Characterization of estuarine sediments by near infrared diffuse reflectance spectroscopy

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/aca

Characterization of estuarine sediments by near infrared diffuse reflectance spectroscopy Javier Moros a , María C. Barciela-Alonso b , Paula Pazos-Capeáns b , b ˜ Pilar Bermejo-Barrera b , Elena Pena-Vázquez , Salvador Garrigues a , Miguel de la Guardia a,∗ a

Department of Analytical Chemistry, University of Valencia, 50 Dr. Moliner Street, 46100 Burjassot, Valencia, Spain Department of Analytical Chemistry, Nutrition and Bromatology, University of Santiago de Compostela, Avenida das Ciencias s/n, Santiago de Compostela E-15782, Spain b

a r t i c l e

i n f o

a b s t r a c t

Article history:

It has been developed a partial least squares near infrared (PLS-NIR) method for the deter-

Received 9 May 2008

mination of estuarine sediment physicochemical parameters. The method was based on

Received in revised form

the chemometric treatment of first order derivative reflectance spectra obtained from

10 June 2008

samples previously lyophilized and sieved through a lower than 63 ␮m grid. Spectra

Accepted 12 June 2008

were scanned from 833 to 2976 nm, averaging 36 scans per spectrum at a resolution

Published on line 26 June 2008

of 8 cm−1 , using chromatographic glass vials of 9.5 mm internal diameter as measurement cells. Models were built using reference data of 31 samples selected through the

Keywords:

use of a hierarchical cluster analysis of NIR spectra of sediments obtained from the

Partial-least-squares

Ria de Arousa estuary and prediction parameters were established from a validation set

Near infrared

of 50 samples of the same area. pH, redox potential (Eh), carbon (C), nitrogen (N) and

Diffuse reflectance

hydrogen (H) content together with Sn, Pb, Cd, As, Sb and total Cr and also acid solu-

Marine sediments

ble, reducible and oxidable Cr fractions were employed as characteristic parameters of

Hierarchical cluster analysis

the studied sediments. Standard error of prediction values for C and N content were of the order of 4 and 1.3 mg g−1 for H. Prediction errors for pH and Eh were 0.15 units and 37 mV, respectively, thus indicating the good prediction capabilities of the method. Regarding trace metal concentrations PLS-NIR provided prediction error levels for unknown samples around 20% for Sn, Pb, As and Sb and root mean square errors of prediction around 40% for concentration levels of 400 ng g−1 Cd and 100 ␮g g−1 Cr. For the different extractable fractions of Cr the residual prediction deviation varied from 1.3 to 1.7 but relative errors found for samples of the validation set were only useful for screening purposes. © 2008 Elsevier B.V. All rights reserved.

1.

Introduction

Estuaries are semi-enclosed coastal bodies of water with one or more rivers and a free connection to the open sea and they

∗

Corresponding author. Tel.: +34 96 354 4838; fax: +34 96 354 4838. E-mail address: [email protected] (M. de la Guardia). 0003-2670/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2008.06.030

are often characterized by sedimentation, carried from terrestrial runoff and from offshore [1]. Estuaries are dynamic marine environments, whose properties, such as pH, salinity, sediment type, biological habitats,

114

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

localized depositional and transport processes and water level vary, depending on the river that feeds the estuary and the ocean but especially on the human activities, which can modify the physical and chemical characteristics of the system [2]. Chemists, oceanographers, and geologists have identified numerous chemical and biological markers or indicators for chemical and physical characterization of sediments. Sediment pH is the master variable controlling the speciation and bioavailability of metals as well as redox potential (Eh) of the sediment is an important factor controlling the partitioning of metals. Both pH and redox potential measurements provide useful process-related information on the nature of sediments under investigation [3]. However, in dynamic areas, it must be taken into consideration that changes on chemical composition may be so rapid that information can considerably change for different sampling times. For this fact, detailed monitoring is needed to investigate seasonal and geographical chemical and physical variations of surface marine sediments [4]. Conventional methods recommended for sediment quality assessment are generally based on the use of spear tip pH electrodes and millivolt reader and graphite or platinum electrodes with combination silver-silver-chloride or calomel reference electrodes, for pH and redox potential measurements, respectively; the use of elemental analyzers to determine particulate carbon (PC) and particulate nitrogen (PN) [5]; and methods commonly involving wet digestion of solid samples in hot concentrated acids followed by inductively coupled plasma (ICP) or atomic absorption spectrometry (AAS) for trace metal determination [6]. The aforementioned methods are in general expensive, tedious, complex and highly time-consuming. Moreover, for monitoring programs that require the analysis of many parameters in a great number of samples with limited resources, fast, cheap and accurate methods are required for the simultaneous screening of as much as possible parameters. In spite of that there are no much references in the literature related to the determination of soils and sediments parameters using fast non-destructive procedures [7]. Near infrared (NIR) spectroscopy provides an unique tool for the determination of many parameters in solid samples and, based on the use of appropriate mathematical models it can be predicted both quality parameters, composition and properties, of samples under study. A bibliographic search on applications of reflectance measurements to determine chemical, physical and biochemical properties of soil [10–17,19–23] or sediment [8,9,18] samples by near infrared (NIR) measurements (see Table 1 ), evidence that visible plus near-infrared regions, compressing between 400 and 2500 nm together with multivariate mathematical treatment, like partial least squares (PLS), modified partial least squares (MPLS), or principal component regression (PCR) [21] and the use of classification and regression trees [16] have been used for the determination of several parameters in this kind of samples. However, the most outstanding inconvenient of the existing literature in this field is the extended use of crossvalidation to evaluate their results [10,11,13,15,18,19,21] which offers guarantees on the model coherence but is unsuitable to

evaluate the capabilities of the developed methods as alternative procedures for the prediction of the aforementioned properties in unknown samples. The aim of this work has been the development of a fast, accurate and reagent free analytical method useful to evaluate the physicochemical characterization of marine surface sediments from the Ria de Arousa employing diffuse reflectance NIR measurements and multivariate calibration.

2.

Experimental

2.1.

Apparatus and reagents

A Fourier transform near infrared (FT-NIR) spectrometer Bruker model Multipurpose Analyzer (MPA) controlled by OPUS® for Windows® software from Bruker Gmbh (Bremen, Germany) and equipped with an integrating sphere, used as measurement accessory, was employed for NIR spectra acquisition. For instrumental and measurement control and data acquisition it was employed the OPUS program (Version 4.2) from Bruker. Spectra treatment and data manipulation were carried out using Omnic 6.1 software from Nicolet (Madison, WI, USA). PLS calibration models were established using the TurboQuant Analyst 6.0 software developed by Thermo Nicolet Corp. To obtain the reference data of samples the following instruments were used. A Model 1100B atomic spectrometer (PerkinElmer Life and Analytical Sciences, Shelton, CT, USA) equipped with and HGA-700 graphite furnace atomizer, AS-40 auto-sampler, and deuterium background correction was used for trace element determination. Sources of radiation were hollow cathode lamps operating at 18.0, 8.0, 12.0, 10.0, 25.0, 12.0 mA for As, Cd, Cr, Pb, Sb and Sn, respectively, with a spectral bandwidth of 0.7 nm for all studied elements. The wavelengths 193.7, 228.8, 357.9, 283.3, 206.8 and 224.6 nm for As, Cd, Cr, Pb, Sb and Sn, respectively, were used. Pyrolytic graphite tubes with platforms were used for all studied elements. More details about these procedures can be checked in [24–27]. Stock standard solutions (1 mg mL−1 ) of As, Cd, Cr, Pb, Sb and Sn (Merck, Darmstadt, Germany) were used. Each test solution was prepared with ultrapure water immediately before use. Magnesium nitrate Suprapur (Merck, Darmstad, Germany). Reference material PACS.-2: Marine sediment, National Research Council Canada (NRCC) (Ottawa, Ont., Canada) was also used. Palladium solution was prepared by dissolving 300 mg of palladium (99.999% purity) (Aldrich, Milwaukee, WI, USA) in 1 mL of concentrated nitric acid, all diluted to 100 mL with ultrapure water. If dissolution was incomplete, 10 ␮L of hydrochloric acid (AnalR, 35% BDH) was added to cold nitric acid and heated to gentle boiling in order to volatilize excess chloride. Argon N-50 (99.999% purity) as sheath gas for the atomizer and for internal purge was used.

Table 1 – State-of-the-art of previously published for the characterization of soil and sediment samples by vibrational procedures Determined properties

Chemometric technique

Freshwater sediment

Fe, Mn, Zn, Cu, Pb, Ni and Cd

Partial least squares regression (PLS)

3 for Pb; 10 for Cd

1100–2500

First derivative

169

Soil from different locations across Uruguay

Silt, sand, clay, Ca, K, Na, Mg, Cu and Fe Total Fe, Zn, Pb, Cd, Cu and Ni

Modified PLS (MPLS)

–

400–2500 (vis–NIR)

First derivative

PLS

3 for Pb; 3 for Cd

400–2500; 2500–25,000 (MIR)

Polluted soils

As, Pb, Cu and Zn

MPLS

–

Forest soil organic horizons

Organic C, total N, and total S content, total and exchangeable Zn and Pb Ni, Cr, Co, Cd and Fe Phosphorous

MPLS

Soil from a metal mining region

Soils Soils

Number of factors

Spectral range (nm)

Spectral data

Total number of samples

Validation procedure

Prediction figures of merit

Reference

Pb: SEP = 3.83 ␮g g−1 , RPD = 2.45; Cd: SEP = 0.32 ␮g g−1 , RPD = 1.74

[9]

332

Odd-numbered spectra as calibration set and even-numbered spectra as prediction set Cross-validation

–

[10]

Raw data

70

Cross-validation

[11]

400–2500

First derivative

100

70 samples for calibration and 30 for an external validation

3 for Pbtotal ; 5 for Pbexchangeable ; 3 for Corg ; 3 for Ntotal

400–2500

Second derivative

74

Group cross-validation

Pb: RMSD = 839 mg kg−1 ; Cd: RMSD = 5.13 mg kg−1 As: SEP = 21.17 mg kg−1 , RPD = 2.25; Pb: SEP = 45.38 mg kg−1 , RPD = 2.13 Pbtotal : RPD = 1.3; Pbexchangeable : RPD = 1.1; Corg : RPD = 1.6; Ntotal : RPD = 1.5

PLS

–

400–2500

–

–

–

–

[14]

PLS

–

400–2500

Raw data

200

Cross-validation

–

[15]

[12]

[13]

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

Sample

115

116

Table 1 (Continued ) Determined properties

Chemometric technique

Soils

Total C, N, and P, organic C, pH, organic matter, various cation and P extractions, extra cellular enzyme activity Humus, organic and total C, total N and several humus substances

Classification and regression trees

–

400–2460

First derivative

MPLS

–

400–2500

Total phosphorus, pH, and total organic carbon Clay, total carbon, silt and sand Soil organic carbon

PLS

–

PLS

PLS

Soils

Lake sediments

Soil from all regions of Denmark Agricultural soils

Number of factors

Spectral range (nm)

Spectral data

Total number of samples

Validation procedure

Prediction figures of merit

Reference

273

Calibration using 67% of the data; validation performed using the remaining data (33%)

Ctotal : RPD = 5.74; Ntotal : RPD = 2.67; pH: RPD = 2.39

[16]

First derivative

127–175

Corg : SEP = 0.107; CDumas : SEP = 0.117; Nkjeldahl : SEP = 0.026; NDumas : SEP = 0.009

[17]

400–2500

Raw data

25 for TOC; 52 for pH

Corg : CAL = 127/VAL = 48; CDumas : CAL = 111/VAL = 24; Nkjeldahl : CAL = 91/VAL = 36; NDumas : CAL = 111/VAL = 24 Internal cross-validation

TOC: SEP = 2.52 mg L−1 ; pH: SEP = 0.4

[18]

12

400–2500

32–471

Cross-validation

Ctotal : SEP 0.42–0.57, RPD 2.4–1.4

[19]

2–9

350–2500

Combinations of pretreatments Combinations of pretreatments

122

Calibration using 3/4 of the data; validation performed using the remaining data 1/4

SEP: 2.4–3.3 g kg−1

[20]

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

Sample

Table 1 (Continued ) Soils from four major land resource areas

Soils

PCR

7 for Ctotal ; 7 for Ntotal ; 13 for pH

1300–2500

First derivative

802

Cross-validation

Ctotal : RPD = 2.79; Ntotal : RPD = 2.52; pH: RPD = 1.43

[21]

PCA-PLS

7 for N; 3 for pH

350–2500

Raw data

165

N: SEP = 3.27 mg kg−1 ; pH: SEP = 0.075

[22]

PLS

7

400–2500

Raw data

125

135 samples for calibration and 30 samples for validation 95 samples for calibration and 30 samples for prediction

N: SEP = 3.28 mg kg−1

[23]

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

Soils

Total C and N, moisture, potentially mineralizable N, respiration rate, biomass C, CEC, pH, aggregation: macro, clay, silt, and sand, extractable metals: Ca, Cu, Fe, K, Mg, Mn, Zn, and P, exchangeable bases: Ca, Mg, Na, and K, sum of extractable bases, exchangeable acidity N, P, K, organic matter and pH Nitrogen and organic matter

Note: Number of factors and prediction figures of merit were only reported for that parameters evaluated under the present study.

117

118

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

Ultrapure water, with resistivity of 18 M cm obtained using a Millipore Milli-Q water purification system (Bedford, MA, USA), was employed for preparing appropriate solutions. Triton X-100 (4-(1,1,3,3-Tetramethylbutyl)phenylpolyethylene glycol), for gas chromatography, was obtained from Merck. All glassware was kept in 10% nitric acid for at least 48 h and washed three times with ultrapure water before use.

2.2.

Samples

A total number of 81 marine sediment samples were taken, at four different sampling periods (winter and summer seasons of years 2002 and 2003). Marine sediment samples were collected, with a grab on board of the R/V Mytilus, from 21 stations, on which the area of the Ria de Arousa (North West of Spain) was divided. Fig. 1 shows the location of sampling stations. Stations 10 and 17 were not sampled due to their inaccessibility by boat, as well as station 7 with a rocky morphology free from sediments. The surface layer of each sample was removed with a polyethylene spoon to avoid contamination, and stored in hermetic polyethylene bottles at −20 ◦ C (previously cleaned with 10% nitric acid for 24 h and rinsed with Milli-Q water). Sediments were lyophilized and sieved through a nylon grid to separate the fraction with particle size lower than 63 ␮m and stored in polyethylene bottles. NIR spectra were obtained directly from untreated samples closed in 2 mL chromatographic glass vials. Data of trace element concentrations were determined experimentally on sample slurries, by using reference procedures based on electrothermal atomic absorption spectrometry. Sediments included in the work contain from 3.5 to 26.2 ␮g g−1 Sn, between 27.4 and 83.3 ␮g g−1 Pb, between 0.13 and 0.97 ␮g g−1 Cd, from 15.0 to 43.3 ␮g g−1 As, from 1.5 to 5.3 ␮g g−1 Sb and between 41.5 and 365 ␮g g−1 Cr, ranging from high concentrations in the inner part of the Ria, near to the port and urban nucleus, such as Vilagarcía or Rianxo, with an important industrial activity, to natural background levels toward the mouth of the Ria. Carbon, nitrogen and hydrogen content of sediments were determined by gas chromatography in an elemental analyzer, Fison model EA 1108 CHNF-O, equipped with a thermal conductivity detector (Beverly, USA), being found a variability range from 36.1 to 65.2, from 7.2 to 15.75, and from 0.23 to 6.30 mg of C, N and H, respectively, per g of sediment. Physical parameters like pH ranged between 7.41 and 8.48 and potential (Eh) varied from 267 to 408 mV and were experimentally measured in the sediment fraction between 200 and 63 ␮m by using a pH-meter, model 8102BN (thermo Orion, Beberly, USA) and an ORP electrode, model Inlab 501 (Mettler Toledo, Greifensee, Switzerland), respectively.

2.3.

Diffuse reflectance near infrared analysis

Samples were placed in the same temperature controlled room where the spectrometer was located before to carry out the analysis. Triplicate diffuse reflectance NIR spectra, for each sample, were obtained directly from the samples inside 2 mL standard

glass chromatography vials (12 mm × 32 mm) of 9.5 mm internal diameter used as measurement cell. Taking into account previous works [28,29], the following instrumental conditions were selected to obtain the appropriate measurements: sample spectra were scanned from 833 to 2976 nm (12,000 to 3360 cm−1 ) by averaging 36 scans per spectrum using a nominal resolution of 8 cm−1 which take 16.8 s as measurement time per spectrum. Spectra were recorded in Kubelka–Munk units and corrected for the different penetration of the NIR radiation as a function of the wavelength. The background spectrum was acquired from the integrating sphere using the same instrumental conditions than those employed for sample measurement. High background stability was found in all the cases. Calibration and validation datasets were established from previously analyzed samples selected as a function of the dendrographic distribution of their NIR spectra. The appropriate PLS models were built using the best pre-processing method, wavelength range and number of factors for physical (pH and Eh) and chemical (C, N, H, Sn, Pb, Cd, As, total Cr, and acid fraction of Cr, oxidable Cr and reducible Cr) characterization of sediments from a calibration set of analyzed samples and applied for the prediction of the aforementioned parameters in a validation set of samples taken from the same estuary but different than those used for calibration.

2.4.

Chemometric data treatment

Hierarchical cluster analysis of the diffuse reflectance NIR spectra of sediments was made in order to evaluate the number of characteristic subsets in which the available samples could be divided. It was carried out using OPUS program (Version 4.2) from Bruker. As already published in previous works [28,30] similar criteria, as those already used for other type of samples, were employed through this work. Appropriate partial least square (PLS) multivariate models were built using the TurboQuant Analyst 6.0 software developed by Thermo Nicolet Corp. Data obtained from Opus were previously exported in JCAMP-DX format. The optimum number of PLS factors was taken from the minimum of the resulting graph of predicted residual error sum of squares (PRESS). Prediction accuracy was established by using the quality coefficient (QC) [31], which gives an indication of the percentage error to be expected for the estimated parameters in samples not used for calibration. Additional figures related to the model’s fit and their predictive power; such as, the root mean square error of calibration (RMSEC), the root mean square error of crossvalidation (RMSECV) and the root mean square error of prediction (RMSEP) were also used through this study. Various quality indicators; such as the mean difference (d(x−y) ) between the predicted diffuse reflectance NIR-PLS values and the reference data (Ci ), and the standard deviation of the mean difference (s(x−y) ), as well as strip and sreg , the deviation between triplicate determinations and the pooled standard error of prediction for validation samples, respectively, were used to evaluate the predic-

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

119

Fig. 1 – Sampling sediment stations in Ria de Arousa.

tion capability of the proposed methodology [32]. In all cases, the aim was to estimate the average deviation of the model from the reference data and the actual possibilities to use NIR spectra of sediments as screening of physical and chemical characteristics of these samples. To evaluate the predictive ability of the PLS models, the residual predictive deviation (RPD), defined as the ratio between the standard deviation of the population (S.D.) and the RMSEP for the NIR calibration, was also calculated [33]. If the error of estimation for a property (RMSEP) is high as compared with the spread in composition of that property in the standard population, and therefore has a relatively small RPD, the NIR calibration model cannot considered as robust. The increase of the value of RPD means an improvement of the power of the model to predict accurately the considered property.

In spite that several spectral windows were tested for evaluating the prediction capabilities of the corresponding models using the validation set, only the most significant information is presented through this paper.

2.5.

Cluster analysis

Cluster analysis encompasses a number of different algorithms and methods for grouping objects of a similar type into respective categories. In hierarchical cluster analysis, the similarity between samples is established using the concept of distance, between samples, which is related to how similar the numerical properties of sample NIR spectra are. Each sample is linked to the closest sample or group of samples and a characteristic distance is used to describe this union. For the present study, the Euclidean normal and Ward methods were used.

120

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

Fig. 2 – Diffuse reflectance NIR spectra of surface marine sediments sampled in Ria de Arousa. Instrumental conditions: 8 cm−1 nominal resolution and 36 cumulated scans per spectrum. Note: Inset shows a detail or the first order derivative spectra of surface marine sediments.

It must be noticed that through this study each one of the replicate spectra obtained for each sample was considered separately in order to evaluate the repeatability of the results and as a guide for the correct classification of samples.

3.

Results and discussion

3.1.

Surface marine sediments NIR spectra

All sediments tested in this study provided similar diffuse reflectance NIR spectra. Fig. 2 shows the spectra of different marine sediment samples in the region compressed between 833 and 2976 nm (12,000 to 3360 cm−1 ). All spectra have three major absorption peaks (around 1400, 1900, and 2200 nm) in the NIR region. Reflectance characteristics of soils are related to chemical groups present in the organic matter, thus chlorophyll, oil, cellulose, pectin, starch, lignin, and humic acids are the spectrally active groups in the NIR (400–2500 nm) region [34]. In general, soil reflectance decreases on decreasing the organic matter content [35]. Bands located at 1414 and 1914 nm could be related with O–H bands (water) and that present at 2208 nm could be related probably to C–H absorption and combination bands. The band at 2080 nm could be associated to absorption of the amine N–H group and CONH2 groups being the bands in this region associated with protein chemical structures. Bands at 2260 nm were related with absorption of C–H and C–H combination tones and that present at 2344 nm to the absorption of C–H and CH2 (cellulose). Additionally few small bands located

between 2200 and 2500 nm, can be also identified in all the samples considered.

3.2.

Diffuse reflectance NIR spectra treatment

To relate the physical and chemical characteristics of marine sediment samples with their NIR spectra PLS regression models were built based on the first derivative spectra in the appropriate selected spectral range. Derivatives were calculated by applying Savitzky–Golay filter being spectra smoothed over segments of 7 data points and defining the shape of the curve through a third order polynomial function. In spite that samples have been previously sieved, to remove the shift effects of possible differences in particle size, standard normal variate (SNV) correction and de-trending (linear removed baseline correction) procedures were applied to the spectra. SNV scales each spectrum to have a standard deviation (S.D.) of 1.0, and de-trending removes linear or curvilinear trends in diffuse reflectance data.

3.3.

Cluster classification of marine sediments samples

In order to evaluate possible groups among samples considered, a clustering method was applied before multivariate data treatment. As previously stated [29,30], this task is an important step to the proper selection of a representative calibration set according the variability of samples under study, thus improving the prediction capabilities of the models. Taking into account our experience [32] we selected dendrogram classification using Euclidean distance with Ward

Table 2 – Characteristics of marine sediment samples classified into clusters after dendrographic treatment of diffuse reflectance NIR data Cluster index

C (mg g−1 )

Number of samples

Mean

±S.D.

N (mg g−1 ) Mean

±S.D.

H (mg g−1 ) Mean

pH

Eh (mV)

±S.D.

Mean

±S.D.

Mean

Samples

±S.D.

8

42

4

12.5

1.1

4.4

1.5

7.71

0.16

308

24

2 3

4 8

39.2 42.4

1.6 1.6

13.5 12.5

0.5 0.5

4.4 4.6

0.4 0.3

7.59 7.82

0.10 0.05

286 295

19 21

4

5

43

4

10.9

1.8

4.5

1.6

7.81

0.09

294

27

5

7

43

6

11

2

4.7

1.0

7.92

0.15

319

42

6

4

44

3

14

2

4.6

1.6

7.74

0.06

282

4

7

13

47

7

10.7

1.5

4.1

1.2

8.0

0.2

314

52

8 9

4 9

49 43.1

7 1.0

10.1 10.3

1.4 1.3

2.8 3.6

1.9 1.1

8.10 7.93

0.15 0.08

283 314

13 34

10

9

45

3

11.2

1.6

4.1

1.5

7.93

0.13

295

18

11

8

45

2

9.6

1.3

3.6

1.0

7.96

0.11

288

20

12

2

54

5

10.0

0.8

4.1

0.3

7.71

0.17

274

2

1 A, 24 B, 24 A, 20 D, 20 B, 21 B, 21 A, 24 C 1 B, 1 C, 1 D, 24 D 2 A, 22 A, 2 C, 3 B, 22 B, 2 B, 2 D, 3 C 11 C, 11 D, 18 A, 18 B, 18 C 6 B, 11 A, 11 B, 20 A, 20 C, 23 B, 21 D 21 C, 22 C, 23 A, 23 D 3 A, 8 B, 3 D, 8 D, 14 C, 5 D, 15 B, 14 B, 6 A, 9 A, 12 B, 12 C, 18 D 4 A, 6 C, 16 D, 12 A 4 B, 19 A, 4 D, 19 B, 14 A, 8 A, 19 D, 8 C, 13 B 4 C, 5 C, 16 B, 15 A, 16 C, 15 C, 5 A, 19 C, 22 D 5 B, 16 A, 13 C, 6 D, 13 A, 13 D, 14 D, 15 D 9 D, 12 D

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

1

121

122

Table 2 (Continued ) Cluster index

Number of samples

[Sn]

[Pb]

[Cd]

[As]

[Sb]

[Cr]total

[Cr]acid soluble

Mean ±S.D. Mean ±S.D. Mean ±S.D. Mean ±S.D. Mean ±S.D. Mean ±S.D. Mean

±S.D.

[Cr]reducible

[Cr]oxidable

Mean ±S.D. Mean ±S.D.

8

17

4

56

7

0.68

0.17

36

5

3.5

1.0

170

70

0.33

0.18

0.9

0.7

50

30

2 3

4 8

22 15

3 3

74 56

10 11

0.8 0.53

0.2 0.15

34 27

4 4

4.0 4.0

1.1 0.6

190 120

120 50

0.37 0.28

0.18 0.10

0.9 0.7

0.2 0.2

90 40

40 14

4

5

12

2

43

5

0.31

0.03

19

3

2.76

0.18

80

30

0.21

0.06

0.48

0.16

24

12

5

7

13

4

50

10

0.5

0.2

29

9

3.3

1.2

100

50

0.21

0.05

0.5

0.2

25

14

6

4

18

3

54

9

0.60

0.11

33

5

4.5

0.6

160

40

0.26

0.10

0.7

0.4

50

20

7

13

10

3

44

8

0.23

0.05

24

3

2.5

0.7

80

30

0.15

0.05

0.22

0.19

19

9

8

4

9

4

39

5

0.19

0.07

25

5

2.2

0.7

70

12

0.15

0.05

0.25

0.14

12

5

9

9

11

3

44

8

0.22

0.04

25

4

2.6

0.6

80

30

0.15

0.03

0.29

0.16

24

14

10

9

11

3

44

10

0.23

0.14

26

4

3.0

1.0

110

60

0.21

0.13

0.4

0.4

20

20

11

8

9

2

44

5

0.20

0.03

25

3

2.2

0.5

71

17

0.13

0.02

0.13

0.10

14

5

12

2

11.1

1.3

48.5

0.5

0.28

0.02

22.8

1.9

2.00

0.14

56

15

0.17

0.02


8.4

0.4

1 A, 24 B, 24 A, 20 D, 20 B, 21 B, 21 A, 24 C 1 B, 1 C, 1 D, 24 D 2 A, 22 A, 2 C, 3 B, 22 B, 2 B, 2 D, 3 C 11 C, 11 D, 18 A, 18 B, 18 C 6 B, 11 A, 11 B, 20 A, 20 C, 23 B, 21 D 21 C, 22 C, 23 A, 23 D 3 A, 8 B, 3 D, 8 D, 14 C, 5 D, 15 B, 14 B, 6 A, 9 A, 12 B, 12 C, 18 D 4 A, 6 C, 16 D, 12 A 4 B, 19 A, 4 D, 19 B, 14 A, 8 A, 19 D, 8 C, 13 B 4 C, 5 C, 16 B, 15 A, 16 C, 15 C, 5 A, 19 C, 22 D 5 B, 16 A, 13 C, 6 D, 13 A, 13 D, 14 D, 15 D 9 D, 12 D

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

1

Note: ±S.D. refers to the standard deviation of the mean. All trace metal concentration values are expressed in ␮g g−1 units.

Samples

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

123

Fig. 3 – Dendrographic classification of marine sediment samples. The Euclidean distance after vector normalization spectra and the Ward linkage method were used. For details about cluster group composition see data in Tables 2 and 3. Samples are identified with a number (corresponding to the sampling site) and a letter (corresponding “A” to Winter 2002, “B” to Summer 2002, “C” to Winter 2003 and “D” to Summer 2003).

linkage upon considering the frequency range between 833 and 2796 nm on the vector normalized spectral data. Fig. 3 shows the dendrographic classification of samples and, as it can be seen, from this figure, 12 different types of samples could be identified for a cut-off value of 0.09.

The main groups of clusters formed (from left to right) are directly correlated with the intensity of the NIR spectra of sediment samples and, thus, samples with high diffuse reflectance level are grouped together. As the diffuse reflectance intensity seems to be mainly related with the carbon, hydrogen and

124

Table 3 – Descriptive statistics of calibration and validation datasets used for diffuse reflectance PLS-NIR analysis of marine sediment samples Set

Number of samples

Trace metals [Sn]

Calibration Validation

Set

[Pb]

Mean

±S.D.

Mean

13 12

5 4

49 49

31 50

[Cd]

±S.D. 11 11

±S.D.

Mean 0.4 0.4

0.2 0.2

[Sb]

Mean

±S.D.

27 27

7 6

[Cr]total

±S.D.

Mean 3.1 3.0

Number of samples

1.0 1.0

Mean 110 104

[Cr]acid soluble

±S.D.

±S.D.

Mean

70 50

0.23 0.20

0.14 0.10

[Cr]reducible Mean 0.5 0.4

[Cr]oxidable

±S.D.

Mean

0.5 0.3

33 29

±S.D. 34 17

Other parameters

±S.D.

Mean 31 50

N (mg g−1 )

43 45

Mean

4 5

11 11.4

±S.D.

H (mg g−1 ) ±S.D.

Mean

2 1.7

4.1 4.1

1.1 1.3

pH Mean

Eh (mV) ±S.D.

Mean

±S.D.

0.19 0.17

293 306

26 35

7.87 7.88

Note: All concentration values for trace metals are expressed in ␮g g−1 units.

Table 4 – Prediction capabilities of PLS-NIR models for the determination of physicochemical parameters of marine sediment samples from Ria de Arousa using diffuse reflectance spectra Wavenumber range (nm) pH Eh (mV) C (mg g−1 ) N (mg g−1 ) H (mg g−1 )

1247–2249 997–2501 1956–2655 1247–2684 1247–2671

PLS factors 3 2 3 4 5

R2 0.85 0.64 0.85 0.97 0.98

RMSEC 0.07 15 1.7 0.3 0.15

RMSECV 0.18 30 4 1.5 1.1

RMSEP 0.13 30 4 1.5 1.1

RRMSEP % 1.7 11 9 13 27

d(x−y) 0.03 −10 −2 −0.8 −0.3

s(x−y) 0.13 30 3 1.3 1.1

QC %

strip

sreg

1.6 11 9 14 29

0.05 9 1.2 0.5 0.4

0.15 37 4 1.8 1.3

RPD 1.46 1.17 1.25 1.13 1.18

Note: RRMSEP is the RMSEP divided by the mean value of each parameter in the validation dataset. strip is the standard deviation of three replicates. d(x−y) and s(x−y) are the mean difference and the standard deviation of mean differences between predicted versus actual values of each parameter, respectively. QC is the quality coefficient. RPD is the residual predictive deviation. For additional details see the text.

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

C (mg g−1 )

Calibration Validation

[As]

is the standard deviation of three replicates. versus actual metal content value, respectively Note: RRMSEP is the RMSEP divided by the mean value of metal content in the validation dataset. strip d(x−y) and s(x−y) are the mean difference and the standard deviation of mean differences between predicted . QC is the quality coefficient. RPD is the residual predictive deviation.

1.6 1.3 1.7 1.7 49 0.13 25 0.4 17 0.05 13 0.12 40 0.10 20 0.3 −3 0.02 −2 0.0 0.83 0.71 0.96 0.89 1080–2253 840–2677 894–2696 1370–2031 Cr Total Acid fraction Oxidable Reducible

2 2 4 3

30 0.08 6 0.15

60 0.13 20 0.4

45 0.11 20 0.3

40 50 70 70

60 50 500 110

1.7 1.2 1.3 1.4 1.7 4 10 0.18 7 0.7 1.4 3 0.06 0.3 0.19 3 9 0.14 5.0 0.60 0 −2 −0.06 −0.9 −0.10 0.99 0.93 0.998 0.43 0.99 1247–2677 995–2696 1256–2620 995–2245 1363–2627 Sn Pb Cd As Sb

5 3 6 1 5

0.6 3 0.009 5 0.10

4 8 0.16 5 0.7

3 9 0.15 5 0.6

20 18 40 19 20

20 18 44 20 19

sreg strip QC % s(x−y) d(x−y) RRMSEP % RMSEP RMSECV RMSEC R2 PLS factors Wavenumber range (nm) Trace metal (␮g g−1 )

Table 5 – Prediction capabilities of PLS-NIR for the determination of trace metals in marine sediment samples from Ria de Arousa using diffuse reflectance spectra

RPD

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

125

nitrogen composition, clusters could basically relate with the similar content of these analytes between the samples. Table 2 details the mean and the standard deviation values of sediment parameters such as C, N, H content as well as pH and redox potential, for the 12 clusters obtained, ordered from the top to the bottom of Fig. 3. In spite that it is clear that metal ions at trace levels do not show absorption features in the NIR region, it was detected a slight correlation between clusters structure and trace metal content of sediments as they may be detectable because of their possible complexation with organic matter or associated with hydroxides, sulphides, carbonates, or oxides that are detectable [36]; or adsorbed to clays that absorb light in this wavelength range [37]. Mean and standard deviation values of trace metal content, for the aforementioned 12 clusters are also detailed in Table 2. As can be seen from sample indicators, no strict correlation was found between the different NIR clusters identified and data available of different sampling points or seasonal period from which sediments were collected.

3.4.

Selection of the calibration set

The number and nature of samples used for calibration are always critical factors in multivariate analysis. Because of that, the dendrogram of Fig. 3 was used to select the calibration and validation data sets. As selection criterion, it was used at least one sample from each cluster for built the calibration set. In the case of clusters containing several samples, the square root of the total number of samples included in the cluster was selected for calibration. The remaining samples were integrated in the validation set. For both sets, samples were randomly selected from each cluster. Based on the aforementioned considerations, we built the calibration models using 31 samples. Predictive capabilities and analytical features of models were established using 50 samples. As shown in Table 3, the average data and data dispersion of all parameters evaluated are very similar for both considered sets.

3.5. Sediment pH, redox potential, carbon, nitrogen and hydrogen content For each one of the properties considered, a PLS calibration model was built and optimized in terms of spectral range and number of factors employed. The main characteristics of models and results obtained for pH and redox potential determination as well as for C, N and H content, are detailed in Table 4. Spectral ranges were chosen using the moving window strategy, thus minimizing the RMSEP values. Table 4 also shows the best properties and prediction capabilities achieved by the PLS-NIR technique. The reproducibility of the determinations, established from the mean standard deviation of each replicate (strip ) and the standard error of prediction (sreg ), which includes the uncertainty in the model, were lower than 1.2 and 4 mg g−1 , respectively for C and N and for H strip and sreg obtained were 0.4 and 1.3 mg g−1 , respectively.

126

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

For electrochemical parameters strip and sreg values of 0.05 and 0.15 pH units and 9 and 37 mV were obtained for pH and Eh, respectively. To evaluate the aforementioned prediction capabilities it must be taken into consideration that acceptable errors for pH and redox potential measurements in sediments are of the order ±0.1 pH units and between 20 and 40 mV, respectively [3].

Acknowledgements Authors acknowledge the financial support of the Ministerio de Educación y Ciencia (Projects CTQ2005-05604 FEDER and AGL2007-64567).

references 3.6.

Trace metal and metal species

NIR data previously obtained for sediment samples were modelized by PLS to predict the total concentration of Sn, Pb, Cd, As, Sb and Cr using the calibration and validation data sets indicated in Table 3. Table 5 shows the best prediction capabilities obtained for elements considered and also, in the case of Cr, for the three fractions of Cr species established by using a sequential extraction scheme [38]. As it can be seen, there are significant differences in the optimum spectral range selected for the different built models. Moreover, the optimum number of extracted factors also varies in a wide range from 1 to 6 as a function of the element considered. Adequate QC values were obtained for Sn, Pb, As and Sb with predicted error levels for unknown samples around 20%. However relative root mean square errors of prediction are around 40% for concentration levels or the order of 400 ng g−1 Cd and around 100 ␮g g−1 of total Cr. These bad prediction results are probably link to the presence of high proportions of Cd and Cr as inorganic forms without a remarkable influence in the NIR spectra and on the basic fact that many of the parameters modelled are obtained indirectly and not based at all on the direct NIR behavior of the elements to be determined. Concerning the different Cr fractions extractable from sediments of Ria de Arousa the RPD values ranged from 1.3 to 1.7 thus indicating an appropriate residual predictive deviation for screening purposes. However the high QC percentages found avoids the use of NIR measurements to quantify Cr species in sediments.

4.

Conclusions

As compared with previously reported procedures (see Table 1) it can be observed that RPD values found through this study are of the same order of those for pH [21], total C [13,19] and total N [19] and also for total Pb [13] and Cd [9]. On the other hand, results found for As and Pb in polluted soils by MPLS [12] were better than those obtained in the present study, probably due to their high content in trace elements and RPD values higher than 2 were found for pH, C and N in soils using first derivative spectra and an extended number of samples [16]. However, it seems clear from results achieved in this work that PLS-NIR could become an alternative methodology for rapid monitoring of estuarine sediment characteristics.

[1] D.W. Pritchard, What is an estuary: physical viewpoint, in: G.H. Lauf (Ed.) Estuaries, A.A.A.S. Publ. No. 83, Washington, DC, 1967, pp. 3–5. [2] G. Branch, TREE 14 (1999). [3] S.L. Simpson, G.E. Batley, A.A. Chariton, J.L. Stauber, C.K. King, J.C. Chapman, R.V. Hyne, S.A. Gale, A.C. Roach, W.A. Maher, Handbook for Sediment Quality Assessment, CSIRO, Bangor, NSW, 2005. [4] P.K.S. Shin, W.K.C. Lam, Environ. Poll. 113 (2001) 281. [5] C.F. Zimmermann, C.W. Keefe, J. Bashe, Method 440.0. Determination of Carbon and Nitrogen in Sediments and Particulates of Estuarine/Coastal Waters Using Elemental Analysis. National Exposure Research Laboratory Office of Research and Development U.S. Environmental Protection Agency Cincinnati, Ohio 45268. [6] T.R. Crompton, Determination of Metals and Anions in Soils, Sediments and Sludges, Taylor & Francis Books Ltd., United Kingdom, 2001. [7] K.D. Shepherd, M.G. Walsh, Soil Sci. Soc. Am. J. 66 (2002) 988. [8] M. Odlarea, K. Svensson, M. Pell, Geoderma 126 (2005) 193. [9] D.F. Malley, P.C. Williams, Environ. Sci. Technol. 31 (1997) 3461. [10] D. Cozzolino, A. Morón, J. Agr. Sci. 140 (2003) 65. [11] G. Siebielec, G.W. McCarty, T.I. Stuczynski, J.B. Reeves III, J. Environ. Qual. 33 (2004) 2056. [12] R. Font, M. del Río, M. Simón, J. Aguilar, A. de Haro, Fresen. Environ. Bull. 13 (2004) 1309. ´ [13] M. Chodak, M. Niklinska, F. Beese, Biol. Fertil. Soils 44 (2007) 171. [14] Y. Wu, J. Chen, J. Ji, P. Gong, Q. Liao, Q. Tian, H. Ma, Soil Sci. Soc. Am. J. 71 (2007) 918. [15] A. Morón, D. Cozzolino, Comm. Soil Sci. Plant Anal. 38 (2007) 1965. [16] M.J. Cohen, J.P. Prenger, W.F. DeBusk, J. Environ. Qual. 34 (2005) 1422. [17] B. Butkuté, A. Slepetiené, Chemija 15 (2004) 12. [18] M.B. Nilsson, E. Dåbakk, T. Korsman, I. Marrenberg, Environ. Sci. Technol. 30 (1996) 2586. [19] L.K. Sørensen, S. Dalsgaard, Soil Sci. Soc. Am. J. 69 (2005) 159. [20] A. Stevens, B. van Wesemael, G. Vandenschrick, S. Touré, B. Tychon, Soil Sci. Soc. Am. J. 70 (2006) 844. [21] C.W. Chang, D.A. Laird, M.J. Mausbach, C.R. Hurburgh Jr., Soil Sci. Soc. Am. J. 65 (2001) 480. [22] Y. He, H. Song, SPIE—The International Society for Optical Engineering, 2006, 10.1117/2.1200604.0164. [23] Y. He, H. Song, A. García Pereira, A. Hernández Gómez, J. Zhejiang Univ. Sci. 11 (2005) 1081. ˜ [24] P. Bermejo-Barrera, M.C. Barciela-Alonso, J. Moreda-Pineiro, C. González-Sixto, A. Bermejo-Barrera, Spectrochim. Acta B 51 (1996) 1235. [25] M.C. Barciela-Alonso, P. Pazos-Capeáns, M.E. Regueira-Miguéns, A. Bermejo-Barrera, P. Bermejo-Barrera, Anal. Chim. Acta 524 (2004) 115. ˜ [26] M.C. Barciela-Alonso, R. Varela Vilela, E. Pena-Vázquez, A. Bermejo-Barrera, P. Bermejo-Barrera, Atom. Spectrosc. 28 (2007) 102.

a n a l y t i c a c h i m i c a a c t a 6 2 4 ( 2 0 0 8 ) 113–127

[27] P. Pazos-Capeáns, M.C. Barciela-Alonso, A. Bermejo-Barrera, P. Bermejo-Barrera, Anal. Chim. Acta 524 (2004) 121. ˜ [28] J. Moros, F.A. Inón, S. Garrigues, M. de la Guardia, Anal. Chim. Acta 584 (2007) 215. [29] J. Moros, S. Armenta, S. Garrigues, M. de la Guardia, Talanta 72 (2007) 72. ˜ [30] J. Moros, F.A. Inón, S. Garrigues, M. de la Guardia, Anal. Chim. Acta 538 (2005) 181. [31] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. de Jong, P.J. Lewi, J. Verbeke, Handbook of Chemometrics and Qualimetrics (Part A and B), Elsevier Science B.V., Amsterdam, 1997. ˜ [32] J. Moros, F.A. Inón, M. Khanmohammadi, S. Garrigues, M. de la Guardia, Anal. Bioanal. Chem. 385 (2006) 708.

127

[33] T. Fearn, in: J.M. Chalmers, P.R. Griffiths (Eds.), Handbook of Vibrational Spectroscopy (volume 3): Discriminant Analysis, Wiley, Chichester, 2002, p. 2086. [34] E. Ben-Dor, Y. Inbar, Y. Chen, Remote Sens. Environ. 61 (1997) 1. [35] L. Kooistra, J. Wanders, G.F. Epema, R.S.E.W. Leuven, R. Wehrens, L.M.C. Buydens, Anal. Chim. Acta 484 (2003) 189. [36] G.R. Hunt, J.W. Salisbury, Mod. Geol. 283 (1970) 1. [37] V.P. Krischenko, S.G. Samokhvalov, L.G. Formina, G.A. Novikova, in: I. Murray, I.A. Cowe (Eds.), Making Light Work: Advances in Near Infrared Spectroscopy, VCH, Weinheim, New York, Basel, Cambridge, 1992, p. 239. [38] P. Pazos Capeáns, M.C. Barciela-Alonso, A. Bermejo-Barrera, P. Bermejo-Barrera, Talanta 65 (2005) 678.