- Email: [email protected]
Quick analysis of organic matter in soil by energy-dispersive X-ray fluorescence and multivariate analysis
Author’s Accepted Manuscript QUICK ANALYSIS OF ORGANIC MATTER IN SOIL BY ENERGY-DISPERSIVE X-RAY FLUORESCENCE AND MULTIVARIATE ANALYSIS Franslley Morona, Felipe R. dos Santos, André M. Brinatti, Fábio L. Melquiades www.elsevier.com/locate/apradiso
PII: DOI: Reference:
S0969-8043(17)30630-9 http://dx.doi.org/10.1016/j.apradiso.2017.09.008 ARI8058
To appear in: Applied Radiation and Isotopes Received date: 11 May 2017 Revised date: 17 August 2017 Accepted date: 7 September 2017 Cite this article as: Franslley Morona, Felipe R. dos Santos, André M. Brinatti and Fábio L. Melquiades, QUICK ANALYSIS OF ORGANIC MATTER IN SOIL BY ENERGY-DISPERSIVE X-RAY FLUORESCENCE AND MULTIVARIATE ANALYSIS, Applied Radiation and Isotopes, http://dx.doi.org/10.1016/j.apradiso.2017.09.008 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
QUICK ANALYSIS OF ORGANIC MATTER IN SOIL BY ENERGY-DISPERSIVE X-RAY FLUORESCENCE AND MULTIVARIATE ANALYSIS Franslley Morona1, Felipe R. dos Santos1, André M. Brinatti2, Fábio L. Melquiades3*
1
Post Graduation Program in Applied Chemistry, Universidade Estadual do Centro Oeste,
UNICENTRO, Zip code: 85.040-080, Guarapuava, Paraná, Brazil 2
Laboratory of Soil Physics and Environmental Sciences, Department of Physics,
Universidade Estadual de Ponta Grossa, UEPG, Zip code: 84.030-900, Ponta Grossa, Paraná, Brazil 3
Applied Nuclear Physics Laboratory, Department of Physics, Universidade Estadual de
Londrina, UEL, Zip code: 86.057-970, Londrina, Paraná, Brazil
*
Corresponding author. Fábio Luiz Melquiades. Department of Physics, Universidade
Estadual de Londrina, PR445, Km 380, Zip Code: 86057-970 Londrina, PR, Brazil, Phone: 55 43 3371-4886, email: [email protected]
Abstract The rapid, simple and accurate determination of soil quality indicators is fundamental for improvements in precision agriculture and consequently in production efficiency. The objectives of this study were to determine the organic matter (OM) and total organic carbon (TOC) concentrations in agricultural soil and to discriminate soil 1
provenance by energy-dispersive X-ray fluorescence (EDXRF) combined with principal component analysis and partial least square regression. The conventional methods used for the determination of OM and TOC concentrations are the gravimetric and Walkley– Black methods, respectively. Figures of merit such as sensitivity, detection and quantification limits, accuracy and precision were evaluated. Samples were differentiated by their provenance, and the quality of the prediction model shows that EDXRF combined with multivariate analysis is a promising methodology to fulfil the lack of rapid and accurate analytical methods for the assessment of OM and TOC concentrations in agricultural soils. Graphical abstract
Abbreviations CEC, Cation Exchange Capacity; EDXRF, Energy-Dispersive X-ray Fluorescence; LV, Latent Variable; OM, Organic Matter; PC, Principal Component; PCA, Principal Component 2
Analysis; PLSR, Partial Least Square Regression; RMSEC, Root Mean Square Error of Calibration; RMSECV, Root Mean Square Error of Cross-Validation; RMSEP, Root Mean Square Error of Prediction; TOC, Total Organic Carbon; VNIR, Visible Near Infrared ; WB, Walkley Black; XRF, X-ray Fluorescence
Keywords: Organic Matter, Total Organic Carbon, multivariate regression, EDXRF, Green Chemistry.
1. INTRODUCTION Organic matter (OM) and total organic carbon (TOC), although present in the range of 1–6% in agricultural soil, play an important role in the fertility of the soil because of their contribution to increasing the cation exchange capacity (CEC) and water retention and decreasing compaction and leaching process, among other properties (Magdoff, 1993). The conventional analytical methods for the determination of OM and TOC concentrations are generally time consuming and cost prohibitive. The most widely used method in several countries is the Walkley–Black (WB) method, which is a wet method that does not follow the green chemistry concept because of the need for concentrated Cr2O72- and H2SO4 use, which could consequently result in the generation of toxic residues (WHO, 1988). The demand for good quality, inexpensive and high-resolution soil information has been growing in areas such as precision agriculture and land planning. Moreover, the several disadvantages of the conventional techniques lead to the need of developing and
3
validating time- and cost-effective quantitative methods in soil analysis (Zhu et al., 2011; Rossel et al., 2009). The application of spectroscopic techniques including visible near infrared (VNIR) (Ge et al., 2011; Kookana et al., 2008; Singh et al., 2012; Li et al., 2015), Raman spectroscopy (Luna et al., 2014) and X-ray fluorescence (XRF) (Weindorf et al., 2012) is currently being explored by several research groups, but it remains to be validated before being used in laboratory soil analyses worldwide. This study proposes the use of energy-dispersive X-ray fluorescence (EDXRF) for soil quality assessment. Recent studies on the non-conventional applications of this technique have come up with good results, for instance, in the texture (Zhu et al., 2012), nutrients (Kaniu et al., 2012), pH (Sharma et al., 2014) and TOC (Sharma et al., 2015, Melquiades et al., 2014). EDXRF analysis is commonly applied for the simultaneous determination of elements from Na to U. The advantages of EDXRF in soil analysis are that the technique facilitates non-destructive and non-invasive soil analysis of a wide range of elements at various concentrations (from ppm to % level) in a short period of time (30–300 s) with acceptable quantitative results; depending on the chemical composition of the metals, the steps in sample preparation, when necessary, can be minimised. In addition, there is a possibility of in situ analysis using a portable equipment (Kalnicky and Singhvi, 2001; Melquiades et al., 2004). In general, in soil samples, the percentage of light elements (Z < 11, for example H, C and O) is high. As a result, X-ray scattering is increased, which is revealed in the spectrum background and in the intensities of Rayleigh and Compton peaks (Bortoleto et al, 2005). The scattering region of the EDXRF spectra carries implicit information about the samples, which can be exploited using appropriate data treatment. 4
For example, the main compounds of OM are C, N and O whose characteristic X-ray photon energies are low, and their peaks do not appear in the EDXRF spectra, but this information is included in the X-ray scattering peaks and could be evaluated by appropriate multivariate analysis methods. By combining EDXRF data with multivariate analysis, qualitative and quantitative conclusive results about implicit sample information were obtained by pattern recognition or multivariate regression methods. Results published in the literature for application in different sets of data is vast. The objectives of this study were to determine the OM and TOC concentrations in agricultural soil and to discriminate the soil depending on its provenance using EDXRF spectroscopy and multivariate analysis. In particular, (a) the principal component analysis (PCA) was used to differentiate the soil depending on its provenance and (b) the partial least square regression (PLSR) was used to quantify OM and TOC based on the results from the WB and gravimetric methods.
2. EXPERIMENTAL 2.1 Sampling The soil samples evaluated in this study were collected by farmers from their own properties and sent to the soil analysis laboratory (Agrotecsolo INC) at Guarapuava, PR, Brazil. A total of 152 surface soil samples were collected at a maximum depth of 20 cm. The samples were from 27 counties of the II and III plateaus (103 and 49 samples, respectively) of Paraná State, Brazil. These counties are located in four hydrographic basins, with the number of samples collected from each basin given in parenthesis: Médio Iguaçú (28), Alto Ivaí (89), Alto Tibagi (27) and Piquiri – Paraná II (8). The map of the study 5
region is shown in Figure 1. Prior to every analytical procedure, the samples were dried at room temperature, ground and sieved to particle size smaller than 1 mm2
2.2 Determination of TOC concentration by the WB Method The principle of WB method is based on the indirect determination of OM from the TOC content (Walkley and Black, 1934). The modified WB method is the most used methodology in Brazilian soil laboratories. The following procedure was performed in duplicate and on different days for each sample: 0.500 ± 0.005 g of the sample was transferred to a 250-mL Erlenmeyer flask, and then 5 mL of 0.167 N K2Cr2O7 and 5 mL of concentrated H2SO4 were added. After 30 min, 50 mL of deionised water, 2 mL of 85% H2PO4 and 1 mL of diphenylamine indicator were added. The ingredients were then slowly titrated with 1 N FeSO4 until the endpoint was reached, i.e. until when the colour changed to green. The TOC value was determined by the following conventional equation (Tomé Junior, 1997): TOC (%) = (5 - Vs) * 0.792
(1)
where 5 is the initial K2Cr2O7 1 N volume in millilitres, Vs is the FeSO4.7H2O 1 N volume in millilitres used in the titration and 0.792 is the factor that comprises the soil mass, reaction stoichiometry, and unit transformations. 2.3 Determination of OM concentration by the Gravimetric method The gravimetric method is based on the mass loss by ignition, which is usually performed in a muffle furnace. It assumes that the organic matter will be eliminated, and the clay minerals will lose light elements, inorganic compounds and water (Dias et al, 6
2013). Several studies have used different heating temperatures in the range of 300–600 °C for OM incineration in the soil (Miyazawa et al., 2000; Conceição, 1999). In this study, the optimal condition for soil analysis by the gravimetric method was 1 g of soil sample placed in a crucible heated at 420ºC for 3 h. A total of 56 samples from Alto Tibagi and Médio Iguaçú basins were analysed by the gravimetric method. Analyses were performed in duplicates and on different days in a muffle furnace. 2.4 Energy-Dispersive X-ray Fluorescence EDXRF analysis is one of the most widely used X-ray techniques not only in environmental science, soil science and agronomy but also in archaeometry, forensic science, chemistry and metallurgy (Weindorf et al., 2014). EDXRF is based on the principle that when a material is exposed to high-energy X-rays, electrons are excited from an inner shell of the atom leaving holes. The higher energy electrons then fill these holes, and the energy difference is emitted as characteristic X-ray photons. Vacancies are produced in different orbitals, and a cascade of photons can be detected. By using the relationship between the emitted photon energy and the atomic number, the element can be identified, and by evaluating the peak intensities in the spectra, the elemental concentrations can be calculated (Jenkins, 1999). In this study, X-ray spectroscopy combined with multivariate analysis was performed to extract the implicit information from different data sets. Three grams of the samples with grain sizes smaller than 1 mm2 were placed in XRF plastic cups (24.6 mm aperture, 22.9 mm height, 7 cm3 volume, Chemplex Inc.) covered with Mylar films (6 m thickness, Chemplex Inc.) containing. The infinite thickness sample condition is guaranteed by the sample height between 5 and 6 mm in 7
the XRF cup. The measurements were made using a benchtop equipment Shimatzu EDX720 in air atmosphere for Ti–U range at 40 kV and 400 µA in the Rh X-ray tube for 100 s, with 10mm focal spot without any filter in the primary beam. The detector used was Si (Li). Three measurements were made for each sample by shaking the XRF cup before each measurement to irradiate different sample areas each time. A mean spectrum value was calculated for each sample. 2.5 Multivariate analysis Multivariate analysis is the key method to evaluate large data sets. The most widely used methods for discriminant analysis and multivariate calibration are the PCA and PLSR, respectively. The objective of the PCA is to reduce the data dimensionality without any loss of information by the construction of successive linear combinations so that each combination accounts for as much of the total variance as possible. After performing the PCA, the original data in the X matrix described by principal components (PCs) were represented in scores and loading matrices. In the PLSR, the PCs of the X matrix were correlated with a dependent variable Y matrix, which comprises a prediction parameter of interest, and a calibration model was constructed (Akbulut, 2014). To evaluate the provenance of the soil samples, a discriminant analysis was performed by the PCA. Several pre-processing steps were tested, and the mean centre resulted in a better sample discrimination. Two data sets were used in the PLSR prediction models: one with 152 TOC results from the WB method and the other with 54 OM results from the gravimetric method. The Kennard Stone algorithm (Kennard and Stone, 1969) was used to divide the TOC data set 8
into calibration and validation groups with 33% of the samples in the validation set. Two regression models were used: one with the complete spectra and the other with only the scattering peak region. By using the OM data set, the PLSR was assessed by the leave-oneout cross-validation. The quality of the models was evaluated by analysing the determination coefficients of the calibration, cross-validation and linear regression validation and the precision considering the root mean square error (RMSE) of the calibration, cross-validation and prediction (Geladi and Kowalski, 1986; Martens and Naes, 1989). In addition, some other figures of merit, such as sensitivity, detection limit (DL) and quantification limit (QL), were determined aiming at the acceptance of the method by regulatory agencies. In multivariate calibration, the net analyte signal (NAS) is a vector containing the values for each sample, and it can be obtained by the Euclidean norm of the regression vector b (‖𝒃‖2 ) using an inverse calibration model such as the PLS (Lorber et al., 1994). The NAS is the basic information for figures of merit calculation. The sensitivity is defined as the signal fraction that increases when the concentration of the specimen of interest is increased by one unit. It is obtained using Eq. (2) (Booksh and Kowalski, 1994): 1
𝑆𝐸𝑁 = ‖𝑵𝑨𝑺‖2 = ‖𝒃‖
(2)
2
The analytical sensitivity () expresses the sensitivity in terms of the concentration magnitude and can be defined as the ratio between SEN and instrumental noise () as shown in Eq (3):
=
𝑆𝐸𝑁 ‖‖2
,
(3)
where the Euclidean norm of the noise ‖‖2 was obtained from the calibration sample set 9
of each PLSR model. In this case, the instrumental noise is the background under the region of interest in the spectrum. As the OM or TOC concentrations were evaluated from the spectrum data, and because there were no peaks for OM and TOC, the final region of the spectra from 24 to 40 keV (847 variables) containing only the background signal was used. The smallest concentration of the analyte that can be identified in the test sample is defined as the DL, and the smallest concentration that can be quantified is the QL. These magnitudes are calculated as follows:
𝐷𝐿 =
3.‖‖2 ‖𝑵𝑨𝑺‖2
and
𝑄𝐿 =
10.‖‖2 ‖𝑵𝑨𝑺‖2
(4)
The t-paired test and F test were used to compare the results obtained from the prediction models and those from the conventional technique. MatLab software was used for the PCA and PLSR modelling.
3. RESULTS AND DISCUSSION 3.1 Results from conventional methods The TOC and OM concentrations determined by the WB and gravimetric methods, respectively, are summarised in Table 1. The range of variation in the values reflects the different soil types. The complete results are available in the supplementary material.
3.2 Energy-Dispersive X-Ray Fluorescence results The 152 mean spectra from EDXRF are presented in Figure 2. K, Ca, Ti, Mn, Fe, Cu, Zn, Rb, Sr and Zr were identified in the spectra. The overlapping spectra highlight their 10
similarity, especially in the background, although differences in Ti, Fe, Zr and Rh scattering peaks (region A) were noted. The region B highlighted in the dashed rectangle is the background used to calculate ‖‖2 and consequently the figures of merit, which are summarised in Table 2.
3.3 PCA of EDXRF spectra The PCA was performed on the whole set of data using mean center in the preprocessing step and two PCs explaining 99.77% of the data variance. Figure 3a shows the score plot that discriminates the three main groups, and the peaks that explain the groupings are shown in the loading plot (Figure 3b). In this case, Fe (Kα: 6.40 keV) and Ti (Kα: 4.51 keV) peaks contributed significantly to PC1 and PC2, respectively. The score plot (Figure 3a) shows that all the samples from the III plateau are in the positive side of PC1 confirming their higher homogeneity than those from the II plateau, which has sedimentary formation with heterogeneous soil texture and composition. Therefore, the variability in soil matrix composition contributed to the dispersion of the samples in the score plot.
The PCA enabled the soil discrimination by hydrographical basins. The Médio Iguaçú basin was grouped in the right side of the score plot with the samples from ParanáPiquiri II in the extreme right as they have higher Fe concentrations. The samples from Alto Ivaí basin were more heterogeneous and border the other basins, thus confirming their spread in the entire score plot. The Alto Tibagi samples were overlapped with the Alto Ivaí samples, and although they are grouped, both are from the II plateau. The Alto 11
Tibagi basin samples, which are characterised by lower Fe concentrations, were in opposite positions to those of the Paraná-Piquiri II samples. The Médio Iguaçú samples were between these two basins, in the score plot and the geographic map (Figure 1). 3.4 TOC concentration prediction by PLSR To obtain the TOC prediction values using EDXRF spectra (X matrix with 152 x 2048), the WB method results (Y matrix with 152 x 1) were used. Several models were tested. First, the whole spectra with all the samples were used to build global models for the studied region. Furthermore, specific models were elaborated restricting the data set by plateaus, hydrographical basins and municipalities to build local models. The trial was to reduce the soil matrix variability to obtain better TOC prediction results. However, the quality of local models were affected by the limited number of samples. Considering the Alto Ivaí, Alto Tibagi and Médio Iguaçú basins, the mean relative deviation values ranged from 13 to 22%, and the R2 values were between 0.53 and 0.90 for the calibration set. Moreover, some models were elaborated by selecting only the scattering peaks region, whose results were very similar to the whole spectra models. However, the best model for TOC prediction using the WB results was the one with 139 samples with the scattering region range, mean center in the pre-processing step and six latent variables (LVs). In this case, as the QL is 1.16%, the samples with values below this limit were not considered. The matrices (93 x 1147) and (46 x 1147) were used in the calibration and validation sets, respectively. Considering the high values of the Hotteling T2 and Leverages, 3 and 4samples were denominated as outliers in the calibration (when high values of the Hotteling T2 and Leverage were present simultaneously) and validation (high Hotteling T2 values only) sets, respectively. The model explained 99.2% of the variance in the X matrix and 97.7% in the Y matrix with R2 values of 0.98 and 0.60 for calibration and 12
validation, respectively. The model robustness could be evaluated by the RMSEC and RMSEP (RMSE of calibration and prediction, respectively), whose values were 0.11 and 0.38, respectively. Figure 4 shows the correlation plot for TOC values. The loadings in each LV are presented in Figure 5. The noise in the loadings is progressive, but it is possible to note that the 6th LV still carries information from Rh (Kα = 20.17 keV).
The t-paired test was applied to the samples to evaluate the accuracy of the method. The t calculated value was 1.58, which is smaller than the t tabled t(41:0.025) = 2.02, thus indicating that for 95% confidence level, the accuracy of the methods are similar. The average relative deviation for the validation samples considering the PLSR prediction model and the WB method was 13%. This value is acceptable when compared to the 17% relative deviation calculated considering the WB and elemental analysis methods from an EMBRAPA study (Conceição and Manzatto, 1999). Kaniu et al, 2012, used 21 LVs for carbon concentration prediction and obtained an average relative deviation of 42%. The overuse of LVs increases the correlation but does not improve the results; therefore, the first variables are the ones that carry more information and the others end up modelling the noise, thereby making the model less robust and more specific. Table 3 presents the external validation results with the respective relative deviation in each measurement. The F test applied to the validation samples resulted in a figure of 1.23 for the ratio between the square sum due to the pure error and the square sum due to the lack of fit. This figure is smaller but is of the same order of magnitude as 13
that of the F tabled (F(41,41) = 1.88), thus indicating that the methods have similar accuracy, but not for all the samples.
3.5 OM concentration prediction by PLSR For OM, a reduced set of 54 samples was considered and the leave-one-out cross-validation was used to evaluate the PLSR model. The complete spectra and the scattering region were tested with some pre-processing steps, and the best model was obtained with the scattering peak region and mean centre in the pre-processing step. The X matrix used was of 54 x 601 size corresponding to the energy range of 18–30 keV. The model with five LVs explained 92.08% for X matrix and 97.94% for the OM values of the Y matrix. The R2 values were 0.98 and 0.82, and the RMSEC and RMSECV (RMSE of cross-validation) were 0.57 and 1.72, respectively. The correlation between the measured values and the predicted values is shown in Figure 6.
The results of the t-paired test and the F test at 95% confidence level confirmed that there were no significant differences between the methods. The t calculated value was 1.00, which is smaller than the t tabled t(53:0.025) = 1.68. The F variance test applied to the cross-validation samples resulted in 1.02, which is smaller than the F tabled (F(53,53) = 1.62). The mean relative deviations were 6% and 16% for the calibration and crossvalidation, respectively. The best results were obtained by correlating the EDXRF spectra with the OM content instead of TOC content, for example, as reported by Kaniu et al, 2012. This is because the information in the EDXRF spectra, especially that in the scattering peaks, is associated with not only carbon but also all the light elements, and the correlation 14
between OM and TOC is not constant. The Van Bemmelen coefficient, which correlates OM and TOC, depends on the soil characteristics and depth and is assumed to have different values, as reported by several authors: 2.13 (Dean, 1974), 1.68 (Giovannini et al., 1985), 2.05 (Ben-Dor and Banin, 1989), 1.724 (Jackson, 1982), 3.18 (Conceição and Manzatto, 1999) and 3.72 (Miyazawa et al., 2000). Furthermore, the OM content affects the soil density because the light elements have smaller density. The higher the OM content, the smaller is the soil density and the greater are the background and scattering peak intensity. Moreover, comparisons of the EDXRF data and NIR studies revealed the same accuracy level (Ge et al., 2011; Li et al., 2015).
4. CONCLUSION The proposed methodology is viable as the prediction model’s performance could provide results at the same confidence level as that in the conventional methods. The advantage is that the EDXRF methodology is fast and nondestructive, and the preparation steps are quite simple when compared to those of the gravimetric, WB and elemental analysis methods. The combination of spectral data with multivariate calibration is a means to extract implicit information from the spectra, in this case, OM and TOC concentrations. The scattering region of the spectra is very important in this calculation. The accuracy and precision of the PLSR results compared to the gravimetric and WB methods were confirmed by the statistical analysis applying the t-paired test and F test. The model’s robustness was evaluated by the RMSEC, RMSECV, RMSEP and QLs, which confirmed the technical potentiality of the proposed methodology. The implementation of the EDXRF combined with PLSR by soil laboratories can 15
solve the problem of carcinogenic reagents and other products as this analytical method is environmentally friendly and could meet the high demand because it is rapid, nonpolluting and cost-effective. Moreover, it meets the needs of large soil laboratories and precision agriculture.
Acknowledgment To the Laboratory of Nuclear Instrumentation of the Centre of Nuclear Energy in Agriculture of the University of São Paulo for the support during statistical analysis and to 16
Agrotecsolo for providing the samples. The authors would like to thank CAPES for the Masters fellowship of the first author. SUPPLEMENTARY MATERIAL A pdf file with the complete data from conventional methods is presented as supplementary material. References Akbulut, S. Validation of Classical Quantitative Fundamental Parameters Method Using Multivariate Calibration Procedures for Trace Element Analysis in ED-XRF. J. Anal. At. Spectrom. 2014, 29 (5), 853 DOI: 10.1039/c3ja50377a. Ben-Dor, E.; Banin, A. Determination of Organic Matter Content in Arid-Zone Soils Using a Simple “loss-on-Ignition” Method. Commun. Soil Sci. Plant Anal. 1989, 20 (15–16), 1675–1695 DOI: 10.1080/00103628909368175. Booksh, K. S.; Kowalski, B. R. Theory of Analytical Chemistry. Anal. Chem. 1994, 66 (15), 782A–791A DOI: 10.1021/ac00087a001. Bortoleto, G. G.; Pataca, L. C. M.; Bueno, M. I. M. S. A New Application of X-Ray Scattering Using Principal Component Analysis - Classification of Vegetable Oils. Anal. Chim. Acta 2005, 539, 283–287 DOI: 10.1016/j.aca.2005.03.025. Conceição, M. da; Manzatto, C. Estudo Comparativo de Métodos de Determinação Do Teor de Matéria Orgânica Em Solos Orgânicos Do Estado Do Rio de Janeiro; EMBRAPA-Empresa Brasileira de Pesquisa Agropecuária: Rio de Janeiro - RJ, 1999. Dean, W. E. J. Determination of Carbonate and Organic Matter in Calcareous Sediments and Sedimentary Rocks by Loss on Ignition: Comparison with Other Methods. J. Sediment. Petrol. 1974, 44 (1), 242–248 DOI: 10.1306/74D729D2-2B21-11D717
8648000102C1865D. Dias, R. D.; De Abreu, C. A.; De Abreu, M. F.; Paz-Ferreiro, J.; Matsura, E. E.; Gonzalez, A. P. Comparison of Methods to Quantify Organic Carbon in Soil Samples from SAo Paulo State, Brazil. Commun. Soil Sci. Plant Anal. 2013, 44 (1–4), 429–439 DOI: Doi 10.1080/00103624.2013.742345. Ge, Y.; Morgan, C. L. S.; Grunwald, S.; Brown, D. J.; Sarkhot, D. V. Comparison of Soil Reflectance Spectra and Calibration Models Obtained Using Multiple Spectrometers. Geoderma 2011, 161 (3–4), 202–211 DOI: 10.1016/j.geoderma.2010.12.020. Geladi, P.; Kowalski, B. R. Partial Least-Squares Regression: A Tutorial. Anal. Chim. Acta 1986, 185 (0), 1–17 DOI: 10.1016/0003-2670(86)80028-9. Giovannini, G.; Riffaldi, R.; Levi-Minzi, R. Determination of Organic Matter in Sewage Sludges. Commun. Soil Sci. Plant Anal. 1985, 16, 775–785. Jackson, M. L. Analisis Quimico de Suelos; 1982. Jenkins, R. X-Ray Fluorescence Spectrometry; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 1999. Kalnicky, D. J.; Singhvi, R. Field Portable XRF Analysis of Environmental Samples. J. Hazard. Mater. 2001, 83 (1–2), 93–122 DOI: 10.1016/S0304-3894(00)00330-7. Kaniu, M. I.; Angeyo, K. H.; Mwala, a. K.; Mangala, M. J. Direct Rapid Analysis of Trace Bioavailable Soil Macronutrients by Chemometrics-Assisted Energy Dispersive XRay Fluorescence and Scattering Spectrometry. Anal. Chim. Acta 2012, 729, 21–25 DOI: 10.1016/j.aca.2012.04.007. Kennard, R. W.; Stone, L. A. Computer Aided Design of Experiments. Technometrics 1969, 11 (1), 137–148. 18
Kookana, R. S.; Janik, L. J.; Forouzangohar, M.; Forrester, S. T. Prediction of Atrazine Sorption Coefficients in Soils Using Mid-Infrared Spectroscopy and Partial LeastSquares Analysis. J. Agric. Food Chem. 2008, 56 (9), 3208–3213 DOI: 10.1021/jf073152n. Li, S.; Shi, Z.; Chen, S.; Ji, W.; Zhou, L.; Yu, W.; Webster, R. In Situ Measurements of Organic Carbon in Soil Profiles Using Vis-NIR Spectroscopy on the Qinghai-Tibet Plateau. Environ. Sci. Technol. 2015, 49 (8), 4980–4987 DOI: 10.1021/es504272x. Lorber, A.; Faber, K.; Kowalski, B. R. Net Analyte Signal Calculation in Multivariate Calibration. Anal. Chem. 1997, 69 (8), 1620–1626 DOI: 10.1021/ac960862b. Luna, A. S.; Lima, I. C. A.; Rocha, W. F. C.; Araújo, J. R.; Kuznetsov, A.; Ferreira, E. H. M.; Boqué, R.; Ferré, J. Classification of Soil Samples Based on Raman Spectroscopy and X-Ray Fluorescence Spectrometry Combined with Chemometric Methods and Variable Selection. Anal. Methods 2014, 6 (22), 8930–8939 DOI: 10.1039/C4AY01967A. Magdoff, F. Building Soils for Better Crops: Organic Matter Management. Soil Sci. 1993, 156 (5), 371 DOI: 10.1097/00010694-199311000-00014. Martens, H.; Naes, T. Multivariate Calibration; John Wiley and Sons: Chichester and New York, 1989. Melquiades, F. L.; Appoloni, C. R. Application of XRF and Field Portable XRF for Environmental Analysis. Journal of Radioanalytical and Nuclear Chemistry. 2004, pp 533–541. Melquiades, F. L.; Santos, F. R. Preliminary Results: Energy Dispersive X-Ray Fluorescence and Partial Least Square Regression for Organic Matter Determination in Soil. Spectrosc. Lett. 2014, No. DOI:10.1080/00387010.2013.874532 DOI: 19
10.1080/00387010.2013.874532. Miyazawa, M.; Pavan, M. A.; Oliveira, E. L. de; Ionashiro, M.; Silva, A. K. Gravimetric Determination of Soil Organic Matter. Brazilian Arch. Biol. Technol. 2000, 43 (5), 475–478 DOI: 10.1590/S1516-89132000000500005. Rossel, R. A. V.; Cattle, S. R.; Ortega, A.; Fouad, Y. In Situ Measurements of Soil Colour , Mineral Composition and Clay Content by Vis – NIR Spectroscopy. Geoderma 2009, 150 (3–4), 253–266 DOI: 10.1016/j.geoderma.2009.01.025. Sharma, A.; Weindorf, D. C.; Man, T.; Aldabaa, A. A. A.; Chakraborty, S. Characterizing Soils via Portable X-Ray Fluorescence Spectrometer: 3. Soil Reaction (pH). Geoderma 2014, 232–234, 141–147 DOI: 10.1016/j.geoderma.2014.05.005. Sharma, A.; Weindorf, D. C.; Wang, D.; Chakraborty, S. Characterizing Soils via Portable XRay Fluorescence Spectrometer: 4. Cation Exchange Capacity (CEC). Geoderma 2015, 239–240, 130–134 DOI: 10.1016/j.geoderma.2014.10.001. Singh, B.; Malley, D. F.; Farenhorst, A.; Williams, P. Feasibility of Using Near-Infrared Spectroscopy for Rapid Quantification of 17 Beta-Estradiol Sorption Coefficients in Soil. J. Agric. Food Chem. 2012, 60, 9948–9953 DOI: 10.1021/jf302492w. om
unio , . . Manual pa a inte p eta o de an lise de solo i a ia e dito a A opecu ia, 1997.
Walkley, A.; Black, I. A. An Examination of the Degtjareff Method for Determining Soil Organic Matter, and Proposed Modification of the Chromic Acid Titration Method. Soil Sci. 1934, 37, 29–38 DOI: 10.1097/00010694-193401000-00003. Weindorf, D. C.; Bakr, N.; Zhu, Y. Advances in Portable X-Ray Fluorescence ( PXRF ) for Environmental , Pedological , and Agronomic Applications; 2014; Vol. 128. Weindorf, D. C.; Zhu, Y.; Chakraborty, S.; Bakr, N.; Huang, B. Use of Portable X-Ray 20
Fluorescence Spectrometry for Environmental Quality Assessment of Peri-Urban Agriculture. Environ. Monit. Assess. 2012, 184 (1), 217–227 DOI: 10.1007/s10661011-1961-6. WHO Task Group on Environmental Health Criteria for Chromium.; United Nations Environment Programme.; International Labour Organisation.; World Health Organization.; International Program on Chemical Safety. Chromium.; World Health Organization, 1988. Zhu, Y.; Weindorf, D. C.; Zhang, W. Characterizing Soils Using a Portable X-Ray Fluorescence Spectrometer: 1. Soil Texture. Geoderma 2011, 167–168, 167–177 DOI: 10.1016/j.geoderma.2011.08.010.
21
Figure 1 – Map of the study region showing basin identification and separation based on plateaus.
22
Figure 2 – EDXRF overlapped spectra of the 152 soil samples. Region A is the scattering peak and region B is the background area used to calculate some figures of merit.
23
Figure 3 – (a) Score and (b) Loading plots of PC1 versus PC2 from the 152 EDXRF soil spectra. In the score plot, the samples are identified depending on plateaus and basins by symbols and colours, respectively.
24
Figure 4. Total organic carbon (TOC) concentration measured by the Walkley Black method versus predicted values obtained from the PLSR model with the whole EDXRF spectra.
25
Figure 5 - Loading plot from the six latent variables that comprise the PLSR model for the scattering range of the EDXRF spectra. 26
.
Figure 6. Organic matter (OM) values measured by the gravimetric method versus predicted values obtained from the PLSR model with scattering range in the EDXRF spectra. The residual plot is given in detail.
Table 1. Descriptive statistical results from conventional methods. N Mean±sd (%) Median (%) Min (%)
Max (%)
OM – Gravimetric
55
9.70±4.36
9.89
2.12
22.26
TOC - Walkley-Black
152
2.42±0.84
2.48
0.35
4.75
27
Table 2. Predicted values for the figures of merit for organic matter and total organic carbon using PLSR models. Analytical Number of calibration Detection Quantification Samples
sensitivity (1/%)
Limit (%)
Limit (%)
OM
54
4.90
0.67
2.04
TOC
92
8.64
0.38
1.16
Table 3. External validation results for total organic carbon (TOC) obtained by the Walkley black (WB) method compared to the PLSR prediction by EDXRF with scattering spectra data. TOC (%) TOC (%) TOC WB / TOC WB / TOC (%) EDXRF TOC (%) EDXRF TOC EDXRF TOC EDXRF WB scattering WB scattering (%) (%) region region 1.75
1.83
-1
3.23
2.69
14
2.29
3.14
-41
1.95
2.47
-24
2.73
2.77
-2
2.63
2.74
-3
2.28
2.11
4
3.85
3.16
17
1.44
1.93
-39
2.36
3.02
-26
2.69
2.32
12
2.22
2.96
-36
2.37
2.41
-3
2.88
2.83
0
2.35
2.69
-10
2.56
2.21
13
1.90
2.07
-13
2.73
3.00
-12
2.73
2.76
-2
2.74
2.52
8
2.53
2.44
3
3.27
3.57
-7
2.75
3.27
-20
1.93
2.66
-38
1.93
1.36
30
2.63
2.67
0
1.71
1.68
2
2.16
2.80
-29
2.45
2.91
-15
1.7
1.93
-12
2.92
2.64
9
2.17
2.05
5
3.27
3.04
7
2.28
2.14
5
3.41
3.41
-1
1.76
1.88
-7 28
2.98
3.18
-6
2.49
2.62
-5
3.45
2.92
14
1.37
1.76
-32
2.76
2.68
2
1.37
1.60
-11
2.54 ± 0.53
13a
Mean 2.45 ± 0.58 (a) Value calculated with absolute relative deviation.
Highlights OM and TOC concentrations in agricultural soil were determined. Samples were differentiated by plateaus and basins using PCA. EDXRF combined with PLSR is a straightforward methodology for OM quantification. The novelty is the direct quantification of OM and TOC from soil EDXRF spectral data.
29
30
PII: DOI: Reference:
S0969-8043(17)30630-9 http://dx.doi.org/10.1016/j.apradiso.2017.09.008 ARI8058
To appear in: Applied Radiation and Isotopes Received date: 11 May 2017 Revised date: 17 August 2017 Accepted date: 7 September 2017 Cite this article as: Franslley Morona, Felipe R. dos Santos, André M. Brinatti and Fábio L. Melquiades, QUICK ANALYSIS OF ORGANIC MATTER IN SOIL BY ENERGY-DISPERSIVE X-RAY FLUORESCENCE AND MULTIVARIATE ANALYSIS, Applied Radiation and Isotopes, http://dx.doi.org/10.1016/j.apradiso.2017.09.008 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
QUICK ANALYSIS OF ORGANIC MATTER IN SOIL BY ENERGY-DISPERSIVE X-RAY FLUORESCENCE AND MULTIVARIATE ANALYSIS Franslley Morona1, Felipe R. dos Santos1, André M. Brinatti2, Fábio L. Melquiades3*
1
Post Graduation Program in Applied Chemistry, Universidade Estadual do Centro Oeste,
UNICENTRO, Zip code: 85.040-080, Guarapuava, Paraná, Brazil 2
Laboratory of Soil Physics and Environmental Sciences, Department of Physics,
Universidade Estadual de Ponta Grossa, UEPG, Zip code: 84.030-900, Ponta Grossa, Paraná, Brazil 3
Applied Nuclear Physics Laboratory, Department of Physics, Universidade Estadual de
Londrina, UEL, Zip code: 86.057-970, Londrina, Paraná, Brazil
*
Corresponding author. Fábio Luiz Melquiades. Department of Physics, Universidade
Estadual de Londrina, PR445, Km 380, Zip Code: 86057-970 Londrina, PR, Brazil, Phone: 55 43 3371-4886, email: [email protected]
Abstract The rapid, simple and accurate determination of soil quality indicators is fundamental for improvements in precision agriculture and consequently in production efficiency. The objectives of this study were to determine the organic matter (OM) and total organic carbon (TOC) concentrations in agricultural soil and to discriminate soil 1
provenance by energy-dispersive X-ray fluorescence (EDXRF) combined with principal component analysis and partial least square regression. The conventional methods used for the determination of OM and TOC concentrations are the gravimetric and Walkley– Black methods, respectively. Figures of merit such as sensitivity, detection and quantification limits, accuracy and precision were evaluated. Samples were differentiated by their provenance, and the quality of the prediction model shows that EDXRF combined with multivariate analysis is a promising methodology to fulfil the lack of rapid and accurate analytical methods for the assessment of OM and TOC concentrations in agricultural soils. Graphical abstract
Abbreviations CEC, Cation Exchange Capacity; EDXRF, Energy-Dispersive X-ray Fluorescence; LV, Latent Variable; OM, Organic Matter; PC, Principal Component; PCA, Principal Component 2
Analysis; PLSR, Partial Least Square Regression; RMSEC, Root Mean Square Error of Calibration; RMSECV, Root Mean Square Error of Cross-Validation; RMSEP, Root Mean Square Error of Prediction; TOC, Total Organic Carbon; VNIR, Visible Near Infrared ; WB, Walkley Black; XRF, X-ray Fluorescence
Keywords: Organic Matter, Total Organic Carbon, multivariate regression, EDXRF, Green Chemistry.
1. INTRODUCTION Organic matter (OM) and total organic carbon (TOC), although present in the range of 1–6% in agricultural soil, play an important role in the fertility of the soil because of their contribution to increasing the cation exchange capacity (CEC) and water retention and decreasing compaction and leaching process, among other properties (Magdoff, 1993). The conventional analytical methods for the determination of OM and TOC concentrations are generally time consuming and cost prohibitive. The most widely used method in several countries is the Walkley–Black (WB) method, which is a wet method that does not follow the green chemistry concept because of the need for concentrated Cr2O72- and H2SO4 use, which could consequently result in the generation of toxic residues (WHO, 1988). The demand for good quality, inexpensive and high-resolution soil information has been growing in areas such as precision agriculture and land planning. Moreover, the several disadvantages of the conventional techniques lead to the need of developing and
3
validating time- and cost-effective quantitative methods in soil analysis (Zhu et al., 2011; Rossel et al., 2009). The application of spectroscopic techniques including visible near infrared (VNIR) (Ge et al., 2011; Kookana et al., 2008; Singh et al., 2012; Li et al., 2015), Raman spectroscopy (Luna et al., 2014) and X-ray fluorescence (XRF) (Weindorf et al., 2012) is currently being explored by several research groups, but it remains to be validated before being used in laboratory soil analyses worldwide. This study proposes the use of energy-dispersive X-ray fluorescence (EDXRF) for soil quality assessment. Recent studies on the non-conventional applications of this technique have come up with good results, for instance, in the texture (Zhu et al., 2012), nutrients (Kaniu et al., 2012), pH (Sharma et al., 2014) and TOC (Sharma et al., 2015, Melquiades et al., 2014). EDXRF analysis is commonly applied for the simultaneous determination of elements from Na to U. The advantages of EDXRF in soil analysis are that the technique facilitates non-destructive and non-invasive soil analysis of a wide range of elements at various concentrations (from ppm to % level) in a short period of time (30–300 s) with acceptable quantitative results; depending on the chemical composition of the metals, the steps in sample preparation, when necessary, can be minimised. In addition, there is a possibility of in situ analysis using a portable equipment (Kalnicky and Singhvi, 2001; Melquiades et al., 2004). In general, in soil samples, the percentage of light elements (Z < 11, for example H, C and O) is high. As a result, X-ray scattering is increased, which is revealed in the spectrum background and in the intensities of Rayleigh and Compton peaks (Bortoleto et al, 2005). The scattering region of the EDXRF spectra carries implicit information about the samples, which can be exploited using appropriate data treatment. 4
For example, the main compounds of OM are C, N and O whose characteristic X-ray photon energies are low, and their peaks do not appear in the EDXRF spectra, but this information is included in the X-ray scattering peaks and could be evaluated by appropriate multivariate analysis methods. By combining EDXRF data with multivariate analysis, qualitative and quantitative conclusive results about implicit sample information were obtained by pattern recognition or multivariate regression methods. Results published in the literature for application in different sets of data is vast. The objectives of this study were to determine the OM and TOC concentrations in agricultural soil and to discriminate the soil depending on its provenance using EDXRF spectroscopy and multivariate analysis. In particular, (a) the principal component analysis (PCA) was used to differentiate the soil depending on its provenance and (b) the partial least square regression (PLSR) was used to quantify OM and TOC based on the results from the WB and gravimetric methods.
2. EXPERIMENTAL 2.1 Sampling The soil samples evaluated in this study were collected by farmers from their own properties and sent to the soil analysis laboratory (Agrotecsolo INC) at Guarapuava, PR, Brazil. A total of 152 surface soil samples were collected at a maximum depth of 20 cm. The samples were from 27 counties of the II and III plateaus (103 and 49 samples, respectively) of Paraná State, Brazil. These counties are located in four hydrographic basins, with the number of samples collected from each basin given in parenthesis: Médio Iguaçú (28), Alto Ivaí (89), Alto Tibagi (27) and Piquiri – Paraná II (8). The map of the study 5
region is shown in Figure 1. Prior to every analytical procedure, the samples were dried at room temperature, ground and sieved to particle size smaller than 1 mm2
2.2 Determination of TOC concentration by the WB Method The principle of WB method is based on the indirect determination of OM from the TOC content (Walkley and Black, 1934). The modified WB method is the most used methodology in Brazilian soil laboratories. The following procedure was performed in duplicate and on different days for each sample: 0.500 ± 0.005 g of the sample was transferred to a 250-mL Erlenmeyer flask, and then 5 mL of 0.167 N K2Cr2O7 and 5 mL of concentrated H2SO4 were added. After 30 min, 50 mL of deionised water, 2 mL of 85% H2PO4 and 1 mL of diphenylamine indicator were added. The ingredients were then slowly titrated with 1 N FeSO4 until the endpoint was reached, i.e. until when the colour changed to green. The TOC value was determined by the following conventional equation (Tomé Junior, 1997): TOC (%) = (5 - Vs) * 0.792
(1)
where 5 is the initial K2Cr2O7 1 N volume in millilitres, Vs is the FeSO4.7H2O 1 N volume in millilitres used in the titration and 0.792 is the factor that comprises the soil mass, reaction stoichiometry, and unit transformations. 2.3 Determination of OM concentration by the Gravimetric method The gravimetric method is based on the mass loss by ignition, which is usually performed in a muffle furnace. It assumes that the organic matter will be eliminated, and the clay minerals will lose light elements, inorganic compounds and water (Dias et al, 6
2013). Several studies have used different heating temperatures in the range of 300–600 °C for OM incineration in the soil (Miyazawa et al., 2000; Conceição, 1999). In this study, the optimal condition for soil analysis by the gravimetric method was 1 g of soil sample placed in a crucible heated at 420ºC for 3 h. A total of 56 samples from Alto Tibagi and Médio Iguaçú basins were analysed by the gravimetric method. Analyses were performed in duplicates and on different days in a muffle furnace. 2.4 Energy-Dispersive X-ray Fluorescence EDXRF analysis is one of the most widely used X-ray techniques not only in environmental science, soil science and agronomy but also in archaeometry, forensic science, chemistry and metallurgy (Weindorf et al., 2014). EDXRF is based on the principle that when a material is exposed to high-energy X-rays, electrons are excited from an inner shell of the atom leaving holes. The higher energy electrons then fill these holes, and the energy difference is emitted as characteristic X-ray photons. Vacancies are produced in different orbitals, and a cascade of photons can be detected. By using the relationship between the emitted photon energy and the atomic number, the element can be identified, and by evaluating the peak intensities in the spectra, the elemental concentrations can be calculated (Jenkins, 1999). In this study, X-ray spectroscopy combined with multivariate analysis was performed to extract the implicit information from different data sets. Three grams of the samples with grain sizes smaller than 1 mm2 were placed in XRF plastic cups (24.6 mm aperture, 22.9 mm height, 7 cm3 volume, Chemplex Inc.) covered with Mylar films (6 m thickness, Chemplex Inc.) containing. The infinite thickness sample condition is guaranteed by the sample height between 5 and 6 mm in 7
the XRF cup. The measurements were made using a benchtop equipment Shimatzu EDX720 in air atmosphere for Ti–U range at 40 kV and 400 µA in the Rh X-ray tube for 100 s, with 10mm focal spot without any filter in the primary beam. The detector used was Si (Li). Three measurements were made for each sample by shaking the XRF cup before each measurement to irradiate different sample areas each time. A mean spectrum value was calculated for each sample. 2.5 Multivariate analysis Multivariate analysis is the key method to evaluate large data sets. The most widely used methods for discriminant analysis and multivariate calibration are the PCA and PLSR, respectively. The objective of the PCA is to reduce the data dimensionality without any loss of information by the construction of successive linear combinations so that each combination accounts for as much of the total variance as possible. After performing the PCA, the original data in the X matrix described by principal components (PCs) were represented in scores and loading matrices. In the PLSR, the PCs of the X matrix were correlated with a dependent variable Y matrix, which comprises a prediction parameter of interest, and a calibration model was constructed (Akbulut, 2014). To evaluate the provenance of the soil samples, a discriminant analysis was performed by the PCA. Several pre-processing steps were tested, and the mean centre resulted in a better sample discrimination. Two data sets were used in the PLSR prediction models: one with 152 TOC results from the WB method and the other with 54 OM results from the gravimetric method. The Kennard Stone algorithm (Kennard and Stone, 1969) was used to divide the TOC data set 8
into calibration and validation groups with 33% of the samples in the validation set. Two regression models were used: one with the complete spectra and the other with only the scattering peak region. By using the OM data set, the PLSR was assessed by the leave-oneout cross-validation. The quality of the models was evaluated by analysing the determination coefficients of the calibration, cross-validation and linear regression validation and the precision considering the root mean square error (RMSE) of the calibration, cross-validation and prediction (Geladi and Kowalski, 1986; Martens and Naes, 1989). In addition, some other figures of merit, such as sensitivity, detection limit (DL) and quantification limit (QL), were determined aiming at the acceptance of the method by regulatory agencies. In multivariate calibration, the net analyte signal (NAS) is a vector containing the values for each sample, and it can be obtained by the Euclidean norm of the regression vector b (‖𝒃‖2 ) using an inverse calibration model such as the PLS (Lorber et al., 1994). The NAS is the basic information for figures of merit calculation. The sensitivity is defined as the signal fraction that increases when the concentration of the specimen of interest is increased by one unit. It is obtained using Eq. (2) (Booksh and Kowalski, 1994): 1
𝑆𝐸𝑁 = ‖𝑵𝑨𝑺‖2 = ‖𝒃‖
(2)
2
The analytical sensitivity () expresses the sensitivity in terms of the concentration magnitude and can be defined as the ratio between SEN and instrumental noise () as shown in Eq (3):
=
𝑆𝐸𝑁 ‖‖2
,
(3)
where the Euclidean norm of the noise ‖‖2 was obtained from the calibration sample set 9
of each PLSR model. In this case, the instrumental noise is the background under the region of interest in the spectrum. As the OM or TOC concentrations were evaluated from the spectrum data, and because there were no peaks for OM and TOC, the final region of the spectra from 24 to 40 keV (847 variables) containing only the background signal was used. The smallest concentration of the analyte that can be identified in the test sample is defined as the DL, and the smallest concentration that can be quantified is the QL. These magnitudes are calculated as follows:
𝐷𝐿 =
3.‖‖2 ‖𝑵𝑨𝑺‖2
and
𝑄𝐿 =
10.‖‖2 ‖𝑵𝑨𝑺‖2
(4)
The t-paired test and F test were used to compare the results obtained from the prediction models and those from the conventional technique. MatLab software was used for the PCA and PLSR modelling.
3. RESULTS AND DISCUSSION 3.1 Results from conventional methods The TOC and OM concentrations determined by the WB and gravimetric methods, respectively, are summarised in Table 1. The range of variation in the values reflects the different soil types. The complete results are available in the supplementary material.
3.2 Energy-Dispersive X-Ray Fluorescence results The 152 mean spectra from EDXRF are presented in Figure 2. K, Ca, Ti, Mn, Fe, Cu, Zn, Rb, Sr and Zr were identified in the spectra. The overlapping spectra highlight their 10
similarity, especially in the background, although differences in Ti, Fe, Zr and Rh scattering peaks (region A) were noted. The region B highlighted in the dashed rectangle is the background used to calculate ‖‖2 and consequently the figures of merit, which are summarised in Table 2.
3.3 PCA of EDXRF spectra The PCA was performed on the whole set of data using mean center in the preprocessing step and two PCs explaining 99.77% of the data variance. Figure 3a shows the score plot that discriminates the three main groups, and the peaks that explain the groupings are shown in the loading plot (Figure 3b). In this case, Fe (Kα: 6.40 keV) and Ti (Kα: 4.51 keV) peaks contributed significantly to PC1 and PC2, respectively. The score plot (Figure 3a) shows that all the samples from the III plateau are in the positive side of PC1 confirming their higher homogeneity than those from the II plateau, which has sedimentary formation with heterogeneous soil texture and composition. Therefore, the variability in soil matrix composition contributed to the dispersion of the samples in the score plot.
The PCA enabled the soil discrimination by hydrographical basins. The Médio Iguaçú basin was grouped in the right side of the score plot with the samples from ParanáPiquiri II in the extreme right as they have higher Fe concentrations. The samples from Alto Ivaí basin were more heterogeneous and border the other basins, thus confirming their spread in the entire score plot. The Alto Tibagi samples were overlapped with the Alto Ivaí samples, and although they are grouped, both are from the II plateau. The Alto 11
Tibagi basin samples, which are characterised by lower Fe concentrations, were in opposite positions to those of the Paraná-Piquiri II samples. The Médio Iguaçú samples were between these two basins, in the score plot and the geographic map (Figure 1). 3.4 TOC concentration prediction by PLSR To obtain the TOC prediction values using EDXRF spectra (X matrix with 152 x 2048), the WB method results (Y matrix with 152 x 1) were used. Several models were tested. First, the whole spectra with all the samples were used to build global models for the studied region. Furthermore, specific models were elaborated restricting the data set by plateaus, hydrographical basins and municipalities to build local models. The trial was to reduce the soil matrix variability to obtain better TOC prediction results. However, the quality of local models were affected by the limited number of samples. Considering the Alto Ivaí, Alto Tibagi and Médio Iguaçú basins, the mean relative deviation values ranged from 13 to 22%, and the R2 values were between 0.53 and 0.90 for the calibration set. Moreover, some models were elaborated by selecting only the scattering peaks region, whose results were very similar to the whole spectra models. However, the best model for TOC prediction using the WB results was the one with 139 samples with the scattering region range, mean center in the pre-processing step and six latent variables (LVs). In this case, as the QL is 1.16%, the samples with values below this limit were not considered. The matrices (93 x 1147) and (46 x 1147) were used in the calibration and validation sets, respectively. Considering the high values of the Hotteling T2 and Leverages, 3 and 4samples were denominated as outliers in the calibration (when high values of the Hotteling T2 and Leverage were present simultaneously) and validation (high Hotteling T2 values only) sets, respectively. The model explained 99.2% of the variance in the X matrix and 97.7% in the Y matrix with R2 values of 0.98 and 0.60 for calibration and 12
validation, respectively. The model robustness could be evaluated by the RMSEC and RMSEP (RMSE of calibration and prediction, respectively), whose values were 0.11 and 0.38, respectively. Figure 4 shows the correlation plot for TOC values. The loadings in each LV are presented in Figure 5. The noise in the loadings is progressive, but it is possible to note that the 6th LV still carries information from Rh (Kα = 20.17 keV).
The t-paired test was applied to the samples to evaluate the accuracy of the method. The t calculated value was 1.58, which is smaller than the t tabled t(41:0.025) = 2.02, thus indicating that for 95% confidence level, the accuracy of the methods are similar. The average relative deviation for the validation samples considering the PLSR prediction model and the WB method was 13%. This value is acceptable when compared to the 17% relative deviation calculated considering the WB and elemental analysis methods from an EMBRAPA study (Conceição and Manzatto, 1999). Kaniu et al, 2012, used 21 LVs for carbon concentration prediction and obtained an average relative deviation of 42%. The overuse of LVs increases the correlation but does not improve the results; therefore, the first variables are the ones that carry more information and the others end up modelling the noise, thereby making the model less robust and more specific. Table 3 presents the external validation results with the respective relative deviation in each measurement. The F test applied to the validation samples resulted in a figure of 1.23 for the ratio between the square sum due to the pure error and the square sum due to the lack of fit. This figure is smaller but is of the same order of magnitude as 13
that of the F tabled (F(41,41) = 1.88), thus indicating that the methods have similar accuracy, but not for all the samples.
3.5 OM concentration prediction by PLSR For OM, a reduced set of 54 samples was considered and the leave-one-out cross-validation was used to evaluate the PLSR model. The complete spectra and the scattering region were tested with some pre-processing steps, and the best model was obtained with the scattering peak region and mean centre in the pre-processing step. The X matrix used was of 54 x 601 size corresponding to the energy range of 18–30 keV. The model with five LVs explained 92.08% for X matrix and 97.94% for the OM values of the Y matrix. The R2 values were 0.98 and 0.82, and the RMSEC and RMSECV (RMSE of cross-validation) were 0.57 and 1.72, respectively. The correlation between the measured values and the predicted values is shown in Figure 6.
The results of the t-paired test and the F test at 95% confidence level confirmed that there were no significant differences between the methods. The t calculated value was 1.00, which is smaller than the t tabled t(53:0.025) = 1.68. The F variance test applied to the cross-validation samples resulted in 1.02, which is smaller than the F tabled (F(53,53) = 1.62). The mean relative deviations were 6% and 16% for the calibration and crossvalidation, respectively. The best results were obtained by correlating the EDXRF spectra with the OM content instead of TOC content, for example, as reported by Kaniu et al, 2012. This is because the information in the EDXRF spectra, especially that in the scattering peaks, is associated with not only carbon but also all the light elements, and the correlation 14
between OM and TOC is not constant. The Van Bemmelen coefficient, which correlates OM and TOC, depends on the soil characteristics and depth and is assumed to have different values, as reported by several authors: 2.13 (Dean, 1974), 1.68 (Giovannini et al., 1985), 2.05 (Ben-Dor and Banin, 1989), 1.724 (Jackson, 1982), 3.18 (Conceição and Manzatto, 1999) and 3.72 (Miyazawa et al., 2000). Furthermore, the OM content affects the soil density because the light elements have smaller density. The higher the OM content, the smaller is the soil density and the greater are the background and scattering peak intensity. Moreover, comparisons of the EDXRF data and NIR studies revealed the same accuracy level (Ge et al., 2011; Li et al., 2015).
4. CONCLUSION The proposed methodology is viable as the prediction model’s performance could provide results at the same confidence level as that in the conventional methods. The advantage is that the EDXRF methodology is fast and nondestructive, and the preparation steps are quite simple when compared to those of the gravimetric, WB and elemental analysis methods. The combination of spectral data with multivariate calibration is a means to extract implicit information from the spectra, in this case, OM and TOC concentrations. The scattering region of the spectra is very important in this calculation. The accuracy and precision of the PLSR results compared to the gravimetric and WB methods were confirmed by the statistical analysis applying the t-paired test and F test. The model’s robustness was evaluated by the RMSEC, RMSECV, RMSEP and QLs, which confirmed the technical potentiality of the proposed methodology. The implementation of the EDXRF combined with PLSR by soil laboratories can 15
solve the problem of carcinogenic reagents and other products as this analytical method is environmentally friendly and could meet the high demand because it is rapid, nonpolluting and cost-effective. Moreover, it meets the needs of large soil laboratories and precision agriculture.
Acknowledgment To the Laboratory of Nuclear Instrumentation of the Centre of Nuclear Energy in Agriculture of the University of São Paulo for the support during statistical analysis and to 16
Agrotecsolo for providing the samples. The authors would like to thank CAPES for the Masters fellowship of the first author. SUPPLEMENTARY MATERIAL A pdf file with the complete data from conventional methods is presented as supplementary material. References Akbulut, S. Validation of Classical Quantitative Fundamental Parameters Method Using Multivariate Calibration Procedures for Trace Element Analysis in ED-XRF. J. Anal. At. Spectrom. 2014, 29 (5), 853 DOI: 10.1039/c3ja50377a. Ben-Dor, E.; Banin, A. Determination of Organic Matter Content in Arid-Zone Soils Using a Simple “loss-on-Ignition” Method. Commun. Soil Sci. Plant Anal. 1989, 20 (15–16), 1675–1695 DOI: 10.1080/00103628909368175. Booksh, K. S.; Kowalski, B. R. Theory of Analytical Chemistry. Anal. Chem. 1994, 66 (15), 782A–791A DOI: 10.1021/ac00087a001. Bortoleto, G. G.; Pataca, L. C. M.; Bueno, M. I. M. S. A New Application of X-Ray Scattering Using Principal Component Analysis - Classification of Vegetable Oils. Anal. Chim. Acta 2005, 539, 283–287 DOI: 10.1016/j.aca.2005.03.025. Conceição, M. da; Manzatto, C. Estudo Comparativo de Métodos de Determinação Do Teor de Matéria Orgânica Em Solos Orgânicos Do Estado Do Rio de Janeiro; EMBRAPA-Empresa Brasileira de Pesquisa Agropecuária: Rio de Janeiro - RJ, 1999. Dean, W. E. J. Determination of Carbonate and Organic Matter in Calcareous Sediments and Sedimentary Rocks by Loss on Ignition: Comparison with Other Methods. J. Sediment. Petrol. 1974, 44 (1), 242–248 DOI: 10.1306/74D729D2-2B21-11D717
8648000102C1865D. Dias, R. D.; De Abreu, C. A.; De Abreu, M. F.; Paz-Ferreiro, J.; Matsura, E. E.; Gonzalez, A. P. Comparison of Methods to Quantify Organic Carbon in Soil Samples from SAo Paulo State, Brazil. Commun. Soil Sci. Plant Anal. 2013, 44 (1–4), 429–439 DOI: Doi 10.1080/00103624.2013.742345. Ge, Y.; Morgan, C. L. S.; Grunwald, S.; Brown, D. J.; Sarkhot, D. V. Comparison of Soil Reflectance Spectra and Calibration Models Obtained Using Multiple Spectrometers. Geoderma 2011, 161 (3–4), 202–211 DOI: 10.1016/j.geoderma.2010.12.020. Geladi, P.; Kowalski, B. R. Partial Least-Squares Regression: A Tutorial. Anal. Chim. Acta 1986, 185 (0), 1–17 DOI: 10.1016/0003-2670(86)80028-9. Giovannini, G.; Riffaldi, R.; Levi-Minzi, R. Determination of Organic Matter in Sewage Sludges. Commun. Soil Sci. Plant Anal. 1985, 16, 775–785. Jackson, M. L. Analisis Quimico de Suelos; 1982. Jenkins, R. X-Ray Fluorescence Spectrometry; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 1999. Kalnicky, D. J.; Singhvi, R. Field Portable XRF Analysis of Environmental Samples. J. Hazard. Mater. 2001, 83 (1–2), 93–122 DOI: 10.1016/S0304-3894(00)00330-7. Kaniu, M. I.; Angeyo, K. H.; Mwala, a. K.; Mangala, M. J. Direct Rapid Analysis of Trace Bioavailable Soil Macronutrients by Chemometrics-Assisted Energy Dispersive XRay Fluorescence and Scattering Spectrometry. Anal. Chim. Acta 2012, 729, 21–25 DOI: 10.1016/j.aca.2012.04.007. Kennard, R. W.; Stone, L. A. Computer Aided Design of Experiments. Technometrics 1969, 11 (1), 137–148. 18
Kookana, R. S.; Janik, L. J.; Forouzangohar, M.; Forrester, S. T. Prediction of Atrazine Sorption Coefficients in Soils Using Mid-Infrared Spectroscopy and Partial LeastSquares Analysis. J. Agric. Food Chem. 2008, 56 (9), 3208–3213 DOI: 10.1021/jf073152n. Li, S.; Shi, Z.; Chen, S.; Ji, W.; Zhou, L.; Yu, W.; Webster, R. In Situ Measurements of Organic Carbon in Soil Profiles Using Vis-NIR Spectroscopy on the Qinghai-Tibet Plateau. Environ. Sci. Technol. 2015, 49 (8), 4980–4987 DOI: 10.1021/es504272x. Lorber, A.; Faber, K.; Kowalski, B. R. Net Analyte Signal Calculation in Multivariate Calibration. Anal. Chem. 1997, 69 (8), 1620–1626 DOI: 10.1021/ac960862b. Luna, A. S.; Lima, I. C. A.; Rocha, W. F. C.; Araújo, J. R.; Kuznetsov, A.; Ferreira, E. H. M.; Boqué, R.; Ferré, J. Classification of Soil Samples Based on Raman Spectroscopy and X-Ray Fluorescence Spectrometry Combined with Chemometric Methods and Variable Selection. Anal. Methods 2014, 6 (22), 8930–8939 DOI: 10.1039/C4AY01967A. Magdoff, F. Building Soils for Better Crops: Organic Matter Management. Soil Sci. 1993, 156 (5), 371 DOI: 10.1097/00010694-199311000-00014. Martens, H.; Naes, T. Multivariate Calibration; John Wiley and Sons: Chichester and New York, 1989. Melquiades, F. L.; Appoloni, C. R. Application of XRF and Field Portable XRF for Environmental Analysis. Journal of Radioanalytical and Nuclear Chemistry. 2004, pp 533–541. Melquiades, F. L.; Santos, F. R. Preliminary Results: Energy Dispersive X-Ray Fluorescence and Partial Least Square Regression for Organic Matter Determination in Soil. Spectrosc. Lett. 2014, No. DOI:10.1080/00387010.2013.874532 DOI: 19
10.1080/00387010.2013.874532. Miyazawa, M.; Pavan, M. A.; Oliveira, E. L. de; Ionashiro, M.; Silva, A. K. Gravimetric Determination of Soil Organic Matter. Brazilian Arch. Biol. Technol. 2000, 43 (5), 475–478 DOI: 10.1590/S1516-89132000000500005. Rossel, R. A. V.; Cattle, S. R.; Ortega, A.; Fouad, Y. In Situ Measurements of Soil Colour , Mineral Composition and Clay Content by Vis – NIR Spectroscopy. Geoderma 2009, 150 (3–4), 253–266 DOI: 10.1016/j.geoderma.2009.01.025. Sharma, A.; Weindorf, D. C.; Man, T.; Aldabaa, A. A. A.; Chakraborty, S. Characterizing Soils via Portable X-Ray Fluorescence Spectrometer: 3. Soil Reaction (pH). Geoderma 2014, 232–234, 141–147 DOI: 10.1016/j.geoderma.2014.05.005. Sharma, A.; Weindorf, D. C.; Wang, D.; Chakraborty, S. Characterizing Soils via Portable XRay Fluorescence Spectrometer: 4. Cation Exchange Capacity (CEC). Geoderma 2015, 239–240, 130–134 DOI: 10.1016/j.geoderma.2014.10.001. Singh, B.; Malley, D. F.; Farenhorst, A.; Williams, P. Feasibility of Using Near-Infrared Spectroscopy for Rapid Quantification of 17 Beta-Estradiol Sorption Coefficients in Soil. J. Agric. Food Chem. 2012, 60, 9948–9953 DOI: 10.1021/jf302492w. om
unio , . . Manual pa a inte p eta o de an lise de solo i a ia e dito a A opecu ia, 1997.
Walkley, A.; Black, I. A. An Examination of the Degtjareff Method for Determining Soil Organic Matter, and Proposed Modification of the Chromic Acid Titration Method. Soil Sci. 1934, 37, 29–38 DOI: 10.1097/00010694-193401000-00003. Weindorf, D. C.; Bakr, N.; Zhu, Y. Advances in Portable X-Ray Fluorescence ( PXRF ) for Environmental , Pedological , and Agronomic Applications; 2014; Vol. 128. Weindorf, D. C.; Zhu, Y.; Chakraborty, S.; Bakr, N.; Huang, B. Use of Portable X-Ray 20
Fluorescence Spectrometry for Environmental Quality Assessment of Peri-Urban Agriculture. Environ. Monit. Assess. 2012, 184 (1), 217–227 DOI: 10.1007/s10661011-1961-6. WHO Task Group on Environmental Health Criteria for Chromium.; United Nations Environment Programme.; International Labour Organisation.; World Health Organization.; International Program on Chemical Safety. Chromium.; World Health Organization, 1988. Zhu, Y.; Weindorf, D. C.; Zhang, W. Characterizing Soils Using a Portable X-Ray Fluorescence Spectrometer: 1. Soil Texture. Geoderma 2011, 167–168, 167–177 DOI: 10.1016/j.geoderma.2011.08.010.
21
Figure 1 – Map of the study region showing basin identification and separation based on plateaus.
22
Figure 2 – EDXRF overlapped spectra of the 152 soil samples. Region A is the scattering peak and region B is the background area used to calculate some figures of merit.
23
Figure 3 – (a) Score and (b) Loading plots of PC1 versus PC2 from the 152 EDXRF soil spectra. In the score plot, the samples are identified depending on plateaus and basins by symbols and colours, respectively.
24
Figure 4. Total organic carbon (TOC) concentration measured by the Walkley Black method versus predicted values obtained from the PLSR model with the whole EDXRF spectra.
25
Figure 5 - Loading plot from the six latent variables that comprise the PLSR model for the scattering range of the EDXRF spectra. 26
.
Figure 6. Organic matter (OM) values measured by the gravimetric method versus predicted values obtained from the PLSR model with scattering range in the EDXRF spectra. The residual plot is given in detail.
Table 1. Descriptive statistical results from conventional methods. N Mean±sd (%) Median (%) Min (%)
Max (%)
OM – Gravimetric
55
9.70±4.36
9.89
2.12
22.26
TOC - Walkley-Black
152
2.42±0.84
2.48
0.35
4.75
27
Table 2. Predicted values for the figures of merit for organic matter and total organic carbon using PLSR models. Analytical Number of calibration Detection Quantification Samples
sensitivity (1/%)
Limit (%)
Limit (%)
OM
54
4.90
0.67
2.04
TOC
92
8.64
0.38
1.16
Table 3. External validation results for total organic carbon (TOC) obtained by the Walkley black (WB) method compared to the PLSR prediction by EDXRF with scattering spectra data. TOC (%) TOC (%) TOC WB / TOC WB / TOC (%) EDXRF TOC (%) EDXRF TOC EDXRF TOC EDXRF WB scattering WB scattering (%) (%) region region 1.75
1.83
-1
3.23
2.69
14
2.29
3.14
-41
1.95
2.47
-24
2.73
2.77
-2
2.63
2.74
-3
2.28
2.11
4
3.85
3.16
17
1.44
1.93
-39
2.36
3.02
-26
2.69
2.32
12
2.22
2.96
-36
2.37
2.41
-3
2.88
2.83
0
2.35
2.69
-10
2.56
2.21
13
1.90
2.07
-13
2.73
3.00
-12
2.73
2.76
-2
2.74
2.52
8
2.53
2.44
3
3.27
3.57
-7
2.75
3.27
-20
1.93
2.66
-38
1.93
1.36
30
2.63
2.67
0
1.71
1.68
2
2.16
2.80
-29
2.45
2.91
-15
1.7
1.93
-12
2.92
2.64
9
2.17
2.05
5
3.27
3.04
7
2.28
2.14
5
3.41
3.41
-1
1.76
1.88
-7 28
2.98
3.18
-6
2.49
2.62
-5
3.45
2.92
14
1.37
1.76
-32
2.76
2.68
2
1.37
1.60
-11
2.54 ± 0.53
13a
Mean 2.45 ± 0.58 (a) Value calculated with absolute relative deviation.
Highlights OM and TOC concentrations in agricultural soil were determined. Samples were differentiated by plateaus and basins using PCA. EDXRF combined with PLSR is a straightforward methodology for OM quantification. The novelty is the direct quantification of OM and TOC from soil EDXRF spectral data.
29
30