Environ. Sci. Technol. 2005, 39, 873-878
Feasibility of Reflectance Spectroscopy for the Assessment of Soil Mercury Contamination Y U N Z H A O W U , * ,†,‡ J U N C H E N , † J U N F E N G J I , * ,† Q I N G J I U T I A N , ‡ A N D XIN MIN WU§ State Key Laboratory of Mineral Deposit Research, Institute of Surficial Geochemistry, Department of Earth Sciences, Nanjing University, Nanjing 210093, China, International Institute for Earth System Science, Nanjing University, Nanjing 210093, China, and Geological Survey of Jiangsu Province, Nanjing 210093, China
Conventional methods for investigating soil Hg contamination based on raster sampling and chemical analysis are timeconsuming and relatively expensive. The objective of this study was to develop a rapid method for investigating Hg concentration in suburban agricultural soils of the Nanjing region using reflectance spectra within the visible-nearinfrared (VNIR) region. Several spectral pretreatments (absorbance, Kubelka-Munk transformations and their derivatives) were applied to the reflectance spectra to optimize the accuracy of prediction. The prediction of Hg concentration was achieved by univariate regression and principal component regression (PCR) approaches. The optimal model (R ) 0.69, RMSEP ) 0.15) for predicting Hg was achieved using the PCR method with the KubelkaMunk transformation as the spectral predictor. Comparison of three wavelength ranges (0.38-1.1, 1.0-2.5, and 0.382.5 µm) on the effect of prediction accuracy showed that the best results were acquired using the 1.0-2.5 µm spectral region. Correlation analysis revealed that Hg concentration was negatively correlated with soil reflectance while positively correlated with the absorption depths of goethite at 0.496 µm and clay minerals at 2.21 µm, suggesting that Hg-sorption by clay-size mineral assemblages in soils was the mechanism by which to predict spectrally featureless Hg. These results indicate that it is feasible to predict Hg levels in agricultural soils using the rapid and cost-effective reflectance spectroscopy. Future study with operational remote sensing techniques and field measurements is strongly recommended.
Introduction Soil is an important component of natural and modified ecosystems, which provide habitats in which terrestrial organisms can thrive. Because of rapid economic development, heavy metal contamination of agricultural soils is be* Authors to whom correspondence should be addressed. Phone: +86-25-83597351 (Y.W.); +86-25-83595795 (J.J.). E-mail: njuessi@ hotmail.com (Y.W.);
[email protected] (J.J.). † Institute of Surficial Geochemistry, Nanjing University. ‡ International Institute for Earth System Science, Nanjing University. § Geological Survey of Jiangsu Province. 10.1021/es0492642 CCC: $30.25 Published on Web 12/30/2004
2005 American Chemical Society
coming increasingly serious in China. Nanjing, the capital of Jiangsu province, is located in the Yangtse Delta long renowned as the “Hometown of fish and rice”. The agricultural activities of the Nanjing area are relatively well developed. During the past 20 years, chemical fertilizers, pesticides, and other substances containing Hg have been continuously applied to the suburban agricultural soils. Coupled with the lack of pollution controls, human activities associated with these developments have caused significant Hg increases (1). Hg contamination of agricultural soils can have longterm environmental implications and pose a threat to longterm sustainable development in this area. Conventional methods of determining soil Hg contamination based exclusively on raster sampling and wet chemical analysis are time-consuming, dangerous, and relatively expensive. Reflectance spectroscopy within the visible-near-infrared (VNIR) region (0.38-2.5 µm) has been widely used to predict spectrally active constituents in soils because of its rapidity, convenience, and accuracy. These include iron oxides (2, 3), carbonates (4), and organic matter (5-7). Moreover, some studies have shown that the intercorrelation between spectrally featureless constituents, which are usually identified by properties such as magnetic susceptibility (8), soil-specific surface area (SSA), and cation-exchange capacity (9), and constituents that are spectrally active, enables even featureless soil constituents to be predicted by reflectance spectra. As trace amounts in soils, heavy metals have no spectral features. The increasing heavy metals from anthropogenic inputs are adsorbed by organic matter, iron oxides, carbonates, and clay minerals (10-12). These minerals are spectrally active within the VNIR region (13, 14); some have specific absorption features (iron oxides, carbonates, and clay minerals), and some absorb light within this wavelength region (organic matter). Therefore, although Hg and other heavy metals at low concentration levels do not have spectral features, predictions for the heavy metal concentrations can be made indirectly. For example, Malley and Williams (15) succeeded in predicting heavy metal concentrations in freshwater sediment using reflectance spectroscopy within the 1.1-2.5 µm region. Wavelengths from partial least-squares regression (PLSR) analysis associated with most of the variance in heavy metals were attributed to organic matter. With reflectance spectroscopy, the Zn and Cd concentrations in flood plains along the river Rhine were predicted on the basis of the correlation between the heavy metals and the organic matter contents (16). Likewise, Kemper and Sommer (17) predicted As, Hg, and Pb concentrations using reflectance spectroscopy in the Aznalcollar Mine area and found that the most important wavelengths for prediction were attributed to absorption features of iron and iron oxides. Until now, examples of the use of reflectance spectroscopy in the prediction of heavy metals in contaminated soils are available only in abandoned mines (17) and in sediments (15, 16). To our knowledge, research on investigating Hg and other heavy metals contamination for agricultural soils has not been reported. The objectives of this study were (i) to explore the feasibility of reflectance spectra for the rapid prediction of Hg concentration in soils of the suburban agricultural fields, and (ii) to investigate the mechanism by which to predict spectrally featureless Hg.
Materials and Methods Soil Sampling. The research area, Jiangning District, is located at the southern edge of Nanjing city. The main soil type is yellow brown soil, with Xiashu loess as its natural parent material. A total of 120 topsoil samples (depth < 10 cm) were VOL. 39, NO. 3, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
873
collected using a systematic sampling scheme. The grid of sampling sites was 2 × 2 km (the territory is about 480 square km). At each sampling site, three subsamples were taken and then mixed to obtain a bulk sample so that a representative estimate of the concentration at that point could be obtained. The collected soil samples were air-dried at 20 °C for 3 days and sieved through a 2 mm polyethylene sieve to remove large debris, stones, and pebbles. They were then ground in a mechanical agate grinder until fine particles were obtained. Each sample was split into two subsamples. One was used for spectral measurements, and the other was analyzed for total Hg concentration. Chemical Analyses. The analytical method for the determination of the total Hg concentration was modified from Rodrigues Filho and Maddock (18). Deionized water and aqua regia were added to 2 g of milled sample, which was dried at 40 °C, and heated to 60 °C for 30 min. After cooling, deionized water and 5% potassium permanganate were added and heated to 60 °C for 30 min. The excess KMnO4 was neutralized with hydroxylamine hydrochlorate 12%. Total Hg was measured using a cold-vapor atomic fluorescence spectrometry (CV-AFS). Data were assessed for accuracy and precision using a quality control program which included reagent blanks, duplicate samples, and certified house reference materials. Spectral Measurements and Transformations. Reflectance spectra of the soil samples were recorded in a PerkinElmer Lambda 900 spectrophotometer. Spectral bandwidth setting of the “Lambda 900” in 0.01 nm increments is possible from 5 nm to a maximum resolution of 0.05 nm in the 0.190.9 µm region and from 20 to 0.2 nm in the 0.9-2.5 µm region. For the current measurements, we used a constant spectral resolution value of 4 nm, at 2 nm increments, which produced 1061 spectral points between 0.38 and 2.5 µm. The sample preparation followed the procedure from Balsam and Deaton (19) and Ji et al. (8). Ground samples were made into a slurry on a glass microslide with distilled water, smoothed, and dried slowly at low temperature (6, respectively. Based on this criterion, 33% of the samples are low contamination, 4% are moderate contamination, 3% are high contamination, and the remaining 60% are unpolluted. Soil Spectra. The reflectance spectrum of a representative soil sample (JNX-1) in the research area is shown in Figure 1. The presence of iron in ferric or ferrous forms results in absorption features at wavelengths in the 0.4-1.3 µm region and rapid falloff of reflectance toward the blue wavelengths, while the absorption features are broad and weak (14). Through continuum removal, the absorption features are enhanced. There are four regions of the spectrum that exhibit distinct absorption features. A doublet absorption feature near 0.496 µm is caused by goethite. As compared to the continuum-removed spectrum of pure goethite, the soil absorption near 0.496 µm is weaker and broader because of the effect of the soil matrix. The two distinct absorption bands at 1.41 and 1.91 µm are attributable to vibrational frequencies of OH groups in the adsorbed water, and the absorption feature around 2.21 µm is related to OH groups in the crystal lattice water. For pure goethite, the two absorption bands, at 0.67 and 0.95 µm, are also characteristics and even stronger than the 0.496 µm band. However, for soil spectra, the two bands cannot be displayed due to the soil matrix effect, and even on the continuum removal curve they still cannot be assigned exactly (Figure 1a). This observation supports the work of
FIGURE 2. Scatterplot of reflectance at 1.144 µm against Hg concentration of 120 soil samples. It can be seen that Hg is significantly negatively correlated with the reflectance. However, the samples with very low Hg concentrations have large deviations from the exponential regression trendline. Balsam and Deaton (19) that for concentrations of goethite less than 2% the NIR absorption band cannot be detected while the absorption band in the VIS is detectable. Therefore, it can be concluded that the diagnostic spectral feature for identifying goethite in soils is the shorter wavelength and not the longer wavelength; that is, the 0.496 µm wavelength is better. This view was strongly supported by the work of Balsam and Wolhart (22) who reported that goethite has 0.535 and 0.435 µm derivative peaks in sediments and that the 0.435 µm derivative peak is a better indicator. Noting that peaks on first derivative curves occur on the shoulders of absorption bands, the 0.435 µm derivative peak corresponds to the 0.496 µm absorption on the reflectance spectra. Prediction of Hg Using Univariate Regression and PCR. The coefficients of linear correlation between reflectance data and Hg concentration of the 120 samples were calculated. Results show that soil reflectance is negatively correlated with Hg concentration, and the largest correlation coefficients are obtained around 1.144 µm. Figure 2 exhibits the scatter plot for the relationship between reflectance at 1.144 µm and Hg concentration. Hg concentration is significantly negatively correlated with the reflectance (r ) -0.64, p < 0.01). However, it can be seen that the samples with very low Hg concentrations have large deviations in the reflectance. These samples can decrease the prediction accuracy. To improve the prediction accuracy, 15 samples VOL. 39, NO. 3, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
875
TABLE 1. Regression Statistics and Parameter Estimates for the Univariate Method calibration (n ) 53) model types exponential a
preprocessing R
validation (n ) 52)
RMSEC
R
constant
weighting coeff
RMSEP
R
0.158
0.731a
21.125
-0.131
0.152
0.61
Correlation is significant at the 0.01 level.
TABLE 2. Regression Statistics and Parameter Estimates for the PCR Method (n ) 105) calibration
validation
preprocessing
factors
RMSEC
R
RMSEP
R
wavelength (µm)
reflectance absorbance K-M
4 4 4
0.135 0.13 0.125
0.759 0.776 0.795
0.153 0.153 0.151
0.679 0.675 0.691
1.0-2.5 1.0-2.5 1.0-2.5
reflectance absorbance K-M
1 2 2
0.147 0.134 0.11
0.153 0.156 0.162
0.672 0.66 0.625
0.38-1.0 0.38-1.0 0.38-2.5
First Derivative 0.7 0.763 0.84
with Hg concentrations lower than 0.05 ppm (which is far lower than background value) were left out (the R value is improved to -0.67). The remaining dataset of 105 samples was used for the univariate and PCR prediction. The visible and near-infrared analysis (VNIRA) techniques require the selection of a representative calibration set, which is chemically matched with its validation set and with the population of a large number of samples (23). Prior to univariate calibration, the samples were sorted from lowest to highest Hg concentration 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 53 samples, and the validation set consists of 52. A summary of the optimal univariate model and statistical parameters from the calibration and test set is given in Table 1. The best prediction equation is inverse exponential using the 1.144 µm band with no spectral preprocessing. It is interesting to note that the results, indicating an inverse exponential model as optimal for predicting Hg, agree well with Galvao’s findings (24). In that study, the relationship between Fe2O3 in Brazilian tropical soils and reflectance acquired from AVIRIS data was also described by an inverse exponential model. The exponential model reflects the physical processes underlying the relationship, that is, Beer’s law for absorption. Hg at very low level has no effect on the photon absorption; however, the coincidence between our results and Galvao’s findings seems to suggest some positive correlation between Hg and iron oxides. The statistical parameters of the final models using the PCR method for the different preprocessing techniques are summarized in Table 2. Because the results of the second derivative spectral data are all unsatisfactory, they are not shown in the table. For each of the spectral responses (R, A, K-M, and their first derivatives), only the wavelength range which gives the minimum RMSEP is shown. For example, the PCR result for the first derivative spectra data of K-M has a very high R value (0.98) and low RMSEC (0.04) for the spectral range of 1.0-2.5 µm. However, for the validation, R is low (0.6) and RMSEP is high (0.17), which is 4 times higher than RMSEC, and so this is not thought to be a good model and hence is not listed in the table. The RMSEP is known as a good criterion to judge the prediction performance (20). Here, the smallest RMSEP value was used to determine the optimum calibration model. In 876
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 39, NO. 3, 2005
that case, the best PCR model for the prediction of Hg is the K-M transformation with four factors using the 1.0-2.5 µm spectral region. Although the absorption features of iron oxides can only be detectable in the VIS region, soil reflectance values beyond 1.0 µm still change with the varying contents of iron oxides. Indeed, the reflectance variation of the 1.02.5 µm region is even larger than the variation of the visible region. What is more, the reflectance decreases of the 1.02.5 µm region as the organic matter increases, another heavy metal absorbant, are also stronger than in the VIS region. The combination effects of iron oxides and organic matter explain the fact that the most important spectral regions selected in the PCR equations were generally within 1.0-2.5 µm. The derivative transformations can reduce spectral interference and remove additive baseline effects; however, the results of the derivative spectra for the validation step are worse than those without derivative transformations. One of the reasons is that the soil samples used in the present study were ground, and hence the effect of particle size is small. The most important reason may be related to the fact that the best wavelength range for prediction is within 1.02.5 µm, where the spectral curves of soils are flat and hence the derivatives are small. The relationship between the predicted and measured Hg values is shown in Figure 3. To study the capability of rapidly classifying soil Hg contamination levels by reflectance spectra, the boundaries of the index of soil Hg pollution are also shown. When the samples of the calibration and test set fall within the boundaries of a certain class depicted by the box on the diagonal, then they are classified correctly. The percentages of correct classification are 60.4% for clean, 64.4% for low contamination, 28.6% for moderate contamination, and 40% for high contamination. Generally speaking, the classification accuracy for the cleanness and low contamination samples is satisfactory; however, for the more highly contaminated classes, it is relatively low. In fact, the required prediction accuracy for metal contamination will depend on the objective of a specific application (16). Besides the determination of the total Hg concentration, investigations on Hg-binding forms are essential for evaluating environmental risks and providing necessary information for drafting/developing soil remediation guidelines. It should be noted that in this study Hg was researched in the form of total concentration, while the chemical forms of Hg were not considered. Besides those adsorbed by iron oxides, clay minerals, and organic matter, there are other chemical
FIGURE 3. Plot of measured against predicted Hg concentration based on the PCR method. The boundaries of Hg pollution classes are also shown.
FIGURE 4. Absorbance depths around 0.496 µm of soils with varying percentages of goethite. It shows that the absorption depth around 0.496 µm increases as goethite increases.
fractions, such as exchangeable, bound to carbonates, and residual phase Hg. The residual phase Hg, which is unavailable for vegetation, is in the crystal lattice of the minerals. The carbonates have spectral features within the VNIR region and can be predicted by reflectance spectra. These investigations can be expected to improve prediction accuracy by studying Hg in other possible chemical fractions using a sequential extraction procedure.
concentration has a higher positive correlation with absorption depth at 0.496 µm than with absorption depth at 2.21 µm. On the contrary, the reflectance values of soils decrease because of absorption of the goethite, clay minerals, and organic matter. Therefore, soil reflectance is negatively correlated with Hg concentration. The increasing Hg in soils resulting from anthropogenic input is absorbed by iron oxides, clay minerals, and organic matter. Hence, the increased Hg concentration is intercorrelated with these matters. Via the intercorrelation with these spectrally active soil constituents, the toxic Hg, although it has no spectral features at low concentration levels, can be predicted. However, it is notable that the levels of Hg in unpolluted soils are naturally accumulated from rocks through pedogenesis. It is not adsorbed by clay-size mineral assemblages but exists in the crystal lattice of the minerals. Therefore, the correlation between the low levels of Hg and soil clay-size mineral assemblages is weak. This explains why the prediction accuracy can be improved by leaving out very low concentration Hg samples (
