1gust
1dsat
atial
ation
m, a
omy
test
ND
[M+
ction
0. In
tural
other
d on
total
order
have
were
cally
CPs
with
the
r the
arest
MS
e to
1sed
very
adar
the
the
). In
such
Iters
the
rest
ach
the
well
tput
n of
the
ning
s to
… In
sed.
and
g an
vere
"nce
x to
specified neighbour (ENVI 1999). The contrast measure
indicates how most elements do not lie on the main diagonal,
whereas, the dissimilarity measure indicates how different the
elements of the co-occurrence matrix are from each other (Lee
et al. 2004). By applying these measures, initially 12 features
have been derived, but after thorough checking of each
individual feature only 2 features, including the result of the
contrast measure applied to the green band and the result of the
dissimilarity measure applied to the near infrared band of
Landsat, were selected.
To define the sites for the training signature selection from the
images, two to three areas of interest (AOT) representing the
selected classes (i.e., coniferous forest, deciduous forest,
grassland, light soil, dark soil and water) have been selected.
As the data sources included both optical and SAR features,
the fused images were very useful for the determination of the
homogeneous AOI as well as for the initial intelligent guess of
the training sites. The separability of the training signatures
was firstly checked in feature space and then evaluated using
transformed divergence (Mather and Koch 2010). After the
investigation, the samples that demonstrated the greatest
separability were chosen to form the final signatures. The final
signatures included about 563-992 pixels. For the
classification, the following feature combinations were used:
1. The original spectral bands of the multitemporal Landsat
data.
2. PALSAR and original bands of the multitemporal Landsat
data.
3. Multiple bands including the original PALSAR and
Landsat images as well as two other derivative bands
obtained from texture measures.
For the actual classification, a standard maximum likelihood
classification has been used assuming that the training samples
have the Gaussian distribution. The maximum likelihood
classification is the most widely used statistical classification
technique, because a pixel classified by this method has the
maximum probability of correct assignment (Erbek ef al.
2004).
To increase the reliability of the classification, to the initially
classified images, a fuzzy convolution with a 3x3 size window
was applied. The fuzzy convolution creates a thematic layer by
calculating the total weighted inverse distance of all the classes
in a determined window of pixels and assigning the centre
pixel the class with the largest total inverse distance summed
over the entire set of fuzzy classification layers, i.e. classes
with a very small distance value will remain unchanged while
the classes with higher distance values might change to a
neighboring value if there are a sufficient number of
neighboring pixels with class values and small corresponding
distance values (ERDAS 1999). The visual inspection of the
fuzzy convolved images indicated that there are some
improvements on the borders of the neighboring classes that
significantly influence the separation of the decision
boundaries in multidimensional feature space. The final
classified images are shown in figure 3(a-c). As seen from
figure 3(a-c), the classification result of the Landsat image
gives the worst result, because there are high overlaps among
grassland and forest classes. However, these overlaps decrease
on other images for the classification of which SAR and optical
bands as well as other derivative features have been used.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
For the accuracy assessment of the classification results, the
overall performance has been used. This approach creates a
confusion matrix in which reference pixels are compared with
the classified pixels and as a result an accuracy report is
generated indicating the percentages of the overall accuracy
(Richards and Jia, 1999). As ground truth information, different
AOIs containing 5,751 purest pixels have been selected. AOIs
were selected on a principle that more pixels to be selected for
the evaluation of the larger classes such as grassland and
coniferous forest than the smaller classes such as deciduous
forest and dark soil. The overall classification accuracies for
the selected classes were 74.35%, 79.18% and 83.09% for the
original Landsat bands, the combined features and multiple
bands, respectively.
As could be seen from the overall classification results,
although the combined use of optical and SAR data sets
produced a better result than the single source image, it is still
very difficult to obtain a reliable forest map by the use of the
standard technique, specifically on decision boundaries of the
statistically overlapping classes.
5. THE REFINED MAXIMUM LIKELIHOOD
CLASSIFICATION
For several decades, single-source multispectral data sets have
been effectively used for a land cover mapping. Unlike single-
source data, multisource data sets have proved to offer better
potential for discriminating between different land cover types
(Amarsaikhan et al. 2012). Generally, it is very important to
design a suitable image processing procedure in order to
successfully classify any RS data into a number of class labels.
The effective use of different features derived from different
sources and the selection of a reliable classification technique
can be a key significance for the improvement of classification
accuracy (Lu and Weng, 2007). In this study, for the
classification of land cover types, a refined statistical
maximum likelihood classification algorithm has been
constructed. As the features, multiple bands that include the
original Landsat and PALSAR images as well as two other
derivative bands obtained from texture measures have been
used.
Unlike the traditional method, the constructed classification
algorithm uses spectral and spatial thresholds defined from the
contextual knowledge. The contextual knowledge was defined
on the basis of the spectral variations of the land surface
features on the fused images as well as the texture information
delineated on the dissimilarity image. For determination of the
spectral thresholds, maximum and minimum values of the
selected training signatures and the pixels falling within 1.0
standard deviation were compared, and the latter was selected.
It is clear that a spectral classifier will be ineffective if applied
to the statistically overlapping classes, because they have very
similar spectral characteristics. For such spectrally mixed
classes, classification accuracies should be improved if the
spatial properties of the classes of objects could be
incorporated into the classification criteria. The idea of the
spatial threshold is that it uses a polygon boundary to separate
the overlapping classes and only the pixels falling within the
threshold boundary are used for the classification. In that case,
the likelihood of the pixels to be correctly classified will
significantly increase, because the pixels belonging to the class
that overlaps with the class to be classified using the threshold
