Assessing soil organic matter of reclaimed soil from a large surface coal mine using a field spectroradiometer in laboratory

Geoderma 288 (2017) 47–55

Contents lists available at ScienceDirect

Geoderma journal homepage: www.elsevier.com/locate/geoderma

Assessing soil organic matter of reclaimed soil from a large surface coal mine using a field spectroradiometer in laboratory Nisha Bao a,⁎, Lixin Wu a, Baoying Ye b, Ke Yang c,d,e, Wei Zhou f a

Institute for Geo-informatics & Digital Mine Research, Northeastern University, Shenyang 110819, China Institute of Geological Survey, China University of Geosciences, Beijing 100083, China School of Earth Sciences and Resources, China University of Geosciences, Beijing 100083, China d Key Laboratory of Geochemical Cycling of Carbon and Mercury in the Earth's Critical Zone, Institute of Geophysical & Geochemical Exploration, Chinese Academy of Geological Sciences, Langfang 065000, China e Institute of Geophysical and Geochemical Exploration, Chinese Academy of Geoscience, Langfang 065000, China f School of Land Science and Technology, China University of Geosciences, Beijing 100083, China b c

a r t i c l e

i n f o

Article history: Received 8 June 2016 Received in revised form 27 October 2016 Accepted 28 October 2016 Available online 13 November 2016 Keywords: Reclamation Spectroscopy Soil organic matter Large surface mining PLS-SVM

a b s t r a c t Soil organic matter (SOM) for topsoil is one of most important indicators to support the success of mine ecological reclamation. SOM varies along the artificial mine landscape characterized by different bench-slopes of dump. Reflectance using field spectroscopy can provide useful information on the assessment of punctual soil variation, and has the advantages of speed and efficiency. The aims of this study were to 1) explore the characteristic spectrum of reclaimed soil of different landforms, 2) develop a key spectral-ratio index for evaluating SOM content, and 3) establish a SOM prediction model using the Partial Least Square Regression-Support Vector Machine (PLS-SVM) method. Based on comprehensive analysis of the relationship between SOM content and corresponding spectral reflectance in soils from different landforms, the results showed a new derived spectral index would be useful for estimating SOM. The ratio spectral index (R2294 nm / R2286 nm), calculated using available wavebands in the 350–2500 nm region, was proposed for use in the reliable estimation of SOM from downslope and midslope. The PLS-SVM calibration model for the raw spectrum, showed a high predictive accuracy for estimating the SOM content, with cross-validated R2 of 0.95, and RMSE of 0.12. These outcomes provide a theoretical basis and technical support for estimations of SOM content using visible/near-infrared spectra in reclamation areas. It is proposed that the spectral difference index and model undergo further testing and optimization prior to wider application for observation of mine-reclamation ecosystems. © 2016 Elsevier B.V. All rights reserved.

1. Introduction The mining and processing of mineral resources, particularly those extracted by surface mining, which require the complete removal of vegetation, surface soil and bedrock, and the mixtures of removed soil and rock are commonly stored in large stockpiles or dumps. The reclamation of the topsoil covering the waste rock dump or tailing dump is a primary strategy used to restore vegetation and for landform stability. Monitoring the progressive changes in the soil conditions during reclamation until the soil properties can support self-sustaining plant growth is critical for successful reclamation programmes. The processes of soil formation over landscapes, along with reclamation-induced soil changes, have created soil variation and areas of varying reclamation age within different bench-slopes of a dump. Multiple, interactive soil properties influence ⁎ Corresponding author. E-mail addresses: [email protected] (N. Bao), [email protected] (L. Wu), [email protected] (B. Ye), [email protected] (K. Yang), [email protected] (W. Zhou).

http://dx.doi.org/10.1016/j.geoderma.2016.10.033 0016-7061/© 2016 Elsevier B.V. All rights reserved.

the suitability of soil for reclamation. These are complex and vary spatially and temporally within the field of activity (Barnes et al., 2003). The traditional methods for the chemical analysis of soil properties are relatively complex, time consuming, and expensive. Reflectance can provide useful information for the assessment of punctual soil variation (Demattê and da Silva Terra, 2014) and has the advantages of speed and efficiency. Soil organic matter (SOM) is one of the main driving forces behind soil property prediction using soil spectroscopy (Vašát et al., 2014). The organic matter present in soils has distinct spectral features in the NIR region because of the relatively strong absorption overtones and combination modes of several functional groups [for example: aliphatic C\\H,\\COOH (carboxyl), \\OH (hydroxyl), and N\\H (in amines and amides)] present in organic compounds (Ben-Dor et al., 1997). Many multivariate techniques have been used to build prediction models to extract quantitative information from spectral features. Partial Least Squares regression (PLS) yields the overall best results and is the most robust with respect to predicting the soil properties of soil types (Ge et al., 2014; Nocita et al., 2011; Viscarra Rossel et al., 2006). A Support Vector Machine (SVM) employs a set of linear equations,

48

N. Bao et al. / Geoderma 288 (2017) 47–55

rather than quadratic programming problems, to obtain support vectors and has attracted attention and gained extensive application in spectral analysis. The integration of PLS-SVM results in better prediction accuracy than using only one model because PLS can remove unreliable information associated with the samples (Chauchard et al., 2004). The prediction accuracy of PLS-SVM models relies on a wide range of spectral bandwidths (Xuemei and Jianshe, 2013). Apart from multivariate techniques, spectral indices, which are calculated based on the sensitive wavebands related to the biophysical attributes, can be used to predict different soil properties. This approach has the advantage of applicability to various spectroradiometer sensors with different wavelength position, bandwidth or number of bands, and it is less complicated to apply in soil evaluation than multivariate techniques (Bartholomeus et al., 2008). Moreover, spectral indices might be applied to remote-sensing sensors to assess the land-surface parameters in arid areas on a landscape or regional scale. It is difficult to use a general model to estimate the SOM content at the scale of local mining because the national soil VIS-NIR library outperforms local-scale models (Gogé et al., 2014). The local calibration of soil spectroscopic models of field sampling sets may be more accurate than national calibration (Wetterlind and Stenberg, 2010). Currently, traditional field sampling and extraction and digestion laboratory methods are the primary techniques used to assess soil properties during mine reclamation. The mining process, particularly for large surface mines, takes a few decades, and is followed by the mine reclamation project. Soil properties are important indicators for assessing reclamation and require long-term, effective, rapid monitoring techniques (Banning et al., 2008). There is strong potential for NIR spectroscopy to be used in the assessment of mine soil quality (Pietrzykowski and Chodak, 2014). Few studies exist on the quantitative estimation of the SOM content in the reclaimed mine soil of the Loess Plain region of China, which is covered with large mining areas. The main aims of this study include: 1) to explore the characteristic spectrum of reclaimed

soil of different landforms, 2) to develop a key spectral-ratio index for evaluating SOM content, and 3) to establish a SOM prediction model using the PLS-SVM method. 2. Materials and methods 2.1. Description of mine reclamation The Pingshuo Surface Coal Mine (PSCM) is located on the Loss Plateau in Central China. As one of the largest coal mines, PSCM has been active since 1987; however, the surface coalmine has been used for N 100 years. Given the environmental rules and regulations affecting mining, a reclamation project was undertaken in addition to the mining exploration. The study area is the west dump, which covers approximately 0.4 km2 and is one of the reclaimed dumps of the PSCM (Fig. 1). The first reclamation project to create conditions for the substantial use of a dump was conducted at the west dump in 1993. Soil substrates on the reclaimed dump consist of rock and coal gangue material from the waste rock dump. This substrate has poor structure, high stone content, and low content of organic matter, nitrogen, phosphorus, and other nutrients. Therefore, over 30 cm of topsoil (natural Loess) is used as a growth medium, to cover the dump and to support revegetation. Artificial revegetation is used to accelerate the revegetation process. The west dump vegetation is dominated by perennial grasses (Medicago sativa L.) on the dump slope and other species (Hippophae rhamnoides L., Robinia pseudoacacia L., Populus simonii Carr) on the dump flat. 2.2. Sampling design and analyses Terrain factors play a role in the selection of an appropriate sample plan. Pennock (2003) developed landform segmentation procedures, including flat, upslope, mid-slope, and downslope categories, to analyse the topography-soil redistribution relationship. This landform segmentation

Fig. 1. Photograph of the mine area showing the open pits, bare soil and reclaimed dumps (left) and digital elevation model (DEM) of the studied dump (right).

N. Bao et al. / Geoderma 288 (2017) 47–55

procedure has been applied in multiple few studies, which indicated a regional-scale pattern of soil redistribution associated with landform elements (Lark, 1999; Pennock, 2003; Young and Hammer, 2000). In particular, the artificial landscape of a dump comprises flats and various side slopes. At a reclaimed dump, the soil in convex slope units dominates the spatial redistribution pattern of soil organic matter (Hansen and Loveland, 2012). The landform classification was conducted using a digital elevation model (DEM) with 1-m resolution. A computer programme based on the ARCGIS 9.2 workstation was developed for automatic landform classification. The terrain attributes were used to place each grid of the DEM into a discrete landform element class using pre-defined ranges of morphological and positional terrain attributes. The morphological attributes of the gradient and profile curvature were used to classify the cells into flat, upslope, mid-slope, and downslope elements. Furthermore, these landform elements provided references for the sampling design. The sampling plots were assigned within the landform categories using grid numbers according to a stratified, random-sample pattern. This design maintains the necessary randomness and avoids the uneven distribution of sample points among the map categories. This approach assigns a specific number of sample points to each category in proportion to the size or significance of the category and with regard to the project objectives. In the end, there were 31, 19, 20, and 23 sample plots were chosen in the flat, upslope, mid-slope, and downslope categories, respectively. Field surveys were conducted at the west dump in July 2013 and August 2015. No extreme weather occurred (e.g., storms or high winds) during the sampling period. Field sampling was conducted in 10 m × 10 m plots, which were laid out with random orientations, where the soil, landform, land cover were uniform based on visual examination upon arrival at the field (Fig. 2). Ninety-three plots were selected in total. At each plot, three 10-cm depth soil cores based on walking a “V” were collected from within a 1-m diameter area. These cores were bulked and cooled in the field and then transported and processed in the laboratory. Each plot was composited as one soil sample. Soil samples were oven-dried and sieved to a size fraction smaller than 2 mm. Organic C was analysed by dry combustion on an LECO1 CHN elemental analyser (Vitti et al., 2016). A factor of 2, as proposed by Pribyl (2010), was applied to estimate the SOM from the organic carbon. ANOVA with post hoc LSD (least-significant difference) test was applied to compare the SOM content from different landforms. 2.3. Laboratory spectral measurement The factors that would influence the accuracy of the measured spectrum include 1) cloudy weather; 2) relief of the dump; 3)

49

background objects (trees and shrubs), and 4) the soil surface. Thus, laboratory spectral measurement was selected to acquire the standard spectrum. Unaltered soil samples were spread on a tray with a sufficiently thick soil layer (3–4 cm) to avoid beam reflectance from the bottom of the tray. Downwelling solar and sky radiation penetrates approximately 1/2 wavelength into the soil, which could have the unwanted effect of modifying the soil spectra (Workman Jr., 2001). A hand-held SVC-HR 1024 was used to record spectra in the range of 350–2500 nm with a 4° field of view at approximately 50 cm above the ground. This process resulted in a ground-sampling diameter of approximately 5 cm. The spectral resolution of the samples was 1.5 nm. The incident radiation was measured by pointing the spectrometer at a white reference panel fabricated from SpectraLon. Reference spectra were the average of 10 spectra per sample. 2.4. Spectral data preprocessing Spectral preprocessing using mathematical functions are commonly used to correct for non-linearity, sample variation, and noisy spectra. These methods can accentuate the spectral features and improve the accuracy of prediction models. Measured diffuse reflectance spectra (R) are generally transformed to absorbance (logarithm of the reciprocal of reflectance, log 1/R) to linearize the relation between the spectra and concentration of a target parameter (Stenberg et al., 2010). R, log 1/R, and the first derivative of reflectance at each waveband are then computed to remove baseline effects (Ben-Dor et al., 2009). The first derivative of absorbance (R′) is computed by dividing the difference of the absorbance values between the absorbance for band number n + 1 and band number n by the difference in the wavelength for band number n + 1 and band number n, as shown in Eq. (1). 0

R ¼

ðRnþ1 −Rn Þ ðλnþ1 −λn Þ

ð1Þ

The spectral absorption index (SAI) (Jacquemoud et al., 1992) is determined based on continuum-removal spectra, which are used to describe the shape of a spectrum. The SAI can also be defined as a straight line joining two local reflectance maxima placed on the shoulders (S1 and S2) of the peak absorption wavelength (M) (Fig. 3). Thus, SAI can indicate the change from a peak to absorption reflectance, where H is the depth of the wave, S1 and S2 are the left and right peak reflectances corresponding to the absorption reflectance, and M is the absorption reflectance, as shown in Eq. (2). This approach removes the effects from non-absorptive

Fig. 2. Landform elements of the west dump and the soil sampling plots.

50

N. Bao et al. / Geoderma 288 (2017) 47–55

Fig. 3. Shape parameters of a spectrum.

materials, which contribute to the absorption features (Clark and Roush, 1984). SAI ¼

HλS1 þ ð1−HÞλS2 Mλ

ð2Þ

2.5. Prediction modelling methodology 2.5.1. Spectral index A spectral index is a non-dimensional measure of spectral reflectance that is obtained using an algebraic operation, such as a ratio, difference, or normalized difference, of two or more bands, to quantify biophysical characteristics (Jackson and Huete, 1991). Many vegetation spectral indices have been developed and established with respect to vegetation canopy and greenness (Ji et al., 2014; Kaufman and Tanre, 1992). For hyper-spectral sensors, VNIR-based indices using longer wavelengths and narrow bands are widely used (Oldeland et al., 2010; Tits et al., 2013). A spectral response for the index in this study was used to establish a correlative model with respect to the SOM. The ratio index (Eq. (3), RI), normalized difference index (Eq. (4), NDI), and difference index (Eq. (5), DI) were calculated using available wavebands in the 350– 2500 nm region. The calculated index generated band-by-band was correlated with the SOM content of the different landform soil types to select the best spectral indices to estimate the SOM content. Here, RλaandRλb are the reflectance values at all combinations of wavelengths from the 350–2500 nm region. A computer programme based on MATLAB was developed to generate the RI with the highest R2. RI ¼

Rλa Rλb

NDI ¼

Rλa −Rλb Rλa þ Rλb

the models were constructed each time by leaving some samples out of the calibration data set for use in the validation process until all samples or groups were tested. The prediction performance was assessed using the samples in the validation set. PLS is a bilinear modelling technique in which information in the original x data is projected onto a small number of underlying (“latent”) variables called PLS components. The y data are actively used to estimate the “latent” variables to ensure that the first components are those that are most relevant for predicting the y variables. In this study, Y is the actual value of SOM measured in the lab, and X is the spectra matrix. Interpretation of the relationship between X and Y data is simplified because this relationship is concentrated on the smallest possible number of components. The aim of the PLSR approach is to develop a model with the optimal number of components with the lowest root squared mean error (RSME) and highest coefficient of determination (R2). SVM is a popular data-mining method and was recently proposed for use in VIS-NIR modelling. SVM is a kernel-based learning method from statistical learning theory. It is possible to derive a linear hyperplane as a decision function for nonlinear problems. The programmes used for the PLS calculations were from the PLS package in R version 3.2.0. The SVM was operated using the “e1071 package”, an R interface to the library for support vector machines (LIBSVM). To integrate PLS-SVM, the component factors generated from the PLS model, which remove unreliable information associated with samples, were considered as the input matrix for developing the SVM model. PLS-SVM is capable of linear and nonlinear multivariate analysis and can solve these problems relatively quickly. The RMSE and adjusted R2 value of the predictions were calculated to verify the predictions of the model. The residual prediction deviation (RPD, the ratio of the standard deviation to RMSE) was used to evaluate the stability and accuracy of the multivariable models (Chang et al., 2002). The best prediction models are characterized by an RPD N 2.0, with an R2 of 0.80–1.0. The RPD value is most meaningful when the validation set is independent of the calibration set; however, with leave-one-out cross-validation, it is still a useful indicator to describe the potential of the technology. Leave-one-out cross-validation was also used for the quantification of prediction biases and errors. 3. Results 3.1. Spectral properties of reclaimed soil The most significant difference in SOM values occurred between the upslope and downslope (P b 0.05) (Table 1). The soil from the downslope had the highest average SOM value (14.16 g/kg). The reflectance spectra for all soil samples had a typical soil-curve shape: relatively low in the blue region and increasing gradually towards the NIR region. The reflectance curve was concave between 450 and 850 nm due to the presence of crystalline iron (Vitorello and Galvão, 1996), which resulted in the first peak reflectance occurring at

ð3Þ Table 1 LSD test for the SOM of different landforms.

ð4Þ

SOM (g/kg) Flat

DI ¼ Rλa −Rλb

ð5Þ Upslope

2.5.2. PLS-SVM model The prediction models in this study were based on the PLS and SVM techniques to extract quantitative information from the spectral features. Cross-validation was used to validate the quality and to prevent overfitting of the calibration model. The data were split into a training set and a test set. We randomly selected 60 samples as the training set and used the remaining 33 samples as the test set to validate the model. Proper fitting was achieved using cross-validation, in which

Downslope

Mid-slope

Mean ± SD Upslope Downslope Mid-slope Flat Downslope Mid-slope Flat Upslope Mid-slope Flat Upslope Downslope

LSD: least-significant difference; SOM: soil organic matter. ⁎ Significant at P b 0.05.

0.105 ± 0.234 −0.456 ± 0.234 −0.182 ± 0.228 −0.105 ± 0.234 −0.561 ± 0.264⁎ −0.287 ± 0.259 0.456 ± 0.234 0.561 ± 0.264⁎ 0.274 ± 0.259 0.182 ± 0.228 0.287 ± 0.259 −0.274 ± 0.259

N. Bao et al. / Geoderma 288 (2017) 47–55

51

Fig. 4. Averaged soil spectra and first derivative spectra for different landforms.

623 nm (visible light, orange to humans), corresponding to the highest value from the first derivative spectrum. The soil spectra were categorized into four groups according to landform elements: flat, upslope, mid-slope, and downslope. As shown in Fig. 4, the level of reflectance corresponded to the SOM of the soils: the higher the level of reflectance, the lower the SOM. As shown in the first derivative spectra, there were several doublet-like features that crossed the x-axis at the original maximum position, particularly near the blue (350–400 nm) and red (620–760 nm) regions. The major spectral feature between 600 and 800 nm was the most significance for the SOM values (Galvão and Vitorello, 1998). Therefore, in this region, the characteristics of the major feature (600–800 nm) were calculated to describe the different soil types (Table 2). The spectral shapes of the four soil types were similar, with a peak at 665 nm. The soil from the upslope had the lowest SOM and the highest SAI (1.5149), which was significantly different from the SAI of the soil from other landforms (Table 2).

3.2. Correlation between the SOM and spectra indices The average SOM content of natural reference soil around the mining area was 10 g/kg. The aim of reclamation is to compatible with the surrounding environmental factors, including the soil properties, vegetation cover and land use. Thus, a limit of 10 g/kg of SOM is considered an indicator of the pre-reclamation stage. Additionally, there was a significant difference (P b 0.05) between the SOM content of the upslope and downslope. Therefore, soil samples were grouped into two classes in terms of SOM content. One class consisted of downslope and mid-

slope samples with SOM content N 10 g/kg. The other class consisted of upslope and flat samples with SOM content b10 g/kg. The spectral indices (NDI, NI, and RI) were then calculated band-by-band and the Pearson correlation coefficient was calculated between the spectral indices and SOM for each sample. The correlation between the SOM content and the calculated ratios are shown in Figs. 5 and 6. The right bar shows the value of the Pearson correlation coefficient by colour, and the x- and y-axes represent the wavelengths 350–2500 nm. A dark orange colour represents a high Pearson correlation coefficient between the spectral index and SOM. There is a strong relationship between the SOM and the indices for the mid-slope and downslope soil with higher SOM compared to the upslope and flat soil. The mid-slope and downslope soil had the highest Pearson correlation coefficient of 0.78 at 2293 nm on the x-axis and 2268 nm on the y-axis, which is referred to as the RI R2293 nm / R2268 nm (Fig. 5A, Table 3). The DI (R642 nm − R631 nm), with a coefficient of 0.76, performed better than NDI {(R1892 nm −R1392 nm) / (R1892 nm + R1392 nm)}, with a Pearson correlation coefficient of 0.68 (Fig. 5B,C). However, when the SOM value was relatively low, such as for soil from the flat (9.6 g/kg) and upslope (8.55 g/kg) areas, the correlations between the spectral indices and SOM content were not strong. The Pearson correlation coefficients between the SOM of the flat and upslope areas and the three indices were b0.60 (Fig. 6, Table 3).

3.3. PLS-SVM SOM prediction model First, the input data, including the raw spectrum and transformed spectrum (derivative and LOG), were prepared for PLS to develop a

Fig. 5. Correlation coefficients between soil organic matter and the ratio index (A), normalized difference index (B), and difference index (C) for mid-slope and downslope soils.

52

N. Bao et al. / Geoderma 288 (2017) 47–55

Fig. 6. Correlation coefficients between soil organic matter and the ratio index (A), normalized difference index (B), and difference index (C) for flat and upslope soils.

multivariate calibration model (Table 4). The PLS model with the raw spectrum input data resulted in the highest prediction capability, with RPD = 2.90 for calibration (R2 = 0.90, RMSE = 0.16) (Fig. 7B). For validation (Fig. 7A), the modelling method was improved according to RPD = 2.9 (R2 = 0.65, RMSE = 0.20) (Fig. 7). Second, considering the good performance off the raw spectrum in the PLS model, 12 component factors generated from the PLS model were considered as the input matrix to develop the SVM model (Table 4). As shown in Fig. 8, the selected 12 components from the entire VIS-NIR spectrum (350– 2500 nm) produced a more accurate prediction, with R2 = 0.95 and RMSE = 0.12 for calibration (Fig. 8B) and R2 = 0.73 and RMSE = 0.16 for validation (Fig. 8A). 4. Discussion 4.1. Reclamation of SOM and spectral properties Hillslope processes affect the distribution of soil organic C, clay and nutrients, which may lead to the various soil attributes at different landform elements (Young and Hammer, 2000). In this particular case, for reclaimed dumps at large surface mines, the dominant features of the artificial landscape can be described as flat, upslope, mid-slope, and downslope. The formation of stable soil organic matter fractions and their downhill transport and deposition in hollow areas affect the

organic matter content, which results in a higher SOM content at the downslope. During reclamation, the flat areas are compressed by large mining equipment, which results in over-compaction of the soil. This over-compaction might reduce the soil volume available for exploitation by young plant roots (Braunack et al., 2006). The relative lack of vegetation would contribute to the relatively low SOM in flat areas. Although the top soil at the dump was reclaimed similarly throughout, after several years, the reclaimed soil of different landforms was spatially smoothed and transformed from simplex space to real space, which suggests that the reclaimed soil responds to environmental factors similarly to natural soil (Ludwig et al., 2003). Thus, the spectral characteristics would be efficient and useful indictors for understanding the progressive changes in the surface soil of various landforms. Our results indicated that the characteristic spectra of soil samples (via spectroscopy) could be used to discriminate soil with different levels of SOM in various landform positions. The SOM concentration was indicated by differences in the absorption area at a peak of 665 nm. The soils varied in absorbance in the visible light region, which was mainly associated with iron-containing minerals and the presence of SOM (St. Luce et al., 2014). Ertlen et al. (2015) found that the NIR spectra of SOM from deep horizons make it possible to recognize the vegetation conditions using these profiles. This study revealed that the characteristic spectra for SOM provide the ability to understand the reconstruction of past soil environments under different landform conditions.

Fig. 7. (A) Calibration between the observed and predicted values of SOM based on the raw spectrum using the PLS model with the calibration set of 60 samples. (B) Validation between the observed and predicted values of SOM based on the raw spectrum using the PLS model with 33 test set samples. SOM: soil organic matter.

N. Bao et al. / Geoderma 288 (2017) 47–55

53

Fig. 8. (A) Calibration between the observed and predicted values of SOM based on the raw spectrum using the PLS-SVM model with the 60 calibration set samples. (B) Validation between the observed and predicted values of SOM based on the raw spectrum using the PLS–SVM model with the 33 test set samples. SOM: soil organic matter.

4.2. Effective VIS-NIS spectroscopy for predicting SOM Multiple spectral indices have been widely proposed to estimate soil nutrient parameters, such as the spectral soil quality index (SSQI) proposed by Paz-Kagan et al. (2014). This method was based on reflectance spectroscopy for characterizing soil function in areas of changed land use. In this study, a spectral index was developed by classifying the soil samples into two types, according to the SOM concentration. The accuracy of the predictions by the spectral index was influenced by the sample-set homogeneity (Brunet et al., 2007). The downslope and mid-slope soil-spectrum index had stronger correlation with SOM. When the SOM value was relatively low, such as for soil from flat (9.6 g/kg) and upslope (8.55 g/kg) areas, there were not very strong correlations between the spectral difference ratio index (RI: R862 nm / R858 nm) and the SOM content. For soil samples from a litchi orchard in South China, DI for 1340 nm–1380 nm produced highly significant correlation coefficients (Li et al., 2012). The sensitive or optimal spectral index response to SOM was uncertain in relation to different soil types and sample sizes. Thus, for reclaimed soil at the local scale, the difference index (RI: R2293 nm / R2286 nm) is recommended for reliable estimation of the SOM content for soil from the downslope and mid-slope. Compared with the PLS modelling method, the use of PLS-SVM is advantageous for spectra because it can perform nonlinear regression efficiently for high-dimensional data sets. According to the optimum scores

from PLS as the input for the SVM model, very reliable cross-validation and external validation of SOM were achieved by establishing the PLSSVM model directly upon the characteristic spectral absorption. The PLS-SVM model performed slightly better than the PLS model. One possible reason was that the PLS-SVM model considered the useful nonlinear information of the spectral data, while PLS only used the linear relationships between the spectra data and SOM (Rossel and Behrens, 2010). Our study provided the optimal spectral bands (R2293 nm / R2268 nm) for predicting the SOM content of reclaimed mine soil of the Loess Plantae mining area. Future research will focus on repeatable observations within predictable levels of error to determine the optimal geometric configuration of the source and error and the optimal time of the year for SOM measurement. In this study, satisfactory SOM content estimations were achieved based on laboratory soil spectral measurements using PLS-SVM and spectral indices. The PLS-SVM method considered many SOM contentrelated bands and exhibited enhanced sensitivity to changes in SOM content. The RMSE of the regression between the measured and predicted values was 0.16 g/kg. The ability of multivariate techniques to associate complicated spectral information with target attributes without sample distribution constraints makes them ideal to describe the intricate and complex nonlinear relationships between spectral signatures and various environmental conditions (Ramirez-Lopez et al., 2013). Spectral indices have merits, especially with respect to the simplicity

Table 2 Characteristic features of the spectrum at 600–800 nm for different soil types. Landform

Peak

Width

Depth

Symmetry

SAI (spectral absorption index)

Flat Upslope Mid-slope Downslope

665 nm 665 nm 665 nm 665 nm

11.10 15.20 8.30 11.10

0.2226 0.4978 0.1408 0.2418

0.3986 0.0611 0.1920 0.1297

1.3517 1.5149 1.3453 1.3337

Table 3 Spectral indices and correlation coefficients for different soil types (RI: ratio index; DI: difference index; NDI: normalized difference index). Soil Flat upslope Down mid-slope

Wavelength Coefficient of correlation Wavelength Coefficient of correlation

RI

NDI

DI

862/858 nm 0.62 2293/2286 nm 0.78

967/964 nm 0.60 1892/1392 nm 0.68

862/858 nm 0.61 642/631 nm 0.76

54

N. Bao et al. / Geoderma 288 (2017) 47–55

Table 4 Input data and output results of the PLS model. Input data

Raw Derivative Derivative 1/LOG LOG

Number of components

External validation

Cross-validation

R2

R2

RSME RPD

RMSE RPD

12 6 10

0.65 0.20 0.15 0.66 0.35 0.46

2.90 0.90 0.16 0.87 0.75 0.27 1.26 0.78 0.25

3.60 2.10 2.32

10

0.51 0.24

2.41 0.87 0.20

2.90

of application. Spectral indices based on spectral differences have the advantage of rapid tracing of soil quality and its changes after longterm mining land reclamation management (Paz-Kagan et al., 2014). Furthermore, the methods are feasible from the laboratory and field scale to the scale of remote-sensing sensors. 5. Conclusions VIS-NIR spectroscopy analysis facilitated the reliable quantitative assessment of soil organic matter in the reclamation soil of a mining area. The advantages of using spectroscopy include: 1) the visible spectrum at approximately 600 nm can be used to discriminate the different soil types found in various landforms; 2) the ratio spectral index (R2294 nm / R2286 nm) calculated using available wavebands in the 350– 2500 nm region was proposed for use in the reliable estimation of SOM from the downslope and mid-slope; 3) the PLS-SVM model based on the raw spectrum produced higher accuracy than the PLS model for predicting the SOM concentration. The results demonstrated that spectroscopy is a reliable, reproducible, rapid, and low-cost technique for the in situ diagnosis of soil properties to support mining reclamation. In addition, this study opens the way for the assessment and monitoring of the reclaimed soil of mining areas at high resolution, without limitations related to the cost and time required to process large numbers of physical samples. However, a large number of samples from various depths should be collected in a practical and timely manner, and the analysis should be repeated. This process should be conducted to determine the optimal sampling design and algorithms for data mining, both as a preliminary to future field surveys and to make the combination of spectroscopic analysis and PLS-SVM more robust and reliable for monitoring and modelling mining environments. Acknowledgments We acknowledge the National Natural Science Foundation of China for young researchers (41401233), the Fundamental Research Funds for the Central Universities (N120801001), and the Geological Survey and Mineral Resources Assessment Project (No. 12120113002600). We appreciate the suggestions from the anonymous two reviewers. References Banning, N.C., Grant, C.D., Jones, D.L., Murphy, D.V., 2008. Recovery of soil organic matter, organic matter turnover and nitrogen cycling in a post-mining forest rehabilitation chronosequence. Soil Biol. Biochem. 40 (8), 2021–2031. Barnes, M.E., Sudduth, A.K., Hummel, W.J., 2003. Remote- and Ground-Based Sensor Techniques to Map Soil Properties. American Society for Photogrammetry and Remote Sensing. Bethesda, MD, ETATS-UNIS . Bartholomeus, H.M., Schaepman, M.E., Kooistra, L., Stevens, A., Hoogmoed, W.B., Spaargaren, O.S.P., 2008. Spectral reflectance based indices for soil organic carbon quantification. Geoderma 145 (1–2), 28–36. Ben-Dor, E., Inbar, Y., Chen, Y., 1997. The reflectance spectra of organic matter in the visible near-infrared and short wave infrared region (400–2500 nm) during a controlled decomposition process. Remote Sens. Environ. 61 (1), 1–15. Ben-Dor, E., Chabrillat, S., Demattê, J.A.M., Taylor, G.R., Hill, J., Whiting, M.L., Sommer, S., 2009. Using imaging spectroscopy to study soil properties. Remote sensing of environment 113. 0 (Supplement 1), S38–S55.

Braunack, M.V., Arvidsson, J., Håkansson, I., 2006. Effect of harvest traffic position on soil conditions and sugarcane (Saccharum officinarum) response to environmental conditions in Queensland, Australia. Soil Tillage Res. 89 (1), 103–121. Brunet, D., Barthès, B.G., Chotte, J.-L., Feller, C., 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), 106–117. Chang, C.-I., Chiang, S.-s., A. Smith, J., 2002. Linear spectral random mixture analysis for hyperspectral imagery. IEEE Trans. Geosci. Remote Sens. 40 (2), 375–391. Chauchard, F., Cogdill, R., Roussel, S., Roger, J.M., Bellon-Maurel, V., 2004. Application of LS-SVM to non-linear phenomena in NIR spectroscopy: development of a robust and portable sensor for acidity prediction in grapes. Chemom. Intell. Lab. Syst. 71 (2), 141–150. Clark, R.N., Roush, T.L., 1984. Reflectance spectroscopy—quantitative analysis techniques for remote sensing applications. J. Geophys. 89, 6329–6340. Demattê, J.A.M., da Silva Terra, F., 2014. Spectral pedology: a new perspective on evaluation of soils along pedogenetic alterations. Geoderma 217–218 (0), 190–200. Ertlen, D., Schwartz, D., Brunet, D., Trendel, J.-M., Adam, P., Schaeffer, P., 2015. Qualitative near infrared spectroscopy, a new tool to recognize past vegetation signature in soil organic matter. Soil Biol. Biochem. 82 (0), 127–134. Galvão, L.S., Vitorello, Í., 1998. Variability of laboratory measured soil lines of soils from southeastern Brazil. Remote Sens. Environ. 63 (2), 166–181. Ge, Y., Morgan, C.L.S., Ackerson, J.P., 2014. VisNIR spectra of dried ground soils predict properties of soils scanned moist and intact. Geoderma 221–222 (0), 61–69. Gogé, F., Gomez, C., Jolivet, C., Joffre, R., 2014. Which strategy is best to predict soil properties of a local site from a national Vis–NIR database? Geoderma 213, 1–9. Hansen, M.C., Loveland, T.R., 2012. A review of large area monitoring of land cover change using Landsat data. Remote Sens. Environ. 122, 66–74. Jackson, R.D., Huete, A.R., 1991. Interpreting vegetation indices. Prev. Vet. Med. 11, 185–200. Jacquemoud, S., Baret, F., Hanocq, J.F., 1992. Modeling spectral and bidirectional soil reflectance. Remote Sens. Environ. 41, 123–132. Ji, L., Zhang, L., Rover, J., Wylie, B.K., Chen, X., 2014. Geostatistical estimation of signal-tonoise ratios for spectral vegetation indices. ISPRS J. Photogramm. Remote Sens. 96, 20–27. Kaufman, Y.J., Tanre, D., 1992. Atmospherically resistant vegetation index (ARVI) for EOSMODIS. IEEE Trans. Geosci. Remote Sens. 30 (2), 261–270. Lark, R.M., 1999. Soil–landform relationships at within-field scales: an investigation using continuous classification. Geoderma 92 (3–4), 141–165. Li, D., Chen, X., Peng, Z., Chen, S., Chen, W., Han, L., Li, Y., 2012. Prediction of soil organic matter content in a litchi orchard of South China using spectral indices. Soil Tillage Res. 123 (0), 78–86. Ludwig, J.A., Hindley, N., Barnett, G., 2003. Indicators for monitoring minesite rehabilitation: trends on waste-rock dumps, northern Australia. Ecol. Indic. 3 (3), 143–153. Nocita, M., Kooistra, L., Bachmann, M., Müller, A., Powell, M., Weel, S., 2011. Predictions of soil surface and topsoil organic carbon content through the use of laboratory and field spectroscopy in the Albany Thicket Biome of Eastern Cape Province of South Africa. Geoderma 167–168, 295–302. Oldeland, J., Dorigo, W., Lieckfeld, L., Lucieer, A., Jurgens, N., 2010. Combining vegetation indices, constrained ordination and fuzzy classification for mapping semi-natural vegetation units from hyperspectral imagery. Remote Sens. Environ. 114 (6), 1155–1166. Paz-Kagan, T., Shachak, M., Zaady, E., Karnieli, A., 2014. A spectral soil quality index (SSQI) for characterizing soil function in areas of changed land use. Geoderma 230–231 (0), 171–184. Pennock, D.J., 2003. Terrain attributes, landform segmentation, and soil redistribution. Soil Tillage Res. 69 (1–2), 15–26. Pietrzykowski, M., Chodak, M., 2014. Near infrared spectroscopy—a tool for chemical properties and organic matter assessment of afforested mine soils. Ecol. Eng. 62, 115–122. Pribyl, D.W., 2010. A critical review of the conventional SOC to SOM conversion factor. Geoderma 156 (3–4), 75–83. Ramirez-Lopez, L., Behrens, T., Schmidt, K., Stevens, A., Demattê, J.A.M., Scholten, T., 2013. The spectrum-based learner: A new local approach for modeling soil vis–NIR spectra of complex datasets. Geoderma 195–196 (0), 268–279. Rossel, R.A.V., Behrens, T., 2010. Using data mining to model and interpret soil diffuse reflectance spectra. Geoderma 158 (1–2), 46–54. St. Luce, M., Ziadi, N., Zebarth, B.J., Grant, C.A., Tremblay, G.F., Gregorich, E.G., 2014. Rapid determination of soil organic matter quality indicators using visible near infrared reflectance spectroscopy. Geoderma 232–234 (0), 449–458. Stenberg, B., Viscarra Rossel, R.A., Mouazen, A.M., Wetterlind, J., 2010. Chapter five - visible and near infrared spectroscopy in soil science. In: Donald, L.S. (Ed.), Advances in Agronomy. Academic Press, pp. 163–215. Tits, L., Somers, B., Stuckens, J., Farifteh, J., Coppin, P., 2013. Integration of in situ measured soil status and remotely sensed hyperspectral data to improve plant production system monitoring: concept, perspectives and limitations. Remote Sens. Environ. 128 (0), 197–211. Vašát, R., Kodešová, R., Borůvka, L., Klement, A., Jakšík, O., Gholizadeh, A., 2014. Consideration of peak parameters derived from continuum-removed spectra to predict extractable nutrients in soils with visible and near-infrared diffuse reflectance spectroscopy (VNIR-DRS). Geoderma 232–234 (0), 208–218. Viscarra Rossel, R.A., Walvoort, D.J.J., McBratney, A.B., Janik, L.J., Skjemstad, J.O., 2006. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131 (1–2), 59–75.

N. Bao et al. / Geoderma 288 (2017) 47–55 Vitorello, Í., Galvão, L.S., 1996. Spectral Properties of Geologic Materials in the 400 to 2500 nm Range: Review for Applications to Mineral Exploration and Lithologic Mapping Photo Interpret 34. pp. 77–99. Vitti, C., Stellacci, A.M., Leogrande, R., Mastrangelo, M., Cazzato, E., Ventrella, D., 2016. Assessment of organic carbon in soils: a comparison between the Springer–Klee wet digestion and the dry combustion methods in Mediterranean soils (Southern Italy). Catena 137, 113–119. Wetterlind, J., Stenberg, B., 2010. Near-infrared spectroscopy for within-field soil characterization: small local calibrations compared with national libraries spiked with local samples. Eur. J. Soil Sci. 61 (6), 823–843.

55

Workman Jr., J., 2001. 12 - review of near-infrared and infrared spectroscopy. In: Workman, J. (Ed.), The Handbook of Organic Compounds. Academic Press, Burlington, pp. 79–129. Xuemei, L., Jianshe, L., 2013. Measurement of soil properties using visible and short wavenear infrared spectroscopy and multivariate calibration. Measurement 46 (10), 3808–3814. Young, F.J., Hammer, R.D., 2000. Soil-landform relationships on a loess-mantled upland landscape in Missouri. Soil Sci. Soc. Am. J. 64 (4), 1443–1455.