Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=lcss20
Communications in Soil Science and Plant Analysis
ISSN: 0010-3624 (Print) 1532-2416 (Online) Journal homepage: http://www.tandfonline.com/loi/lcss20
Chemometric Methods to Predict of Pb in Urban
Soil from Port Pirie, South Australia, using
Spectrally Active of Soil Carbon
Ali Al Maliki, Gary Owens, Hussain M Hussain, Saadi Dahaan & Nadhir
Al-Ansari
To cite this article: Ali Al Maliki, Gary Owens, Hussain M Hussain, Saadi Dahaan & Nadhir Al-Ansari (2018): Chemometric Methods to Predict of Pb in Urban Soil from Port Pirie, South Australia, using Spectrally Active of Soil Carbon, Communications in Soil Science and Plant Analysis, DOI: 10.1080/00103624.2018.1464178
To link to this article: https://doi.org/10.1080/00103624.2018.1464178
Published with license by Taylor & Francis Group, LLC © 2018 [Ali Al Maliki, Gary Owens, Hussain M Hussain, Saadi Al-Dahaan, and Nadhir Al-Ansari] Published online: 07 May 2018.
Submit your article to this journal
Article views: 26
View related articles
Chemometric Methods to Predict of Pb in Urban Soil from Port
Pirie, South Australia, using Spectrally Active of Soil Carbon
Ali Al Malikia, Gary Owensb, Hussain M Hussainc, Saadi Al-Dahaand, and Nadhir Al-Ansarie aEnvironment Research Centre, Ministry of Since and Technology, Naghdad, Iraq;bUniversity of South Australia,
Adelaide, Australia;cRemote Sensing Centre, University of Kufa, Al-Najf, Iraq;dEnvironment and Geomorphology
Department, University of Kufa, Al-Najf, Iraq;eLulea Tekniska Universitet, Luleå, Sweden
ABSTRACT
A total of 73 soil samples were initially analyzed for lead (Pb) concentration as an indicator of the environment impact of smelter activity in the Port Pirie, South Australia. Chemometric techniques were used to assess the ability of near-infrared (NIR) reflectance spectroscopy to predict soil Pb using spectrally active soil characteristics such as soil carbon (C). The result indicated a strong linear relationship between log-transformed data of soil Pb and spectral reflectance in the range between 500 and 612 nm with R2= 0.54 and a low root-mean-square error (RMSEv = 0.38) for the validation
mode with an acceptable ratio of performance to deviation and ratio of error range (1.6 and 7.7, respectively). This study suggested that NIR spec-troscopy based on auxiliary spectrally active components is a rapid and noninvasive assessment technique and has the ability to determine Pb contamination in urban soil to be useful in environmental health risk assessment.
ARTICLE HISTORY
Received 11 March 2017 Accepted 26 March 2018
KEYWORDS
Chemometric model; lead contamination; multivariate analysis; NIR; PLSR model; soil Pb
Introduction
While soil may naturally contain varying amounts of lead (Pb) depending on its geological source, soil and dust are believed to be the two main anthropogenic sources and pathways of Pb exposure due primarily to depositions from leaded industrial and manufacturing processes or from smelter emissions (McMichael et al.1994; Tong et al.2000; Martley, Gulson, and Pfeifer2004). Evaluating the environmental impact of soil Pb contamination in urban areas is a complex task that requires a good understanding of mechanisms operating at the interfaces between environmental health, urban geochemistry, historical geography, and spatial distributions (Al Maliki, Bruce, and Owens 2015). This would facilitate reducing the uncertainties of prediction and minimizes investigation (Al Maliki, Bruce, and Owens2012a).
There have also been some attempts to detect Pb concentration from reflectance spectra directly; Choe et al. (2008) found a significant correlation between the ratios of reflectance at 610–500 nm and Pb concentration. Zhang, Wen, and Zhao (2010) found a strong spectral signature for soil Pb content using bands selected from hyperspectral imaging laboratory measurements, where the soil spectral curve absorption values or reflectance peaks occurred at 838, 1930, and 2148 nm. These bands associated with complementary compounds that relate to Pb. A significant linear relationship between Pb and reflectance was observed, with coefficients of determination (R2) of 0.691 and 0.683 at 800 and 1300 nm, respectively (Pandit, Filippelli, and Li2010).
CONTACTNadhir Al-Ansari alyay004@mymail.unisa.edu.au;nadhir.alansari@ltu.se Lulea Tekniska Universitet, Luleå, Sweden.
Color versions of one or more of the figures in the article can be found online atwww.tandfonline.com/lcss.
Published with license by Taylor & Francis Group, LLC © 2018 [Ali Al Maliki, Gary Owens, Hussain M Hussain, Saadi Al-Dahaan, and Nadhir Al-Ansari] This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http:// creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way.
Heavy metals (HMs) do not generally absorb in the visible-near-infrared (VNIR) region; however, they can be detected via regression analysis due to covariation (relation) with moieties such as hydroxides, sulfides, carbonates, oxides, or organic matter which are detectable or adsorption to clays that absorb light in the VNIR-short-wave infrared region (Schwartz2012).
The absorption feature for soil Pb could be attributed to the absorption features of soil constituents such as iron, organic carbon, and clay mineral (Summers 2009; Stenberg et al. 2010; Ji et al.2010). Therefore, the correlation between Pb and other soil constituents that have spectrally active signatures can be indicative of the Pb contamination levels in soil (Al Maliki, Bruce, and Owens2012a). The spectral predictive mechanism using organic carbon as a key for an indirect spectroscopy method to detect Pb concentration has been applied in several previously published studies. For example, Ji et al. (2010) found that Pb bound to organic carbon was useful for the prediction of Pb concentration in the same spectral detection regions as organic carbon. Yaolin et al. (2011) obtained the reliable estimates for Pb concen-tration, from spectral parameters using a correlation between soil organic matter and Pb concenconcen-tration, where the prediction of Pb concentration was derived via the ratio between the first-order derivative for wavelengths between 564 and 624 nm. Moreover, there was a significant relationship between some metal concentrations [cadmium (Cd), mercury (Hg), Pb] within the spectral domain 500–700 nm, which was related to organic carbon (Bray, Rossel, and McBratney2009; Ji et al.2010; Pandit, Filippelli, and Li 2010). This spectral predictive mechanism was also found in some other studies (e.g., Malley and Williams1997), but differs from the results of Wu et al. (2007). They found that soil Pb contamination can be estimated indirectly using reflectance spectroscopy (RS) via correlations with iron oxide contents. It is likely that both the aforementioned direct and indirect methods used for estimating spectral reflectance will need to be employed to remotely detect the presence and concentration of Pb. The general objectives of this study are to develop and demonstrate the utility of VNIR techniques to detect soil constituents that help to estimate Pb contamination in areas of Port Pirie, South Australia. This study is the first attempt to use spectral reflectance signatures to predict the Pb concentration in urban soil and to find suitable spectrum for the detection of Pb in soil. This study thus supports ongoing investigation and monitoring of Pb contamination for studying environmental and health impact of Pb contamination which can substantially reduce the time and costs associated with traditional soil contaminant analysis.
Methods and materials Study area and sampling
The town of Port Pirie is located on the eastern shore of the Spencer Gulf, 230 km north of Adelaide, in South Australia. The study area at Port Pirie included sample sites of occupied houses, vacant lots, abandoned houses, common areas (playground sites), kindergartens, schools, or other nonresidences such as the land east of the Port Pirie river about 500 m east of smelter (Figure 1).
The study area extended to a buffer zone of 5 km radius from the Nyrstar smelter, which was large enough to encompass all residential areas as well as covering a risk area that contained soil and dust with highest levels of lead contamination. This area was divided into a regular 1 × 1-km grid, and initial soil sample locations were taken from the corners of each grid cell, together with one sample in the center of each grid cell. A map of the combined survey area with the sampling grid and sites is shown inFigure 1. A total of 73 samples were collected from urban areas surrounding the large lead smelter as well as two additional background samples from outside the buffer zone of 5 km from the smelter. Prior to field sampling, proper, potential sampling sites were initially selected from digital aerial ortho-photographs. Aerial color ortho-photographs of 10 cm spatial resolution with Map Grid of Australia zone 54 for Port Pirie location were obtained from AEROmetrex Pty Ltd. and were acquired on 25 February 2011 (http://aerometrex.com.au) and displayed using ArcMap 10.2. At each sampling location, a total of three discrete bare surface soil samples were collected using a hand auger at a depth of 0–5 cm, as solar radiation in the VNIR has limited penetration capabilities. Samples were collected from within a 1 m2area and were thoroughly mixed onsite to generate a single composite soil sample
for further analysis. A composite soil sample allows for a more representative soil sample to be obtained (Bickel2010). The objective of this technique was to combine a number of discrete samples collected from each site into a single homogenized sample which was considered to represent
average conditions in the sampled body for the purpose of analysis (EPA 2005). This sampling
was chosen to characterize soils that had varying levels of Pb (Body et al.1991; Taylor2012), making it an ideal and suitable site to assess urban atmospheric deposition of Pb contaminants (Al Maliki, Bruce, and Owens2012b). Soil samples (n = 73) were transported to the laboratory and desiccated in a fan assisted oven at 60°C until visually dried. As the soil dried, a porcelain mortar was used to gently disaggregate any lumps that had formed. Prior to any analysis, each sample was then sieved to pass through a 2-mm stainless steel sieve.
Soil physiochemical properties
HM contents of the soils including soil Pb and total Fe were determined using Inductively Coupled Plasma-Mass Spectrometry following microwave-assisted digestion with aqua regia. Total carbon (C) and nitrogen (N) contents were determined via dry oxidation (automated combustion) on a LECO CNS analyzer (LECO Corp., St Joseph, MI). Soil pH and electrical conductivity (EC) were measured in 1:5 soils:water suspension using a pH/conductivity meter (Smart CHEM-LAB, TPS, Australia). Soil texture was determined using a micropipette method (Miller and Miller1987). The summary of descriptive statistics for above soil attributes is displayed inTable 1.
NIR spectroscopy for soil samples
NIR reflectance spectra were obtained between 350 and 2500 nm using a HyLogger™ system (model
VSTR-1) with a high intensity halogen lamp as a light source (Mason and Huntington2012), available at the Glenside Core Library of the South Australian Department for Manufacturing, Innovation, Trade,
Resources and Energy (DMITRE). The HyLogger™ system was developed by CSIRO to extract the
mineralogy of drill cores using reflectance spectroscopy (Mason and Huntington2012). As the system South Australia
State
Port Pirie
was designed to determine spectral reflectance caused by molecular vibrations indicative of the chemical bonds in crystalline minerals, RS obtained from a range of wavelengths can be used to identify common minerals present in a soil sample. Semiquantitative mineralogy and associated mineralogical parameters such as alteration intensity‘crystallinity’ and chemistry can also be extracted or inferred from the HyL ogging data (Mason and Huntington2012). Here the HyLogger system was adapted to obtain reflectance of soils contained in plastic Petri dishes, rather than intact soil cores. Ten Petri dishes were placed side by side in one line for each spectral collection and processing episode. For each soil sample, 20 individual reflectance spectra were extracted from points across the surface of the soil within the Petri dish. Subsequently, the individual spectra were averaged to obtain representative reflectance spectra for transformation and processing. Spectral reflectance was measured at a 4-nm bandwidth from 380 to 2500 nm so that a total of 537 data points per spectrum were measured for each sample location in the soil. The Spectral Geologist (TSG) software (www.thespectralgeologist.com) was used for visualization. TSG collected reflectance spectra for hundreds of samples that included rich information for soil constituents. HyLogger data were subjected to multivariate analysis using the Unscrambler X software package version 10.3 (http://www.camo.com) to extract evidence of the underlying qualities and attributes within the spectral information.
Spectral preprocessing
Many preprocessing techniques have been applied successfully in previous studies to spectral reflectance raw data in order to assist in the interpretation of overlapping weak spectral overtones, to remove spectral noise originating from the effects of nonhomogeneous distributions of particle sizes, and to provide optimum predictions. These techniques included averaging, smoothing, absorption (log (1/R), where R is reflectance), transformations, normalization, and derivative processing. In this study, the spectral reflec-tance at a wavelength of 537 nm as normalized to characterize the soils variables. Normalization was applied to remove spectral noise and to improve the detection of absorption feature at spectral reflectance (Kooistra et al. 2001; CAMO Software 2012). In this work, the mean normalization wasre applied to improve prediction of the reflectance spectra and compared for efficiency of estimation of soil parameters. Normalize data
The distribution patterns of soil Pb and other variables were assessed using statistical analysis of a histogram and normal plots. These were used to assess whether a data set was normally distributed (i.e., for a QQ-plot if it was linear, or asymmetric (i.e., far from normal) if the QQ-plot was nonlinear). These statistical tools found that raw Pb concentrations across Port Pirie were not normally distributed. Consequently, Pb was log transformed to normalize the data and reduce the skew of the raw data with the aim to reduce error in prediction models. Thus, all subsequent regression processes for predicted soil Pb across the study area were based on log-transformed data. This technique was used in some studies to predict metal concentra-tions; for example, Choe et al. (2008) in studying pollution in stream sediments indicated the ability to detect HM spectral absorption features by combining geochemistry log scale with field spectroscopy and hyperspectral remote sensing (Choe et al.2008). Other studies suggested that data transformation to a log scale is an effective way to reduce the effect of outliers (Dao, Morrison, and Kiely2013) and is a commonly used transform for soil data (Choe et al.2009; Hengla, Heuvelinkb, and Stein2004). HM measurements in general conform to normal distributions when natural log transformations are applied, while lognormal
Table 1.Descriptive statistics of heavy metals (mg kg−1), pH, EC (µS/cm), carbon (%) of soil samples (n = 73) and background samples (n = 2). Id As (mg kg−1) Cd (mg kg−1) Cu (mg kg−1) Mn (mg kg−1) Fe (mg kg−1) Zn (mg kg−1) Pb (mg kg−1) pH EC µS/cm C% Min 2 2 5 46 7371 32 4 5.62 78.6 0.65 Max 137 26 2495 30168 261386 9704 5002 9.5 21000 8.38 Mean 25 5 155 1652 31284 1110 592 8.26 3370 2.55 St. Dev 27 4 353 4818 37354 2005 790 0.71 5625 1.80
distributions have been used to describe soil Pb with a strongly positive skew (Griffith2002). The results of these studies indicated that a log scale provides a better opportunity to successfully predict HM concentra-tions compared with non-transformed data.
In addition to spectral preprocessing, described earlier, outlier removing was applied to develop the regression model.
Outlier values may be defined as when the sample is spectrally different than the rest of the samples, and the residual difference is found to be anomalously much greater than found in the other samples. Principal component (PCs) analysis (PCA) was used with a default P-value of 0.05 to identify and remove calibration outliers in the spectral data, and new calibrations were then developed. Primarily, PCA is a feature reduction mathematical procedure that can be useful to provide the basis for the other multivariate linear regression analyses (e.g., partial least-squares regression (PLSR)) (Niazi2011).
Multivariate analysis and regression techniques Calibration
The PLSR calibration technique was applied in this study to handle all the spectroscopy data sets in the field of NIRS to get and display chemically relevant information out of measured chemical data. The spectral information was transformed to PLSR components (factors) similarly to the PCs in a PCA, compressing the data and at the same time dealing with the correlations within the reflectance spectral data. A PLSR technique was implemented to establish the relationship between the reflec-tance spectra and measured soil variables including soil Pb concentration.
The PLSR technique was based on latent variable decomposition of two variable blocks (matrices) X– spectral parameter (predictors) and Y – measured soil variables (dependants), resulting in the smallest potential number of significant components which allowed for the simplification of the
relationship between matrices X and Y (Gannouni, Rebai, and Abdeljaoued2012; Hong-Yan et al.
2009; Kooistra et al.2001; Vohland, Bossung, and Fründ2009; Wu et al.2005). The purpose of the method is to find a small number of latent factors that are predictive for Y, and which also use X efficiently (Bilgili et al.2010; Hong-Yan et al.2009).
The specific statistical methods used to process and analyze the data can significantly alter the final prediction capabilities of the models. Thus, the efficiency of the predictive models developed was evaluated by comparing correlations of predicted and actual Pb concentrations.
In correlation line loadings plot for PLSR model, the values between dotted lines within the upper and lower bounds represented the components (factors) in the spectra. These factors and the associated scores are used to model the covariance matrix between spectra data and reference measurement. Calibration models in most NIR applications have some samples that fall outside the properties of the main sample range, and present much greater or lower than the rest of the data set. These are known as calibration outliers values, which are from samples that have poorly determined values during the one-out cross-validation process (Siebielec et al.2004).
Validation
Two validation methods were applied in order to validate the prediction results including
Internal validation. A full cross-validation of the leave-one-out method was used to verify the prediction capability of the PLSR models.
External validation. An extra seven samples with a variety of Pb concentration values (low, medium, high) and not used in the calibration model were used to evaluate the prediction accuracy of the model through many statistical parameters. The use of an external data set for validation is appropriate to check whether the model predicts new samples well, as using cross-validation only can give overly optimistic estimates of prediction ability.
Evaluation
Several prediction accuracy parameters were applied in this work including, the coefficient of determination (R2), the root-mean-square error (RMSE) for both calibration (RMSEC) and
valida-tion based on the cross validavalida-tion method (leaving one out) (RMSECV), coefficient of regression or
the slope of the linear regression, the absolute estimation error or standard error (SE) for both calibration (SEC) and cross validation (SECV), as well as the intercept of regression and the bias,
which were all taken as relative predictors of the best model (Nduwamungu et al. 2009). Other
statistical parameters for prediction accuracy were the ratio of performance to deviation (RPD) which was computed as the ratio of the standard deviation (SD) of reference values to the RMSECV
(Equation 1) and the Ratio of Error Range (RER) computed by dividing the range in the reference value (Max-Min) by the RMSECVused in the validation set (Eq. 2).
RPD¼ SD=RMSEcv (1)
RER¼ Range=RMSEcv (2)
In general, the lowest RMSE, SE, bias, offset (intercept) together with higher RPD≥ 2, RER> 6, and R2 coefficient of regression close to 1.0 were used as indicators of the most accurate regressions (Bilgili et al.2010; Brunet et al. 2007; Nduwamungu et al. 2009; Wu et al.2007).
Results and discussion Spectra analysis of soil data
The reflectance spectra measured with the HyLogger system from air-dried soil samples exhibited three specific absorption features (Figure 2), two spectral absorption features (1400 and 1900 nm) are common features for water and the other strong spectral feature (around 2200 nm) is related to clay or hydroxyl mineral. Correlation coefficients between soil measurements and spectral data were used to evaluate the spectral response for soil properties. The relationship between the normalized spectral data and measured values for some soil properties in the VNIR region are shown inTable 2. Correlation coefficients results of the above soil properties were in agreement with the often superior predictions of these properties in the literature. For example, the highest correlation with reflectance spectra in about 400 nm of wavelength region indicated the spectral features for soil total C, similar absorption feature due to total organic carbon was found in the VNIR region by Pastor et al.2008; (Wetterlind 2009). The correlation between total iron and spectral measurements with r = 0.38 in region of spectra of 512 nm can be referred to absorption band for iron content, similar information was found by Ben-Dor, Irons, and Epema (1999) who pointed, that the absorption feature of iron
0 0.1 0.2 0.3 0.4 0.5 400 650 900 1150 1400 1650 1900 2150 2400 Reflecance Wavelength nm iron-oxide hydroxide carbonates clay water mineral organic carbon content
Figure 2.Average reflectance spectra for Port Pirie soil samples (n = 73) indicating three specific absorption features at 1400, 1900, and 2200. The other main spectral of soil constituents represented most common features as mentioned in literatures.
and transition metals can be exhibited at wavelengths near 500 nm (Ben-Dor, Irons, and Epema 1999). Both of C and iron content can be directly related to absorption in spectral reflectance through a number of fundamental groups such as hydroxyl, carboxyl, and amine group which can be detectable in VNIR (Wetterlind 2009). The carbonate and salt minerals also include many of the molecular bonds absorbing in region about 2300 nm (Clark et al.1990).
NIR to detect soil Pb using prediction models
The summary of some prediction models for soil Pb, total C, iron content, and soil texture (sand and clay fractions) are shown inTable 3, and the confidence level of P =0.005 was used in all prediction models. From the result of using the entire set of soils (calibration set n = 73), it was observed that Pb concentrations predicted from spectral reflectance measurements were in poor agreement with the known concentrations of Pb in the soils obtained from soil Pb references (Table 3). However, a better prediction result was found when using log scale of Pb concentration via reflectance. In addition to this result, the PLSR prediction for the Pb concentration model in Figure 3showed a
good correlation (R2 = 0.54) and a low RMSE (RMSEv = 0.38) for the validation mode with
acceptable RPD and RER (1.6 and 7.7, respectively).
All of these measures of accuracy indicated a suitable prediction model (Bilgili et al. 2010;
Nduwamungu et al. 2009). This result agreement with Hayes 2012 who recommended that the
numbers obtained for the RER typically be around four to five times larger than those for the RPD. This indicated that spectral reflectance analysis could successfully predict soil Pb concentrations for a set of Port Pirie soils (n = 73) on a log scale.
The NIR is based on molecular vibrations, which need to be robust to have strong bands and to be strong enough in the fundamentals to still have overtone and combination bands (Csorba et al. 2012). Therefore, the Pb concentration predicted values in this experiment may depend on the level
of interactions with the spectrally active components throughout the range of 350–2500 nm. In
Table 2.Correlation coefficient (R) between Pb concentration and wavelength (nm) and other elements that have distinct spectral signatures.
Variables SR regions with best correlation (nm)
Correlation coefficient (R) R Pb (ppm) Log Pb Pb (mg kg−1) 512 0.334 1 Log Pb 512 0.385 NA 1 N (%) 1404 0.623 0.22 0.36 C (%) 400 0.592 0.29 0.44 Clay (%) 1972 0.332 −0.01 −0.05 Fe (mg kg−1) 548 0.384 0.08 0.16
SR: Spectral reflectance, NA: Not applicable
Table 3.Summary of statistics for chemometric analysis represented prediction models for soil Pb and other spectrally active soil properties
PLSR model
with normalized treatment for: No. of samples No. of components
Calibration Validation R2 RMSEC R 2 RMSEV RPDV a RERV b Pb mg kg−1 73 7 0.19 598 0.08 646 1.0 5.7 Pblogc 73 7 0.63 0.32 0.54 0.38 1.6 7.7 Felog c 73 7 0.49 0.21 0.22 0.26 1.1 5.9 % Carbon 73 7 0.76 0.86 0.67 1.05 1.6 7.0 % Sand texture 73 7 0.60 11.5 0.49 13 1.4 6.1 % Clay texture 73 7 0.29 7.87 0.20 8.9 1.1 3.7
The confidence level ofP = 0.005 was used in all prediction models. a
RPDV is the ratio of the standard deviation (SD) of the reference value to the RMSE in validation, suitable values of 1.4–2 bRERV is the ratio of error range in the reference value (Max-Min) by the RMSE in validation, with suitable values of >6 c
AB Figure 3. Relationship between predicted and measured Pb concentrations (log mg kg − 1) for the PLSR models. Construction model (A) and Validation model (B).
general, HMs are not considered spectrally active because they may represent much weaker large-order overtones of the soil on the level of interactions with the spectrally active components throughout the range of 350-2500 nm.
In general, Pb is a spectrally inactive metal; however, Pb concentration can be measured indirectly by correlation with another soil constituent that has the responsibility of RS. Correlation analysis between soil Pb concentration with spectral parameters and soil constituents can be used to detect significant factors for prediction models. In the present study, the correlation coefficient values between Pb concentrations with spectral parameters and some soil components which were expected to have a relationship to spectral data was showed in Table 2. The Pearson correlation coefficient between Pb and the corresponding absorption band values (γ) showed a best correlation of R = 0.33 on a natural scale and R = 0.33 (same) on a log scale at a wavelength of 512 nm. While the best correlation for Pb concentration was found with total C, where R = 0.29 on a natural scale and R = 0.44 on a log scale. The result inTable 4andFigure 4shows that the removal of nine outlier samples from the regression data set (n = 73) was a good process to improve the accuracy of Pb concentra-tion predicconcentra-tions.
The correlation line for PLSR model highlighted the regions of high importance in the spectra
to detect soil Pb (Figure 5). The component values between dotted lines within the upper and
lower bounds evident in the loadings plot are highly correlated with the measured data. It is clear from this figure that the wavelength between 500 and 612 nm may be referred to important spectra to detect Pb concentration which was in agreement with the mentioned Pearson correlation.
Table 4.Comparison between prediction accuracy factors for cross-validation of PLS regression for Pb concentration before and after removable outliers values.
No. of samples R R2 RMSE SLOP SE
CV All samples
73 0.70 0.54 0.38 0.58 0.39
After removal of nine outlier values
64 0.75 0.57 0.36 0.64 0.36
SECV: Standard error cross-validation.
Reference Pb (log, mg/kg)
Predicted Pb (log, mg/kg)
To validate the regression model, the predicted values should be equal to the reference values. Regression analysis between predicted model and actual concentrations for seven extra samples (which were not used in the data set) indicated that the PLSR model had the predictive power to assess Pb
contamination to a good degree as evidenced by the linear regression correlation of R2 = 0.60
(Figure 6). The final regression model is the one that, in general, exhibited the greatest goodness of fit between predicted and actual Pb concentrations (Al Maliki, Bruce, and Owens2012a).
The final models developed here clearly demonstrated that spectral reflectance could potentially be used to predict soil Pb concentrations in soils contaminated by long-term atmospheric smelter Pb emissions. Nevertheless, there are two critical points to consider. First, the results can be quite site-specific and are unlikely to transfer to another site for two reasons: the calibration set of less than 100 samples was not enough to be generalized and the prediction models generated here do not include independent validation data. Second, Pb prediction models are not standardized because they are based on the feature space of the study area, which may not be applicable to another area. Therefore, the use and interpretation of models by other users requires some additional factors to be addressed such as soil type, particle size, soil texture, and organic matter in order to be truly effective.
Using spectrally active of soil constituents as indicator to detect soil Pb
This research showed that soil C regression model has the best detection in wavelengths (Table 3), and it has also shown that the best correlation was found between soil Pb and soil C. The spectral predictive mechanism using C fraction as a key for an indirect spectroscopy method to detect Pb concentration was applied in the VNIR region. High correlation coefficients for C contents were obtained in the visible spectral range (500–700 nm) (Ji et al.2010). While in the current study, both Pearson correlation result of R > 0.50 and correlation loading for regression model observed that the wavelengths between 380 and 550 nm were important to detect soil C. Another study mentioned that the suitable correlation between organic carbon (spectrally active content) with a set of HMs such as Cu, Pb, and Zn (spectrally nonactive)
Figure 5.Correlation line loading plot for the PLSR model highlighting important wavelengths between 500 and 612 nm which may be important regions for soil Pb detection; factor 1 represented the most important sources of information.
Figure 6. Evaluation of Pb concentration for seven extra samples (out dataset) using predicted PLSR model.
(with R = 0.67, 0.66, and 0.65, respectively) was helpful to reveal coherences between these variable and spectral bands than that identified (predicted) of HMs in similar spectral regions of organic carbon (Vohland, Bossung, and Fründ 2009). The organic matter present in soil which constitutes a large fraction of the C content of a soil is considered to be the more spectrally active due to strong molecular vibrations in the NIR region. As a result, the spectral band of Pb at 512 nm was in therange of spectral bands for total C. So, soil C contents may be useful to assist in predicting Pb concentration in aspecific spectral range of C due to strong bond that exists between them.
Conclusion
This study has shown that NIRS may potentially estimate the total Pb of soil around the Port Pirie smelter. Although Pb is a spectrally inactive metal, a good predictive model for Pb in VNIR was achieved by combining soil RS with soil measurements using a PLSR model. The regression model showed that Pb concentration when converted to a log scale correlated optimally with the modified reflectance spectra. In addition to cross-validation, the accuracy and precision of the PLSR model were determined using an extra set of data that was not used in the original prediction. This data set was evaluated through many prediction accuracy parameters such as R2, RMSE, RPD, and RER. The chemometric methods also used correlation loading function to highlight the regions of high spectral importance between 500 and 612 nm for Pb content determination. Total C was more spectrally active and more correlated with Pb than other elements in soil such as iron oxides and clays. In addition to an acceptable correlation between total C and soil Pb, the region associated with Pb detection was also coincident with the region where soil C could be detected. Thus, these results indicated that the absorption features for C content can be utilized to predict soil Pb. While the regression model for Pb concentration developed here was indeed quantitative and accurate, this model was developed for a specific soil and may not be applicable to another soil that has different feature space.
Acknowledgments
All authors acknowledge the support of the future Industries Institute (FII), the University of South Australia, while undertaking this research. The first author gratefully acknowledges the financial support of the Iraqi government, Ministry of Higher Education and Scientific Research for a Ph.D. stipend.
References
Al Maliki, A., D. Bruce, and G. Owens.2012a. Capabilities of remote sensing hyperspectral images for the detection of
lead contamination: A review. Paper presented at ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 55. Melbourne.
Al Maliki, A., D. Bruce, and G. Owens.2012b. Combining AHP and TOPSIS approaches to support site selection for a
lead pollution study. Paper presented at 2nd International Conference on Environmental and Agriculture
Engineering, 1–8. South Korea.
Al Maliki, A., D. Bruce, and G. Owens.2015. Spatial distribution of Pb in urban soil from Port Pirie, South Australia.
Environmental Technology & Innovation 4:123–36. doi:10.1016/j.eti.2015.05.002.
Ben-Dor, E., J. Irons, and G. Epema.1999. Soil reflectance. In Manual of remote sensing: Remote sensing for the earth
sciences, Vol. 3, 3ed ed. 111–89. New York: John Wiley and Sons Inc.
Bickel, M. 2010. Spatial and temporal relationships between blood lead and soil lead concentrations in Detroit,
Michigan. M.S diss., Civil and Environmental Engineering, Wayne State University, USA.
Bilgili, A., H. Van Es, F. Akbas, A. Durak, and W. Hively.2010. Visible-near infrared reflectance spectroscopy for
assessment of soil properties in a semi-arid area of Turkey. Journal of Arid Environments 74:229–38. doi:10.1016/j.
jaridenv.2009.08.011.
Body, P., G. Inglis, P. Dolan, and D. Mulcahy.1991. Environmental lead: A review. Critical Reviews in Environmental
Control 20 (5–6):299–310. doi:10.1080/10643389109388403.
Bray, J., R. Rossel, and A. McBratney.2009. Diagnostic screening of urban soil contaminants using diffuse reflectance
Brunet, D., B. Barthes, J. Chotte, and C. Feller.2007. Determination of carbon and nitrogen contents in alfisols, oxisols and ultisols from Africa and Brazil using NIRS analysis: Effects of sample grinding and set heterogeneity. Geoderma
139 (1–2):106–17. doi:10.1016/j.geoderma.2007.01.007.
CAMO Software.2012. The Unscrambler 10.2. AS, NSW, Australia.http://www.camo.com.
Choe, E., K. Kim, S. Bang, I. Yoon, and K. Lee. 2009. Qualitative analysis and mapping of heavy metals in an
abandoned Au–Ag mine area using NIR spectroscopy. Environmental Geology 58 (3):477–82. doi:
10.1007/s00254-008-1520-9.
Choe, E., F. Meer, F. Ruitenbeek, H. Werff, B. De Smeth, and K. Kim.2008. Mapping of heavy metal pollution in
stream sediments using combined geochemistry, field spectroscopy, and hyperspectral remote sensing: A case study
of the Rodalquilar mining area, SE Spain. Remote Sensing of Environment 112:3222–33. doi:10.1016/j.
rse.2008.03.017.
Clark, N., V. King, A. Swayze, and N. Vergo.1990. High spectral resolution reflectance spectroscopy of minerals.
Journal of Geophysical Research-Solid Earth and Plant 95 (B8):12653–80. doi:10.1029/JB095iB08p12653.
Csorba, A., V. Lang, K. Szalay, E. Micheli, and L. Fenyvesi. July 8-12,2012. Prediction of soil organic carbon and
calcium carbonate content based on reflectance spectroscopy and multivariate modelling methods. Infomation technology, automation and precision farming. Paper presented at International Conference of Agricultural Engineering - CIGR-AgEng 2012: Agriculture and Engineering for a Healthier Life: C-1306ref.10., Valencia, Spain, .
Dao, L., L. Morrison, and G. Kiely.2013. Spatial distribution of potentially bioavailable metals in surface soils of a
contaminated sports ground in Galway, Ireland. Environmental Geochemistry and Health 35 (2):227. doi:10.1007/
s10653-012-9478-7.
EPA.2005. Composite soil sampling in site contamination assessment and management, 5. Adelaide, South Australian:
South Australian Environment Protection Authority.
Gannouni, S., N. Rebai, and S. Abdeljaoued.2012. A Spectroscopic approach to assess heavy metals contents of the
mine waste of Jalta and Bougrine in the north of Tunisia. Journal of Geographic Information System 4 (3):242–53.
doi:10.4236/jgis.2012.43029.
Griffith, D.2002. The geographic distribution of soil lead concentration: Descriptionand concerns. Urisa 14 (1):5–15.
Hayes, D. J. M. 2012. Development of near infrared spectroscopy models for the quantitative prediction of the
lignocellulosic components of wet Miscanthus samples. Bioresource Technology 119:393–405. doi:10.1016/j.
biortech.2012.05.137.
Hengla, T., G. Heuvelinkb, and A. Stein.2004. A generic framework for spatial prediction of soil variables based on
regression-kriging. Geoderma 120:75–93. doi:10.1016/j.geoderma.2003.08.018.
Hong-Yan, R., Z. Da-Fang, A. Singh, P. Jian-Jun, Q. Dong-Sheng, and S. Run-He.2009. Estimation of As and Cu
contamination in agricultural soils around a mining area by reflectance spectroscopy: A case study. Pedosphere 19
(6):719–26. doi:10.1016/S1002-0160(09)60167-3.
Ji, J., Y. Song, X. Yuan, and Z. Yang.2010. Diffuse reflectance spectroscopy study of heavy metals in agricultural soils
of the Changjiang River Delta, China. Paper presented at 19th World Congress of Soil Science, Soil Solutions for a Changing World, Brisbane, Australia.
Kooistra, L., R. Wehrens, R. Leuven, and L. Buydens.2001. Possibilities of visible–Near-infrared spectroscopy for the
assessment of soil contamination in river floodplains. Analytica Chimica Acta 446 (1–2):97–105. doi:
10.1016/S0003-2670(01)01265-X.
Malley, D., and P. Williams.1997. Use of near-infrared reflectance spectroscopy in prediction of heavy metals in
freshwater sediment by their association with organic matter. Environmental Science and Technology 31 (12):3461–
467. doi:10.1021/es970214p.
Martley, E., B.Gulson, and H. Pfeifer.2004. Metal concentrations in soils around the copper smelter and surrounding
industrial complex of Port Kembla, NSW, Australia. Science of the Total Environment 325:113–27. doi:10.1016/j.
scitotenv.2003.11.012.
Mason, P., and J. Huntington. 2012. HyLogger 3 components and pre-processing: An overview. Technical Note
2012-002. Northern Territory Geological Survey.
McMichael, A., P. Baghurst, G. Vimpani, N. Wigg, E. Robertson, and S. Tong.1994. Tooth lead levels and iq in
school-age children: The port pine cohort study. American Journal of Epidemiology 140 (6):489–99. doi:
10.1093/oxford-journals.aje.a117275.
Miller, W., and D. Miller.1987. A micro-pipette method for soil mechanical analysis. Communications in Soil Science
and Plant Analysis 18 (1):1–15. doi:10.1080/00103628709367799.
Nduwamungu, C., N. Ziadi, L. Parent, G. Tremblay, and L. Thuriès.2009. Opportunities for, and limitations of, near
infrared reflectance spectroscopy applications in soil analysis: A review. Canadian Journal of Soil Science 89 (5):531–
41. doi:10.4141/CJSS08076.
Niazi, N. K.2011. Variability, speciation and phytoremediation of soil Arsenic at Cattle Dip sites in NSW, Australia.
Pandit, M., M. Filippelli, and L. Li.2010. Estimation of heavy-metal contamination in soil using reflectance
spectro-scopy and partial least-squares regression. International Journal of Remote Sensing 31 (15):4111–23. doi:10.1080/
01431160903229200.
Pastor, I., J. Pedreño, I. Gómez, and M. Koch.2008. Identifying optimal spectral bands to assess soil properties with
VNIR radiometry in semi-arid soils. Geoderma 147:126–32. doi:10.1016/j.geoderma.2008.08.004.
Schwartz, G.2012. Reflectance spectroscopy as a rapid tool for qualitative mapping and classification of hydrocarbons
soil contamination. PhD Thesis, Porter School of environmental studies, Tel Aviv University.
Siebielec, G., G. McCarty, T. Stuczynski, and J. N. Reeves. 2004. Mid-infrared diffuse reflectance spectroscopy for
measuring soil metal content. Journal of Environmental Quality 33 (6). doi:10.2134/jeq2004.2056.
Stenberg, B., A. Raphael, R. Viscarra, M. Abdulmounem, and W. Juhanna.2010. Visible and near infrared
spectro-scopy in soil science. Advances in Agronomy 107: 163–215. doi:10.1016/S0065-2113(10)07005-7.
Summers, D. 2009. Discriminating and mapping soil variability with hyperspectral reflectance data. PhD Thesis,
Faculty of science, School of Earth and Environmental Science, Adelaide University.
Taylor, M.2012. Lead poisoning of Port Pirie children: A long history of looking the other way. The conversation.
Accessed August 20, 2012.http://theconversation.com/lead-poisoning-of-port-pirie-children-a-long-history-of-look
ing-the-other-way-8296.
Tong, S., E. Yasmin, V. Schirnding, and T. Prapamontol.2000. Environmental lead exposure: A public health problem
of global dimensions. Bulletin of the World Health Organization 78 (9): 1068–77.
Vohland, M., C. Bossung, and H. Fründ.2009. A spectroscopic approach to assess trace–Heavy metal contents in
contaminated floodplain soils via spectrally active soil components. Plant Nutrition and Soil Science 172 (2):201–09.
doi:10.1002/jpln.200700087.
Wetterlind, J.2009. Improved farm soil mapping using near infrared reflection spectroscopy. PhD thesis., Department
of Soil and Environment, Swedish University of Agricultural Sciences.
Wu, Y., J. Chen, J. Ji, P. Gong, Q. Liao, Q. Tian, and H. Ma.2007. A Mechanism study of reflectance spectroscopy for
investigating heavy metals in soils. Soil Science Society of America Journal 17 (3): 918–926.
Wu, Y., J. Chen, X. Wu, Q. Tian, J. Ji, and Z. Qin.2005. Possibilities of reflectance spectroscopy for the assessment of
contaminant elements in suburban soils. Applied Geochemistry 20 (6):1051–59. doi:10.1016/j.
apgeochem.2005.01.009.
Yaolin, L., L. Wei, W. Guofeng, and X. Xinguo. 2011. Feasibility of estimating heavy metal contaminations in
floodplain soils using laboratory-based hyperspectral data—A case study along Le’an River, China. Geo-Spatial
Information Science 14 (1):10–16.
Zhang, Z., J. Wen, and D. Zhao. 2010. Band selection method for retrieving soil lead content with hyperspectral
Remote sensing data. Paper presented at the Society of Photo-Optical Instrumentation Engineers (SPIE) Remote
