The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
348 
2. FNN FOR REMOTE SENSING INFORAMTION 
EXTRACITON 
The FNN system should be trained in an iterative training 
process using the obtained training datasets. After updating this 
model for several times, the topological structure of this 
network and all the weighting indices describing the 
interconnection strengths between neighboring neurons are 
fixed. Then this model will be able to map input variables to an 
estimated output promptly and accurately. 
Much of the previous FNN classification work in remote 
sensing has used multilayer feed-forward networks that are 
trained according to backpropagation algorithm (Martin 
Hellmann et al., 2002; Jiao Yunqing et al., 2007). But this 
training procedure is sensitive to the choice of initial network 
parameters. To overcome this problem, adaptive resonance 
theory is used to network development (Sucharita Gopal et 
al., 1999; Hasi Bagan et al., 2003).Another problem is that 
ancillary information applied in FNNs, which is useful to 
eliminate spectral ambiguities, may cause class proliferation, 
that is to say , too many clusters are created or many pixels are 
left unclassified.To address this problem,other methods,such as 
ID3 learning algorithm, are used to postprocess results obtained 
from fuzzy neural-networks(Jesus Favela et al. ,1998). 
Despite the recent progress, use of FNN technology in 
remotely sensed data processing is still in primary stage. The 
previous studies indicate that network learning algorithm, 
network topology structure, initialization of parameters and 
input signal presentations are important influencing factors to 
the performance of a FNN system (Wu Yifan, 2004; Zhang 
Qiang et al., 2006). 
In the application of FNNs, training time and the accuracy of 
information extraction are important standards of system 
performance evaluation. The accuracy of network can be 
improved by increasing the number of nodes. But as the 
number of neurons increases, more time will be spent on 
network training and data processing. Therefore, the selection 
of appropriate network architecture and learning algorithm is of 
great importance. Besides, the output error of a FNN model is 
greatly depended on the completeness of the training datasets, 
which should cover all the possible cases that the system may 
be encountered (H. Noh et al., 2006). So, it’s vital to prepare a 
typical field site for training data collection. 
3. A DFNN MODEL FOR CROP HEAVY METAL 
STRESS ASSESSMENT 
In this paper, a DFNN model is presented in order to extract 
crop heavy metal stress information form MODIS data. Quite 
unlike traditional FNNs, the architecture of this model is 
determined by training datasets instead of being predefined. As 
a result, it’s unnecessary to apply expert knowledge in this 
model. Furthermore, all the fuzzy rules are generated or deleted 
according to the network performance and the significance of 
each rule during training procedure. So the amount of fuzzy 
rules generated by this model will not increase exponentially as 
the number of input variables increases. 
It is remarked that this DFNN system is equivalent to a Takagi- 
Sugeno-Kang model (Min Han et al., 2008). And it can be 
described by the following formula: 
f:K r : 
y(x) = co 0 + I coiRiHX-Ql) (1) 
i=l 
y 
Where X ( X e 91 ) = r-dimensional state vector with 
x. = the input variable of the DFNN system 
r = the number of input variables 
s = the number of output variables 
u = the number of total fuzzy rules 
o). = the significance of each rule 
R. (•) = activation function of the hidden units 
|| • . || = the Euclidean norm 
Y 
C. e 5K = the centre of this system 
<Wq = the excursion value. 
3.1 DFNN Structure Initialization 
3.1.1 DFNN layers: This DFNN model includes one input 
layer, multiple hidden layers and one output layer. Hidden 
layers can be divided into fuzzifier and inference engine 
according to their functional aims. In fuzzifier layer, input 
values will be transformed into fuzzy values based on 
membership functions. Then, they will be analyzed by 
inference engine according to fuzzy reasoning rules obtained 
from training process. The output of hidden layers is crop 
heavy metal stress information described in the form of fuzzy 
values. Ultimately, in output layer, or defuzzifier layer, all 
these fuzzy output values will be transformed into certain 
values which represent the levels of crop stress induced by 
heavy metal contamination. Generally speaking, fuzzifier layer 
and inference engine layer compose the antecedent network 
which corresponds to the "IF" parts of rules. Consequent 
network consist of defuzzifier layer corresponding to the 
"THEN" parts of rules. The structure of this model is presented 
in Figure 1.