ul 2004
<<
©. Oo
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
classifier ISODATA. The maximum likelihood
classifier using the patch mean resulted in a relatively
high kappa value of 0.735. Maximum likelihood
classifier with pdf produced the overall best accuracy,
0.783.
Looking at the results in more detail, the unsupervised
classifier resulted in many isolated pixels and small
clusters, as expected (Figure 5 a). The Water class in
the region of the stream was almost completely
misclassified as Building with this method. The
stream has exposed and shallow covered rock that is
apparently spectrally similar to the materials from
which buildings are constructed. Building was also
misclassified as Road, and consequently the Building
omission error was relatively high (Table 1). Pixel-
based supervised classification (Figure 5 b), like the
unsupervised classification, resulted in a rather noisy
classification. The classes of Buildings and Roads
were extensively confused, resulting in high errors of
commission and omission for both classes. However,
compared to the unsupervised classification, the
confusion between Building and Water was
dramatically reduced for the pixel-based maximum
likelihood classification.
The maximum likelihood classifier using the patch
mean (Figure 5 c) yielded a visually pleasing
classification, and the second best overall accuracy.
The higher classification accuracy of the maximum
likelihood classification with patch pdf is most likely a
result of the incorporation of differences in the kurtosis
of classes through the variance-covariance matrix data.
When only the patch mean is used in the classification,
such differences are suppressed. The particular
classes that were less well classified in the maximum
likelihood using the patch mean, compared to the patch
pdf, were the Building and Road classes. But the
computing cost for classification with the mean was
much lower than with the pdf. Thus, the classifier
with the patch mean is an efficient alternative to
classification with pdf.
The maximum likelihood classification with pdf
produced higher accuracy than any other classifier
(Table 1). The segmentation suppresses isolated
pixels and small clusters (Figure 5 d), and thus
classification error resulting from high within object
variance was efficiently controlled by this method.
However, a number of cases of confusion arose
between Building and Road, and Lawn and Forest.
The confusion between Lawn and Forest can be related
to segmentation. Although these two classes
generally had sufficient spectral difference between
them for good classification, in some cases the low
Legend
Building Forest ME Shadowed vegetation
Road Lawns ME Water ME Other shadow
Figure 5. Results of the classifications. (Above)
Legend. (Right) a): ISODATA from ERDAS
Imagine. b): Maximum likelihood classification from
ERDAS IMAGINE. c): Maximum likelihood classifier
with patch mean. d): Maximum likelihood classifier
with patch pdf.