Estimating Soil Organic Carbon Content with Visible–NearInfrared (Vis-NIR) Spectroscopy Yin Gao,a,b Lijuan Cui,c Bing Lei,d Yanfang Zhai,e Tiezhu Shi,b Junjie Wang,b Yiyun Chen,b Hui He,f Guofeng Wug,* a

National Geomatics Center of China, 100830, Beijing, China School of Resource and Environmental Science and Key Laboratory of Geographic Information System of the Ministry of Education, Wuhan University, 430079, Wuhan, China c Institute of Wetland Research, Chinese Academy of Forestry, 100091, Beijing, China d Satellite Surveying and Mapping Application Center, National Administration of Surveying, Mapping and Geoinformation of China, 100830, Beijing, China e Chongqing Institutes of Surveying and Mapping, 400014, Chongqing, China f Star Map Press, 100088, Beijing, China g Key Laboratory for Geo-Environment Monitoring of Coastal Zone of the National Administration of Surveying, Mapping and GeoInformation and Shenzhen Key Laboratory of Spatial Smart Sensing and Services and College of Life Sciences, Shenzhen University, 518060, Shenzhen, China b

The selection of a calibration method is one of the main factors influencing measurement accuracy with visible-near-infrared (VisNIR, 350–2500 nm) spectroscopy. This study, based on both airdried unground (DU) and air-dried ground (DG) soil samples, used nine spectral preprocessing methods and their combinations, with the aim to compare the commonly used partial least squares regression (PLSR) method with the new machine learning method of support vector machine regression (SVMR) to find a robust method for soil organic carbon (SOC) content estimation, and to further explore an effective Vis-NIR spectral preprocessing strategy. In total, 100 heterogeneous soil samples collected from Southeast China were used as the dataset for the model calibration and independent validation. The determination coefficient (R2), root mean square error (RMSE), residual prediction deviation (RPD), and ratio of performance to interquartile range were used for the model evaluation. The results of this study show that both the PLSR and SVMR models were significantly improved by the absorbance transformation (LOG), standard normal variate with wavelet detrending (SW), first derivative (FD), and mean centering (MC) spectral preprocessing methods and their combinations. SVMR obtained optimal models for both the DU and DG soil, with R2, RMSE, and RPD values of 0.72, 2.48 g/kg, and 1.83 for DU soil and 0.86, 1.84 g/kg, and 2.60 for DG soil, respectively. Among all the PLSR and SVMR models, SVMR showed a more stable performance than PLSR, and it also outperformed PLSR, with a smaller mean RMSE of 0.69 g/kg for DU soil and 0.50 g/kg for DG soil. This study concludes that PLSR is an effective linear algorithm, but it might not be sufficient when dealing with a nonlinear relationship, and SVMR turned out to be a more suitable nonlinear regression method for SOC estimation. Effective SOC estimation was obtained based on the DG soil samples, but the accurate estimation of SOC with DU soil samples needs to be further explored. In addition, LOG, SW, FD, and MC are valuable spectral preprocessing methods for Vis-NIR optimization, and choosing two of them (except for FD þ SW and LOG þ FD) in a simple combination is a good way to get acceptable results. Index Headings: Soil organic carbon; Visible-near-infrared, Vis-NIR; Support vector machine regression; Partial least squares regression; Spectral preprocessing.

Received 1 February 2013; accepted 15 January 2014. * Author to whom correspondence should be sent. E-mail: guofeng. [email protected]. DOI: 10.1366/13-07031

712

Volume 68, Number 7, 2014

Introduction Soil organic carbon (SOC) is one of the key soil properties because of its influence on plant growth, water holding capacity, soil structure and fertility, and it governs many soil processes.1 The evaluation of SOC variability using a fast, robust, and inexpensive tool is one of the key steps toward the successful implementation of environmental monitoring, modeling, and precision agriculture.2 Compared to traditional laboratory analyses, visible near-infrared (Vis-NIR) spectroscopy provides a fast, cost-effective, and relatively accurate alternative method for SOC content estimation.3,4 Over the past two decades, three spatial scales of Vis-NIR spectroscopic techniques have been applied for SOC content determination: laboratory spectroscopy, in situ or field spectroscopy, and imaging spectroscopy. In the case of laboratory spectroscopy, the illumination conditions and sample pretreatments can be controlled, which results in the most stable model calibrations.2–10 Field spectroscopy has been used for rapid in situ monitoring of soil; however, in this situation, the illumination conditions of the sun can vary and no sample pretreatment is applied.6,11,12 Regional studies have most often relied on imaging spectroscopy, mainly based on airborne imaging spectrometers.1,4,13 Spectroradiometers usually provide a large set of potential predictor variables that cover the complete VisNIR spectrum with a high spectral resolution, but they are compromised by multi-collinearity and noise.14 The selection of a multivariate calibration method, and its ability to handle the data, is therefore important for calibration success.5 Previous studies have reviewed or compared the most commonly used multivariate methods for spectroscopic techniques. For instance, Rossel et al.15 reviewed multivariate methods, including multiple linear regression and stepwise multiple linear regression (SMLR), multivariate adaptive regression splines, principal component regression (PCR), partial least squares regression (PLSR), and boosted regression trees, and they concluded that PLSR is the most appropriate technique for spectral calibration and estimation. Vasques et al.16 compared five multivariate

0003-7028/14/6807-0712/0 Q 2014 Society for Applied Spectroscopy

APPLIED SPECTROSCOPY

methods, namely, SMLR, PCR, PLSR, regression tree, and committee trees, and they also found that PLSR performed better than the other techniques. To date, a large number of PLSR models have been built for SOC estimation, from regional to global scales,8,13 from cropland or grassland soil, to quite heterogeneous soils with variable soil types or variable vegetation cover,13,15,17,18 and from both collected samples and spectral libraries.3,5,14 It is well known that the relationship between SOC and Vis-NIR reflectance data is complex and is often nonlinear.13 However, PLSR is a linear approach that may fail to represent such a complex and nonlinear relationship.14 For accurate and robust calibration, and the estimation of SOC content based on Vis-NIR reflectance, other effective nonlinear techniques are urgently required. Support vector machine regression (SVMR) is the latest powerful nonlinear regression technique based on statistical learning theory.6 Through a kernel function, SVMR nonlinearly maps date into higher dimensional feature spaces and facilitates the finding of an interpolation function.11 SVMR was designed to handle nonlinear relationships with large input spaces and noisy patterns, and it is therefore expected to be an effective approach for Vis-NIR reflectance calibration. Neural networks are also considered to be good nonlinear regression methods for spectral data, and some studies have applied them to SOC estimation, such as Mouazen et al.5 However, SVMR has outperformed neural networks and has the advantage of being a global model that is capable of efficiently dealing with highdimensional input vectors.11 To the best of our knowledge, the potential of SVMR for SOC content estimation has only been reported in the last three years. Stevens et al.13 first applied SVMR for SOC content estimation using AHS-160 imaging data (430–2450 nm); Rossel and Behrens3 applied SVMR to estimate SOC content based on Vis-NIR reflectance data of 1104 dried and ground soil samples; Vohland et al.14 estimated the SOC content of floodplain grassland topsoil in Germany by using SVMR; and Stevens et al.7 recently tested SVMR for continentalscale SOC estimation of cropland, grassland, and woodland. However, regardless of the complex and nonlinear reflectance behavior of soil, the characteristics of soil composition and the influence of environmental factors, such as surface roughness, ambient light, texture, color, and temperature, can also be possible sources of error.5 The successful estimation of SOC content further depends on sample constitution, soil particle sizes, and the spectral preprocessing methods used. Considering the current state of SOC content estimation with SVMR using the Vis-NIR technique, three questions emerge: (1) SVMR has mostly been tested on cropland, grassland, or woodland soil, and the applicability of SVMR for natural soils with more heterogeneous characteristics remains questionable; (2) previous studies have confirmed the potential of SVMR for soil with a good pretreatment (dried and 2 mm sieved), but the performance of SVMR on soil with larger particles and a coarse surface is unknown; and (3) several spectral preprocessing methods and wavelet transformation have been applied for SVMR optimization, but a more systematic test and

comparison of the spectral preprocessing methods for SVMR optimization has not been carried out to date. Accordingly, the scope of this study is to investigate the potential of SVMR for SOC content estimation based on heterogeneous natural soils, with different soil types, land-use types, and vegetation cover. In addition, dried unground (DU) and dried ground (DG) soil samples were used to compare the performance of SVMR on different particle-size soil. Nine spectral preprocessing methods and their combinations were applied to optimize SVMR and to further test the robustness of SVMR across various transformed spectra. PLSR was applied for comparison.

Materials and methods Field Sampling. The Yixing region (119831 0 –120803 0 N, 31807 0 31837 0 E) of China was selected as the study area because of its variation in topography and soil characteristics over short distances. The fieldwork was carried out on 11–14 August 2010. In total, 30 sampling sites were selected, considering different topographies, different soil parent materials, and different vegetation types. Ten sites were retained for intensive soil sampling (4–12 samples per site), and 35 additional samples were collected from the other 20 sites (one to three samples per site). The site areas ranged from 200 m2 (one sample) to 2 km2 (12 samples). At each sampling site, samples were selected with a distance of more than 100 m between each other. Finally, 100 heterogeneous topsoil samples were taken. The location of each sampling site was recorded with a global positioning system receiver. For each soil sample, five soil subsamples taken within a square of 1 m2 were collected from a large flat plot with almost the same vegetation cover and soil type and were then thoroughly mixed to get a representative sample.2 The soil from the surface to a depth of about 20 cm was collected using an iron shovel after removing the plant material and debris from the soil surface. About 500 g of soil was collected for each sample, and the samples were preserved in a labeled sample bag. When the fieldwork was finished each day, the soil samples were taken to the laboratory for air drying.

Laboratory analysis Vis-NIR Spectral Measurement. The soil Vis-NIR reflectance spectra were measured using an ASD FieldSpec Pro FR portable spectroradiometer (ASD Inc.) that consists of three individual sensors covering the spectral ranges of 350–1000, 1000–1800, and 1800– 2500 nm, with a sampling interval and spectral resolution of 1.4 and 3 nm for the 350–1000 nm region and 2 and 10 nm for the 1000–2500 nm region, respectively. The measured values were interpolated, and a 2150-band spectrum with a uniform 1 nm resolution was finally provided. The Vis-NIR spectral measurement was carried out in dark laboratory conditions. Considering the optimal values of beam angle, lamp distance, and sensor distance recommended by Jarmer et al.,19 the spectra of the soil samples were measured with a 50 W quartz halogen lamp positioned 30 cm away from the samples, with a 308 zenith angle as the light source,

APPLIED SPECTROSCOPY

713

whereas the spectroradiometer probe was mounted about 10 cm above the sample. Before Vis-NIR spectral measurement, each soil sample was divided into two equal parts, one of which was used as the unground sample, and the other was carefully ground with an agate mortar and sieved using a 2 mm nylon-fiber sieve to obtain the ground sample. The spectra of the two different types of pretreated soil samples were then obtained: DU (fresh soil samples were spread out on dishes and air-dried for two days at room temperature) and DG (air-dried samples were ground and passed through a 2 mm sieve). For each measurement, the soil sample was packed onto a 10 cm diameter Petri dish with a thickness of about 15 mm and gently shaken to get a flat surface. A white Spectralon panel was used to optimize the signal and to calibrate the spectroradiometer response, and its radiance was collected before the first scan and after every six samples. The radiance of the soil samples was obtained with 10 consecutive scans. By dividing the mean radiance of the soil sample by the radiance over the Spectralon panel, the spectral reflectance of each soil sample was derived. For each sample, the average reflectance of 10 repeated measurements was calculated and considered as the final reflectance spectrum. Laboratory SOC Measurement. After the spectral measurements, all the dried ground soil samples were sent to the laboratory for SOC content measurement. The Walkley–Black method,20 a commonly used chemical SOC measurement method,1,15 was used to determine the SOC content.

Model development Outlier Detection. In general, the distribution of univariate data is described by a location-scale model in which the n univariate observations xi are independent and identically distributed with the distribution function F ((x – lx/r).21 For the distribution function, F is typically the known standard Gaussian distribution function U, l denotes the location parameter (commonly known as the sample mean x¯ or median), and r denotes the scale parameter (classically, the standard deviation (s) or the median absolute deviation (MAD)). The classical outlier detection rule is based on the standardized residuals of the observations, and sets the location parameter as x¯ and the scale parameter as s in the distribution function F. Flagged as an outlier is xi, if jxi  x¯j/s exceeds a certain value, such as 2.5. However, the location and scale parameters, namely, the x¯ and s, are non-robust because removing only one out of n observations of a large value can completely change the mean x¯ and, similarly, one outlier can make the standard deviation s value arbitrarily large. It is possible that some outliers will not be detected and/or some regular observations will be incorrectly flagged as outliers.21 Thus, Rousseeuw et al.21 recommended the application of robust estimators for the location and scale, such as the median and the MAD, which yield jxi – medianj=1,an(xj)j/MAD as a much more reliable outlier detection tool. We applied this outlier detection method to the Vis-NIR reflectance and SOC content data by the use of the LIBRA toolbox.22 Two of the 100 soil samples that had an SOC content of 0.82 and

714

Volume 68, Number 7, 2014

TABLE I. The four categories of preprocessing transformations applied to the spectral curves of the soil samples.

Category Transformation

Normalization

Differentiation Center data

Preprocessing transformation technique Raw reflectance Absorbance transformation: R to log (1/R)16 Multiplicative signal correction12 Standard normal variate24 SNV detrending24 Wavelet detrending25 SNV with wavelet detrending25 First derivative Second derivative Mean centering15

Abbreviationa R LOG MSC SNV SNVD WD SW FD SD MC

a The preprocessing transformations are presented as an abbreviation in this study, and their combinations are expressed as the order of the abbreviation; for example, ‘‘LOG þ MC’’ means that the raw reflectance was first transformed to log (1/R) and was then processed with mean centering.

17.07 g/kg, but had unusual spectra, were identified as outliers and excluded from the following analyses. Spectral Interpretation and Preprocessing. Splice corrections were carried out using ViewSpec Pro version 5.6 software (ASD Inc.) to remove the splicing noise near 1000 and 1800 nm between the three individual sensors. Before further preprocessing, the reflectance values between 350–410 and 2451–2500 nm were removed because of their low signal-to-noise ratios.5 The remaining spectra with a 1 nm interval over 411–2450 nm were averaged with a 10 nm window to reduce the data dimensionality and to smooth the raw curves.12,16 The average raw reflectance curves of the remaining DU and DG soil samples, as well as their derivatives, were visualized and analyzed. The autocorrelation of the raw reflectance spectra of the DU and DG soils was calculated using full wavelength data.23 By comparing the autocorrelations of the DU and DG soil spectra, the effects of the grinding and 2 mm sieving pretreatments on the Vis-NIR diffuse reflectance of the air-dried soil samples were explored. To reduce the spectral noise and optimize the spectral data, four categories in total consisting of nine kinds of preprocessing transformations were applied (see the specific categories, preprocessing transformation techniques, and related abbreviations in Table I). Because the raw curves were previously smoothed by a 10 nm window, no other smoothing techniques were applied. The spectral normalizations were used to deal with the spectra affected by light scattering and background,12,24,25 which likely occurred in our study because the DU samples contained many large particles. Therefore, as many as five kinds of different normalization techniques were tested in our study. For full details of those preprocessing methods, please refer to the corresponding references cited in Table I. More than one category of preprocessing transformation has usually been used in previous research;4,8,15,26 thus, we expected that the different preprocessing methods would hold particular benefits for certain spectra and that the combinations of several useful categories of preprocessing methods might produce an

additional improvement. Therefore, the above-mentioned nine preprocessing methods were applied alone first, and then the method with the best performance in each category was chosen as a representative method for the further combinations. In total, ten preprocessing combinations were thus obtained from the four categories. In addition, the best-performing method of each category was chosen as a representative method for the combination, instead of an all-subsets combination, for the following reasons: (1) an all-subsets approach for all the four categories of preprocessing transformation combinations would produce 280 combinations in total, which is a large workload and would not be easy to process; (2) to compare the performance of SVMR and PLSR in SOC estimation, a large number of preprocessing procedures would make a more stable and convincing result, but not necessarily an exhaustive one; and (3) a previous all-subsets combination experiment with fewer preprocessing methods showed that the specific combination approach, as used in this study, performed better than the all-subsets combinations. All the preprocessing techniques were carried out using ParLeS software version 3.1.25 Model Calibration. After two outliers were removed, the remaining dataset with 98 soil samples was randomly divided into two sub-datasets. One sub-dataset with 66 samples (calibration dataset) was used for the model calibrations, and the other with 32 samples (validation dataset) was used to independently validate the calibrated models. PLSR and SVMR were applied to perform the calibrations between the Vis-NIR spectra and the SOC content. The PLSR model calibration was implemented with ParLeS version 3.1 software. An important issue for PLSR calibration is to determine the proper number of latent variables. To avoid over- or under-fitting problems, the minimum number of latent variables that explains most of the variations in both the independent and dependent variables should be used in the model development.15 In this study, the number of latent variables was determined by using leave-one-out cross-validation13 with the calibration dataset. The optimal number of latent variables is the one that produces the minimum RMSE of the cross-validation13 and the lowest Akaike information criterion (AIC)27 values. The AIC was calculated by Nln(RMSE) þ 2m, where N is the sample size and m is the number of latent variables used.25 The basic concept of SVMR is the kernel function that nonlinearly maps the input dataset into a higher dimensional feature space and then derives a linear hyperplane in this feature space as a decision function for the regression problem.3 The kernel function is identified by fitting a tube with radius epsilon to training data only using boundary samples, called support vectors (SVs), instead of the whole training dataset.28 As a result, SVMR has the ability to model nonlinear relationships, with a high level of performance. The details of the SVM regression algorithm can be found in Thissen et al.11 There are two different SVMR algorithms (e-SVMR and t-SVMR) and four commonly used kernel functions (linear kernel, polynomial kernel, radial basis function

(RBF) kernel, and sigmoid kernel).9 e-SVMR is the most commonly used SVMR algorithm. RBF is a nonlinear kernel function that is able to handle nonlinear relationships and reduce the computational complexity of the calibration procedure.9 Thus, the e-SVMR algorithm and RBF kernel function were chosen for the model calibration in this study. To calibrate e-SVMR, two parameters, namely, the cost parameter (C) and the c parameter in the RBF kernel, are required.9 In this study, the combination of (C, c) was optimized by a two-step grid search technique with five-fold cross-validation.11 Finally, the (C, c) combination with the smallest mean square error (MSE) value was chosen. The ranges of C and c were set to (210–210), based on experience and previous research.9 The SVMR model calibration was carried out using the LibSVM toolbox version 3.19 implemented in Matlab 2010 (The Mathworks, Natick, MA). Model Validation. To assess the predictive abilities of the calibrated models, several parameters were calculated using the validation dataset. The determination coefficients (R2) of the regression line between the measured and estimated values and the root mean square error (RMSEv) were used.16 The residual prediction deviation (RPD) (Eq. 1), the ratio of the standard error of performance to the standard deviation of the reference data, was used to evaluate the estimation ability of the calibrated models.29,30 Veronique et al.31 proposed the ratio of performance to inter-quartile range (RPIQ) (Eq. 2) as a new metric, which is based on quartiles, to represent the spread of population. The bias is the difference between the estimated mean and the mean measured values, and it indicates the error of the means. RPD ¼ SDv =RMSEv RPIQ ¼ ðQ3  Q1Þ=SEP

ð1Þ ð2Þ

where SDv denote the sample number and the standard deviation of the measured values of the validation dataset, respectively. Q1 is the value below which we can find 25% of the samples, Q3 the value below which we find 75% of the samples, and SEP is the standard error of the estimation. The best estimation model was considered to be that with the largest R2, RPD, and RPIQ values, and the smallest error (RMSEv) in the validation.32 As is generally accepted, five levels of prediction accuracy were considered: a value for the RPD below 1.5 indicates that the calibration is not usable; a value for the RPD between 1.5 and 2.0 reveals a possibility to distinguish between high and low values; a value between 2.0 and 2.5 makes approximate quantitative predictions possible; and for values between 2.5 and 3.0, and above 3.0, the prediction is classified as good and excellent, respectively.29

Results Descriptive Statistics. The descriptive statistics of the SOC content measured using the Walkley–Black method (Table II) showed that the SOC content of the study area consisted of low values (¯ x = 13.53 g kg1) with moderate

APPLIED SPECTROSCOPY

715

TABLE II. Statistics describing the SOC content (g/kg) for the whole dataset and for the calibration and validation dataset. Dataset

SNa

Mean

Min

Q1b

Median

Q3c

Max

SDd

SEe

Whole Calibration Validation

98f 66 32

13.53 13.34 13.92

0.79 0.79 5.30

9.23 9.20 10.90

14.34 14.34 14.47

17.58 17.72 17.33

30.73 30.73 21.26

5.43 5.83 4.54

5.40 5.79 4.47

a

SN is the sample number. Q1 is the value below which 25% of the samples can be found. c Q3 is the value below which 75% of the samples can be found. d SD is the standard deviation. e SE is the standard error. f Two outliers were removed, and only the 98 remaining samples were statistically described. b

variation (standard deviation = 5.43 g/kg, standard error = 5.79 g/kg). The distribution analysis of the whole dataset showed that the SOC content of the 98 remaining samples was close to a normal distribution, with a skewness value of 0.46. The similarities in the range and variation of the SOC content between the calibration and validation datasets provide the essential foundation for the effective calibration and validation of the estimation models. Interpretation of Vis-NIR Spectra. Figure 1 shows the average raw reflectance curves and the first derivatives (FDs) of the 98 remaining DU and DG samples. The average reflectance curves have an increasing pattern over the range of 410–1400 nm and then fluctuate between 1400 and 2500 nm (Fig. 1a), consistent with the typical characteristics of soil spectra. The effects of water absorption near 1400, 1900, and 2300 nm are apparent in the raw reflectance and are strengthened in the FD curves (Figs. 1a and 1b). For the raw reflectance, the DU and DG curves have a similar pattern, but with a large difference in the values, and the average reflectance of the DU soil samples was much lower than that of the DG soil samples (Fig. 1a). The differences were not linear across the spectrum because the visible part was less affected than the NIR region. When looking at the FD, apart from the minor differences over the range of 410–750 nm, the DU and DG spectra yielded more or less the same curves (Fig. 1b). The spectral autocorrelations of the DU and DG soil samples are given in Fig. 2. The values of the scale bar in Fig. 2 represent the values of the correlation coefficients. The darker color near the diagonal lines of the plots indicates that adjacent reflectances have a

stronger correlation. Specifically, the left sides (around 410–600 nm) of the two figures are lighter than the other parts, which might indicate that these regions are less correlated with the full wavelength. Three diverse horizontal and vertical lines near 1400, 1900, and 2200 nm can be observed, which might be caused by water absorption. In general, the color of Fig. 2b is lighter than that of Fig. 2a, especially the correlation between band 4101400 and 19002500 nm, which might indicate that the multi-collinearity problem in Vis-NIR soil diffuse reflectance is reduced by the grinding and 2 mm sieving pretreatments. Performance of the Vis-NIR Spectral Estimation Models. The validation results of the PLSR and SVMR models using the raw reflectance and their nine single preprocessing transformed spectra (Table I), based on the DU and DG soil spectra, are shown in Fig. 3. The raw reflectance produced poor validation accuracies for the PLSR models (DU, R2 = 0.27; DG, R2 = 0.50) and moderate validation accuracies for the SVMR models (DU, R2 = 0.65; DG, R2 = 0.70). According to the single preprocessing results, SW, FD, absorbance transformation (LOG), and mean centering (MC) were chosen as the representative methods of the normalization, differentiation, transformation, and centering data categories, respectively, and were then further combined. The validation results of the 10 preprocessing combinations are also shown in Fig. 3. For the PLSR model calibrations, the optimal number of latent variables varied from four to eight, with a mean of seven. The accuracies of the PLSR models were highly variable in both calibrations and validations, considering the raw reflectances and the 19 preprocess-

FIG. 1. (a) Average raw absolute reflectance curves of the 98 remaining soil samples. The solid and dashed gray curves show the 95% confidence limits of the DG and DU average curves, respectively. (b) Average FDs of the reflectance curves.

716

Volume 68, Number 7, 2014

FIG. 2. Spectral autocorrelation coefficients of (a) DU and (b) DG soil samples. The values of the scale bar represent the values of the correlation coefficients. A point with a value of 0.8 in the plot means that the correlation coefficient of the corresponding band of the x-axis and y-axis is 0.8.

ing transformed spectra. Compared with the raw reflectance (DU , R2 = 0.27; DG , R2 = 0.50), all the single preprocessing methods improved the PLSR models for both the DU and DG spectra, except for standard normal variate (SNV) detrending (DU, R2 = 0.03; DG , R2 = 0.45) and second derivative (SD) (DU, R2 = 0.23; DG , R2 = 0.34), whereas LOG (DU, R2 = 0.65; DG , R2 = 0.65) and MC (DU, R2 = 0.64; DG , R2 = 0.67) performed the best. In addition, SW (DU, R2 = 0.52; DG , R2 = 0.61) was the most suitable normalization method and FD (DU, R2 = 0.45; DG , R2 = 0.51) was the most suitable differentiation method for PLSR. The preprocessing combinations produced additional improvements compared with the single preprocessing transformations. The best preprocessing combination was FD þ MC (DU, R2 = 0.69; DG , R2 = 0.79), which produced the optimal PLSR models for both the DU and DG soil samples. Nearly 70% of the training samples (between 33 and 47 samples out of 66) were retained as support vectors for the SVMR model calibrations. Compared with the raw reflectance (DU , R2 = 0.65; DG, R2 = 0.70), the SVMR models benefited less from the spectral preprocessing than the PLSR models, and their validation results were less variable than those of the PLSR models, especially for the DG models. The best single preprocessing transformation for SVMR was FD (DU, R2 = 0.72; DG , R2 = 0.86). The preprocessing combinations produced no significant additional improvement, and the best one was FD þ MC (DU, R2 = 0.72; DG , R2 = 0.86), which produced a similar accuracy to that of FD for both the DU and DG soil spectra. A statistical description of the validation results of the DU and DG models produced by PLSR and SVMR is shown in Table III. Among the 20 DU models and 20 DG models, SVMR outperformed PLSR, with larger R2, RPD, and RPIQ values and a smaller RMSEv value. In addition, SVMR was more stable than PLSR, with smaller SD values of the four assessment parameters (R2, RMSEv, RPD, and RPIQ) for both the DU and DG soil spectra across the tested preprocessing transformations. The worst model among the SVMR models obtained a higher accuracy (DU, R2 = 0.41; DG, R2 = 0.61) than the PLSR models (DU, R2 = 0.03; DG, R2 = 0.34), which might indicate that SVMR tends to be less negatively affected by spectral preprocessing transformations, and usually produces reasonable models. When comparing the

performance between DU and DG soil, with the ground and 2 mm sieved pretreatment, the SOC estimation accuracies of the PLSR models were improved from a mean R2 of 0.47 for the DU soil to a mean R2 of 0.60 for the DG soil, and the SOC estimation accuracies of the SVMR models were improved from a mean R2 of 0.62 for the DU soil to a mean R2 of 0.73 for the DG soil. According to Fig. 3, the optimal preprocessing methods in this study were FD þ MC for PLSR and FD for SVMR. The optimal preprocessing methods were then applied to PLSR and SVMR, respectively, to estimate the SOC contents of the validation dataset. The detailed calibration and validation results and the statistics describing the SOC content estimated by the optimal PLSR and SVMR models of the DU and DG soil samples are given in Table IV. In addition, the estimated versus observed SOC content scatter plots for the calibration and validation datasets of the four optimal models is shown in Fig. 4.

Discussion Many studies have applied the Vis-NIR spectroscopic technique to the estimation of SOC content (Table V), and our study obtained estimation accuracies (R2 = 0.86, RMSE = 1.84 g/kg, RPD = 2.60) that were comparable to most of the other studies; for example, Vohland et al.14 (R2 = 0.89), Fystro18 (R2 = 0.87), Aichi et al.17 (R2 = 0.83), and Rossel and Behrens3 (R2 = 0.84). However, the estimation accuracy of this study was lower than the study by Kuang and Mouazen33 (R2 = 0.96), which obtained excellent accuracies based on farmland soil. The lower accuracy in this study could be attributed to the more heterogeneous soils used, because Udelhoven et al.26 and Bricklemyer and Brown10 demonstrated that models generally have a higher estimation accuracy when they are developed using a dataset with more homogeneous soil. In addition, compared with the other studies based on heterogeneous soils, such as Morgan et al.34 (R2 = 0.81) and Vasques et al.8 (R2 = 0.74), our study obtained a better result. This might be a result of the spectral preprocessing and calibration methods we adopted being more effective, because the results of this study (Fig. 3) showed that the FD þ MC preprocessing method and the SVMR calibration method were significantly more effective than FD and PLSR, respectively,

APPLIED SPECTROSCOPY

717

FIG. 3. Performances of the SOC content estimation models with the validation dataset (32 samples), using raw reflectance and their 19 preprocessing transformations of the DU and DG soil samples. (a) PLSR models of DU soil spectra. (b) SVMR models of DG soil spectra. (c) PLSR models of DG soil spectra. (d) SVMR models of DG soil spectra. R2, RMSEv (g/kg) of the validation and RPD are as described in the text. The long horizontal dashed line at the top of each plot represents an RPD of 2.60 for clear visualization. The abbreviations are the same as those used in Table I

which were commonly used in the above-mentioned studies. The spectral autocorrelation analysis of the DU and DG soil confirmed that the spectral reflectance of soil across the Vis-NIR domain was highly multi-collinear (Fig. 2). To deal with the Vis-NIR spectra, multivariate methods that could adapt to the high-dimensional and

multi-collinear data structure, and select variables to eliminate redundancy, were expected to perform effectively.35 PLSR statistically rotates input data to derive a small number of latent variables for calibration and has been considered an effective method in handling highdimensional and multi-collinear data.11 However, from the results across 19 kind of preprocessing transformed

TABLE III. Statistics describing the validation results of the PLSR and the SVMR models for the DU and DG soil samples. PLSR

SVMR

Statisticsa

R2

RMSEv, g/kg

RPD

RPIQ

Bias

R2

RMSEv, g/kg

RPD

RPIQ

Bias

DU spectra Min Mean Max SD

0.03 0.47 0.69 0.18

2.87 3.98 13.90 2.31

0.33 1.29 1.58 0.30

0.46 1.87 2.24 0.42

13.17 0.35 0.87 2.96

0.41 0.62 0.72 0.08

2.48 3.29 9.87 1.50

1.25 1.55 1.84 0.14

0.65 2.15 2.59 0.38

9.19 0.23 2.02 2.15

DG spectra Min Mean Max SD

0.34 0.60 0.79 0.12

2.32 2.95 4.04 0.47

1.04 1.56 1.95 0.24

1.59 2.28 2.77 0.34

12.90 0.10 1.67 3.00

0.61 0.73 0.86 0.05

1.84 2.45 2.81 0.26

1.60 1.92 2.60 0.24

2.29 2.69 3.50 0.30

0.89 0.29 1.47 0.60

a

Min is the minimum; Max is the maximum; SD is the standard deviation.

718

Volume 68, Number 7, 2014

TABLE IV. Calibration and validation results as well as SOC (g/kg) estimation statistics of the best PLSR and SVMR models with the DU and DG soil samples. Calibration (66a)

Validation (32a)

Statistics of estimated SOC (g/kg)b

Multivariate method Preprocessing

Fc

R2

RMSE

R2

Dried underground soil samples PLSR FD þ MCh SVMR FD

6 42

0.71 0.74

2.94 2.29

0.69 0.72

2.89 2.48

1.57 1.83

2.23 2.59

0.63 0.40

14.55 3.05 11.88 14.32 1.30 12.38

15.81 15.24

18.11 22.35 5.07 4.99 17.54 19.88 4.54 4.47

Dried ground soil samples PLSR FD þ MC SVMR FD

7 42

0.78 0.82

2.58 2.13

0.79 0.86

2.33 1.84

1.95 2.60

2.76 3.50

0.90 0.64

14.82 4.86 13.82 14.55 5.41 12.19

15.81 15.44

18.33 22.37 4.73 4.66 17.65 22.01 4.54 4.47

RMSEv RPD RPIQ Bias Mean Min

Q1d

Median

Q3e

Max

SDf

SEg

a

Number of samples in the calibration/validation dataset. Number of samples is 32. c Number of latent variables used in PLSR and support vectors used in SVMR. d Bias is the difference between the estimated mean and the mean measured values, and it indicates the error of the means. e Q1 is the value below which we can find 25% of the samples. f Q3 is the value below which we find 75% of the samples. g SD is standard deviation. h RPD . 3.0 = excellent estimations; 3.0 . RPD . 2.5 = good estimations; 2.5 . RPD . 2.0 = approximate quantitative estimations; 2.0 . RPD . 1.5 = a possibility to distinguish between high and low values; RPD , 1.5 = not usable estimations. b

spectra of the thoroughly ground (2 mm) DG soil, SVMR outperformed PLSR, with more accurate estimation and better robustness (Fig. 3; Tables III and IV). In detail, SVMR produced a smaller RMSE of 1.84 g/kg with DG soil, as against 2.33 g/kg for PLSR. For the 40 DG models, SVMR, with an average RMSE of 2.45 g/kg, provided a more robust performance than PLSR, which had a larger average RMSE of 2.95 g/kg. Together with

some other studies, such as Stevens et al.,13 Vohland et al.,14 Rossel and Behrens,3 and Stevens et al.,7 our study confirmed the potential of the SVMR method for estimating SOC content from thoroughly ground soil Vis-NIR reflectance measurements. In addition, this study first implemented SVMR based on unground rough soil with a maximum particle size of nearly 10 cm. Even though only moderate accuracy was obtained (R2 = 0.72,

FIG. 4. Estimated versus observed SOC content for the calibration dataset (66 samples represented as  in Fig. 4) and validation dataset (32 samples represented as  in Fig. 4) using: (a) PLSR and DU spectra, (b) SVMR and DU spectra, (c) PLSR and DG spectra, and (d) SVMR and DG spectra. Solid lines represent the regression lines of the validation; the dashed lines represent the 1:1 correlation lines. R2 and RMSE are as described in the text.

APPLIED SPECTROSCOPY

719

TABLE V. A list of the latest literature comparing quantitative estimations of SOC content (g/kg) using multivariate analyses and Vis-NIR spectroscopy. Samplea

SNb

Cropland Heterogeneous Heterogeneous Heterogeneous Heterogeneous Farms Heterogeneous Grassland Heterogeneous Grassland Severald

90 64 146 540 765 408 7120 149 1104 123 20 000

Preprocessingc SG R R FD FD SG SG VN CR R SG

þ FD

þ FD þ FD

þ FD þ SNV

Method

R2

RPD

Reference

PLSR PLSR PLSR PLSR PLSR PLSR PLSR PLSR/SVMR SVMR PLSR SVMR

0.86 0.83 0.67 0.81 — 0.96 0.74 0.89 0.84 0.87 0.67–0.87

2.66 2.35 1.75 2.08 1.1 4.95 1.96 2.77 — 2.70 1.74–2.88

Kusumo et al.32 Aichi et al.17 Gomez et al.2 Morgan et al.34 Bricklemyer and Brown10 Kuang and Mouazen33 Vasques et al.8 Vohland et al.14 Rossel and Behrens3 Fystro18 Stevens et al.7

a

Samples are the types of soil samples. SN, sample number. c R, reflectance; VN, vector normalization39; CR, continuum removal40; SNV24; SG þ FD, first Savitzky–Golay derivative.39 d Models were established individually in several different soil types of cropland, grassland, woodland, and mineral, and the R2, RMSE, and RPD are given in a range across those soil types. b

RMSE = 2.48 g/kg, RPD = 1.83), the better estimation and robustness of SVMR compared with PLSR further confirmed that SVMR is a more suitable multivariate method for Vis-NIR spectral calibration in cases with rough surface conditions, such as field spectroscopy or imaging spectroscopy. The nonlinear relationship between the Vis-NIR reflectance data and SOC content inevitably emerged,13,14 and some other studies have reported that the experimental conditions and absorption characteristics of the analyzed component might enhance the nonlinear relationship.11 Thissen et al.11 pointed out that the nonlinear relationship between the soil component content and spectral reflectance could only be modeled by PLSR in a limited way by adding more latent variables. Thus, the remarkable improvements obtained by the SVMR model can be explained by its ability to model the nonlinearity between the reflectance data and SOC content. Moreover, SVMR simplifies high-dimensional data by compressing the training samples to a smaller subset of significant support vectors, then uses the support vectors for further model development.6 Thus, another explanation might be that the simplification mechanism of support vectors is more effective than the latent variables of PLSR for calibration. Zhai et al.36 found that six latent variables of PLSR accounted for more than 50% of the variation of plant biochemical components, whereas 43 support vectors accounted for more than 90% of the variation of the same plant biochemical components in their research. In this study, the DG soil (maximum particle diameter of nearly 2 mm) performed better than the DU soil (maximum particle diameter of nearly 1 cm) in SOC estimation across all the 80 models. Such a result is understandable, because the DG sample was reduced to a uniform size, moisture was reduced, and any physical heterogeneity of the soil was homogenized. A further reason, as Gomez et al.2 and Udelhoven et al.26 reported, is that the high roughness of the soil will add microshadows to the measured surface and might lead to absorption features being hidden in the Vis-NIR region. The field sampling method, which compounded five subsamples to get a representative soil sample for each

720

Volume 68, Number 7, 2014

site, might be another reason. Even though the sample sites were always large flat plots with almost the same vegetation cover and soil type, and the composite samples were thoroughly mixed in the fieldwork, this sample collection approach could lead to the DG samples being more completely mixed than the DU samples. Despite this slight difference, to further improve the performance of the DU soil, the following aspects could be helpful: (1) the reflectance signal of the soil sample was greatly enhanced by the grinding pretreatment (Fig. 1), and (2) the autocorrelation analysis indicated that the multi-collinearity of the DG spectra was less than that of the DU spectra (Fig. 2) (Hill et al.28 revealed that data with a lower multi-collinearity related to more useful information). It was shown in this study that the preprocessed spectra obtained better results (Fig. 3) compared with the raw spectral reflectance for both the PLSR and SVMR methods when estimating the SOC content of DU and DG soil. Similar results have been reported in many studies,16–18,36 which also confirmed the effects of combined preprocessing methods in improving SOC estimation.5,19 These preprocessing methods improve the relationship between the SOC content and spectral reflectance of soil by reducing the noise, and they strengthen the useful information. The large improvement resulting from the FD method can be explained by its effectiveness in eliminating background signals (Figs. 1 and 3). Tsai and Philpot37 confirmed that the FD method can explore information that was often suppressed by other analysis methods. Compared with the poor performance of LOG in the SVMR models (Fig. 3), the excellent performance in the PLSR models can be explained by their ability to reduce data nonlinearity.38 Vasques et al.16 demonstrated that LOG was the best preprocessing method for PLSR among 14 tested methods. The improvement of MC in the PLSR and SVMR models indicated that it is a good method for VisNIR spectra. Because the specific combination approach of this study is not an all-subsets approach, it might have missed methods that would have worked well in combination, though not necessarily individually. According to the 80 models in this study, choosing a single

method or two of the above-mentioned LOG, FD, and MC (except LOG þ FD) in a simple combination is an advisable way to get acceptable results. In addition, the remarkable performance of SW proved it to be a more useful normalization technique than the traditional multiplicative signal correction (MSC) and SNV in either compensating for the particle size in spectral reflectance or resolving scatter problems. This study focused on testing SVMR using nine kinds of common preprocessing methods and their representative combinations, based on laboratory pretreated DU and DG soil. For a scientific comparison, an all-subsets approach of these preprocessing methods combined with some other useful methods, such as vector normalization,39 continuum removal,40 and wavelet transform, all of which have shown potential in SVMR optimization,3,14 is recommended. Many studies have estimated SOC content using in situ Vis-NIR spectroscopy or airborne and spaceborne hyperspectral reflectance data.1,4 Although environmental factors, such as the changing sunshine, vegetation cover, soil moisture, and atmosphere effect, make the estimation of SOC more complex at the outdoor level than in laboratory conditions, we believe that the spectral autocorrelation, preprocessing, and SVMR methods suggested in this study have the potential to retrieve SOC content in situ or even at a landscape level, and we are therefore investigating to see whether this also applies for heterogeneous soil of nitrogen and phosphorus.

Conclusions Using DU and DG pretreated soil samples, we compared 19 kinds of spectral preprocessing methods, and we investigated the potential of the PLSR and SVMR methods in SOC content estimation. The principal results obtained are summarized as follows: (1) Based on heterogeneous soils, SVMR obtained good SOC estimation accuracy and robustness with an RPD of 2.60 on dried ground samples and moderate SOC estimation accuracy and robustness with an RPD of 1.92 on dried unground samples. The SVMR model with dried ground soil is capable of accurate SOC estimation, but accurate estimation based on unground soil needs to be further explored. (2) SVMR outperformed PLSR in SOC content estimation for both ground and unground pretreated soil. Across 20 kinds of Vis-NIR spectral preprocessing methods, the SVMR models were more accurate and robust for SOC estimation than PLSR. (3) FD þ MC performed much better than FD for PLSR, and FD turned out to be a good preprocessing method for SVMR. SW is a more suitable normalization technique than MSC and SNV. LOG and MC are two kinds of valuable transformations in Vis-NIR preprocessing. When dealing with dried soil Vis-NIR spectra, choosing a single method, or two of the above-mentioned LOG, FD, and MC (except FD þ SW and LOG þ FD) in a simple combination is an advisable way to get acceptable results; however, a

more detailed Vis-NIR preprocessing strategy is needed for further specialization. ACKNOWLEDGMENTS This study was supported by grants from the Special Foundation of the Ministry of Finance of China for Nonprofit Research of Forestry Industry (200904001) and the National Natural Science Foundation of China (41171290). 1. H. Bartholomeus, L. Kooistra, A. Stevens, M. van Leeuwen, B. van Wesemael, E. Ben-Dor, B. Tychon. ‘‘Soil Organic Carbon Mapping of Partially Vegetated Agricultural Fields with Imaging Spectroscopy. Int. J. Appl. Earth Obs. Geoinf. 2011. 13(1): 81-88. 2. C. Gomez, R.A. Viscarra Rossel, A.B. McBratney. ‘‘Soil Organic Carbon Prediction by Hyperspectral Remote Sensing and Field VisNIR Spectroscopy: An Australian Case Study’’. Geoderma. 2008. 146(3-4): 403-411. 3. R.A. Viscarra Rossel, T. Behrens. ‘‘Using Data Mining to Model and Interpret Soil Diffuse Reflectance Spectra’’. Geoderma. 2010. 158(1-2): 46-54. 4. A. Stevens, B. van Wesemael, H. Bartholomeus, D. Rosillon, B. Tychon, E. Ben-Dor. ‘‘Laboratory, Field and Airborne Spectroscopy for Monitoring Organic Carbon Content in Agricultural Soils’’. Geoderma. 2008. 144(1-2): 395-404. 5. A.M. Mouazen, B. Kuang, J. De Baerdemaeker, H. Ramon. ‘‘Comparison Among Principal Component, Partial Least Squares and Back Propagation Neural Network Analyses for Accuracy of Measurement of Selected Soil Properties with Visible and Near Infrared Spectroscopy’’. Geoderma. 2010. 158(1-2): 23-31. 6. V.N. Vapnik. ‘‘The Nature of Statistical Learning Theory’’. In: M. Jordan, S.L. Lauritzen, J.F. Lawless, V. Nair, editors. Statistics for Engineering and Information Science. New York: Springer Verlag, 2000. Pp. 1-314. 7. A. Stevens, M. Nocita, G. To´th, L. Montanarella, B. Van Wesemael. ‘‘Prediction of Soil Organic Carbon at the European Scale by Visible and Near Infrared Reflectance Spectroscopy’’. PLoS One. 2013. 8(6): 1-13. 8. G.M. Vasques, S. Grunwald, W.G. Harris. ‘‘Spectroscopic Models of Soil Organic Carbon in Florida, USA’’. J. Environ. Qual. 2010. 39(3): 923-934. 9. C.-C. Chang, C.-J. Lin. ‘‘LIBSVM: A Library for Support Vector Machines’’. Environ. Sci. Technol. 2011. 2(3): 1-27. 10. R.S. Bricklemyer, D.J. Brown. ‘‘On-the-Go VisNIR: Potential and Limitations for Mapping Soil Clay and Organic Carbon’’. Comput. Electron. Agric. 2010. 70(1): 209-216. 11. U. Thissen, M. Pepers, B. Ustun, W.J. Melssen, L.M.C. Buydens. ‘‘Comparing Support Vector Machines to PLS for Spectral Regression Applications’’. Chemom. Intell. Lab. Syst. 2004. 73(2): 169-179. 12. K.D. Shepherd, M.G. Walsh. ‘‘Development of Reflectance Spectral Libraries for Characterization of Soil Properties’’. Soil Sci. Soc. Am. J. 2002. 66(3): 988-998. 13. A. Stevens, T. Udelhoven, A. Denis, B. Tychon, R. Lioy, L. Hoffmann, B. van Wesemael. ‘‘Measuring Soil Organic Carbon in Croplands at Regional Scale Using Airborne Imaging Spectroscopy’’. Geoderma. 2010. 158(1-2): 32-45. 14. M. Vohland, J. Besold, J. Hill, H.C. Fruend. ‘‘Comparing Different Multivariate Calibration Methods for the Determination of Soil Organic Carbon Pools with Visible to Near Infrared Spectroscopy’’. Geoderma. 2011. 166(1): 198-205. 15. R.A. Viscarra Rossel, D.J.J. Walvoort, A.B. McBratney, L.J. Janik, J.O. Skjemstad. ‘‘Visible, Near Infrared, Mid Infrared or Combined Diffuse Reflectance Spectroscopy for Simultaneous Assessment of Various Soil Properties’’. Geoderma. 2006. 131(1-2): 59-75. 16. G.M. Vasques, S. Grunwald, J.O. Sickman. ‘‘Comparison of Multivariate Methods for Inferential Modeling of Soil Carbon Using Visible/Near-Infrared Spectra’’. Geoderma. 2008. 146(1-2): 14-25. 17. H. Aichi, Y. Fouad, C. Walter, R.A. Viscarra Rossel, Z.L. Chabaane, M. Sanaa. ‘‘Regional Predictions of Soil Organic Carbon Content from Spectral Reflectance Measurements’’. Biosyst. Eng. 2009. 104(3): 442-446. 18. G. Fystro. ‘‘The Prediction of C and N Content and Their Potential Mineralisation in Heterogeneous Soil Samples Using Vis-NIR Spectroscopy and Comparative Methods’’. Plant Soil. 2002. 246(2): 139-149.

APPLIED SPECTROSCOPY

721

19. T. Jarmer, M. Vohland, H. Lilienthal, E. Schnug. ‘‘Estimation of Some Chemical Properties of an Agricultural Soil by Spectroradiometric Measurements’’. Pedosphere. 2008. 18(2): 163-170. 20. A. Walkley, I.A. Black. ‘‘An Examination of the Degtjareff Method for Determining Soil Organic Matter, and a Proposed Modification of the Chromic Acid Titration Method’’. Soil Sci. 1934. 37(1): 29. 21. P.J. Rousseeuw, M. Debruyne, S. Engelen, M. Hubert. ‘‘Robustness and Outlier Detection in Chemometrics’’. Crit. Rev. Anal. Chem. 2006. 36(3-4): 221-242. 22. S. Verboven, M. Hubert. ‘‘LIBRA: a MATLAB Library for Robust Analysis’’. Chemom. Intell. Lab. Syst. 2005. 75(2): 127-136. 23. R.A. Viscarra Rossel, A. Chappell, P. de Caritat, N.J. McKenzie. ‘‘On the Soil Information Content of Visible-Near Infrared Reflectance Spectra’’. Eur. J. Soil Sci. 2011. 62(3): 442-453. 24. R.J. Barnes, M.S. Dhanoa, S.J. Lister. ‘‘Standard Normal Variate Transformation and De-Trending of Near-Infrared Diffuse Reflectance Spectra’’. Appl. Spectrosc. 1989. 43(5): 772-777. 25. R.A. Viscarra Rossel. ‘‘ParLeS: Software for Chemometric Analysis of Spectroscopic Data’’. Chemom. Intell. Lab. Syst. 2008. 90(1): 7283. 26. T. Udelhoven, C. Emmerling, T. Jarmer. ‘‘Quantitative Analysis of Soil Chemical Properties with Diffuse Reflectance Spectrometry and Partial Least-Square Regression: A Feasibility Study’’. Plant Soil. 2003. 251(2): 319-329. 27. H. Akaike. ‘‘Fitting Autoregressive Models for Prediction’’. Ann. Inst. Stat. Math. 1969. 21(1): 243-247. 28. J. Hill, T. Udelhoven, M. Vohland, A. Stevens. ‘‘The Use of Laboratory Spectroscopy and Optical Remote Sensing for Estimating Soil Properties’’. In: E.-C. Oerke, R. Gerhards, G. Menz, R.A. Sikora, editors. Precision Crop Protection - the Challenge and Use of Heterogeneity. Berlin, Germany: Springer, 2010. Pp. 67-85. 29. P.C. Williams. ‘‘Implementation of Near-Infrared Technology’’. In: P. Williams, K.H. Norris, editors. Near-Infrared Technology in the Agricultural and Food Industries. Minnesota, USA: American Association of Cereal Chemists, Inc, 2001. 2nd ed. Pp. 145–169. 30. D.F. Malley, P.C. Williams. ‘‘Use of Near-Infrared Reflectance Spectroscopy in Prediction of Heavy Metals in Freshwater

722

Volume 68, Number 7, 2014

31.

32.

33.

34.

35.

36.

37. 38.

39.

40.

Sediment by Their Association with Organic Matter’’. Environ. Sci. Technol. 1997. 31(12): 3461-3467. V. Bellon-Maurel, E. Ferna´ndez Ahumada, B. Palagos, J.-M. Roger. ‘‘Critical Review of Chemometric Indicators Commonly Used for Assessing the Quality of the Prediction of Soil Attributes by NIR Spectroscopy’’. Trends Anal. Chem. 2010. 29(9): 1073-1088. B.H. Kusumo, M.J. Hedley, C.B. Hedley, M.P. Tuohy. ‘‘Measuring Carbon Dynamics in Field Soils Using Soil Spectral Reflectance: Prediction of Maize Root Density, Soil Organic Carbon and Nitrogen Content’’. Plant Soil. 2011. 338(1-2): 233-245. B. Kuang, A.M. Mouazen. ‘‘Calibration of Visible and Near Infrared Spectroscopy for Soil Analysis at the Field Scale on Three European Farms’’. Eur. J. Soil Sci. 2011. 62(4): 629-636. C.L.S. Morgan, T.H. Waiser, D.J. Brown, C.T. Hallmark. ‘‘Simulated In Situ Characterization of Soil Organic and Inorganic Carbon with Visible Near-Infrared Diffuse Reflectance Spectroscopy’’. Geoderma. 2009. 151(3-4): 249-256. ˜ J.D. Munoz, A. Kravchenko. ‘‘Soil Carbon Mapping Using On-the-Go Near Infrared Spectroscopy, Topography and Aerial Photographs’’. Geoderma. 2011. 166(1): 102-110. Y.F. Zhai, L.J. Cui, X. Zhou, Y. Gao, T. Fei, W.X. Gao. ‘‘Estimation of Nitrogen, Phosphorus, and Potassium Contents in the Leaves of Different Plants Using Laboratory-Based Visible and Near-Infrared Reflectance Spectroscopy: Comparison of Partial Least-Square Regression and Support Vector Machine Regression Methods’’. Int. J. Remote Sens. 2013. 34(7): 2502-2018. F. Tsai, W. Philpot. ‘‘Derivative Analysis of Hyperspectral Data’’. Remote Sens. Environ. 1998. 66(1): 41-51. D.M. Haaland, E.V. Thomas. ‘‘Partial Least-Squares Methods for Spectral Analyses. 1. Relation to Other Quantitative Calibration Methods and the Extraction of Qualitative Information’’. Anal. Chem. 1988. 60(11): 1193-1202. A. Rinnan, F. Van den Berg, S.B. Engelsen. ‘‘Review of the Most Common Pre-Processing Techniques for Near-Infrared Spectra’’. Trends Anal. Chem. 2009. 28(10): 1201-1222. R.N. Clark, T.L. Roush. ‘‘Reflectance Spectroscopy: Quantitative Analysis Techniques for Remote Sensing Applications’’. J. Geophys. Res. Solid Earth. 1984. 89(B7): 6329-6340.

Estimating soil organic carbon content with visible-near-infrared (vis-NIR) spectroscopy.

The selection of a calibration method is one of the main factors influencing measurement accuracy with visible-near-infrared (Vis-NIR, 350-2500 nm) sp...
457KB Sizes 1 Downloads 3 Views