Screening analysis of river seston downstream of an effluent discharge point using near-infrared reflectance spectrometry and wavelet-based spectral region selection.

Abstract

A methodology for screening analysis of river seston downstream of an industry effluent by using near-infrared reflectance spectrometry was developed. A wavelet transform (WT)-based strategy is used to select a spectral region in which the effect of the effluent on the optical properties of the seston is more evident. The methodology was applied to samples from the River Mumbaba in northeast Brazil. Four sites were monitored: two upstream (1 and 2), one at the discharge point of the effluent (3), and another downstream (4). Soft Independent Modelling of Class Analogies (SIMCA) models were built for site 1 and were then applied to the classification of samples from sites 2 and 4. The results reveal that the WT-based spectral region selection is essential to ensure good sensitivity and specificity with respect to the detection of events associated to the effluent discharges at site 3. In fact, the changes in site 4 caused by the effluent are masked by other environmental factors when the full spectrum is employed.

Full text

Page 1

Water Research 39 (2005) 3089–3097
Screening analysis of river seston downstream of an effluent
discharge point using near-infrared reflectance spectrometry
and wavelet-based spectral region selection
Va ˆ nia Maria de Medeirosa, Ma ´ rio Ce ´ sar Ugulino Arau ´ joa,?, Roberto Kawakami
Harrop Galva ˜ ob, Edvan Cirino da Silvaa, Teresa Cristina Bezerra Saldanhaa,
Ilda Antonieta Salata Toscanoa, Maria do Socorro Ribeiro de Oliveiraa,
Sueny Ke ˆ lia Barbosa Freitasa, Manoel Mariano Netoa
aUniversidade Federal da Paraı ´ba, CCEN, Departamento de Quı ´mica, Caixa Postal 5093, 58051-970–Joa ˜o Pessoa, PB, Brazil
bInstituto Tecnolo ´gico de Aerona ´utica, Divisa ˜o de Engenharia Eletro ˆnica 12228-900–Sa ˜o Jose ´ dos Campos, SP, Brazil
Received 12 May 2004; received in revised form 16 May 2005; accepted 18 May 2005
Available online 5 July 2005
Abstract
A methodology for screening analysis of river seston downstream of an industry effluent by using near-infrared
reflectance spectrometry was developed. A wavelet transform (WT)-based strategy is used to select a spectral region in
which the effect of the effluent on the optical properties of the seston is more evident. The methodology was applied to
samples from the River Mumbaba in northeast Brazil. Four sites were monitored: two upstream (1 and 2), one at the
discharge point of the effluent (3), and another downstream (4). Soft Independent Modelling of Class Analogies
(SIMCA) models were built for site 1 and were then applied to the classification of samples from sites 2 and 4. The
results reveal that the WT-based spectral region selection is essential to ensure good sensitivity and specificity with
respect to the detection of events associated to the effluent discharges at site 3. In fact, the changes in site 4 caused by the
effluent are masked by other environmental factors when the full spectrum is employed.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Industrial effluents; River Seston; Screening analysis; Near-infrared reflectance spectrometry; Spectral range selection;
Wavelet transform
1. Introduction
The increasing concern with environmental problems
in water resources has brought the need for the
implementation of rigorous monitoring programs by
regulatory agencies. However, many of the classic
analytical methods adopted for monitoring are costly,
laborious, and time-consuming, thus restricting the
ability of the agencies to take appropriate actions in a
timely manner. In view of this problem, much effort has
been placed on the development of alternate optical
techniques that might be able to detect relevant
environmental events in real time. In this context,
near-infrared reflectance (NIRR) spectrometry is an
attractive analytical tool owing to its properties of high
ARTICLE IN PRESS
www.elsevier.com/locate/watres
0043-1354/$-see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.watres.2005.05.018
?Corresponding author. Tel.: +55832167438;
fax: +55832167437.
E-mail address: laqa@quimica.ufpb.br
(M. Ce ´ sar Ugulino Arau ´ jo).

Page 2

sample throughput, non-destructiveness and non-che-
mical consumption.
Several studies have demonstrated the usefulness of
such an instrumental technique for monitoring biologi-
cal, physical and chemical parameters associated to
water quality. Nilsson et al. (1996), established relation-
ships between sediment properties and measured lake
water chemistry (total phosphorus, pH, total organic
carbon) by using partial-least-squares (PLS) modelling
applied to NIRR spectra. Malley et al. (1996) applied
NIRR spectroscopy and multiple linear regression to the
determination of C, N, and P in suspended, particulate
matter in lake waters. Han and Rundquist (1997)
demonstrated the existence of correlation between
algal-chlorophyllconcentration
NIRR spectrum in a case study involving a turbid
reservoir. Dabakk et al. (1999) employed NIRR spectra
of seston material from lakes in Sweden and concluded
that PLS models could be used to predict total organic
carbon, total phosphorus, 420nm absorption and pH.
Christy and Egeberg (2000) used NIRR spectrometry to
predict biopolymer input (carbohydrates, N-acetyl
amino sugars, proteins and polyhydroxy aromatics) of
the natural organic matter from different water sources
in southern Norway. In that work, principal component
analysis (PCA) revealed a cluster structure in the
samples that was in agreement with earlier studies using
classic analytical techniques.
It is worth noting that several works in the field of
multivariate analysis of infrared spectrometric data have
suggested that better prediction (Jiang et al., 2002;
Dantas Filho et al., 2004) and classification (Andrade et
al., 2003; Smith and Gemperline, 2000; Kramer and
Ebel, 2000) results might be achieved through the use of
conveniently selected spectral regions containing rele-
vant information for the problems under consideration,
instead of using the full spectrum. In this context, the
main inconvenient for NIRR spectrometry consists of
the difficulty of finding such regions. This problem is
particularly noticeable for environmental matrices,
owing to complexity of the sample compositions. In
fact, aquatic systems are strongly affected by several
environmental factors, such as geochemical and weather
conditions, as well as antropic activities, which cannot
be easily controlled.
This paper proposes a methodology for screening
analysis (Valca ´ rcel et al., 1999; Trullols et al., 2004) of
changes in the properties of river seston caused by
industrial effluent discharges. The analysis of seston in
aquatic systems is very important to study the fate and
bioavailability of many contaminants that adsorb to the
surface of the particulate matter (Malley et al., 1996).
In order to increase the sensitivity and specificity of
the screening analysis, a strategy for the selection of
regions in the NIRR spectrum based on the wavelet
transform (WT) is used (Jetter et al., 2000; Alsberg et al.,
and thederivative
1998; Walczak, 2000; Galva ˜ o et al., 2004). The strategy
is aimed at identifying regions in which the effect of the
effluent on the optical properties of the seston are more
evident. WT is a local transformation in that each
transformed variable (wavelet coefficient) is related to a
limited spectral range, instead of the full spectrum.
Thus, by using a convenient selection algorithm, it is
possible to isolate wavelet coefficients that are related
only to informative spectral regions. In the present
paper, the relevant regions are considered to be those
associated to changes caused by an effluent discharge
from a textile industry. In the wavelet analysis such
changes are evaluated by comparing the spectra of
samples from sites located upstream and downstream
with respect to the discharge point. After the selection of
relevant spectral regions, a Soft Independent Modelling
of Class Analogies (SIMCA) model (Beebe et al., 1998)
was applied to discriminate seston samples with regard
to changes caused by the effluent. The results demon-
strate that the spectral region selection strategy is an
essential step for a good classification.
2. Materials and methods
This work focused on the River Mumbaba, located in
the industrial district of the city of Joa ˜ o Pessoa (Paraı´ba,
Brazil). Four sampling sites were considered to investi-
gate possible associations between alterations in the
river seston and effluent discharges from a textile
industry: sites 1 and 2 (1100 and 300m upstream of
the discharge point, respectively), site 3 (discharge
point), and site 4 (500m downstream of the discharge
point). The samples were collected monthly from
September 2002 to June 2003.
2.1. Collecting and processing samples
The samples were collected in triplicate for each site
and were stored in 5-L plastic flasks. Seston was
collected from 500ml of each sample on Whatman
GF/C filters (diameter 47mm, pore size 1.0mm) by
vacuum filtration. The filters were placed in Petri dishes
and heated at 501C during 6h, following the procedure
employed by Dabakk et al. (1999), then dried in
desiccators at room temperature (241C). A 10mm-
diameter circle was randomly cut from each filter and
then subjected to a NIRR measurement. The circle
diameter was chosen to be compatible with the size of
the diffuse reflectance accessory specified below.
2.2. NIR measurements
The spectra were recorded using a FT-NIR/MIR
spectrometer Perkin Elmer GX with a spectral resolu-
tion of 4cm?1and 8 scans. The NIRR region in the
ARTICLE IN PRESS
V. Maria de Medeiros et al. / Water Research 39 (2005) 3089–3097
3090

Page 3

range 4014–7398cm?1was used. The blank was taken as
the spectrum of the Whatman GF/C filter after filtration
of distilled water and application of the procedures
described above. The measurements were carried out at
241C and 46% relative humidity by using a Perkin
Elmer diffuse reflectance accessory.
2.3. Software
In order to circumvent the problem of systematic
variations in the baseline, derivative spectra were
employed after smoothing by a Savitzky-Golay filter
(Beebe et al., 1998) with a 2nd-order polynomial and a
29-point window. Spectrum differentiation, Savitzky-
Golay smoothing, and SIMCA modelling procedures
were carried out in the CAMO-AS Unscrambler soft-
ware. The WT-based spectral region selection was
carried out in the Mathworks Matlab platform by using
a lab-made software.
3. WT-based strategy for spectral region selection
WT is a signal processing tool that has found several
applications in the pre-treatment of NIR spectra as
indicated by Depczynski et al. (1999) and Trygg and
Wold (1998). The transform has a multiscale nature in
that it divides the spectrum into intervals of varying
widths that can be treated separately (Strang and
Nguyen, 1996; Coelho et al., 2003). The interval widths
are related to the scale of the analysis: small (large)
scales are associated to the processing of small (large)
intervals. In this sense, WT is a useful tool to determine
the relevance of individual regions in the NIRR
spectrum.
The WT of a spectrum x ¼ ½xðn1Þxðn2Þ...xðnJÞ?, where
nk is the kth wavenumber, can be obtained in a
computationally efficient manner by using a digital filter
bank structure of the form depicted in Fig. 1 (Strang and
Nguyen, 1996; Coelho et al., 2003). The low-pass (H)
and high-pass (G) filters correspond to moving average
and moving difference operations, respectively (Beebe et
al., 1998). Thus, at the first scale level (k ¼ 1), the width
of the intervals analyzed by the WT equals the size of the
moving window chosen by the analyst. The down-
sampling (k2) amounts to moving the window by two-
point steps. The downsampled results of the low-pass
and high-pass filtering operations are termed approx-
imation and detail coefficients, respectively, and are
stored sequentially in vectors c(1) and d(1), as shown in
Fig. 1. The approximation coefficients correspond to
slowly varying features (such as smooth slopes), whereas
the detail coefficients correspond to sharp peaks and/or
oscillations.
The wavelet coefficients at the second scale level
(k ¼ 2) are obtained by re-applying the moving average
and difference operations to the approximation coeffi-
cients c(1) obtained at the first level. The resulting
approximation and detail coefficients, which are sequen-
tially stored in vectors c(2) and d(2), are associated to
intervals whose width is roughly twice the width of the
intervals considered at the first scale level. Such a
procedure is repeated by subjecting the approximation
coefficients to successive filtering iterations, where the
number of iterations, Nitis chosen by the analyst. The
final result of the decomposition of x is a vector
t ¼ [c(Nit)|d(Nit)|d(Nit?1)|y|d(1)], where vectors c(k)
and d(k) are termed, respectively, approximation and
detail coefficients at the kth scale level. Coefficients at
larger scales are associated to wider intervals in the
spectrum, whereas coefficients at smaller scales are
associated to narrower intervals.
3.1. Spectral region selection
Let x1, x3, and x4be the spectra of samples (triplicate
averages) from sites 1 (upstream), 3 (effluent discharge
point), and 4 (downstream), respectively, at a given
month. The WT strategy starts by using the sample from
site 1 as a blank signal for the selection procedure, that
is, by considering difference spectra Dx3and Dx4defined
as
Dx3¼ x3? x1, (1)
Dx4¼ x4? x1.
In this manner, the selection focuses on the changes
with respect to the natural physical and chemical
features of the river.
Moreover, let t3 and t4 be the vectors of wavelet
coefficients obtained by applying a filter bank to Dx3
and Dx4, respectively. The relevance index r(i) of the ith
wavelet coefficient is then assessed according to the
following expression:
(2)
rðiÞ ¼
X
j¼1
M
½tj
3ðiÞ ?¯ t3ðiÞ?½tj
4ðiÞ ?¯ t4ðiÞ?
?????
?????, (3)
where superscript j denotes the jth month, M ¼ 10
months (September 2002 to June 2003), | | stands for
ARTICLE IN PRESS
Fig. 1. Filter bank implementations of the wavelet transform.
V. Maria de Medeiros et al. / Water Research 39 (2005) 3089–3097
3091

Page 4

absolute value, and¯ t3ðiÞ;¯ t4ðiÞ are sample means of the ith
wavelet coefficient for Dx3and Dx4, respectively, that is
¯ t3ðiÞ ¼1
M
X
j¼1
M
tj
3ðiÞ, (4)
¯ t4ðiÞ ¼1
M
X
j¼1
M
tj
4ðiÞ. (5)
In this manner, the wavelet coefficients (and their
associated spectral regions) can be ranked with respect
to their ability to correlate effluent discharges (site 3)
with changes downstream (site 4).
4. Results
4.1. NIRR spectra
Fig. 2a shows the NIRR spectra of the triplicate
average for the four monitored points in December. As
can be seen, the spectra displays a large amount of noise,
as well as baseline drifts, which are typical problems,
associated to NIRR measurements of Seston samples.
To minimize such problems, the first derivative of the
spectra was used after smoothing by a Savitzky-Golay
filter (Beebe et al., 1998) as shown in Fig. 2b. In order to
illustrate the variability of the spectrum within each
triplicate, Fig. 2c presents an enlargement of the
4014–4700cm?1region with twice the standard devia-
tion around the average derivative spectra.
4.2. SIMCA modelling employing the full spectrum
The screening analysis for the seston samples (tripli-
cate averages of the full spectrum) at site 4 was carried
out by using a SIMCA model built for the upstream
region of the river (site 1). The model was also applied to
the samples from site 2, in order to assess the robustness
of the classification. It is expected that samples from site
2, which is also upstream of the discharge point, should
be classified as belonging to site 1, because both sites
correspond to the natural state of the river. On the other
hand, if the screening is to be sensitive to changes caused
by the effluent, samples from site 4, which is down-
stream of the discharge point, should be classified as not
belonging to site 1. In this sense, the reference
information for the SIMCA analysis is the location of
the sites with respect to the discharge point.
Fig. 3 presents a three-dimensional SIMCA score plot
that illustrates the scattering of the modelling (site 1)
and test (sites 2 and 4) objects. The construction of the
SIMCA model, which was based on samples from site 1,
was carried out by using all the triplicate spectra, rather
than the triplicate averages, in order to convey more
information into the modelling process. By using the
ARTICLE IN PRESS
Fig. 2. (a) Original and (b) derivative NIRR spectra of the
triplicate average for each of the four sites in December. (c)
Enlargement of the 4014–4700cm?1region with twice the
standard deviation (thin lines) around the average derivative
spectra (thick lines). (An offset was added for clarity).
V. Maria de Medeiros et al. / Water Research 39 (2005) 3089–3097
3092

Page 5

default settings of the chemometrics software (CAMO-
AS Unscrambler), the optimum number of principal
components to be included in the model was found to be
eight, which accounts for 87.1% of the modelling data
variance. The test objects from sites 2 and 4 were
represented by their triplicate averages.
As a result of the SIMCA classification, all samples
from site 2 were classified as belonging to site 1, as
expected. However, the full-spectrum SIMCA model
was not sensitive to changes in the site-4 Seston (after
the discharge point). In fact, all site-4 samples were
classified as belonging to site 1, with the exception of the
March sample (indicated by a circle in Fig. 3).
4.3. SIMCA modelling employing the WT-selected
spectral region
The Symlet 6 WT (Daubechies, 1992) with three
decomposition levels was employed for the selection of
the most relevant region for monitoring the changes
in the seston samples after the effluent discharge,
site 4. This WT corresponds to a moving window
size of 12 points at the first scale level. By taking the
wavelet coefficient with the largest relevance index, the
selection resulted in the region between 4271 and
4348cm?1.
Fig. 4 presents a new score plot resulting from the
construction of a SIMCA model for site 1 by using only
the WT-selected region. Eight principal components
were employed, as in the previous case, in order to
enable a direct comparison with the model obtained
from the full spectral data. As before, all site-2 samples
were correctly classified as belonging to site 1. However,
in addition to the event that was detected by the full-
spectrum model (March), three other events were
detected in site 4 (November, December, and January).
These results suggest that restricting the screening
analysis to the WT-selected region makes the model
more sensitive to the occurrence of events (indicated by
circles in Fig. 4).
In order to gain a better understanding of the
classification carried out by the new SIMCA model, an
enlargement of the derivative spectra in the WT-selected
region (4271–4348cm?1) is presented in Figs. 5a–d
for the four months in which events were detected
(November, December, January, March). In these
ARTICLE IN PRESS
Fig. 3. SIMCA score plot obtained from the full spectral data. Objects from sites 1, 2, and 4, are indicated by the number of the
respective sites. Principal components PC1, PC2, and PC3 explain 65.8%, 5.2%, and 3.7% of the modelling data (site 1) variance,
respectively. The entire model comprises eight principal components. The circle indicates an object from site 4 (March) that, according
to the SIMCA model, is significantly different from the site-1 objects.
V. Maria de Medeiros et al. / Water Research 39 (2005) 3089–3097
3093