**3. Results and discussion**

#### **3.1 Laboratory analyses**

Means and distributions of the reference values of total nitrogen (TN), total carbon (TC), organic carbon (OC), inorganic carbon (IC), the ratio of TC to TN, and the ratio of OC to TN in samples are summarized in Table 2. The averaged content (±s.d.) for TN, TC and OC are 0.2(±0.06)%, 2.11(±0.57)% and 1.98(±0.54)%, respectively. The IC concentration for the studied samples is very low with an average value of 0.12%, although its coefficient of variance (c.v.) is 1.08. This leads to large skewness of IC content distribution in samples. The ratio of OC to TN is as nearly constant as 10 with very low c.v. value of ~0.05. The intercorrelation coefficients among these properties are summarized in Table 3. TN, TC and OC were strongly correlated to each other (*r*=0.97~0.99), while they have weak correlation with IC (*r*=0.13~0.35). The TC/TN or OC/TN was poorly correlated to TN, TC and OC, however, they had good correlation with IC (*r*=0.53 or -0.49).


a one sample outlier (A02) detected by PCA was not included.

b standard deviation

c coefficient of variance(=s.d./Mean)

Table 2. Laboratory reference statistics for soil total nitrogen (TN), total carbon (TC), organic carbon (OC), inorganic carbon (IC), TC/TN and OC/TN


Table 3. Correlation matrix between soil properties and the first two principal component scores of the Vis-NIR spectra, ATR-FTIR spectra and DRIFT spectra of 121 samples

#### **3.2 Vis-NIR spectral analysis and model calibrations**

192 Infrared Spectroscopy – Life and Biomedical Sciences

Means and distributions of the reference values of total nitrogen (TN), total carbon (TC), organic carbon (OC), inorganic carbon (IC), the ratio of TC to TN, and the ratio of OC to TN in samples are summarized in Table 2. The averaged content (±s.d.) for TN, TC and OC are 0.2(±0.06)%, 2.11(±0.57)% and 1.98(±0.54)%, respectively. The IC concentration for the studied samples is very low with an average value of 0.12%, although its coefficient of variance (c.v.) is 1.08. This leads to large skewness of IC content distribution in samples. The ratio of OC to TN is as nearly constant as 10 with very low c.v. value of ~0.05. The intercorrelation coefficients among these properties are summarized in Table 3. TN, TC and OC were strongly correlated to each other (*r*=0.97~0.99), while they have weak correlation with IC (*r*=0.13~0.35). The TC/TN or OC/TN was poorly correlated to TN, TC and OC, however,

Reference valuea

Mean Median Range s.d.b c.v.<sup>c</sup>

TN(%) 0.20 0.19 0.09-0.31 0.06 0.30 TC(%) 2.11 2.00 0.95-3.41 0.57 0.27 OC(%) 1.98 1.84 0.85-3.02 0.54 0.27 IC(%) 0.12 0.09 0.00-0.64 0.13 1.08 TC/TN 10.57 10.47 9.56-13.12 0.59 0.06 OC/TN 9.95 9.92 8.88-12.47 0.52 0.05

Table 2. Laboratory reference statistics for soil total nitrogen (TN), total carbon (TC), organic

All TN(%) TC(%) OC(%) IC(%) TC/TN OC/TN TN(%) 1 0.97 0.99 0.23 -0.25 -0.24 TC(%) 0.97 1 0.97 0.35 -0.07 -0.18 OC(%) 0.99 0.97 1 0.13 -0.19 -0.07 IC(%) 0.23 0.35 0.13 1 0.53 -0.49 TC/TN -0.25 -0.07 -0.19 0.53 1 0.41 OC/TN -0.24 -0.18 -0.07 -0.49 0.41 1 PC-1(Vis-NIR) 0.77 0.74 0.76 0.13 -0.22 -0.16 PC-2(Vis-NIR) -0.48 -0.48 -0.46 -0.23 0.08 0.22 PC-1(ATR-FTIR) 0.90 0.87 0.88 0.19 -0.29 -0.28 PC-2(ATR-FTIR) 0.15 0.16 0.21 -0.17 0.05 0.30 PC-1(DRIFT) 0.56 0.55 0.52 0.26 -0.12 -0.24 PC-2(DRIFT) -0.71 -0.68 -0.71 -0.05 0.24 0.16 Table 3. Correlation matrix between soil properties and the first two principal component scores of the Vis-NIR spectra, ATR-FTIR spectra and DRIFT spectra of 121 samples

**3. Results and discussion** 

they had good correlation with IC (*r*=0.53 or -0.49).

a one sample outlier (A02) detected by PCA was not included.

carbon (OC), inorganic carbon (IC), TC/TN and OC/TN

**3.1 Laboratory analyses** 

Soil property

b standard deviation

c coefficient of variance(=s.d./Mean)

Different wavelength bands respond to different chemical compositions or molecular groups in soil. However, this response is strongly influenced by soil texture classes. Figure 4 shows the representative Vis-NIR absorption spectra of samples from each soil texture class, e.g. sandy loam (field #D), clay loam (field #B) and clay (fields #C and #E), with high and low TC content. The shift in overall baseline in the Vis-NIR spectra is likely caused by the overall difference in the particle size distribution (Madari*, et al.*, 2006). The clay soil has a finer texture compared to the others, which results in a higher baseline. In general, smaller particle size results in higher reflectance or lower absorbance, but for our case, the higher absorption coefficients for the clay fraction apparently dominate the particle size effect resulting in higher absorbance. Within the clay texture class, the samples with higher TC content tend to exhibit stronger absorption in Vis-NIR spectra than those with lower TC contents (Fig.4). This observation seems true for sandy loam soils but not for clay loam soils, which might be attributed to the effect of soil colour. In Fig.4, the sample from clay loam class with low TC content of 1.59% shows higher absorbance than that with high TC content of 2.44%. This is probably due to particle size effect. The Vis-NIR spectra are characteristic of absorption bands associated with colour (400-760 nm), the bending (1413 nm) and stretching (1916 nm) of the O-H bonds of free water and lattice minerals at around 2210 nm (Madari*, et al.*, 2006).

Fig. 4. Vis-NIR absorption spectra of samples from each soil texture class with high and low TC content.

#### **3.2.1 PCA analysis for Vis-NIR spectra**

Figure 5 shows all raw Vis-NIR spectra, PCA scores plot for the spectra, and residual Xvariance for PC-1 and PC-2. PC-1 and PC-2 explains 96% and 3% of total variance, respectively. The PC-1 may explain variation related to SOM of the samples, as the PC-1 was better correlated to TN, TC and OC (*r*=0.74~0.76) than to PC-2 (*r* =-0.46~-0.48) (Table 3). Samples originated from different fields can be divided into two clusters: one for Luvisol soils (Showground Field, #D) and another for Cambisol soils (Orchard, #B; Ivy Ground, #C;

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 195

None 5 0.90 3.33 5 0.86 2.73 5 0.90 3.16 **5 0.42 1.34**  SNV **3 0.91 3.75** 3 0.87 2.74 **4 0.91 3.44** 3 0.37 1.29 MSC 4 0.90 3.33 4 0.87 2.74 4 0.90 3.20 3 0.39 1.31 BOC 4 0.90 3.33 4 0.86 2.70 5 0.91 3.27 5 0.37 1.29 Center & Scale 5 0.91 3.53 **5 0.87 2.78** 5 0.90 3.21 6 0.42 1.34

Maximum- 4 0.89 3.33 4 0.86 2.69 4 0.89 3.05 5 0.38 1.29 Range- 5 0.90 3.33 5 0.87 2.74 5 0.90 3.23 4 0.34 1.26 Mean- 4 0.89 3.16 4 0.85 2.61 4 0.89 3.00 4 0.39 1.31 Quantile- 5 0.91 3.53 5 0.87 2.78 5 0.91 3.25 6 0.42 1.34

1st- 4 0.90 3.53 4 0.87 2.75 4 0.90 3.20 4 0.38 1.30 2nd- 3 0.89 3.16 3 0.85 2.57 4 0.89 3.05 3 0.39 1.31

1st- 1 0.85 2.73 1 0.80 2.26 1 0.85 2.63 3 0.17 1.12 2nd- 4 0.61 1.71 1 0.56 1.51 4 0.63 1.65 1 0.14 1.11

Table 4. Cross validation result of PLSR models calibrated for raw and various transformed

algorithms, model performance was improved for a certain degree. For examples, PLSR model developed for TN after SNV-transformed spectra resulted in *R*2 of 0.91 and RPD of 3.75. By the same pre-processing technique, prediction of soil OC was improved with *R*2 of 0.91 and RPD of 3.44. The best calibration model for TC was obtained when the spectra were transformed by Center-&-Scale technique. It is worth noting that these optimized PLSR models need less latent variables (3-5) than those for raw spectra. In general, the fewer the latent variables used, the better is the model developed, as the calibration is more apt to be applicable to new samples (Madari*, et al.*, 2006). For IC, PLSR calibration with raw and various transformed spectra failed to produce useful models with *R*2≤0.42 and RPD≤1.34. Figure 6 shows the correlation between the measured and PLSR-predicted values of each soil property. The linear fitting slopes for TN, TC and OC are close to 1, which suggests that PLSR models developed with Vis-NIR absorption spectra for predicting soil TN, TC and OC

B-coefficients curves of the PLSR models calibrated for the best transformed Vis-NIR spectra against each soil property are shown in Fig.7. The B-coefficients for TN, TC and OC exhibit strong similarity among them. This is mainly due to their high correlation obtained for

**3.2.3 B-coefficients analysis of PLSR models for Vis-NIR spectra** 

PLSR models calibrated for Vis-NIR spectra

Total N Total C Organic C Inorganic C

LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD

Spectral pretreatment method

Normalization

De-trending

Derivative

are successful.

Vis-NIR spectra with 121 samples

Fig. 5. Vis-NIR absorption spectra of all samples (a), principal components analysis scores plot for the Vis-NIR spectra (b), and residual X-variance for PC-1 (c) and PC-2 (d).

Copse Field, #E). Samples from Avenue Field (#A) exhibit partially mixing with the two clusters (Fig.5b). This might be attributed to the nature of this field as it is a mixture of Cambisol and Luvisol types. Although being of the same soil type of Cambisol, samples from field #B are clearly separated from those from fields #C and #E (Fig.5b). This is mainly due to different soil textures, namely, clay loam vs. clay (Table 1). Several samples located outside of the Hotelling T2 ellipse become the candidates of samples outliers (Fig.5b). However, apart from sample A02, other outliers locate close to their member samples. In addition, sample A02 exhibits large residual X-variance for PC-1 (Fig.5c) and PC-2 (Fig.5d). The raw spectrum of sample A02 also shows distinct color features (lowest absorbance in visible range) from other samples (Fig.5a). Thus, sample A02 was considered as an outlier and excluded from further investigation.

#### **3.2.2 Vis-NIR calibration models**

Table 4 summarizes the cross-validation results of the PLSR models developed with raw and various transformed Vis-NIR spectra against each soil property. For raw spectra, PLSR models produced good or excellent prediction accuracy with *R*2 of 0.86~0.90 and RPD of 2.73~3.33 for soil TN, TC and OC. Coupled with appropriate spectral pre-processing


(a) (b)

(c) (d)

and excluded from further investigation.

**3.2.2 Vis-NIR calibration models** 

Fig. 5. Vis-NIR absorption spectra of all samples (a), principal components analysis scores

Copse Field, #E). Samples from Avenue Field (#A) exhibit partially mixing with the two clusters (Fig.5b). This might be attributed to the nature of this field as it is a mixture of Cambisol and Luvisol types. Although being of the same soil type of Cambisol, samples from field #B are clearly separated from those from fields #C and #E (Fig.5b). This is mainly due to different soil textures, namely, clay loam vs. clay (Table 1). Several samples located outside of the Hotelling T2 ellipse become the candidates of samples outliers (Fig.5b). However, apart from sample A02, other outliers locate close to their member samples. In addition, sample A02 exhibits large residual X-variance for PC-1 (Fig.5c) and PC-2 (Fig.5d). The raw spectrum of sample A02 also shows distinct color features (lowest absorbance in visible range) from other samples (Fig.5a). Thus, sample A02 was considered as an outlier

Table 4 summarizes the cross-validation results of the PLSR models developed with raw and various transformed Vis-NIR spectra against each soil property. For raw spectra, PLSR models produced good or excellent prediction accuracy with *R*2 of 0.86~0.90 and RPD of 2.73~3.33 for soil TN, TC and OC. Coupled with appropriate spectral pre-processing

plot for the Vis-NIR spectra (b), and residual X-variance for PC-1 (c) and PC-2 (d).

Table 4. Cross validation result of PLSR models calibrated for raw and various transformed Vis-NIR spectra with 121 samples

algorithms, model performance was improved for a certain degree. For examples, PLSR model developed for TN after SNV-transformed spectra resulted in *R*2 of 0.91 and RPD of 3.75. By the same pre-processing technique, prediction of soil OC was improved with *R*2 of 0.91 and RPD of 3.44. The best calibration model for TC was obtained when the spectra were transformed by Center-&-Scale technique. It is worth noting that these optimized PLSR models need less latent variables (3-5) than those for raw spectra. In general, the fewer the latent variables used, the better is the model developed, as the calibration is more apt to be applicable to new samples (Madari*, et al.*, 2006). For IC, PLSR calibration with raw and various transformed spectra failed to produce useful models with *R*2≤0.42 and RPD≤1.34. Figure 6 shows the correlation between the measured and PLSR-predicted values of each soil property. The linear fitting slopes for TN, TC and OC are close to 1, which suggests that PLSR models developed with Vis-NIR absorption spectra for predicting soil TN, TC and OC are successful.

#### **3.2.3 B-coefficients analysis of PLSR models for Vis-NIR spectra**

B-coefficients curves of the PLSR models calibrated for the best transformed Vis-NIR spectra against each soil property are shown in Fig.7. The B-coefficients for TN, TC and OC exhibit strong similarity among them. This is mainly due to their high correlation obtained for

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 197

Fig. 7. B-coefficients curves obtained from PLSR analysis with the SNV-transformed Vis-NIR spectra for TN and OC, Center-&-Scale-transformed Vis-NIR spectra for TC, and raw Vis-

mineralogy of the data set originated from the same parent material in the current study, as compared to those using larger scale data sets (e.g. Mouazen*, et al.*, 2006; Vohland & Emmerling, 2011). Although the prediction accuracy of IC did not satisfy with lowest quantification standard (RPD>2.0), the characteristic bands of carbonate in 2300 and 2500 nm (Gaffey, 1986; Reeves III*, et al.*, 2002; Viscarra Rossel & Behrens, 2010) can be clearly

The procedure of analyzing DRIFT spectra was the same as that presented for Vis-NIR

Figure 8 shows the representative DRIFT absorption spectra of samples from each soil texture class, e.g. sandy loam (field #D), clay loam (field #B) and clay (fields #C and #E), with high and low TC content. Obviously, the absolute magnitude and range of the DRIFT absorptions are much greater than those found for corresponding Vis-NIR spectrum (Fig.4). The DRIFT spectrum generally shows more distinctive spectral features than the Vis-NIR spectrum. The peaks between 3698 and 3620 cm-1 and various peaks below 1100 cm-1 are characteristic for kaolinite (Madari*, et al.*, 2006). Soil organic matter also has characteristic absorption bands in the Mid-IR range (Table 5), however, due to its low concentration in the soil samples, and their overlapping with mineral peaks, most of these could not readily be identified by simple visual analysis of spectra. For example, the small bands around 2940- 2935 cm-1 and 2886-2877 cm-1 indicate the presence of organic aliphatic C-H stretching (Table 5). This band is more evident in the spectra of soil samples having higher concentrations of organic carbon (Madari*, et al.*, 2006). Other organic peaks overlap with the mineral peaks. Viscarra Rossel *et al*. (2006) has compared the usefulness of visible, NIR and Mid-IR diffuse

NIR spectra for IC.

spectral analysis.

identified in its B-coefficients curve.

**3.3.1 DRIFT spectral response** 

**3.3 DRIFT spectral analysis and model calibrations** 

Fig. 6. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on Vis-NIR absorption spectra.

reference values (*r*=0.97-0.99, Table 3). It is worth noting that the visible range (400-760 nm) associated with soil colour shows huge influence on model accuracy, which is in line with other reports (Stenberg*, et al.*, 2010; Viscarra Rossel*, et al.*, 2006). The absorption feature in the visible and short-wave NIR (400-1000 nm) might be due to the Fe oxides in soil, mainly haematite and goethite (Viscarra Rossel & Behrens, 2010). The influential wavelengths located between 1000 and 2500 nm can be attributed to water, clay minerals and organic matter (Viscarra Rossel & Behrens, 2010). Using samples collected from Belgium and Northern France, Mouazen *et al*. (2006) compared the performance of two commercially available spectrophotometers with different wavelength ranges for the measurement of selected soil attributes including TC and TN. They found that the best accuracy was obtained when using a full wavelength range of 451-2459 nm, as compared to a short wavelength range of 401-1770 nm. Using samples collected from two depths along 11 km section of floodplain, Vohland and Emmerling (2011) reported that genetic algorithm (GA) allocated significant wavelengths most frequently to the range of 1970-2490 nm for soil OC, which is not particularly in line with those bands shown in Fig.7. One reason might be the use of farm-scale local set in our case in which the visible range associated with soil colours has the most influence on model calibration, whereas the NIR range has a smaller influence. This contradictory results against those published by others might be explained by the same

Fig. 7. B-coefficients curves obtained from PLSR analysis with the SNV-transformed Vis-NIR spectra for TN and OC, Center-&-Scale-transformed Vis-NIR spectra for TC, and raw Vis-NIR spectra for IC.

mineralogy of the data set originated from the same parent material in the current study, as compared to those using larger scale data sets (e.g. Mouazen*, et al.*, 2006; Vohland & Emmerling, 2011). Although the prediction accuracy of IC did not satisfy with lowest quantification standard (RPD>2.0), the characteristic bands of carbonate in 2300 and 2500 nm (Gaffey, 1986; Reeves III*, et al.*, 2002; Viscarra Rossel & Behrens, 2010) can be clearly identified in its B-coefficients curve.

#### **3.3 DRIFT spectral analysis and model calibrations**

The procedure of analyzing DRIFT spectra was the same as that presented for Vis-NIR spectral analysis.

#### **3.3.1 DRIFT spectral response**

196 Infrared Spectroscopy – Life and Biomedical Sciences

(a) (b)

(c) (d)

NIR absorption spectra.

Fig. 6. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on Vis-

reference values (*r*=0.97-0.99, Table 3). It is worth noting that the visible range (400-760 nm) associated with soil colour shows huge influence on model accuracy, which is in line with other reports (Stenberg*, et al.*, 2010; Viscarra Rossel*, et al.*, 2006). The absorption feature in the visible and short-wave NIR (400-1000 nm) might be due to the Fe oxides in soil, mainly haematite and goethite (Viscarra Rossel & Behrens, 2010). The influential wavelengths located between 1000 and 2500 nm can be attributed to water, clay minerals and organic matter (Viscarra Rossel & Behrens, 2010). Using samples collected from Belgium and Northern France, Mouazen *et al*. (2006) compared the performance of two commercially available spectrophotometers with different wavelength ranges for the measurement of selected soil attributes including TC and TN. They found that the best accuracy was obtained when using a full wavelength range of 451-2459 nm, as compared to a short wavelength range of 401-1770 nm. Using samples collected from two depths along 11 km section of floodplain, Vohland and Emmerling (2011) reported that genetic algorithm (GA) allocated significant wavelengths most frequently to the range of 1970-2490 nm for soil OC, which is not particularly in line with those bands shown in Fig.7. One reason might be the use of farm-scale local set in our case in which the visible range associated with soil colours has the most influence on model calibration, whereas the NIR range has a smaller influence. This contradictory results against those published by others might be explained by the same

Figure 8 shows the representative DRIFT absorption spectra of samples from each soil texture class, e.g. sandy loam (field #D), clay loam (field #B) and clay (fields #C and #E), with high and low TC content. Obviously, the absolute magnitude and range of the DRIFT absorptions are much greater than those found for corresponding Vis-NIR spectrum (Fig.4). The DRIFT spectrum generally shows more distinctive spectral features than the Vis-NIR spectrum. The peaks between 3698 and 3620 cm-1 and various peaks below 1100 cm-1 are characteristic for kaolinite (Madari*, et al.*, 2006). Soil organic matter also has characteristic absorption bands in the Mid-IR range (Table 5), however, due to its low concentration in the soil samples, and their overlapping with mineral peaks, most of these could not readily be identified by simple visual analysis of spectra. For example, the small bands around 2940- 2935 cm-1 and 2886-2877 cm-1 indicate the presence of organic aliphatic C-H stretching (Table 5). This band is more evident in the spectra of soil samples having higher concentrations of organic carbon (Madari*, et al.*, 2006). Other organic peaks overlap with the mineral peaks. Viscarra Rossel *et al*. (2006) has compared the usefulness of visible, NIR and Mid-IR diffuse

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 199

(cm-1) Assignments Vis-NIR wavelength

3300 N-H stretching 1500,1000,751 3030 Aromatic C-H strecthing 1650,1100,825

2930,2850 Alkyl asymmetric-symmetric C-H stretching 1706,1754,1138,

1725-1720 C=O stretching of –COOH and ketones 1930,1449

C=O stretching of amide groups (amide I band), quinine C=O and/or C=O of H-bonded conjugated

1610 N-H stretching of Amine 2060

1460-1450 Aliphatic C-H 2275,1706

H deformation of CH2 and CH3 groups, COO-

1270 C-OH stretching of phenolic OH 1961

1050 C-O stretching of carbohydrates 2381

2006) and corresponding wavelengths in the Vis-NIR range (Stenberg*, et al.*, 2010)

OH deformation and C-O stretching of phenolic OH, C-

like substances <sup>2137</sup>

Table 5. Absorption bands of C and N in organic bonds in the Mid-IR range (Madari*, et al.*,

1620-1600 Aromatic C=C stretching and/or asymmetric –COO

1590-1517 COO- symmetric stretching, N-H deformation+C=N

<sup>1350</sup>Symmetric COO- stretching and/or –CH bending of

1280-1200 C-O stretching and OH deformation of COOH, C-O

1170-950 C-O stretching of polysaccharides or polysaccharide-

<sup>1170</sup>C-OH stretching of aliphatic OH, C-C stretching of

1225 C-O stretching and OH deformation of COOH

stretching (amide II band)

asymmetric stretching

stretching of aryl ethers

aliphatic groups

830 Aromatic CH out of plane bending 775 Aromatic CH out of plane bending (nm)

1170,853,877

2033,1524

Mid-IR band

1660-1630

1400-1390

3380 O-H stretching of phenolic OH

2940-2900 Aliphatic C-H stretching

ketones

stretching

1525 Aromatic C=C strecthing

aliphatics

3400-3300 O-H stretching (H bonded OH groups),

2600 O-H stretching of H-bonded -COOH

Fig. 8. DRIFT absorption spectra of samples from each soil texture class with high and low TC content. Some dominant soil components and absorption peaks are shown for quartz (Q), organic compounds (OC), calcite (Ca), kaolinite (K), and the (OH) features of free water and lattice minerals (Viscarra Rossel*, et al.*, 2006).

reflectance spectroscopy by examining PLSR factor loadings weights. They showed that frequencies in the Mid-IR range corresponding to the absorption of organic compounds, like organic acids, or acidic functional groups, like alkyl, amide and aromatic groups, were indicators of correlation between organic carbon concentrations in the bulk soil samples and the Mid-IR spectra.

#### **3.3.2 PCA analysis for DRIFT spectra**

Figure.9a shows all Mid-IR spectra obtained by FT-IR spectrometer with DRIFT accessory. Figure 9 also shows the PCA scores plot for the DRIFT spectra, and residual X-variance for PC-1 and PC-2. PC-1 and PC-2 explains 83% and 14% of total variance contained in the spectra, respectively. The PC-1 axis may explain differences attributed to SOM content of the samples as the PC-1 was better correlated to TN, TC and OC (*r*=0.87~0.90) than to PC-2 (*r*=0.15~0.21) (Table 3). Samples originated from different fields can be separated into two clusters according to soil types: one for Luvisol soils from Showground Field (#D) and another for Cambisol soils from Orchard (#B), Ivy Ground (#C) and Copse Field (#E),

(c) Fig. 8. DRIFT absorption spectra of samples from each soil texture class with high and low TC content. Some dominant soil components and absorption peaks are shown for quartz (Q), organic compounds (OC), calcite (Ca), kaolinite (K), and the (OH) features of free water

reflectance spectroscopy by examining PLSR factor loadings weights. They showed that frequencies in the Mid-IR range corresponding to the absorption of organic compounds, like organic acids, or acidic functional groups, like alkyl, amide and aromatic groups, were indicators of correlation between organic carbon concentrations in the bulk soil samples and

Figure.9a shows all Mid-IR spectra obtained by FT-IR spectrometer with DRIFT accessory. Figure 9 also shows the PCA scores plot for the DRIFT spectra, and residual X-variance for PC-1 and PC-2. PC-1 and PC-2 explains 83% and 14% of total variance contained in the spectra, respectively. The PC-1 axis may explain differences attributed to SOM content of the samples as the PC-1 was better correlated to TN, TC and OC (*r*=0.87~0.90) than to PC-2 (*r*=0.15~0.21) (Table 3). Samples originated from different fields can be separated into two clusters according to soil types: one for Luvisol soils from Showground Field (#D) and another for Cambisol soils from Orchard (#B), Ivy Ground (#C) and Copse Field (#E),

(a) (b)

and lattice minerals (Viscarra Rossel*, et al.*, 2006).

**3.3.2 PCA analysis for DRIFT spectra** 

the Mid-IR spectra.


Table 5. Absorption bands of C and N in organic bonds in the Mid-IR range (Madari*, et al.*, 2006) and corresponding wavelengths in the Vis-NIR range (Stenberg*, et al.*, 2010)

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 201

None 5 0.95 4.62 5 0.93 3.93 5 0.94 4.19 5 0.70 1.86 SNV 4 0.94 4.29 5 0.94 3.93 4 0.94 4.06 3 0.68 1.82 MSC 4 0.94 4.29 4 0.92 3.63 4 0.94 4.15 3 0.68 1.81 BOC 5 0.95 4.62 **6 0.94 4.10** 5 0.95 4.25 4 0.71 1.88 Center & Scale 5 0.95 4.62 5 0.93 3.88 5 0.94 4.25 **4 0.71 1.89** 

Maximum- 6 0.94 4.29 6 0.93 3.80 6 0.94 4.15 5 0.68 1.80 Range- 6 0.94 4.62 6 0.93 3.85 6 0.94 4.22 5 0.69 1.84 Mean- 5 0.94 4.29 6 0.93 3.85 5 0.94 4.12 5 0.67 1.78 Quantile- **5 0.95 5.00** 5 0.94 4.01 **5 0.95 4.43** 4 0.70 1.88

1st- 4 0.94 4.62 4 0.94 4.04 4 0.95 4.35 4 0.70 1.86 2nd- 5 0.94 4.62 5 0.93 3.83 5 0.95 4.32 3 0.67 1.79

1st- 2 0.83 2.61 2 0.79 2.18 2 0.79 2.21 5 0.30 1.22 2nd- 2 0.69 1.94 2 0.66 1.70 2 0.65 1.71 1 0.10 1.01

Table 6. Cross validation result of PLSR models calibrated for raw and various transformed

For raw spectra, PLSR models produced excellent prediction accuracy with *R*2 of 0.93~0.95 and RPD of 3.93~4.62 for TN, TC and OC. These models were calibrated with 5 latent variables. Coupled with proper spectral pre-processing techniques, model prediction was improved. For examples, the best model for soil TN produced prediction accuracy with *R*2 of 0.95 and RPD of 5.00 using spectral pre-processing of quantile normalization. It was also effective for OC model improvement with *R*2 of 0.95 and RPD of 4.43. For TC, the best model was built using spectral transformation of BOC, although this model needed 6 latent variables. Figure 10 shows the correlation between the measured and PLSR-predicted values of each soil property. The linear fitting lines for TN, TC and OC are closer to 1:1 as compared to the corresponding plots for Vis-NIR spectra (Fig.6). DFIRT spectra outperformed Vis-NIR spectroscopy for the prediction of these three properties (Table 4 & 6). Even for IC, almost all PLSR models developed with raw and various transformed DRIFT spectra outperformed their counterparts for Vis-NIR spectra. Accuracy obtained for IC can be used to distinguish high and low content with RPD>1.5 (Table 6). These results suggest that PLSR models developed with DRIFT absorption

B-coefficients curves obtained after PLSR analyses with the best transformed DRIFT spectra for each soil property are combined in Fig.11. The B-coefficients for TN, TC and OC exhibit

PLSR models calibrated for DRIFT spectra

Total N Total C Organic C Inorganic C

LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD

Spectral pretreatment method

Normalization

De-trending

Derivative

DRIFT spectra with 121 samples

spectra for predicting these soil properties are more accurate.

**3.3.4 B-coefficients analysis of PLSR models for DRIFT spectra** 

Fig. 9. DRIFT absorption spectra of all samples (a), principal components analysis scores plot for the DRIFT spectra (b), residual X-variance for PC-1 (c) and PC-2 (d).

although several samples from the field #B were mixed with former cluster. Samples from the field #A completely mixed together with the former cluster, although some of them belong to latter cluster. It seems that the soil type has no effect on samples separation into different classes but the texture diversity within the sample population used for calibration has clear effect (Madari*, et al.*, 2006). Samples from the fields #C and #E are totally separated although they are of same clay texture. This is mainly due to their distinct SOM-related soil properties concentrations as different vegetations have been growing in both fields (soybean in the field #C with TN of 0.28±0.02%, TC of 2.94±0.23% and OC of 2.73±0.15%; wheat in the field #E with TN of 0.25±0.02%, TC of 2.62±0.15% and OC of 2.56±0.16%). No sample locates outside of the Hotelling T2 ellipse. Although sample A02 exhibits low residual X-variance for PC-1 (Fig.9c) which accounts for the most amount of variance in the DRIFT spectra, it displays large residual X-variance for PC-2 (Fig.9d). Also, in the raw spectrum, sample A02 shows quite different from the others (Fig.9a). Thus, sample A02 was considered as a sample outlier and excluded from further investigation.

#### **3.3.3 DRIFT calibration models**

Table 6 summarizes the cross-validation results of the PLSR models developed with raw and various transformed DRIFT spectra against each soil property, e.g. TN, TC, OC and IC.


(a) (b)

A02

A11

A03A04

DRIFT

(c) (d)

outlier and excluded from further investigation.

Leverage (PC-1) 0 0.01 0.02 0.03 0.04 0.05

A08 A09 A10

B01 B02

A14 A15 A17 A16 A18 A19 A20

D35

C21

C28 C29 C30 C31 C32 C33 C34 C35 C36 C37 C38 C39

A01

D25D27D26 D24D23D28 D30 D31 D29 D32D33 D34

A07

B22

C22 C23 C24 C25 C26 C27

B03 B04 B05 B06

B07

D11 D12 D13 D14 D15D16 D17 D18 D19 D20

D03

D05

E17

D06 D07 D08 D09

B09 B0811 B10B12 B13 B15 B14 B16 B17 B18 B19 B20 B21

E06 E07 E08 E09 E10 E11 E12 E13

> B23 B24 B25 C15

A12

E15 E16

E18

D10

D04

D21 D22

D01 D02

E19

E14

E01 E02 E04E03 E05

E20 E21 E22 E23

0

0.01

0.02

**3.3.3 DRIFT calibration models** 

plot for the DRIFT spectra (b), residual X-variance for PC-1 (c) and PC-2 (d).

A05 A06

Fig. 9. DRIFT absorption spectra of all samples (a), principal components analysis scores

although several samples from the field #B were mixed with former cluster. Samples from the field #A completely mixed together with the former cluster, although some of them belong to latter cluster. It seems that the soil type has no effect on samples separation into different classes but the texture diversity within the sample population used for calibration has clear effect (Madari*, et al.*, 2006). Samples from the fields #C and #E are totally separated although they are of same clay texture. This is mainly due to their distinct SOM-related soil properties concentrations as different vegetations have been growing in both fields (soybean in the field #C with TN of 0.28±0.02%, TC of 2.94±0.23% and OC of 2.73±0.15%; wheat in the field #E with TN of 0.25±0.02%, TC of 2.62±0.15% and OC of 2.56±0.16%). No sample locates outside of the Hotelling T2 ellipse. Although sample A02 exhibits low residual X-variance for PC-1 (Fig.9c) which accounts for the most amount of variance in the DRIFT spectra, it displays large residual X-variance for PC-2 (Fig.9d). Also, in the raw spectrum, sample A02 shows quite different from the others (Fig.9a). Thus, sample A02 was considered as a sample

0 0.001 0.002 0.003 0.004 0.005 0.006

B15

B16 B17 B18 B19 B20

D10

D26 D27

B21 B22 B23B24 B25C15 C22 C21 C23C24 C25 C26 C28 C27

D30 D31D32

D16

D34

0.007 DRIFT

A01

E11 E12

D21 D22 D23 D24D25

A07

Leverage (PC-2) 0.01 0.02 0.03 0.04 0.05

A03

D06 D08 D07

C34 C35 C37 C36 C38 C39

D28 D29

E20 E21 E22 E23

A20 B01 B02 B03 B04 B05 B06 B07 B08 B09B10 B11B12 B13 B14

E13E16 E15 E14

D11 D13 D12 D15 D14

C30 C31 C32 C33

D33

E04 E05 E06 E07 E09 E08

E18

D09

A08 A09 A10 A11 A12 A14 A15 A16 A17 A18

D17 D18 D19

D35 E01 E02 E03

E10

A02

A04 A05 A06

E17

D20

C29

E19

A19

D02 D01 D03 D04 D05

Table 6 summarizes the cross-validation results of the PLSR models developed with raw and various transformed DRIFT spectra against each soil property, e.g. TN, TC, OC and IC. Table 6. Cross validation result of PLSR models calibrated for raw and various transformed DRIFT spectra with 121 samples

For raw spectra, PLSR models produced excellent prediction accuracy with *R*2 of 0.93~0.95 and RPD of 3.93~4.62 for TN, TC and OC. These models were calibrated with 5 latent variables. Coupled with proper spectral pre-processing techniques, model prediction was improved. For examples, the best model for soil TN produced prediction accuracy with *R*2 of 0.95 and RPD of 5.00 using spectral pre-processing of quantile normalization. It was also effective for OC model improvement with *R*2 of 0.95 and RPD of 4.43. For TC, the best model was built using spectral transformation of BOC, although this model needed 6 latent variables. Figure 10 shows the correlation between the measured and PLSR-predicted values of each soil property. The linear fitting lines for TN, TC and OC are closer to 1:1 as compared to the corresponding plots for Vis-NIR spectra (Fig.6). DFIRT spectra outperformed Vis-NIR spectroscopy for the prediction of these three properties (Table 4 & 6). Even for IC, almost all PLSR models developed with raw and various transformed DRIFT spectra outperformed their counterparts for Vis-NIR spectra. Accuracy obtained for IC can be used to distinguish high and low content with RPD>1.5 (Table 6). These results suggest that PLSR models developed with DRIFT absorption spectra for predicting these soil properties are more accurate.

#### **3.3.4 B-coefficients analysis of PLSR models for DRIFT spectra**

B-coefficients curves obtained after PLSR analyses with the best transformed DRIFT spectra for each soil property are combined in Fig.11. The B-coefficients for TN, TC and OC exhibit

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 203

Fig. 11. B-coefficients curves obtained from PLSR analyses with the transformed DRIFT spectra by quantile normalization for TN and OC, baseline offset correction for TC, and

Figure12 shows the representative ATR-FTIR spectra of samples from each soil texture class with high and low TC content. Obviously, the absolute magnitude and range of the ATR-FTIR light intensity are much lower than those for corresponding DRIFT spectra. Besides, the ATR-FTIR spectra of the samples present quite different shape together with absence of many characteristic peaks between 3000 and 1700 cm-1, compared to the corresponding DRIFT spectra of these samples. The peaks between 3696 and 3620 cm-1 and various peaks

Fig. 12. ATR-FTIR spectra of samples from each soil texture class with high and low TC

Center & Scale for IC.

content.

**3.4.1 ATR-FTIR spectral response** 

Fig. 10. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on DRIFT absorption spectra.

strong similarity among them. This is mainly due to their high correlation obtained with reference values (*r*=0.97-0.99, Table 3). Overall, the most influential frequencies for predicting these soil properties locate in the wavenumber range from 2100 to 1200 cm-1. As indicated in Table 5, the peaks at around 1620-1558 cm-1 are for aromatic C=C stretching and/or asymmetric –COO stretching; the peaks at 1460 cm-1 and 1229 cm-1 are corresponding to Aliphatic C-H stretching and C-O stretching/OH deformation of COOH. Other significant wavebands can be found at round 2930 cm-1, corresponding to Aliphatic C-H stretching. Interestingly, although the DRIFT-calibrated models are not accurate enough for IC quantification, the corresponding B-coefficients curve exhibits several distinctive frequencies for model calibration. The most influential frequency locates at around 1474 cm-1 with two subordinate ones at 1795 cm-1 and 2513 cm-1, which are characteristic of existence of calcium carbonate (Reeves III*, et al.*, 2002). Viscarra Rossel and Behrens (2010) found the carbonate of clay minerals in the NIR band of 2336 nm to be associated with the third overtones of CO� ��in MIR of 1415 cm-1.

#### **3.4 ATR-FTIR spectral analysis and model calibrations**

The procedure of analyzing ATR-FTIR spectra was the same as that presented for Vis-NIR and DRIFT spectral analysis.

Fig. 11. B-coefficients curves obtained from PLSR analyses with the transformed DRIFT spectra by quantile normalization for TN and OC, baseline offset correction for TC, and Center & Scale for IC.

#### **3.4.1 ATR-FTIR spectral response**

202 Infrared Spectroscopy – Life and Biomedical Sciences

(a) (b)

(c) (d)

DRIFT absorption spectra.

overtones of CO�

and DRIFT spectral analysis.

��in MIR of 1415 cm-1.

**3.4 ATR-FTIR spectral analysis and model calibrations** 

Fig. 10. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on

strong similarity among them. This is mainly due to their high correlation obtained with reference values (*r*=0.97-0.99, Table 3). Overall, the most influential frequencies for predicting these soil properties locate in the wavenumber range from 2100 to 1200 cm-1. As indicated in Table 5, the peaks at around 1620-1558 cm-1 are for aromatic C=C stretching and/or asymmetric –COO stretching; the peaks at 1460 cm-1 and 1229 cm-1 are corresponding to Aliphatic C-H stretching and C-O stretching/OH deformation of COOH. Other significant wavebands can be found at round 2930 cm-1, corresponding to Aliphatic C-H stretching. Interestingly, although the DRIFT-calibrated models are not accurate enough for IC quantification, the corresponding B-coefficients curve exhibits several distinctive frequencies for model calibration. The most influential frequency locates at around 1474 cm-1 with two subordinate ones at 1795 cm-1 and 2513 cm-1, which are characteristic of existence of calcium carbonate (Reeves III*, et al.*, 2002). Viscarra Rossel and Behrens (2010) found the carbonate of clay minerals in the NIR band of 2336 nm to be associated with the third

The procedure of analyzing ATR-FTIR spectra was the same as that presented for Vis-NIR

Figure12 shows the representative ATR-FTIR spectra of samples from each soil texture class with high and low TC content. Obviously, the absolute magnitude and range of the ATR-FTIR light intensity are much lower than those for corresponding DRIFT spectra. Besides, the ATR-FTIR spectra of the samples present quite different shape together with absence of many characteristic peaks between 3000 and 1700 cm-1, compared to the corresponding DRIFT spectra of these samples. The peaks between 3696 and 3620 cm-1 and various peaks

Fig. 12. ATR-FTIR spectra of samples from each soil texture class with high and low TC content.

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 205

shows variation related to SOM of the samples as the PC-2 was better correlated to TN, TC and OC (*r*=-0.68~0.71) than to PC-1 (*r*=0.52~0.56) (Table 3). Samples originated from different fields can be divided into one cluster for Luvisol type soils (field #D) and another for Cambisol type soils (fields #B, #C, and #E). Although samples from the field #A are the mixture of Cambisol and Luvisol types, most of them locate within the former cluster. Samples from the fields #C and #E are hardly separated, although they are of the same soil texture of clay (Table 1). Compared to the samples D28, B04 and B10 located outside of the Hotelling T2 ellipse (Fig.13b), sample A02 exhibits large residual X-variance for PC-1 (Fig.13c) and PC-2 (Fig.13d). Thus, sample A02 was considered as a sample outlier and

Table 7 summarizes the cross-validation results of PLSR models developed with the raw and various transformed ATR-FTIR spectra against each soil property. For raw spectra, PLSR models produced excellent prediction accuracy for TN, TC and OC with *R*2 of 0.89~0.92 and RPD of 3.02~3.75. These models were developed with 6 latent variables. None of spectral pre-processing techniques adopted can effectively improve the prediction accuracy of these models. This might be due to absence of the baseline offset of spectra (Fig.12). For IC, all models can be used to distinguish high and low concentration with RPD

None **6 0.92 3.75 6 0.89 3.02 6 0.89 3.10** 6 0.71 1.91 SNV 5 0.90 3.33 5 0.88 2.89 3 0.87 2.74 4 0.69 1.82 MSC 5 0.90 3.33 4 0.87 2.75 3 0.87 2.74 4 0.69 1.82 BOC 7 0.91 3.53 7 0.88 2.95 7 0.89 3.07 4 0.70 1.87 Center & Scale 6 0.91 3.53 6 0.89 2.97 6 0.89 2.97 4 0.70 1.88

Maximum- 5 0.90 3.33 5 0.87 2.79 5 0.87 2.78 5 0.70 1.85 Range- 5 0.90 3.53 4 0.86 2.71 6 0.88 2.93 5 0.70 1.85 Mean- 6 0.91 3.53 4 0.86 2.69 5 0.88 2.87 5 0.70 1.87 Quantile- 7 0.91 3.75 6 0.88 2.89 6 0.89 3.00 4 0.70 1.87

1st- 6 0.90 3.53 6 0.87 2.79 7 0.89 3.09 **5 0.72 1.93**  2nd- 6 0.91 3.53 6 0.87 2.79 5 0.88 2.92 5 0.72 1.91

1st- 4 0.88 3.16 4 0.85 2.60 4 0.86 2.69 5 0.71 1.91 2nd- 3 0.78 2.31 3 0.75 1.99 2 0.74 1.98 7 0.60 1.61

Normalization

De-trending

Table 7. Cross validation result of PLSR models developed with raw and various

transformed ATR-FTIR spectra with 121 samples

PLSR models calibrated for ATR-FTIR spectra

Total N Total C Organic C Inorganic C

LVs *R*<sup>2</sup> RPD LVs *R*<sup>2</sup> RPD LVs *R*<sup>2</sup> RPD LVs *R*2 RPD

excluded from further investigation.

**3.4.3 ATR-FTIR calibration models** 

Spectral pretreatment method

Derivative

below 1100 cm-1 might be characteristic for kaolinite (Madari*, et al.*, 2006). However, most characteristic wavebands of soil organic matter presented in Table 5, due to their low concentrations in the soil samples and their overlapping with mineral peaks, could not readily be identified by simple visual observation. For example, the bands around 2940-2935 cm-1 and 2886-2877 cm-1, indicating the presence of organic aliphatic C-H stretching (shown in DRIFT spectra, Fig.8), are nearly vanished. Although the ATR-FTIR spectra present severe deviations from corresponding DRIFT spectra, the light intensity in ATR-FTIR spectra exhibits strong correlation with particle size distribution. For example, samples of clay texture with fine particles present strongest light intensity, whereas samples with sandy loam texture with coarse particle size have lowest light intensity. The clay loam texture samples with middle size particles correspond to intermediate light absorption. For each soil texture class, the spectral difference due to high and low TC content is not apparent. All these spectra appear to be free from baseline offset.

#### **3.4.2 PCA analysis for ATR-FTIR spectra**

Figure 13 shows all Mid-IR spectra obtained by FI-IR spectrometer with ATR accessory together with the PCA scores plot for the spectra and residual X-variance for PC-1 and PC-2. PC-1 and PC-2 explained 98% and 1% of total variance of the spectra, respectively. The PC-2

Fig. 13. All ATR-FTIR spectra (a), principal components analysis scores plot for the ATR-FTIR spectra (b), and residual X-variance for PC-1 (c) and PC-2 (d).

shows variation related to SOM of the samples as the PC-2 was better correlated to TN, TC and OC (*r*=-0.68~0.71) than to PC-1 (*r*=0.52~0.56) (Table 3). Samples originated from different fields can be divided into one cluster for Luvisol type soils (field #D) and another for Cambisol type soils (fields #B, #C, and #E). Although samples from the field #A are the mixture of Cambisol and Luvisol types, most of them locate within the former cluster. Samples from the fields #C and #E are hardly separated, although they are of the same soil texture of clay (Table 1). Compared to the samples D28, B04 and B10 located outside of the Hotelling T2 ellipse (Fig.13b), sample A02 exhibits large residual X-variance for PC-1 (Fig.13c) and PC-2 (Fig.13d). Thus, sample A02 was considered as a sample outlier and excluded from further investigation.

#### **3.4.3 ATR-FTIR calibration models**

204 Infrared Spectroscopy – Life and Biomedical Sciences

below 1100 cm-1 might be characteristic for kaolinite (Madari*, et al.*, 2006). However, most characteristic wavebands of soil organic matter presented in Table 5, due to their low concentrations in the soil samples and their overlapping with mineral peaks, could not readily be identified by simple visual observation. For example, the bands around 2940-2935 cm-1 and 2886-2877 cm-1, indicating the presence of organic aliphatic C-H stretching (shown in DRIFT spectra, Fig.8), are nearly vanished. Although the ATR-FTIR spectra present severe deviations from corresponding DRIFT spectra, the light intensity in ATR-FTIR spectra exhibits strong correlation with particle size distribution. For example, samples of clay texture with fine particles present strongest light intensity, whereas samples with sandy loam texture with coarse particle size have lowest light intensity. The clay loam texture samples with middle size particles correspond to intermediate light absorption. For each soil texture class, the spectral difference due to high and low TC content is not apparent. All

Figure 13 shows all Mid-IR spectra obtained by FI-IR spectrometer with ATR accessory together with the PCA scores plot for the spectra and residual X-variance for PC-1 and PC-2. PC-1 and PC-2 explained 98% and 1% of total variance of the spectra, respectively. The PC-2

(a) (b)

ATR

(c) (d)

E18

E20

FTIR spectra (b), and residual X-variance for PC-1 (c) and PC-2 (d).

A01

E07

E03E02E04 E05

C32 C33

E22

Leverage (PC-1) 0 0.01 0.02 0.03 0.04

A04A03 A05 A06

E23

D30

E13 E12 E14 E15 E16 E17

A10 A09 A12 A11 A16 A18 A15 A17A14 A19 A20 B02 B01

D32 D31 D33 D34D35 E01

D29

B23 B25C22C21 C15 C23 C24 C25 C26C28 C27 C29 C30 C31

C34 C36 C35 C37 C38 D01 C39 D02 D06 D05D03 D04D08 D07D09

D10 D12D11 D13D14 D15 D16D20D19 D21D18D17 D22 D24D23 D26D27 D25

E09 E10 E11

A02

A08A07

E06

B05 B06 B08B07B09

B10 B11 B13 B12 B14 B15 B16 B17 B18B19 B20 B21 B22 B24

E08

B03

B04

D28

E19

E21

0

0.0001

0.0002

Fig. 13. All ATR-FTIR spectra (a), principal components analysis scores plot for the ATR-

0

B05 B07B08 B06

0.0001

0.0002

Leverage (PC-2) 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

B09 B10 B11 B12 B13 B14 B15

E12E13 E14 E15

E20E21 E22E23

E17 E18

D24 D20D15 D23 D22D17 D25D16 D28 D32D31D33D29D35 D34D30

E16

A02

A03 A04 A05 A06 A07 A10 A09 A08 A12 A11 A16A17A18A14 A15 A19 A20 B02B01

E02E03E04E07E05E06

D12D13 D11D14 D19 D27 D21 D26 D18

B16 B17 B18 B19 B20 B21 B23 B24B22 B25 C15 C21C22 C24 C23 C25 C26 C27 C28 C29 C30 C31 C32 C33

C34 C36 C35 C37 C38 C39 D08 D09 D07D06 D05D01 D03 D04D02

E08 E10E09E11

E01

E19

A01

D10

B03 B04

ATR

these spectra appear to be free from baseline offset.

**3.4.2 PCA analysis for ATR-FTIR spectra** 

Table 7 summarizes the cross-validation results of PLSR models developed with the raw and various transformed ATR-FTIR spectra against each soil property. For raw spectra, PLSR models produced excellent prediction accuracy for TN, TC and OC with *R*2 of 0.89~0.92 and RPD of 3.02~3.75. These models were developed with 6 latent variables. None of spectral pre-processing techniques adopted can effectively improve the prediction accuracy of these models. This might be due to absence of the baseline offset of spectra (Fig.12). For IC, all models can be used to distinguish high and low concentration with RPD


Table 7. Cross validation result of PLSR models developed with raw and various transformed ATR-FTIR spectra with 121 samples

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 207

Fig. 15. B-coefficients curves obtained from PLSR analysis with the raw spectra for TN, TC

distinctive bands, similar to those obtained with DRIFT spectroscopy. The most significant band locates at around 1430 cm-1, which corresponds to calcium carbonate (CaCO3). However, compared to the B-coefficients curve for IC based on DRIFT spectra (Fig.11), two characteristic bands of CaCO3 at 1795 cm-1 and 2513 cm-1 disappear completely, which is mainly due to the weaker light response of ATR-FTIR spectra compared to DRIFT spectra.

The raw Vis-NIR and Mid-IR spectra were combined to develop PLSR calibration models of soil TN, TC, OC and IC concentrations. Table 8 shows the cross-validation results of the PLSR models. For the Vis-NIR-ATR spectra, PLSR models produced excellent prediction accuracy with *R*2 of 0.89~0.91 and RPD of 3.00~3.75 for TN, TC and OC. By comparison, PLSR models calibrated for the Vis-NIR-DRIFT spectra achieved higher prediction accuracy than those with Vis-NIR-ATR spectra, with *R*2 of 0.93~0.95 and RPD of 3.83~4.62 for TN, TC and OC. For IC, PLSR model for Vis-NIR-ATR spectra was developed with 10 latent variables and validated

Vis-NIR-ATR 5 0.91 3.75 5 0.89 3.00 6 0.91 3.42 10 0.69 1.83 Vis-NIR-DRIFT 6 0.95 4.62 6 0.93 3.83 6 0.95 4.32 5 0.70 1.88

Table 8. Cross validation result of PLSR models developed with raw Vis-NIR-MIR spectra

PLSR models calibrated for raw combinational Vis-NIR-MIR spectra

Total N Total C Organic C Inorganic C

LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD LVs *R*2 RPD

and OC, and the 1st-detrending-transformed spectra for IC.

**3.5 Model calibrations for combinational Vis-NIR-MIR spectra** 

Combinational spectra

with 121 samples

Fig. 14. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on the ATR-FTIR spectra.

of 1.61~1.93. Figure 14 shows the correlation between the measured and predicted values of each soil property. The linear fitting lines for TN, TC and OC are closer to the 1:1 lines as compared to the corresponding plots for Vis-NIR spectra (Fig.6a-c), but not so well as those for DRIFT spectra (Fig.10a-c). However, the linear fitting for IC is better than that for Vis-NIR spectra (Fig.6d) and that for DRIFT spectra (Fig.10d).

#### **3.4.4 B-coefficients analysis of PLSR models for ATR-FTIR spectra**

B-coefficients curves of the PLSR models developed with the best transformed ATR-FTIR spectra for each soil property are compared in Fig.15. The B-coefficients for TN, TC and OC exhibit strong similarity among them. This is mainly due to the high correlation obtained with the reference values (*r*=0.97-0.99, Table 3). Overall, the most significant wavebands for predicting these soil properties locate in the wavenumber range between 1700 and 400 cm-1. As indicated in Table 5, the peaks at 1612-1551 cm-1 are for aromatic C=C stretching and/or asymmetric –COO stretching and the peaks at around 1050 cm-1 correspond to Polysaccharides. Other significant wavebands can be found at round 2925 cm-1, which correspond to Aliphatic C-H stretching. Interestingly, although the ATR-calibrated models are not accurate enough for IC quantification, its B-coefficients curve for IC exhibits several

(a) (b)

(c) (d)

**3.4.4 B-coefficients analysis of PLSR models for ATR-FTIR spectra** 

NIR spectra (Fig.6d) and that for DRIFT spectra (Fig.10d).

ATR-FTIR spectra.

Fig. 14. Measured vs. predicted values of soil TN (a), TC (b), OC (c) and IC (d) based on the

of 1.61~1.93. Figure 14 shows the correlation between the measured and predicted values of each soil property. The linear fitting lines for TN, TC and OC are closer to the 1:1 lines as compared to the corresponding plots for Vis-NIR spectra (Fig.6a-c), but not so well as those for DRIFT spectra (Fig.10a-c). However, the linear fitting for IC is better than that for Vis-

B-coefficients curves of the PLSR models developed with the best transformed ATR-FTIR spectra for each soil property are compared in Fig.15. The B-coefficients for TN, TC and OC exhibit strong similarity among them. This is mainly due to the high correlation obtained with the reference values (*r*=0.97-0.99, Table 3). Overall, the most significant wavebands for predicting these soil properties locate in the wavenumber range between 1700 and 400 cm-1. As indicated in Table 5, the peaks at 1612-1551 cm-1 are for aromatic C=C stretching and/or asymmetric –COO stretching and the peaks at around 1050 cm-1 correspond to Polysaccharides. Other significant wavebands can be found at round 2925 cm-1, which correspond to Aliphatic C-H stretching. Interestingly, although the ATR-calibrated models are not accurate enough for IC quantification, its B-coefficients curve for IC exhibits several

Fig. 15. B-coefficients curves obtained from PLSR analysis with the raw spectra for TN, TC and OC, and the 1st-detrending-transformed spectra for IC.

distinctive bands, similar to those obtained with DRIFT spectroscopy. The most significant band locates at around 1430 cm-1, which corresponds to calcium carbonate (CaCO3). However, compared to the B-coefficients curve for IC based on DRIFT spectra (Fig.11), two characteristic bands of CaCO3 at 1795 cm-1 and 2513 cm-1 disappear completely, which is mainly due to the weaker light response of ATR-FTIR spectra compared to DRIFT spectra.

#### **3.5 Model calibrations for combinational Vis-NIR-MIR spectra**

The raw Vis-NIR and Mid-IR spectra were combined to develop PLSR calibration models of soil TN, TC, OC and IC concentrations. Table 8 shows the cross-validation results of the PLSR models. For the Vis-NIR-ATR spectra, PLSR models produced excellent prediction accuracy with *R*2 of 0.89~0.91 and RPD of 3.00~3.75 for TN, TC and OC. By comparison, PLSR models calibrated for the Vis-NIR-DRIFT spectra achieved higher prediction accuracy than those with Vis-NIR-ATR spectra, with *R*2 of 0.93~0.95 and RPD of 3.83~4.62 for TN, TC and OC. For IC, PLSR model for Vis-NIR-ATR spectra was developed with 10 latent variables and validated


Table 8. Cross validation result of PLSR models developed with raw Vis-NIR-MIR spectra with 121 samples

Vis/Near- and Mid- Infrared Spectroscopy for Predicting Soil N and C at a Farm Scale 209

on its correlation with C if a high C-to-N correlation exists. In the current study, the excellent prediction of TN might be due to its high correlation with TC or OC (*r*=0.97-0.99) (Table 3), which can be proved by the similar distribution of their B-coefficients curves for

The Mid-IR spectroscopy, including ATR and DRIFT, and Vis-NIR spectroscopy were implemented for the prediction of soil TN, TC, OC and IC. Results proved that both Vis-NIR and Mid-IR when combined with chemometric methods have great potential to quantify soil N and C at the field scale. It was also shown that DRIFT is more robust than Vis-NIR or ATR in terms of prediction accuracy. Although the Mid-IR spectra holds more information and usually easier to interpret as compared to Vis-NIR spectra with overtones and combinations features, until recently the MIR instruments are less portable and born to easier damage of optical materials. In contrast, the Vis-NIR has some advantages related to portability, mobile (on-line) measurement, remote sensing and others. This study suggests that Vis-NIR spectroscopy, if coupled with proper spectral pre-processing techniques, has the potential for successful prediction of soil N and C, although the combination and overtone peaks in

We gratefully thank the Soil Labs in Cranfield University for the provision of spectrophotometer. Sincere thanks are given to Dr. Boyan Kuang for assisting soil sampling and laboratory reference measurement, and Dr. Robert J. A. Jones for assisting soil description. Financially supported by Natural Science Foundation of Zhejiang Province, P.R.China (Project

Chang, C.-W. & Laird, D. A. (2002). Near-Infrared Reflectance Spectroscopic Analysis of Soil

Chang, C.-W.; Laird, D. A.; Mausbach, M. J. & Hurburgh, C. R. (2001). Near-Infrared

Gaffey, S. J. (1986). Spectral reflectance of carbonate minerals in the visible and near infrared

Haché, C.; Shibusawa, S.; Sasao, A.; Suhama, T. & Sah, B. P. (2007). Field-derived spectral

Reflectance Spectroscopy–Principal Components Regression Analyses of Soil Properties. *Soil Sci. Soc. Am. J.*, Vol.65, No.2, (2001/3), pp.480-490, ISSN 0361-5995 Fystro, G. (2002). The prediction of C and N content and their potential mineralisation in

heterogeneous soil samples using Vis-NIR spectroscopy and comparative methods.

(0.35-2.55 microns):calcite, aragonite, and dolomite. *American Mineralogist*, Vol.71,

characteristics to classify conventional and conservation agricultural practices. *Computers and Electronics in Agriculture*, Vol.57, No.1, pp.47-61, ISSN 0168-1699 Ludwig, B.; Khanna, P. K.; Bauhus, J. & Hopmans, P. (2002). Near infrared spectroscopy of

forest soils to determine chemical and biological properties related to soil

No.Y1090885) and by the State Scholarship Fund of China (Grant No.[2009]3004).

C and N. *Soil Science*, Vol.167, No.2, pp.110-116, ISSN 0038-075X

*Plant and Soil*, Vol.246, pp.139-149, ISSN 0032-079X

Vis-NIR spectra (Fig.7), DRIFT spectra (Fig.11), and ATR-FTIR spectra (Fig.15).

**4. Conclusions** 

**5. Acknowledgment** 

**6. References** 

the Vis-NIR spectral range are usually weak.

pp.151-162, ISSN 0003-004X

with *R*2 of 0.69 and RPD of 1.83, whereas PLSR model for Vis-NIR-DRIFT spectra calibrated with 5 latent variables performed better than Vis-NIR– ATR with higher prediction accuracy (*R*2 of 0.70 and RPD of 1.94). These models for combinational spectra performed slightly better than those for Vis-NIR spectra (Table 4). However, these models did not produce better performance than those for DRIFT spectra (Table 6) and ATR spectra (Table 7) with only exception of Vis-NIR-ATR models for OC with *R*2 of 0.91 and RPD of 3.42.

#### **3.6 Comparison of PLSR model performance among Vis-NIR, ATR-FTIR, DRIFT and combinational spectra**

As shown in Tables 4, 6 and 7, model performance is not only a function of wavelength ranges used during PLS regression analysis, but also a function of spectral pre-processing techniques. Overall, for TN, TC and OC, PLSR models calibrated for DRIFT spectra outperformed those for Vis-NIR spectra and ATR-FTIR spectra. For IC, both ATR-FTIR and DRIFT models outperformed Vis-NIR models no matter what spectral pre-processing techniques were applied. However, if coupled with appropriate spectral pre-processing techniques, Vis-NIR models for TN and OC can produce competitive prediction performance (R2>0.90 and RPD>3.0) with less number of latent variables (3 or 4) as compared to best ATR-PLSR models calibrated with 6 latent variables. For TC, ATR-FTIR models performed slightly better than Vis-NIR models. The lower accuracy for the calibration of IC compared to TN, TC and OC may be attributed to errors in the reference method for IC determination, since IC is calculated by difference between TC and OC.

Researchers have reported that the particle size distribution within the soil sample population and also within each sample of the calibration set affects the accuracy of calibration for TC and OC both in Mid-IR and Vis-NIR (Madari*, et al.*, 2006; Mouazen*, et al.*, 2005, Yang*, et al.*, 2011b). However, Vis-NIR proved to be more sensitive to particle size effects than the Mid-IR range (Madari*, et al.*, 2006). Vis-NIR spectroscopy performed very well for a very homogenous sample population, even slightly better than Mid-IR, but with increasing heterogeneity among and within the soil samples the accuracy decreased drastically. By contrast, the particle size distribution had less effect in the Mid-IR range. For the very homogeneous sample population, the accuracy was slightly lower than Vis-NIR, but with the increase in the heterogeneity of the sample population the accuracy did not diminish drastically and was higher than using with Vis-NIR (Madari*, et al.*, 2006). Thus Mid-IR spectroscopy coupled with appropriate chemometrics can be considered to be more robust than Vis-NIR.

#### **3.7 Fundamentals of predicting N in soil**

Soil N content is often highly correlated with C (Martin*, et al.*, 2002). For example, Chang, *et al.* (2001) reported *r* of 0.95 between TC and TN. In this study, the mean (±s.d.) values of TC/TN and OC/TN are 10.6(±0.59) and 9.95(±0.52), respectively (Table 2). It is an interesting point to explore whether there is an independent spectral basis for the determination of N in soil by infrared (IR) spectroscopy or whether N is predicted through high correlation with C. In the work by Chang and Laird (2002), in which C and N were added to a soil resulting in a wide range of C-to-N ratios, N was proved to be predicted in soil independently of C. Although the N absorbers are present in the soil spectra, their absorbance is not as strong as that of C bonds, as the mass of C in soil is generally an order of magnitude higher than that of N (about 10 times in our case). Thus, Martin *et al*. (2002) explained that N is predicted best
