Satellite Image Classification Using Cascaded Architecture of Neural Fuzzy Network - Advanced Techniques in Knowledge Discovery and Data Mining - page 211

classification with arbitrary patterns, not just targets of selected blocks with dif-
ferent textures. Others have applied the Bayesian classifier [8] and Markov ran-
dom field [9], [10] to obtain relative frequencies of individual and neighbors
among a pixel. With these structures it is hard to obtain the required results when
samples are insufficient. Moreover, the consumption of time to do the classifica-
tion should also be one of the considerations. Some people use neural networks
[11], [12], [13] to proceed the classification. Their results show that neural net-
works are likely to get a satisfying result and this is why the neural network is
used as our reference. Because the satellite images usually contain many complex
factors and mix-up samples, a high recognition rate is not easy to attain. Espe-
cially for a nonhomogeneous region, the gray values of its satellite image vary
greatly, and thus the direct use of gray-level statistics fails to do the categorization
task satisfactorily. The goal of this chapter is to develop a cascaded architecture of
a neural fuzzy network with feature mapping (CNFM) to help the clustering of
satellite images.
In this chapter, the dimension of inputs is first reduced by Kohonen's
self-organizing feature map (SOM). It is an unsupervised neural network whose
inputs of each channel are composed as follows. First, gray values are selected as
the reference values. Second, statistical features computed from co-occurrence
matrices [7] such as contrast, inverse difference moment, angular second moment,
and entropy are used. Third, energies and entropies from wavelet decomposition
[16], [17] are served as spectral features. No matter how many features and how
many channels we use, each group of features in high dimension can be trans-
formed into 2D coordinates by Kohonen's SOM. In addition to the benefit of re-
duction in dimension, it can remove some noisy areas and avoid the training proc-
ess being overoriented to the training patterns. After the inputs are transformed by
Kohonen's SOM, further classification will be performed by a neural fuzzy-net-
work (called SONFIN [14]). It is a supervised neural network that can classify de-
sired outputs delicately. This cascaded architecture, named CNFM, is a general
and powerful structure that can give very promising results in terms of accuracy
and performance. Experimental results show that the CNFM can reach an accu-
racy of 96.5% with respect to all feature domains.
Figure 8.1 shows the system architecture of CNFM. There are three types of
input, which are spatial features of gray values, statistical features from an occur-
rence matrix, and spectral features from wavelet decomposition with N channels.
Suppose there are M features in total. If we do not reduce our dimension of inputs,
our network inputs will be of ( M * N) dimension. It must be noted that if the
number of features increases, the input space of the multichannel satellite-image
classification problem grows. However, our system can solve this problem grace-
fully; the input dimension is first reduced by Kohonen's SOM, and further classi-
fication is performed by a neural fuzzy network (SONFIN). Figure 8.2 shows the
details of our CNFM. Given a center pixel, we select different neighborhood sizes
for different feature domains. We use co-occurrence matrix and wavelet decompo-
sition to extract the required features. After that, we pass the features, such as gray
values, angular second moment (ASM), inverse difference moment (IDM), con-
trast (CON), entropies (ENT) and energies to Kohonen's SOMs. Finally, we use a
neural fuzzy network, SONFIN, to train the outputs, in 2D coordinate format, of