Analytica Chimica Acta 537 (2005) 331–338

Comparing radial basis function and feed-forward neural networks

assisted by linear discriminant or principal component analysis for
simultaneous spectrophotometric quantification of mercury and copper
Y. Akhlaghi, M. Kompany-Zareh∗
Institute for Advanced Studies in Basic Sciences (IASBS), GavaZang, Zanjan 45195-159, Iran

Received 7 June 2004; received in revised form 28 December 2004; accepted 28 December 2004
Available online 16 February 2005

Abstract

Copper(II) and mercury(II) were analyzed simultaneously employing a spectrophotometric method based on application of murexide
solution as a chromogenic reagent. A full factorial six level design was used for the construction of calibration and prediction data sets
consisting of absorption spectra recorded in 350–700 nm range from solution mixtures. A control data set, from a random design, was
applied for validation of the calibration models. The presence of non-linearities was confirmed by a recently discussed methodology based on
augmented partial residual plots (APaRPs). Combinations of principal component analysis (PCA) or linear discriminant analysis (LDA) with
radial basis function networks (RBFNs) or feed-forward neural networks (FFNNs) were built and investigated, as four calibration models.
Number of inputs and hidden nodes for each of the networks were optimized. Performances of methods were tested with relative standard
error of prediction (RSEP%), using synthetic solutions of two metal ions as prediction set. Linear discriminant analysis assisted networks
(LDRBNN) resulted in preferred models, using only one latent variable for each of the analytes. All of the methods were applied for the
analysis of a number of synthetic samples and a dental alloy sample and satisfactory results were obtained.
© 2005 Elsevier B.V. All rights reserved.

Keywords: Artificial neural networks; PCA; Linear discriminant analysis; Metal ions; Spectrophotometry

1. Introduction The ratio of the number of samples to the number of

Recently, chemometrics methods based on artificial neu-
ral networks (ANNs) have found increasing applications for
multicomponent determination, specially in the non-linear
calibration systems [1–5]. Among neural networks, the most
popular is the multi-layer feed-forward networks (FFNNs)
with the back-propagation learning algorithm. Recently,
radial basis function networks (RBFNs) as potential alter-
native approach has been described [6,7]. This offers some
advantages such as robustness to noisy data, compared with
FFNN [8]. The theory for RBFN and application to chemical
problems are found in the literature [9–11].
∗ Corresponding author. Tel.: +98 241 4242235.
E-mail address: kompanym@iasbs.ac.ir (M. Kompany-Zareh).
adjustable parameters in the ANN should be kept as large
as possible. One way of overdetermining the problem is to
compress the input data, especially when they consist of
absorbances recorded at several hundred wavelengths. In ad-
dition to reducing the size of input data, compression allows
one to eliminate the irrelevant information such as noise
or redundancies present in the data matrix [12]. Successful
data compression can result in increased training speed, a
reduction of memory storage, better generalization ability of
the model, and enhanced robustness with respect to noise.
The most popular method for data compression in chemo-
metrics is principal component analysis (PCA). Most ANN
applications in quantitative analysis with spectral data use
PC scores, obtained from PCA, as input variables [5,13–19].
Linear discriminant analysis (LDA) is also a data compres-

0003-2670/$ – see front matter © 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.aca.2004.12.079
332 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338
sion procedure. Similar to PCA, LDA summarizes almost all
variances in the X-matrix on only a few axis, which are mu-
tually orthogonal [20]. The main difference between the two
compression methods is that LDA analysis focuses on the
dissimilarity between classes [21–22]. This work is among
the first studies [23] on the application of LDA as a data com-
pression method, before performing calibrations using ANN.
Due to its direct and continuous contact to saliva, com-
position of dental amalgams is of medical significance. Par-
ticularly, heavy metals contents of these kinds of alloys are
important factor. Copper and mercury are among the impor-
tant constituents of these alloys. In a number of recent stud-
ies, simultaneous determination of copper and mercury based
on atomic absorption spectrometry [24] and electrochemical
methods [25,26] are reported. The aim of this study is to
profit by the advantages of spectrophotometry, as a low cost
and simple analysis method, in coincident determination of
the two metals. There are also some reports on recent spec-
trophotometric quantification of Cu and Hg, but they are not
simultaneous determinations [27,28].
In this paper, murexide was added to mixtures of mercury
and copper cations solutions to form colored complexes for
spectrophotometric measurements. The reason for selecting
murexide as coloring agent is the not complete complex
formation of Hg(II) with this ligand which results to a non-
linear relation between the spectral data and concentration
of analytes. The main goal in this article is the comparison
of four artificial neural networks based methods (PCFFNN,
PCRBFN, LDFFNN, and LDRBFN) for performing
non-linear calibration. The metal ions were estimated and
methods were validated using synthetic and a dental amalgam
samples.
2. RBFN theory
The structure of radial basis function network (RBFN) is
comprised of three node layers of input, hidden and output.
The input layer serves only to distribute input to the hidden
layer. Each neuron of the hidden layer represents a kernel
or a basis function, with equal dimensions to the input data.
RB networks generally use a Gaussian function to account
for the non-linearity of the hidden layer processing elements.
The Gaussian function responds only to a small region of the
input space where the Gaussian is centered. The successful
implementation of these networks is to find suitable centers
for such a Gaussian functions, which is characterized by two
parameters, i.e. center (cj ), and peak width (σ j ). The output
for the jth Gaussian neuron for an input object xi can be
calculated by the following equation:
outj = Φj (xi − cj ) = exp(xi − cj /σj )2 (1)
where xi − cj  is the calculated Euclidean distance between
xi and cj , and σ j determines the portion of the input space
where the jth RBF will have a non-significant zero response.
The input value to each output node is the weighted sum
of all the outputs of the hidden nodes. Finally, the response
of each output node is calculated by a linear function of its
input (including the bias wk0 ), that is, the output of hidden
layer (outk ). The relation between the value outk and the input
variable xi can be represented by:


outk = wk0 + wkj Φj (xi − cj ) (2)
j
The weights wkj are adjusted to minimize the mean square
error of the net output. Two sets of parameters (the centers
and the widths) in the hidden layer and a set of weights in the
output layer are adjusted. Therefore, the adjustment of the
output layer is simple and RBFN has a guaranteed learning
procedure for convergence. However, in back-propagation
FFNN, the parameters of transfer functions both in hidden
and output layers should be adjusted by using the Sigmoid
transfer functions and generally it is time-consuming.
A considerable point is that many of authors call the
RBF system of linear equations the “neural” networks, which
seems not correct. The RBF system lacks all the attributes that
go with the natural-inspired methods, such as ANN, genetic
algorithm (GA), and simulated annealing (SA). All natural-
inspired methods have in common (a) any form of random-
ness influencing the outcome of the procedure (sample order,
initialization of weights, etc.), (b) similarity to some natural
phenomenon (functioning of biological neuron, survival of
the fittest, movement of atoms during cooling, etc.), and (c)
an iterative learning procedure, with the number of iteration
steps influencing the results. RBFN has none of these features
and can be solved straight by the multiple linear regression
(MLR) method.
3. Experimental
3.1. Apparatus
A Pharmacia Ultraspec 4000 spectrophotometer with
Swift II (Amersham Pharmacia Biotech, Piscataway, NJ,
USA), as instrument control and data acquisition soft-
ware, was employed for spectrophotometric measurements.
A model 713 Metrohm pH meter (Metrohm Ltd., Herisau,
Switzerland) was used for the measurement of pH of the
solutions. A Pentium III computer and programs written in
MATLAB 6.0 (Mathworks Inc., Natick, MA, USA) were uti-
lized for processing of data.
3.2. Reagents
All solutions were prepared with analytical reagent
grade reagents. Stock solutions of copper (1.600 mmol L−1 ),
and mercury (1.600 mmol L−1 ) were prepared from
Cu(NO3 )2 ·3H2 O, and Hg(NO3 )2 ·6H2 O, respectively. A bo-
rate buffer pH = 12 solution was prepared using 0.05 mol L−1
 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338 333

sodium tetraborate and 0.10 mol L−1 NaOH solutions [29].

A fresh 1.600 mmol L−1 of murexide was prepared by dis-
solving 0.1130 g of murexide in 250.0 mL of doubly distilled
water, daily.

3.3. Procedure

Suitable volumes of stock of sample solutions of Cu(II)

and Hg(II) were transferred to 10 mL volumetric flasks, fol-
lowed by the addition of 3.0 mL of borate buffer solution
and 3.5 mL of murexide solution. The solution then diluted
to the mark with doubly distilled water. The absorbance of
each solution contained in 1 cm cell was measured with re-
spect to a blank (distilled water) every 1 nm between 350 and Fig. 1. The absorbance spectra of the murexide (Mu, light black) and its
700 nm. complexes with Cu(II) (Cu–Mu, heavy gray) and Hg(II) (Hg–Mu, heavy
For analysis of the real samples, about 1.00 g of the alloy black) at pH = 12. Analytical concentrations for the ligand and metal ions
were 80 ␮mol L−1 .
was precisely weighted, dissolved in a 2 mL aliquot of con-
centrated 1:3 (v/v) H2 SO4 :HNO3 solution, transferred to a
Table 1
100.0 mL volumetric flask and diluted to the mark with dis- Concentrations of copper and mercury ions in the 36 standards prepared
tilled water. The rest of the procedure was the same as the
Mercury Copper (␮mol L−1 )
previous paragraph. (␮mol L−1 )
0.0 4.0 7.5 11.4 15.4 19.5
0.0 1A 2Aa 3Aa 4A 5Aa 6A
23.6 1Ba 2B 3B 4Ba 5B 6Ba
4. Results and discussion
46.1 1Ca 2C 3Ca 4C 5C 6Ca
68.4 1D 2Da 3D 4Da 5Da 6D
4.1. Experimental conditions 90.1 1E 2E 3E 4E 5E 6E
111.3 1F 2Fa 3Fa 4F 5F 6F
Considering the unstability of murexide in acidic solu- a Prediction set.
tions, the overlapping spectra of two metal cation complexes
with murexide were investigated at pH values higher than
7.0. Based on the obtained selectivity and sensitivity in the concentrations of two metal ions in the mention ranges were
spectra from the ligand and its metal complexes, pH = 12 was applied for construction of control set (Fig. 2). The predic-
chosen for our study. It was also assumed that addition of con- tive ability of each method for each metal ion was described
stant volume of borate buffer pH = 12 made a similar ionic in term of relative standard error for estimation of concen-
strength for all solutions. The concentration of murexide dur- trations in the control set (RSE%) and/or the prediction set
ing the experiments was 250 ␮M which was about two times (RSEP%) [31].
as maximum total metal concentrations in the calibration, or
Cmu /max(CCu + CHg ) = 2. Pure absorption spectra of metal
ion complexes, which were estimated using beta-correction
procedure [30], and ligand (Fig. 1) illustrate the significant
spectral overlap in the system in the considered conditions.

4.2. Mixtures design

To maximize the statistical information content in the

spectra, a six-level full factorial design was utilized for as-
signing the concentration of two metal ions in the considered
mixtures in the ranges 0.0–20.0 and 0.0–120.0 ␮mol L−1 for
copper(II) and mercury(II), respectively. In assigning the up-
per limits for the concentration ranges, it was considered that
the resulting absorbance values would be less than 2.000,
for all mixtures. Calibration and prediction spectral matrices
were built from the 36 mixtures in Table 1 and were sub-
mitted to four calibration procedures: PCFFNN, LDFFNN,
PCRBFN, and LDRBFN. Thirteen mixtures with selected Fig. 2. Concentrations of Cu(II) and Hg(II) ions in the 13 control samples.
334 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338

Fig. 3. Augmented partial residual plots (APaRPs) for Cu (considering PC1) (a), Hg (considering PC2) (b), Hg (considering PC3) (c), and Cu (considering
PC4) (d) when all selected PCs were included in the model. The selected PCs for plots (a–c) were PC1 to PC3, and for plot (d) were PC1 to PC4.

4.3. Detection of non-linearities
Application of ANNs in multivariate calibrations was pro-
posed when a significant non-linearity is observed in the data.
Apparent partial residual plot (APaRP) [32], as the most uni-
versal diagnostic tool, was the applied statistical test for dis-
tinguishing the non-linearity of the relation between some
of the first factors and the concentrations in this study. This
procedure was implemented in the following way: individual
analyte concentrations contained in the vector cm were first
regressed against the first A PCs of the data matrix and the
square of the first PC [32]:
cm = b0 + b1 PC1 + · · · + bA PCA + baa (PCa )2 + eAPaRP ,
a = 1, . . . , A
where eAPaRP was a vector collecting the APaRP fitting resid-
uals. The relevant plot was obtained by plotting the sum
[eAPaRP + ba PCa + baa (PCa )2 ] as a function of PC1 [32].
Fig. 3 shows the APaRPs for the calibration of the two
analytes, using three or four PCs for copper and mercury. As
can be found from the plots, for the first PC a linear relation
with the variation of mixtures components was obtained. For
the second and third PCs a significant non-linearity was the
result of the test. The fourth PC seems to contain no consid-
erable information. Totally, the test illustrates the significant
non-linearity in the considered data, and need to non-linear
calibration models like ANNs.
4.4. Reducing the number of data
Decreasing the data volume before using ANNs for non-
linear multivariate calibration was suggested as a preprocess-
ing step in many of the previous studies [12,33]. Two common
methods for reducing data are linear discriminant analysis
(LDA) and principal component analysis (PCA). In this work,
before being reduced, the absorbance data were autoscaled
(using the mean and standard deviation of each column). For
(3) estimation of first significant score vectors using PCA, con-
centrations of calibration mixtures were not considered and
one set of principal components for both metals was esti-
mated. The plot of first two score vectors versus the sample
codes, which are in correlation with concentrations, is illus-
trated in Fig. 4a. No correlation between the values of scores
and the concentrations can be seen, for any of the PCs. Es-
timation of first significant discrimination factor (DF) using
LDA included a classification step, based on concentration
matrix. The members (samples) for each class were differ-
 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338
Fig. 4. (a) Estimated first discrimination factors (DF1s) from LDA on spec-
tral calibration data, once based on copper concentrations (white) and then
based on mercury contents (gray). (b) Calculated scores for the first two
PCs from PCA on spectral calibration data (without considering the analyte
contents of solutions).
ent when the classification was on Hg contents of solutions,
compared to when it was based on Cu contents. The first DFs
from applying LDA on spectral data for Hg and Cu are shown
in Fig. 4b. Considerable correlation can be observed between
the values of DFs and the sample codes (concentrations). The
eigenvalues from PCA and LDA calculations on the calibra-
tion data are in Table 2 and well illustrates that more than
99.9% of variations are in the first DFs.
4.5. Optimizing the number of factors and the network
variables
One layer of hidden nodes and one node in the output layer
are common topological suggestions, when using an artificial
neural network as a model for the non-linear calibration [12].
In this way, separate networks were trained for copper and
Table 2
First five eigenvalues estimated from applying LDA and PCA on calibration set spectra
λ1 λ2
LDA
Cu 23362 6.42 × 10−11
Hg 12835 3.33 × 10−11
PCA 106.4147 57.4301
335
mercury ions. Relative standard error percent (RSE%) es-
timated for the prediction of analytes concentrations in the
control set was the criteria for illustration of the performance
of the trained networks as calibration models. The control set
was applied to prevent the overitting of the networks to the
noises from the calibration set, during training.
4.5.1. PCFFNN
Utilizing the Levenberg–Marquardt algorithm, feed-
forward neural networks (FFNNs) including one to six nodes
in the hidden layer, and one to seven PCs as input were trained.
To observe the extent of reproducibility of the results, and
the robustness of the network, the training process in each
of the conditions was repeated 10 times. From the 10 ob-
tained results, the one with minimum RSE% was considered
for investigations. As shown in Fig. 5(a) and (b), the low-
est RSE% values were obtained with five input factors and
four hidden nodes for Cu(II) and with five input factors and
two hidden nodes for Hg(II), which were included in the fi-
nal PCFFNN models. It can be seen that in the most of the
cases resulting RSE%s for 10 times repetition of trainings
in each condition are not similar. This shows the degree of
robustness of FFNNs when using PCA for data compression.
The predicted concentrations of two analyte ions using the
final obtained PC-FFNN models for seven test samples are
in Table 3.
4.5.2. LDFFNN
Discrimination factors (DFs) were the obtained com-
pressed data from LDA for this part of work. For each of the
analytes, the first DF included the majority of the variations in
the spectral data (Table 2). Anyway, in addition to the effect of
number of hidden nodes on the values of RSE%s for the con-
trol samples, the influence of utilizing more than one DF as
the input for the calibration model was investigated. Applica-
tion of more than one node in the hidden layer made consider-
able improvements in the obtained RSE%s. But, application
of more than one DF decreased the repeatability of the results,
in addition to increasing the error values. This is due to the
classification ability of LDA. The results from application of
1–10 hidden nodes and one DF for each of two analytes are
illustrated in Fig. 5(c) and (d). One DF (one node in the input
layer) and nine hidden nodes were selected for both Cu and
Hg in the final LDFFNN models. The estimated concentra-
tions for the two cations using the final obtained LDFFNN
models in seven test samples are in Table 3. Compared to
the condition of using PCA prior to FFNN, application of
λ3 λ4 λ5
4.92 × 10−11 4.77 × 10−11 4.59 × 10−11
2.61 × 10−11 2.22 × 10−11 1.78 × 10−11
23.8994 15.6024 10.6913
336 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338

Fig. 5. The relation of RSE% and the number of input PCs in PCFFNN using four nodes for Cu (a), and two nodes for Hg in the hidden layer (b). RSE% values
vs. the number of hidden nodes in LDFFNN, using one DF for both copper (c) and mercury (d).

LDA resulted in a considerable improvement in the obtained
RSEs. The second point is that the input variable of the net-
work for LDFFNN was only one DF, whereas in the case of
PCFFNN the number of input PCs were >4 for both analyte
ions.
4.5.3. PCRBFN
The exact type of radial basis function networks (RBFNs),
from ANN Toolbox of MATLAB version 6.0, was the choice
for this part. In the exact RBFNs the number of hidden nodes
are equal to the number of nodes in the input layer. In this
way, the adjustable parameters in this part were the number of
input variables and SPREAD. The latter parameter was in re-
lation with the spread of radial basis functions in the network.
Input variables for this part were the PCs from application of
PCA on spectral data. RSE%s for the prediction of copper
and mercury ions (in the control samples) at the number of
input PCs from 3 to 6, and SPREAD values of 1–100 were
investigated. Error values for number of PCs less than 3 and
more than 6, and SPREAD values less than 1 were very high.
Introduction of four to six PCs resulted in significantly im-
proved RSEs. In this way, four first PCs as input variables
and SPREAD = 21 were chosen for copper, and same four
PCs and SPREAD = 45 were the choice for mercury in the

Table 3
Results for metal ions analysis in synthetic samples
Calculated conc. (␮mol L−1 ) Actual conc. (␮mol L−1 )

LDRBFN PCRBFN LDFFNN PCFFNN Cu Hg No.

Cu Hg Cu Hg Cu Hg Cu Hg
0.00 23.64 0.25 19.49 0.00 23.62 0.00 20.83 0.00 23.64 1
3.99 0.00 3.86 0.08 3.98 0.02 3.75 0.07 3.99 0.00 2
3.70 111.34 3.72 112.34 3.69 111.32 3.61 112.46 3.71 111.36 3
7.51 90.14 7.79 88.85 7.51 90.31 7.38 93.52 7.51 90.14 4
11.40 68.40 11.71 72.15 11.40 68.29 11.97 69.60 11.40 68.41 5
15.38 46.15 15.50 44.24 15.40 46.14 15.52 50.76 15.38 46.15 6
19.46 23.57 19.34 21.65 19.45 23.87 19.26 21.39 19.46 23.36 7
0.03 0.12 1.85 3.69 0.11 0.35 2.38 3.99 RSE%
 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338 337

In the case of using RBNNs, the obtained results from

repetition of training procedure for each condition were the
same. It was due to high reproducibility of modeling that
is a main advantage for the RBFNs. In this way, 10 times
repetition of training in each of experimental conditions was
not performed.

4.5.4. LDRBFN
This part included LDA as a compression procedure,
before modeling the data with RBNN. The adjustable pa-
rameters for the utilized exact RB network were SPREAD
and the number of input DFs, calculated using LDA. RSEs
for prediction of the concentrations of copper and mercury
(in the control set) at the number of DFs from 1 to 3,
and log(SPREAD) values of −1.3 to 2 were investigated.
The results showed that the minimum RSE% values for Cu
and Hg were at log(SPREAD) values of −1.0 and −1.2
(or SPREAD values of 0.10 and 0.07), respectively, and
only one DF as input (Fig. 6b), which were the choice for
training the final LDRBFNs. The specific RSE% values of
0.03 and 0.12% in the analysis of test samples (Table 3)
illustrates the effect of gathering the abilities of LDA as
a classification method and RBFN as a flexible and pre-
cise calibration model. In this part also, due to high repro-
ducibility of RBNN, 10 times repetition of training was not
necessary.

4.6. Application to dental alloy sample

Fig. 6. The relationship of RSE% and the SPREAD parameter of radial basis
networks in PCRBFN using four PCs (a), and in LDRBFN using only one
DF (b) for both analyte ions.
final trained PCRBFN networks plotted in Fig. 6a. Estimated
concentrations of the two metals at the mentioned settings
are in Table 3. Results from the table indicate that, specially
in the case of Hg(II), there was no considerable improvement
in the estimated RSEs when PCRBFN was applied in place
of PCFFNN. It shows the importance of data compression
procedure.
Dental alloy samples were obtained from a dentistry and
treated as described in the procedure. The obtained spectra
were used to calculate the concentration of each metal ion by
the different proposed calibration methods (Table 4). Com-
paring the results with that obtained from the analysis of
sample using atomic absorption spectroscopy, LDFFNN and
LDRBFN methods performed equally well, while PCFFNN
and PCRBFN showed slightly less performance. Presence of
impurities in the alloy sample decreased the accuracy of pre-
diction of Hg concentration, when the applied compression
method was PCA. It shows the more desirability of LDA com-

Table 4
Accuracy and precision for the metal ions analysis in dental alloy samples
Sample PCFFNNa LDFFNNa PCRBFNa LDRBFNa AAS
Hg (␮mol L−1 )
1 16.84 ± 0.15 16.38 ± 0.19 16.75 ± 0.20 16.36 ± 0.15 16.46
2 23.93 ± 0.34 24.28 ± 0.26 24.71 ± 0.29 24.38 ± 0.24 24.33
3 19.90 ± 0.17 20.21 ± 0.21 20.42 ± 0.17 20.14 ± 0.24 20.21
4 13.86 ± 0.15 14.06 ± 0.11 13.99 ± 0.13 14.10 ± 0.14 14.09
5 23.72 ± 0.25 23.55 ± 0.23 23.90 ± 0.20 23.52 ± 0.26 23.52
Cu, ␮mol L−1
1 14.78 ± 0.13 14.77 ±0.09 14.83 ± 0.06 14.76 ± 0.10 14.75
2 15.92 ± 0.21 15.79 ±0.12 15.74 ± 0.17 15.87 ± 0.15 15.84
3 18.28 ± 0.15 18.30 ±0.17 18.21 ± 0.17 18.30 ± 0.12 18.24
4 12.51 ± 0.11 12.50 ±0.10 12.57 ± 0.08 12.55 ± 0.10 12.54
5 17.69 ± 0.16 17.60 ±0.18 17.59 ± 0.17 17.65 ± 0.13 17.63
a With three times replication of the experiments.
338 Y. Akhlaghi, M. Kompany-Zareh / Analytica Chimica Acta 537 (2005) 331–338
pared to PCA for data compression in non-linear calibrations
by NNs.
5. Conclusion
Satisfactory precision and accuracy was obtained with all
of four investigated methods, although, because of surpris-
ingly lower RSE% values, LDFFNN and LDRBFN were the
preferred methods. These two methods as well as PCFFNN
and PCRBFN were employed for the analysis of results from
a dental alloy. All of the four procedures performed well in the
case of prediction of copper in the alloy sample, although, in
the case of mercury low accuracy results using the two PCA
based methods were obtained.
When applying LDA only the first DF was satisfactory
for preparation of a proper calibration model, although, in
the case of using PCA more than two PCs were required.
It is due to classification ability of LDA. High reproducibil-
ity of the training procedure and considerably lower train-
ing period in the RBFNs, are among the main advantages of
these networks compared to FFNNs. Totally, application of
RBFN assisted with LDA data compression seems the most
proper for performing the non-linear multivariate calibra-
tions.
Acknowledgment
The authors are grateful to the Institute for Advanced Stud-
ies in Basic Sciences (IASBS) for the supports.
References
[1] J. Zupan, J. Gasteiger, Anal. Chim. Acta 248 (1991) 1.
[2] G. Kateman, Chemom. Intel. Lab. Syst. 19 (1993) 135.
[3] J.R.M. Smits, W.J. Melssen, L.M.C. Buydens, G. Kateman,
Chemom. Intel. Lab. Syst. 22 (1994) 165.
[4] J. Zupan, J. Gasteiger, Neural Networks for Chemists: An Introduc-
tion, VCH, Weinheim, 1993.
[5] M. Kompany-Zareh, A. Massoumi, Sh. Pezeshk-Zadeh, Talanta 48
(1999) 283.
[6] B. Walczak, D.L. Massart, Chemom. Intel. Lab. Syst. 50 (2000) 179.
[7] Q.F. Li, X.J. Yao, X.G. Chen, M.C. Liu, R.S. Zhang, X.Y. Zhang,
Z.D. Hu, Analyst 125 (2000) 2049.
[8] E.P.P.A. Derks, M.S. Sanchez Pastor, L.M.C. Buydens, Chemom.
Intel. Lab. Syst. 28 (1995) 49.
[9] B. Walkzak, D.L. Massart, Anal. Chim. Acta 331 (1996) 177.
[10] A. Pulido, I. Ruisanchez, F.X. Ruis, Anal. Chim. Acta 388 (1999)
273.
[11] Y. Ni, Sh. Chen, S. Kokot, Anal. Chim. Acta 463 (2002) 305.
[12] F. Despagne, D.L. Massart, Analyst 123 (1998) 157R.
[13] C. Borggaard, H.H. Thodberg, Anal. Chem. 64 (1992) 545.
[14] T.B. Blank, S.D. Brown, Anal. Chem. 65 (1993) 3081.
[15] P.J. Gemperline, Chemom. Intel. Lab. Syst. 39 (1997) 29.
[16] J.H. Jiang, J.H. Wang, X.H. Song, R.Q. Yu, J. Chemom. 10 (1996)
253.
[17] T.B. Blank, S.D. Brown, Anal. Chim. Acta 277 (1993) 273.
[18] L. Zhang, J.H. Jiang, P. Liu, Y.Z. Liang, R.Q. Yu, Anal. Chim. Acta
344 (1997) 29.
[19] F. Despagne, D.L. Massart, Chemom. Intel. Lab. Syst. 40 (1998)
145.
[20] M. Otto, Chemometrics: Statistics and Computer Application in An-
alytical Chemistry, Wiley-VCH, Weinheim, 1999.
[21] W. Wu, Y. Mallet, B. Walczak, W. Penninckx, D.L. Massart, S.
Heuerding, F. Erni, Anal. Chim. Acta 329 (1996) 257.
[22] Y. Mallet, D. Coomans, O. de Vel, Chemom. Intel. Lab. Syst. 35
(1996) 157.
[23] J. Irudayaraj, F. Xu, J. Tewari, J. Food Sci. 68 (2003) 2040.
[24] Y. Guo, B. Din, Y.W. Liu, X.J. Cheng, S.M. Meng, M.Z. Tian, Anal.
Chim. Acta 504 (2004) 319.
[25] S.A.M. Marzouk, W.T. Al-Ariqui, S.S.M. Hassan, Anal. Bioanal.
Chem. 375 (2003) 1186.
[26] G. Sanna, M.I. Pilo, P.C. Piu, N. Spano, A. Tapparo, R. Seeber,
Electroanalysis 14 (2002) 1512.
[27] S. Suresha, M.F. Silwadi, A.A. Syed, Int. J. Environ. Anal. Chem.
82 (2002) 275.
[28] B. Coulomb, F. Theraulaz, C. Brach-Papa, M. Carbonel, J.L.
Boudenne, Quimica Anal. 20 (2001) 99.
[29] N. Bobrov, Ju LURIE Handbook of Analytical Chemistry, Mir,
Moscow, 1978.
[30] H.-W. Gao, P.-F. Zhang, Anal. Proc. 31 (1994) 85.
[31] M. Blanco, J. Coello, H. Ituriaga, S. Maspoch, M. Redon, Appl.
Spectrosc. 48 (1994) 37.
[32] V. Centner, O.E. de Noord, D.L. Massart, Anal. Chim. Acta 376
(1998) 153.
[33] P.J. Gemperline, Chemom. Intel. Lab. Syst. 15 (1992) 115.