Satellite Image Classification Using Cascaded Architecture of Neural Fuzzy Network - Advanced Techniques in Knowledge Discovery and Data Mining - page 228

Number of epoch
Fig. 8.13. The MSE of backpropagation network against SONFIN.
8.5 Conclusions
Data mining has become widely recognized as an important concept by research-
ers in many areas. The use of valuable information “mined” from data is recog-
nized as necessary to maintain competitiveness in today's business environments.
This chapter proposed a cascaded architecture of neural fuzzy network with fea-
ture mapping (CNFM) for a pragmatic approach to solving pattern recognition,
classification problems, and feature analysis, which are all important for data
mining and knowledge extraction. The proposed CNFM signifies a novel alge-
braic system identification method, which can be used for knowledge extraction in
general and for satellite image analysis in particular.
Although the neural network is a useful black-box tool that can do good pre-
diction for many problems, it still cannot handle the problem of “garbage in, gar-
bage out.” We must provide some auxiliary information to handle the texture
problem and the variation of neighborhood in satellite images. However, the addi-
tional features may increase the computation complexity and mislead the training.
To overcome this problem, we presented a new cascaded system that uses Koho-
nen's self-organizing feature map as the reduction mechanism of input dimension.
After we have extracted the lower-dimensional inputs, the classification task is
performed by a neural fuzzy network (SONFIN).
The important contribution of the proposed system is that it can solve the
problem of scaling and selection of features gracefully. Also, we extended the
mechanism of Kohonen's SOM to reduce the high input dimension rather than fil-
tering the training sets. Our CNFM combines spatial, statistical, and spectral fea-
tures into one system resulting in a classifier with improved performance.
Furthermore, we improved the computation complexity of the co-occurrence
matrix by the symmetric linked list (SLL) structure. Our future work is to enhance