The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
3.1.2 DFNN input variables: Vegetation spectral
reflectance is associated with the biochemical composition in
leaves, such as structure of mesophyll cells, chlorophyll
content and water content. In different wavelengths, vegetation
spectrum reflectance curves show different patterns and
characteristics (Yunzhao Wu et al., 2005). A number of studies
have demonstrated that shifts in vegetation spectra due to
heavy metal contamination occurred both the visible and the
near-infrared part of the spectrum. These studies used spectral
vegetation indices to investigate changes in plant stress, for
they can combine two or more spectral bands to enhance the
vegetative signal while minimizing background effects
(Lehmann, F. et al., 1991; Sommer, S. et al., 1998; Mohammed
et al., 2000; L. Kooistr et al., 2004).Heavy metal contamination
will affect the status of plants, such as pigment contents,
photosynthetic efficiency, nitrogen contents in canopy, and
carbon contents of leaf. Values of NDVI, EVI and NDVIg can
indicate the changes of these factors listed above. Therefore,
they were chosen as the input variables in this model.
3.2 DFNN Training Algorithm
The aim of the training procedure is to minimize the error, i.e.
the difference between the calculated output values and the
target output values, and to generate fuzzy rules. The
adaptation of the weights during the training process can lead
to a so called over training problem. This means that the neural
network can reproduce the training data quite well but has lost
its ability to generalize. The phenomenon is especially severe
when only a few training patterns are available.
Therefore, this model starts with a simple network structure
which contains no fuzzy rules and goes over stepwise to more
complicated structure. During training process, fuzzy rules will
be generated according to the system performance. And in the
meantime, insignificant rules will be deleted.
3.2.1 DFNN structure determination: In fuzzifier layer,
input data are fuzzified and membership grades are calculated
according to Gaussian membership function. The membership
grade of each input value X i (i=l, 2...r) is given by:
p..(x ) = exp[ ' 2 V ] (i=l ...r; j=l, 2... u) (2)
a v
Where p.. = the membership grade of x. according to the j th
V t
membership function
= the centre of the j th membership function
<Tj = the importance of the j th membership function.
The output results obtained from this layer are then used as
input values to the inference engine where T-norm product
operator is applied to calculate the trigger weight of each rule.
The output value of the j th rule is computed according to the
following function:
r (Xi-ca)
<f> fx v x 2 ,...,x r ) = exp[-X -J-—] (3)
/=1 a ÿ
The single node in defuzzifier layer is a fixed node labeled X ,
which computes the overall output as the summation of all
incoming signals:
u
y(Xj, x 2 ,..., x r ) = jL^ (û. ■ (f). (4)
Where y = the output results of this system
cùj = the weight of the j th rule
3.2.2 Rule extraction standards: Two standards, output
error and the width of a Gaussian membership function, are
introduced to determine whether a new rule should be added
into current system.
To the i th training data (X., ) , the output error can be
computed as follows:
Ikll = Ik-T/ll (5)
Where X. = the input vector
t. = the target output value
y. = the calculated output value resulted from current
system
In comparison with the predefined precision k £ , if
Then a new rule should be generated.
In this model, the input variables are classified into several
fuzzy sets according to Gaussian membership function. The
amount of overlap between data sets is controlled by the widths
of Gaussian membership functions. An input training data can
be present by a Gaussian membership function, if its
membership grade is within the accommodation range.
To the i^ training data (X., t.), the distance between input
value X. and the center of Gaussian membership function can
be computed as follows:
350