted in bold).
Z 19 21 23 27 31 33 35 37 39 41 47 49 55 59 61 Ave GRNIR
15 0.823 0.421 0.249 0.816 1.550 0.843 0.420 1.194 0.335 1.196 0.592 0.084 0.761 1.069 0.843 0.747
9.673
19
1.266 0.577 0.000 0.732 1.685 1.262
2.043
1.171
2.046
1.435 0.744 1.604 0.246 1.685 1.178
8.886
21
0.678 1.255
2.009
0.436 0.000 0.794 0.085 0.795 0.177 0.511 0.351 1.518 0.436 0.696
10.300
23
0.572 1.307 1.101 0.676 1.455 0.588 1.457 0.849 0.166 1.019 0.824 1.101 0.926
9.463
27
0.726 1.671 1.251
2.025
1.161
2.028
1.423 0.738 1.590 0.244 1.671 1.321
8.822
31
2.423 2.003 2.783
1.907
2.788 2.175
1.476
2.343
0.486
2.423 2.081 8.138
33
0.434 0.353 0.515 0.354 0.259 0.935 0.085 1.936 0.000 0.541
10.673
35
0.791 0.084 0.792 0.176 0.509 0.350 1.513 0.434 0.581
10.272
37
0.870 0.000 0.614 1.289 0.439
2.295
0.353 0.837
11.073
39
0.871 0.259 0.422 0.431 1.420 0.515 0.653
10.131
41
0.615 1.292 0.440
2.299
0.354 1.000
11.090
47
0.682 0.174 1.686 0.259 0.700
10.443
49
0.853 0.991 0.935 0.926
9.648
55
1.855 0.085 0.970
10.604
59
1.936 1.354
8.643
61
0.869
10.673
*Ave: the average z scores per window size.
706
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
