Environ. Sci. Technol. 2002, 36, 2742-2747
Estimate of Heavy Metal Contamination in Soils after a Mining Accident Using Reflectance Spectroscopy THOMAS KEMPER* AND STEFAN SOMMER Joint Research Centre of the European Commission, Institute for Environment and Sustainability, Soil and Waste Unit, T.P. 262, I-21020 Ispra (Va), Italy
The possibility to adapt chemometrics approaches for the quantitative estimation of heavy metals in soils polluted by a mining accident was explored. In April 1998, the dam of a mine tailings pond in Aznalco´ llar (Spain) collapsed and flooded an area of more than 4000 ha with pyritic sludge contaminated with high concentrations of heavy metals. Six months after the end of the first remediation campaign, soil samples were collected for chemical analysis and measurement of visible to near-infrared reflectance (0.352.4 µm). Concentrations for As, Cd, Cu, Fe, Hg, Pb, S, Sb, and Zn were well above background values. Prediction of heavy metals was achieved by stepwise multiple linear regression analysis (MLR) and an artificial neural network (ANN) approach. It was possible to predict six out of nine elements with high accuracy. Best R 2 between predicted and chemically analyzed concentrations were As, 0.84; Fe, 0.72; Hg, 0.96; Pb, 0.95; S, 0.87; and Sb, 0.93. Results for Cd (0.51), Cu (0.43), and Zn (0.24) were not significant. MLR and ANN both achieved similar results. Correlation analysis revealed that most wavelengths important for prediction could be attributed to absorptions features of iron and iron oxides. These results indicate that it is feasible to predict heavy metals in soils contaminated by mining residuals using the rapid and cost-effective reflectance spectroscopy.
The conventional method of estimating the spatial distribution of heavy metals is by a raster sampling and a time-consuming laboratory analysis followed by geostatistical interpolation (7, 8). This paper shall discuss the possibilities to detect qualitatively and quantitatively heavy metal contamination using reflectance spectroscopy. The approach is based on the direct link between the heavy metal contamination and the spill of contaminated mine flotation mud after the break of a mine tailings dam in Spain. The sludge exhibits distinct mineralogical features in contrast to the soil mineralogy of the affected lower parts of the catchment. This contrast enables the detection using spectroscopy data and a quantitative mapping of heavy metals. A successful quantification based on spectral measurements could open the possibilities for a quantitative airborne mapping based on new hyperspectral sensor techniques, which could overcome the qualitative mapping of minerals of interest (9). Reflectance spectroscopy is known as a fast nondestructive method for the characterization of different materials. In agriculture and food research, it became known as nearinfrared analysis (NIRA) and is used extensively as quantitative technique (10-12). It is based on the assumption that the concentration of a constituent is proportional to a combination of several absorption features. Because it is an empirical method, it does not need other physical or chemical factors or assumptions. Relatively few researchers have extended the methods for the extraction of soil/sediment information. Ben-Dor and Banin (13, 14) have used this method for the prediction of different soil constituents. Malley and Williams (15) determined the heavy metal contamination in lake sediments using multivariate methods. Udelhoven and Schu ¨ tt (16) used artificial neural networks for the chemical characterization of sediments. The most common methods of calibrating the chemometrics systems are by multivariate statistical approaches such as multiple linear regression (MLR), principal components regression (PCR), or partial least-squares regression (PLS). Alternatives to these statistical approaches are artificial neural networks (ANN), whose applications in the field of chemometrics increase strongly (17). The main advantage of ANN is its ability to model any nonlinear relation (18).
Materials and Methods Introduction Acid mine drainage (AMD) from mine waste and the contamination of water and soils with heavy metals are considered major problems in mining areas (1). In addition, natural mineralized areas, even when not mined, can affect the environment (2). Acid water is produced by the oxidation of the common iron disulfide mineral pyrite. Heavy metals can be leached from rocks that are exposed to the acid water. This process is also significantly enhanced substantially by bacterial action (3). In mining areas, release of heavy metals is accelerated due to the increased oxidation rates, which are caused by providing greater accessibility of air through mine workings, waste rock, and tailings by mineral processing (4). Furthermore, there is the added risk of mining accidents such as in Aznalco´llar (Spain 1998) or Baia Mare/Borsa, (Romania 2000) (5, 6). During these accidents, huge amounts of mine waste and toxic substances (heavy metals, cyanide) were set free in one go, contaminating rivers and alluvial soils. * Corresponding author phone: +39-0332-789773; fax: +39-0332789469; e-mail:
[email protected]. 2742
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 36, NO. 12, 2002
Study Area. The study area is situated approximately 40 km west of Seville (Spain, Figure 1). After a collapse of a tailings dam at the Aznalco´llar Mine in April 1998, an area of 4286 ha, mainly flood plains of the Agrio and Guadiamar rivers and agricultural land, was covered with toxic sludge, endangering the wetlands of the Don ˜ ana National Park. Sludge and the contaminated topsoils had been removed within 6 months after the accident by using heavy machinery. Monitoring activities have been started by Spanish institutions to control the possible mobility of residual contaminants due to oxidation processes (5). Soil Sampling and Processing. Within the contaminated area, six test sites were selected in zones, where soil sampling had already taken place by other research groups (19, 20), which could be used for validation and detection of changes in the contamination. At each site, sampling transects, consisting of 5-12 points, were selected following possible contamination gradients. At each point, soil samples were taken at four different levels (0-2, 2-20, 20-40, 40-60 cm) resulting in 214 soil samples. In addition to the collected soil samples, a mixture series with artificially contaminated soils was produced. For this 10.1021/es015747j CCC: $22.00
2002 American Chemical Society Published on Web 05/10/2002
TABLE 1. Heavy Metal Concentration (Average, Range) of Contaminated Soils Compared to Background Soils and the Ranges of Concentration for Normal Soils background
experiment, three noncontaminated soils from the area, but not affected by the accident, were used to add sludge in increasing weight percentage from 0% to 100% (32 samples). Chemical Analyses. From the 214 soil samples, major and trace elements were estimated directly through X-ray fluorescence analysis (XRF). Mercury was analyzed directly with the AMA 254, an atomic absorption spectrometer (AAS), which is constructed specific for mercury detection. Arsenic was analyzed in solution, obtained by microwave acid digestion, by AAS with hydride generation. Cadmium was analyzed in solution, obtained by microwave acid digestion, by AAS in a graphite furnace. The contents of the elements differ substantially in absolute values as well as in data range; therefore, the data were range-scaled between 0 and 1. Laboratory Spectroscopy. Laboratory reflectance measurements of the samples were taken using a high spectral resolution ASD Fieldspec II spectroradiometer. This instrument covers the spectral range between 350 and 2500 nm with a spectral resolution between 3 and 10 nm, interpolated to 1 nm. Illumination was provided by an ASD high reflectance probe, and a calibrated Spectralon panel was used as a white reference. Each sample was measured four times and averaged afterward. Samples were air-dried and sieved through a 2 mm sieve prior to measurement. The spectral processing was done using IDL/ENVI software. The spectra were degraded from 1 nm to 5, 10, and 20 nm using a Gaussian model that takes into account the band center and the full width half-maximum (fwhm). The resampling reduces the number of wavelengths from 2151 to 108, which smoothes the spectra and reduces the problems of over-fitting. The spectral channel degradation technique has been found effective for prediction of different soil properties (14). Furthermore, several data treatments were applied to enhance the spectral features. The reflectance data were transformed to absorption (log 1/R, where R is reflectance) to account for scattering effects. This method is widely applied in chemometrics (21). From these absorption spectra, first- and second-order derivatives were calculated using the Savitzky-Golay method (22). In addition, the reflectance data were standardized in such a way that all spectra have a mean reflectance of 0 and a standard deviation of 1. This method allows for the comparison of the data independent of the intensity scale and enhances the absorption features and curve shape (23). Model Calibration and Validation. In the calibration stage, the chemometrics models are developed and tested. First, the chemical information was combined with the spectral information. Prior to calibration, the spectra were
normala
[ppm]
mean
min
max
mean
min
max
min
max
As Cd Cu Hg Pb Sb Zn
16.75 0.43 50.65 0.06 52.65 275.25 191.69
8 0.05 18 0.01 18 196 98
27 1.88 178 0.29 221 382 748
61.26 1.26 120.44 0.45 202.21 438.91 380.44
7 0.05 17.5 0.01 17.5 196 94
442 14.8 521 13.9 3331.5 3362 3887
1 0.1 2 0.02 2 0.2 10
20 0.6 40 0.2 80 10 80
a
FIGURE 1. Study area.
contaminated
Alloway (26).
sorted from highest to lowest contamination and divided into two data sets by selecting odd- and even-numbered samples for calibration and validation sets. This method ensures that in each set the full range of concentrations is represented. The resulting calibration set consists of 119 samples and the validation set of 118. For calibration based on multivariate statistics, a stepwise regression procedure was used in order to limit the number of input variables. Variables were entered or removed from the model depending on the significance of the F value. Constituents were modeled one at a time. The derived calibration equations were used to predict the concentrations of the validation set from the spectra. The building and training of the ANN was accomplished using the Stuttgart Neural Network Simulator (SNNS) using a feed-forward network topology. In contrast to MLR, ANN are able to model all constituents at a time. In a previous test series, the optimum size of the network was defined with only one hidden layer and a ratio between hidden and input neurons set to 0.15. Resilient back-propagation (Rprop) (24) was used exclusively as learning algorithm. The learning phase was terminated after 30-60 learning cycles, when the error of the validation set did not improve any more. For cross-validation purposes, the validation and calibration sets were exchanged in a second calibration run. The results from these reciprocal calibrations were similar to the previous ones. The predicted concentration values of the 118 samples validation set were correlated to the reference data of the chemical analysis. Correlations are expressed as coefficients of determination (R2). Because of the skewed distribution of concentration values, the coefficient of determination would overestimate the quality of the results. Therefore, additional error measures were calculated. The standard error of prediction (SEP) was calculated as standard deviation of differences between reference values and predicted values. With this SEP, two other measures were used for the evaluation of the calibration. The RPD is the ratio of the standard deviation of a reference chemistry set to the SEP; it should be as high as possible (RPD > 3 for prediction purposes). The RER is the ratio of the range in the prediction set to the SEP. It should be greater than 10 (15).
Results and Discussion Concentrations of Heavy Metals in Soil Samples. The first studies after the accident have shown that the main contaminants, in descending order of concentrations, were Zn, Pb, Cu, As, Sb, and Cd (19, 20). Additionally, we also analyzed Hg, Fe, and S in this study. Hg was included because it is one of the most toxic elements when released into the environment (25). Fe and S were included because they are the driving variables for the acidification process. Table 1 shows mean heavy metal (HM) concentrations and ranges in the samples compared to normal soil ranges (26). After VOL. 36, NO. 12, 2002 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2743
FIGURE 2. Average contamination profiles of contaminated (continuous) and uncontaminated (dashed) soils for arsenic and lead.
FIGURE 3. Reflectance spectra of a soil artificially contaminated with sludge in wt % (left). Reflectance spectra of soil samples with different contamination levels (right). Pb concentrations for comparison of the contamination level. the end of the first cleanup operation, the contamination is partly still high, which corresponds to the analysis of Simo´n et al. (20). The average values of contaminated soils for all HM are still above the average values of background soils, and the range of contamination values is very wide. The HM concentrations are rapidly decreasing with increasing depth, but even the levels in 40-60 cm are still clearly above the background values. This is demonstrated for selected HM in Figure 2 in comparison to uncontaminated soils of the area, which shows that the HM have already affected deeper horizons. Spectral Measurements. Already, the visual inspection of the measured soil spectra showed a significant difference between contaminated and uncontaminated soils. Figure 3 shows the reflectance change for a soil that was contaminated artificially and for some examples of variably contaminated samples collected in the study area. In both cases, the spectral signatures change significantly and in the same characteristic way as function of their level of contamination with material from the Aznacollar tailing pond. The most evident change is a strong decrease of overall albedo. In particular, at longer wavelengths, the spectrum is leveling off strongly. Besides the general decrease in albedo, some changes in the absorption features are apparent. With the building up of a 2744
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 36, NO. 12, 2002
very wide absorption feature between 700 and 1400 nm, the absorption features at 1400, 1900, and 2200 nm are diminishing and are almost extinguished at sludge concentrations above 50 wt %; smaller absorption features disappear even earlier. The same behavior can be observed for the soil samples collected in the study area. Prediction of HM using MLR and ANN. It was possible to detect six contaminants out of nine with high precision with both methods (Table 2). Coefficients of determination for Hg, Pb, and Sb are above 0.9. Also, RPDs > 4 and RER > 20 are indicative of very good models. The validation results for As, Fe, and S gave R2’s between 0.7 and 0.9, RPDs between 2 and 4, and RERs between 10 and 16, which are also satisfactory. The results would be even better if some outliers would have been left out. Figure 4 shows the scatterplots of predicted versus measured concentrations. For Cd, Cu, and Zn, the models did not provide a significant prediction. The poor prediction quality can be attributed to a different geochemical behavior. According to Blume and Bru ¨ mmer (27) and Ainsworth et al. (28), the solubility of HM decreases in the following order: Cd > Zn > Cu > As > Hg ∼ Pb ∼ Sb. Accordingly, Simo´n et al. (20) found that the bigger part of the Cd, Cu, and Zn penetrated the soil and precipitated from the solution phase of the spill, while the other elements were
TABLE 2. Validation Results for ANN and MLRa ANN
As Cd Cu Fe Hg Pb S Sb Zn
MLR
R2
SEP
RPD
RER
λ
type
R2
SEP
RPD
RER
λ
type
0.858 0.494 0.446 0.714 0.929 0.940 0.845 0.927 0.220
0.0073 0.0850 0.0965 0.0447 0.0332 0.0282 0.0410 0.0282 0.0861
3.28 0.96 0.84 1.98 4.30 5.89 2.66 4.55 0.51
15.88 4.32 4.57 9.83 24.21 33.43 13.05 22.72 2.17
20 20 1 10 10 10 20 10 20
der2 der1 der1 der1 std std std der1 der1
0.837 0.510 0.540 0.721 0.957 0.944 0.839 0.929 0.234
0.0070 0.0794 0.0858 0.0569 0.0276 0.0312 0.0681 0.0331 0.0827
3.83 1.23 1.23 1.92 5.77 5.30 2.38 4.66 0.59
16.98 5.95 6.95 9.53 33.21 30.00 13.04 25.88 2.79
20 5 10 20 20 10 10 20 5
abs std abs std std std std std abs
a Results of ANN calculated all at a time, MLR based on single predictions. λ denotes the wavelength interval yielding the best results; type refers to the type of transformation (abs ) absorption, der1 ) 1st derivative, der2 ) 2nd derivative, std ) standardization).
FIGURE 4. Plot of measured versus predicted concentrations of elements (validation). The predicted values are based on the methods and preprocessing combinations described in Table 3. rather sedimented as part of the pyrite dominated solid phase. Hence, a great part of the mobile HM Cd, Cu, and Zn are distributed in the soil profile by an independent process and, thus, not directly linked to the pyritic sludge, which most significantly influences the spectral response of the contaminated soils. Alastuey et al. (29) and Galan et al. (30) found equal results in their studies of the mobility of HM in sludge and soils in the study area. The prediction quality strongly depends on the preprocessing steps applied to the reflectance spectra. For ANN, best results could be achieved using first- and secondorder derivatives and standardized spectra; with MLR, standardized and absorption spectra obtained the best results. The standardized spectra and derivatives gave the best results because they are independent of baseline effects, which might be caused, for example, by differences in grain size rather than in chemical composition. The second-order derivatives could eliminate baseline effects even better and enhance minor absorption features more, but they are more
sensitive to noise and thus produced only once the best results. Generally, broader resampling intervals (10 and 20 nm) yielded better results, which prove the usefulness of the spectral degradation technique. As can be seen in Figure 3, the sludge contamination does not produce any narrow absorption features but broader absorption bands, which can still be resolved with broader wavelength intervals. Besides the noise reduction effect, the reduction of variables stabilizes the variable selection in stepwise MLR. In fact, the selection based on the F-value encountered problems with 1 and 5 nm resampling intervals. Because of the large number of variables and the high collinearity, more wavelengths than samples had been selected inhibiting the matrix inversion. However, because broader wavelengths were preferred, the differences between ANN and stepwise MLR are minor; the MLR performs even slightly better. This might indicate a relatively linear behavior of well-predicted HM with regard to contamination-induced changes of the reflectance spectra. VOL. 36, NO. 12, 2002 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2745
TABLE 3. Correlation Matrix of Elementsa As As Cd Cu Fe Hg Pb S Sb Zn a
Cd
Cu
Fe
Hg
Pb
S
0.79 0.85 0.91 0.88
0.97 0.93 0.96
0.95 0.99
0.96
Sb
Zn
0.78 0.84 0.86 0.86 0.85 0.88
0.60 0.70 0.61 0.63 0.70 0.64 0.83
0.68
Significant coefficients (R 0.01) > 0.6.
FIGURE 5. Correlogram of elements and single wavelengths (standardized spectra, 10 nm). This shows that the choice between the application of statistical models or ANN methods is strongly dependent on the situation considered (31). When the problem is known to be nonlinear, the application of ANN would be advantageous, because they are not based on the assumption of a statistical distribution and are, thus, capable of modeling any measurable linear or nonlinear function (18). However, it has to be taken into consideration that the application of ANN is more complex and more experience is necessary for the correct selection of network architecture, learning algorithm, and parameters. Correlation Analysis. For a better understanding of the functioning of multivariate statistics and ANN, which are often regarded as black boxes, the correlation structure of HM and spectral information was examined because both prediction methods make use of the correlation structure among intercorrelated variables (Table 3). Most heavy metals, iron, and sulfur (As, Fe, Hg, Pb, S, Sb) are strongly intercorrelated (r > 0.8); only copper and zinc are less correlated with the other elements. Cadmium plays an intermediate roll, as it is correlated significantly with As, Fe, Hg, Pb, S, and Sb and relatively high (r ) 0.83) with Zn. The correlation structure reflects clearly the prediction results: all elements that are highly correlated were predicted most precisely. The separation of different subgroups is even more clearly visible in the correlogram (Figure 5) illustrating the correlation between the reflectance at single wavelengths and respective elements. The structure is similar for all elements, but there are strong differences in the strength of the correlation. The areas of strongest correlations can be attributed mainly to absorption features of iron and iron oxides. The high positive correlation at about 550 nm displays the change from opaque to transparent behavior, which is caused by an intense chargetransfer band in the UV region of the spectrum (32). The broad area with the highest negative correlations (700-1400 nm) is linked to the strong absorption band of the ferrous 2746
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 36, NO. 12, 2002
ion centered at about 1000 nm (33). Strong molecular water bands at 1400 and 1900 nm in combination with hydroxyl absorptions, centered at 1400 and 2200 nm, are related to the three peaks in the central part of the correlogram. In particular, secondary clay minerals show distinct absorption features at these wavelengths (34). In this data set, higher amounts of secondary clay minerals represent less contaminated soils. Another feature at 2350 nm is caused by carbonate absorption (33); like the ones related to OH-bearing secondary minerals, its depth is negatively correlated with the HM concentration. Most of the spectral wavelengths selected by stepwise MLR are closely related to the active spectral features described previously. The study demonstrated the feasibility of applying MLR and ANN to build reliable chemometric models linking field spectral measurements and geochemical variables from laboratory analyses. The sampling campaign after the first remediation phase with the mechanical removal of the sludge and surface soil layer has confirmed that considerable contamination with residual pyritic tailings material and associated HM is remaining a problem. The remaining sludge will proceed to react and may cause problems of acidification. Thus, monitoring activities of the affected area are indispensable. In July 2000, a second sampling campaign was carried out in order to investigate the potential of the presented methodology for monitoring purposes. The new data are used for further validation and refinement of the models derived from the 1999 data. After this step, the method should be further tested directly “on site” as a fast screening method for the evaluation of the contamination status of a larger area. Transfer of the derived models to other problem areas of HM contamination (e.g., sewage sludge) is not intended at this stage as this would require new calibration activities. However, it is planned to extend the models to available data of airborne hyperspectral imaging spectrometers such as HyMap and DAIS, which would allow for a fully synoptic estimation of the remaining contamination in the entire affected area. This move from field/laboratory point sample spectroscopy to image spectra recorded at an altitude of approximately 3000 m above ground implies that a large number of additional factors and constraints have to be investigated. The following factors are the most important to be taken into account in the data processing. The signalto-noise ratio (SNR) of image spectra is usually significantly lower, which has a clear impact on the modeling strategy and choice of methodology. An accurate radiometric and atmospheric correction that provides reflectance spectra, which are comparable to the field/laboratory spectra, is an indispensable, not at all trivial, prerequisite. Although the spatial resolution of the available data is reasonably good (approximately 5 × 5 m2), most of the extracted pixel spectra represent a mixture of different surface materials containing also nonsoil material (e.g., green and dry vegetation and water). As the laboratory models are developed for pure soil spectra, methods such as spectral mixture modeling have to be applied, which allow for the modeling of the pure soil spectral information from the mixed spectrum.
Acknowledgments We express our gratitude to the head of the IES-Land Management Unit (IES-LM) Mr. Meyer-Roux and to the head of the IES Soil and Waste Unit (IES-SW) Mr. Bidoglio for supporting this study as part of the JRC institutional work program. Mr. M. D’Alessandro and the members of the IESSW inorganic laboratory team provided invaluable support during the field campaign and through the laboratory analysis of the soil samples. Furthermore, we thank the Consejeria de Medioambiente of the Junta de Andalucia in Seville for their support.
Literature Cited (1) U.S. Bureau of Mines. Proceedings of the International Land Reclamation and Mine Drainage Conference and 3rd International Conference on the Abatement of Acidic Drainage; U.S. Bureau of Mines Special Publication SP 06D-94; 1994; pp 1-4. (2) Runnels, D. D.; Shepherd, T. A.; Angino, E. E. Environ. Sci. Technol. 1992, 26 (12), 2316-2323. (3) Nordstrom, D. K. Acid sulfate weathering; Soil Science Society of America Special Publication No. 10; 1979; pp 37-56. (4) Nordstrom, D. K.; Alpers, C. N. Proc. Natl. Acad. Sci. U.S.A. 2000, 96, 3455-3462. (5) Grimalt, J. O.; Ferrer, M.; Macpherson, E. Sci. Total Environ. 1999, 242 (1-3), 3-11. (6) European Commission 2000. Communication from the Commission, Comm 2000 664 final, 23.10.2000. http://europa.eu.int/ eur-lex/en/com/cnc/2000/com2000_664en01.html (accessed 03/04/02). (7) Leenaers, H.; Okx, J. P.; Burrough, P. A. Soil Use Manage. 1990, 6, 105-114. (8) Von Steiger, B.; Webster, R.; Schulin, R.; Lehmann, R. Environ. Pollut. 1996, 94 (2), 205-215. (9) Swayze, G. A.; Smith, K. S.; Clark, R. N.; Sutley, S. J.; Pearson, R. M.; Vance, J. S.; Hageman, P. L.; Briggs, P. H.; Meier, A. L.; Singleton, M. J.; Roth, S. Environ. Sci. Technol. 2000, 34, 47-54. (10) Williams, P. C., Norris, K., Eds. Near-infrared analysis in agriculture and food industries; American Association of Food Chemists Inc.: St. Paul, MN, 1987. (11) Massart, D. L.; Vandeginste, B. G. M.; Deming, S. N.; Buydens, L. M. C.; De Jong, S.; Lewi, P. J.; Smeyers-Verebeke, J. Handbook of chemometrics and qualimetrics; Elsevier: Amsterdam, The Netherlands, 1998; Part A and B. (12) Hildrum, K. I.; Isaksson, T.; Naes, T.; Tandberg, A. Near-infrared spectroscopy, bridging the gap between data analysis and NIR applications; Ellis Horwood: New York, 1992. (13) Ben-Dor, E.; Banin, A. Appl. Spectrosc. 1990, 44 (6), 1064-1069. (14) Ben-Dor, E.; Banin, A. Remote Sens. Environ. 1994, 48, 261274. (15) Malley, D. F.; Williams, P. C. Environ. Sci. Technol 1997, 31, 3461-3467. (16) Udelhoven, T.; Schu ¨ tt, B. Chemom. Intell. Lab. Syst. 2000, 51, 9-22. (17) Brown, S. D.; Sum, S. T.; Despagne, F. Anal. Chem. 1996, 68, 21R-61R.
(18) Naes, T.; Kvaal, K.; Isaksson, T.; Miller, C. J. Near Infrared Spectrosc. 1993, 1, 1-11. (19) Cabrera, F.; Clemente, L.; Diaz Barrientos, E.; Lo ´ pez, R.; Murillo, J. M. Sci. Total Environ. 1999, 242 (1-3), 117-130. (20) Simo´n, M.; Martin, F.; Ortiz, I.; Garcı´a, I.; Ferna´ndez, E.; Ferna´ndez, J.; Dorronsoro, C.; Aguilar, J. Sci. Total Environ. 2001, 279 (1-3), 63-74. (21) Hruschka, W. R. In Near-infrared analysis in agriculture and food industries; Williams, P. C., Norris, K., Eds.; American Association of Food Chemists Inc.: St. Paul, MN, 1987. (22) Savitzky, A.; Golay, M. J. E. Anal. Chem. 1964, 36 (8), 16271639. (23) Mackin, S.; Settle, J.; Drake, N.; Briggs, S. Proc. 3rd JPL Airborne Earth Science Workshop, 1991. http://popo.jpl.nasa.gov/docs/ workshops/91_docs/20.pdf (accessed 03/04/02). (24) Riedmiller, M.; Braun, H. ICCN’93 Proceedigs, IEEE International Conference on Neural Networks, San Francisco, CA, 1993; pp 586-591. (25) Steinnes, E. In Schwermetalle in Bo¨den; Alloway, B. J., Ed.; Springer: Berlin, Germany, 1999. (26) Alloway, B. J. Schwermetalle in Bo¨den, Analytik, Konzentrationen, Wechselwirkungen; Springer: Berlin, Germany, 1999. (27) Blume, H. P.; Bru ¨ mmer, G. Ecotoxicol. Environ. Saf. 1991, 22, 164-174. (28) Ainsworth, N.; Cooke, J. A.; Johnson, M. S. Water, Air, Soil Pollut. 1991, 57/58, 193-199. (29) Alastuey, A.; Garcı´a-Sa´nchez, A.; Lo´pez, F.; Querol, X. Sci. Total Environ. 1999, 242 (1-3), 41-55. (30) Galan, E.; Gonzalez, I.; Fernandez-Caliani, J. C. Sci. Total Environ. 2002, 286, 167-179. (31) Despagne, F.; Massart, D. L. Analyst 1998, 123, 157R-178R. (32) Hunt, G. R.; Salisbury, J. W.; Lenhoff, C. J. Mod. Geol. 1971, 2, 195-205. (33) Clark, R. N. In Manual of Remote Sensing, 3rd ed.; Rencz, A. N., Ed.; Wiley and Sons: New York, 1999; pp 3-58. (34) Hunt, G. R.; Salisbury, J. W. Mod. Geol. 1970, 1, 283-300.
Received for review October 15, 2001. Revised manuscript received March 27, 2002. Accepted April 8, 2002. ES015747J
VOL. 36, NO. 12, 2002 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2747
