Satellite Image Classification Using Cascaded Architecture of Neural Fuzzy Network - Advanced Techniques in Knowledge Discovery and Data Mining - page 224

three channels, and represents the features of spatial domain. The second set of
inputs consists of angular second-moment, contrast, inverse difference moment,
and entropy that come from co-occurrence matrices. This set of inputs represents
the features of statistical domain. The last set of inputs includes energy and en-
tropy that come from wavelet decomposition and represents the features of spec-
tral domain. After the transformation of Kohonen's SOMs, these three sets of in-
puts are reduced into 2D coordinates. However, we do not expand a gray value
into 2D coordinates because its dimension is low enough in our test examples. If
we have transformed the gray values of all channels, we may remove the informa-
tion of differences among channels. Because the statistical and spectral features
focus on local varieties, we can apply the transformation directly and do not need
to care for the information of differences among channels. Then we reduce the
dimension from 39 features to 15. As shown in Fig. 8.10, the three types of inputs
are: three gray values obtained from each channel, four statistical features from
each of the three channels, and eight spectral features obtained by wavelet de-
composition from each of the three channels. Hence, if we do not reduce the di-
mension of the inputs, there will be 39 input features. If the number of features in-
creases, a large number of inputs are to be used for classification. However, our
system can solve this problem gracefully. Input dimension is first reduced by Ko-
honen's SOM, then further classification is performed by SONFIN.
After the features have been reduced and transformed by Kohonen's SOM, we
pass them to a supervised neural fuzzy network, SONFIN. This network can per-
form online input space partitioning, which creates only the significant member-
ship functions on the universe of discourse of each input variable using a fuzzy
measure algorithm and the orthogonal least square (OLS) method. Our objective is
to use the ability of SONFIN to obtain lower mean square error (MSE) and higher
learning rate. The result will be compared to that of normally used statistical
methods and backpropagation network in Section 8.4.

SONFIN

SOM SOM

12 statistical features
from
co-occurrence matrix

3 gray values
from
each channel

24 spectral features
from
wavelet decomposition

Fig. 8.10 . System architecture of CNFM.