Besides, low density residential has the most change and thus
the greatest improvement, especially at the 31 – 33 windows,
followed by commercial and industrial, medium residential,
forest, and cropland and pasture that give a moderate varia-
tion. Subsequently, the use of fractal feature is most promis-
ing for classifying the group of low density residential.
The Fractal Texture’s Potential on Image Classification
When comparing to the spectral classification, most texture
classifications can increase the separability between indi-
vidual
LULC
classes, especially for commercial and industrial,
except at the windows of 31 – 39, and 61. Yet these improve-
ments do not guarantee a good classification results, suggest-
ing other information are needed in addition to considering
the separability condition among various classes for a suc-
cessful
ML
image classification. Generally speaking, with the
highest overall accuracy of only 25.75% and the best overall
Khat of only 15.00%, the fractal-based texture classification
is not satisfactory to map an urban environment when only
the
ETM+
image’s multispectral bands were applied and its
outputs are mostly reported to be much poorer than that by
the pure spectral classification. These results are also signifi-
cantly lower than those reported by Emerson
et al.
(2005) that
conducted the urban characterization of Atlanta, Georgia with
a similar approach (with the same dataset, the same
FD
algo-
rithm that computed using the same software, and the same
classifier), except that the fractal texture was extracted from a
panchromatic band using the 27 window, which was selected
by a visual interpretation, and less number of urban classes
were identified. In their research, they reported an overall
accuracy of 77.30% and a Khat of 69.9%. Among the eight
LULC
categories considered in this study, the fractal features
are found to be useless for classifying those that tend to be ho-
mogeneous and/or with a limited overall spatial coverage that
cannot occupy enough space in a moving window (e.g., water
and high density residential), heterogeneous and with a great
details that cannot be capture by image’s coarse resolution
(e.g., high density residential and commercial and industrial),
and/or extremely rich with spectral information (e.g., water,
grass, and forest), while they could make a limited contribu-
tion in identifying others that are both heterogeneous and
homogeneous (e.g., cropland and pasture). Nevertheless, the
texture information is found most applicable for classifying
medium and low density residential classes and the latter in
particular where the greatest enhancement of 74%, 20%, and
100% in producers’ accuracy, user’s accuracy, and Khat were
calculated as compared to the pure spectral classification.
This finding confirms Emerson
et al.
(2005)’s results that also
indicated the fractal information is particularly valuable in
delineating low density residential areas.
Overall, with an image having similar spatial, spectral,
and radiometric characteristics to the Landsat data is used
for urban classification by the
ML
classifier, the following
recommendations can be made regarding the window-based
fractal texture analysis: first, it is only suggested for classes
like medium- or low-density residential that is heterogeneous
and with a spatial extent equaling to the window sizes of
consideration. Second, there is no need to use window-based
texture analysis for classes like forest and grass since they can
be easily identified even by the traditional
ML
algorithm. Fi-
nally, the employment of this texture analysis for water, high
density residential, commercial, and industrial, and cropland
and pasture should be reserved, unless other data with higher
spatial, spectral, and radiometric resolutions and/or advance
classification algorithms (e.g., object-based, artificial neural
network, support vector machines, random forests, or deci-
sion tree algorithms) that offers the flexibility of extracting
textures differently are also involved in the process. Classify-
ing texture features by the
ML
logic is likely to suffer more
from the “mixed-texel” problem that is similar to the “mixed-
pixel” issue in the per-pixel classification, since the texture
features extracted from moving windows commonly contain
information for more than one class (Shaban and Dikshit,
2001). This is particularly true to classes that have a small
spatial extent (e.g., ponds, rivers, and roads) and even with a
heterogeneous nature (e.g., high density residential and com-
mercial and industrial), at class boundaries that are extremely
complex, and when a large window size is applied (Shaban
and Dikshit, 2001; Zhou and Lam, 2008). This so-called “edge
effect” is another well-known problem, in addition to the
window effect, with the window-dependent textural clas-
sification, leading to a poor-oriented classification for these
categories. Therefore, to better capture the local structures for
these
LULC
classes, one can consider using a window smaller
than those that are used for the medium- to low-residential
categories that often span enough spatially, which can be eas-
ily achieved by the use of high spatial resolution data. With
the
ML
classifier, previous studies have shown that the use of
the 15 m
ETM+
panchromatic band or even the 20 m
SPOT
data
could result in much promising outcomes with an overall
textural classification accuracy of 86% (Shaban and Dik-
shit, 2001) and 77.30% (Emerson
et al.
, 2005), respectively.
Other solutions may be to enhance the high-frequency local
variance of a given class by applying images with a coarse
radiometric resolution and/or by incorporating extra textural
features derived differently (e.g., using different spectral
bands, the same geospatial algorithm but with a different
setting, or different algorithms). Shaban and Dikshit (2001)
have reported that, the 4 bits data were better than 5, 6, and
7 bits for the
SPOT
image involved for texture extraction and
the use of two texture features and three spectral bands could
improve an overall accuracy result by 9% than a combination
of one texture feature and three spectral bands. Analyzing
cropland and pasture can be tricky as it seems to represent
a balance between spatial (regular pattern) and spectral (like
vegetation) and between homogenous and heterogeneous and
its local structure can thus share similar characteristics with
other classes varying from grass to commercial and industrial.
Probably because of this, the classification results of this land
use remain to be poor with either the textural or spectral clas-
sification in this research. As a result, if a texture analysis is
chosen for this category, the use of an advance classification
algorithm may be desirable. Examples include the object-
based texture that allows to extract textures from objects
whose irregular shape may better approximate this class’s
spatial feature at changing scales and the artificial neural net-
work, support vector machines, random forests, or decision
tree algorithms that are able to integrate various data sources
and can learn by itself to achieve the best result.
Conclusions
This paper provided a comprehensive examination of the ap-
plication of different moving windows in generating fractal-
based texture features using the triangular prism approach
and their potential on urban characterization by the
ML
super-
vised classification with a Landsat
ETM+
multispectral image.
The window size effect on the textural feature extraction and
the employment of these features in characterizing eight
LULC
categories were fully evaluated by a total of fourteen measures
performed at both preprocessing and post-processing stages of
image analysis and interpretation. Overall, all measures used
prior to classification are found to be useful to filter out fractal
features that could facilitate the following
ML
classification.
Among these measures, the fractal analysis, and Moran’s I in-
dex had rarely been applied in previous literature. This study
shows they can also provide indicative information about the
quality of texture images, and hence can serve as alternatives
708
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
