Quantifying soil organic carbon fractions by infrared-spectroscopy

Quantifying soil organic carbon fractions by infrared-spectroscopy

ARTICLE IN PRESS Soil Biology & Biochemistry 39 (2007) 224–231 www.elsevier.com/locate/soilbio Quantifying soil organic carbon fractions by infrared...

369KB Sizes 2 Downloads 119 Views

ARTICLE IN PRESS

Soil Biology & Biochemistry 39 (2007) 224–231 www.elsevier.com/locate/soilbio

Quantifying soil organic carbon fractions by infrared-spectroscopy M. Zimmermann, J. Leifeld, J. Fuhrer Agroscope ART Reckenholz, Swiss Federal Research Station for Agroecology and Agriculture, Air Pollution/Climate Group, Reckenholzstrasse 191, 8046 Zurich, Switzerland Received 20 April 2006; received in revised form 26 June 2006; accepted 21 July 2006 Available online 31 August 2006

Abstract Methods to quantify organic carbon (OC) in soil fractions of different stabilities often involve time-consuming physical and chemical treatments. The aim of the present study was to test a more rapid alternative, which is based on the spectroscopic analysis of bulk soils in the mid-infrared region (4000–400 cm1), combined with partial least-squares regression (PLS). One hundred eleven soil samples from arable and grassland sites across Switzerland were separated into fractions of dissolved OC, particulate organic matter (POM), sand and stable aggregates, silt and clay particles, and oxidation resistant OC. Measured contents of OC in each fraction were then correlated by PLS with infrared spectra to obtain prediction models. For every prediction model, 100 soil spectra were used in the PLS calibration and the residual 11 spectra for validation of the models. Correlation coefficients (r) between measured and PLS-predicted values ranged between 0.89 and 0.97 for OC in different fractions. By combining different fractions to one labile, one stabilized and one resistant fraction, predictions could even be improved (r ¼ 0:98, standard error of prediction ¼ 16%). Based on these statistical parameters, we conclude that mid-infrared spectroscopy in combination with PLS is an appropriate and very fast tool to quantify OC contents in different soil fractions. r 2006 Elsevier Ltd. All rights reserved. Keywords: Carbon fractions; Mid-infrared spectroscopy; Partial least-squares regression

1. Introduction Organic carbon (OC) in agricultural soils is of increasing interest. This is not only because of its well-known beneficial effect on nutrient dynamics and soil structure, but also because of its potential role as a sink for atmospheric carbon dioxide (CO2) (IPCC, 2000). Changes in the type and intensity of land management, or in land use such as conversion of cropland to grassland, can positively influence the sequestration of atmospheric carbon (Lal, 2004; DeGryze et al., 2004), and thus could potentially play an important role in the mitigation of climate change. Both land-use and soil management influence the amount of OC in soils in many ways, positively as well as negatively. But these influences are not of equal concern to all parts of the soil system. Importantly, they differ Corresponding author. Tel.: +41 44 377 7125; fax: +41 44 377 7201.

E-mail address: [email protected] (M. Zimmermann). 0038-0717/$ - see front matter r 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.soilbio.2006.07.010

between soil fractions of different physical and chemical stabilities. Soil samples can be separated into fractions that are either (i) easily decomposable, (ii) stabilized by physical–chemical mechanisms or (iii) biochemically recalcitrant. Fresh plant residues are only partly incorporated into the soil matrix. Thus, they are easily available for microorganisms and, consequently, rapidly decomposed. By further incorporation into the soil matrix, OC becomes stabilized either through physical protection inside aggregates, where access by microorganisms is restricted, or physico-chemically by binding onto mineral surfaces of silt and clay particles (Six et al., 2002a). Biomolecules of degraded SOM are even biochemically recalcitrant (Krull and Skjemstad, 2003). Various methods have been proposed to separate soil samples into fractions with distinct chemical and physical characteristics corresponding to different stabilizing mechanisms and soil functions. Typically, these methods were used to determine the OC content of the different soil fractions in order to calculate turnover times, and to identify relationships between the distribution of OC

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

between the different fractions and soil management or land-use (Hassink et al., 1997; Sohi et al., 2001; Six et al., 2002b; John et al., 2005). Most of the proposed procedures included dispersion of aggregates and isolation of a light particulate organic matter (POM) fraction. This latter fraction has been shown to play a crucial role in the formation of aggregates (Six et al., 2001), and to be most sensitive to changes in soil management (Franzluebbers and Stuedemann, 2002). To obtain a chemically resistant fraction, treatments like hydrolysis or oxidation are applied to plant-free soil fractions (Paul et al., 1997; Siregar et al., 2005). These fractionation procedures have in common that they are very time-consuming. For routine investigations, more rapid methods are needed. A combination of spectroscopic measurement and partial least-squares regression (PLS) was introduced by Haaland and Thomas (1988) to quantify different chemical compounds. PLS is a chemometric method to derive prediction models for specific compounds from spectroscopic data. Later, Janik and Skjemstad (1995) quantified different properties of bulk soils by combining PLS with mid-infrared spectra. Since then, PLS was applied in various studies to estimate bulk soil characteristics from visible, near- and mid-infrared spectroscopic measurements (Viscarra-Rossel et al., 2006). Most recently, Cozzolino and Moron (2006) quantified OC of particle-size fractions by near-infrared spectroscopy and PLS, thus indicating the potential of this approach for rapid analysis of OC in soil size-fractions. In the present study, we aimed to go one step further. The objective was to quantify OC from differently stabilized soil fractions by use of mid-infrared spectra obtained from bulk samples of agricultural topsoils. The approach was to first fractionate the samples by a combination of different treatments to obtain two labile, two stabilized and one chemically resistant fraction, and then to determine the amounts of OC in these fractions and to correlate OC contents with data obtained from infrared spectra of bulk soil samples using PLS. Finally, these correlations were used as a set of predictive models, which could be tested using independent data from OC measurements. 2. Materials and methods 2.1. Soil samples We fractionated 111 archived soil samples from two different research projects in Switzerland. One archive was the collection of the Swiss national soil survey, from which we analysed 41 samples representing agriculturally managed (arable land, temperate and alpine permanent grassland) sites across Switzerland. Sampling sites varied in altitudes from 265 to 2400 m above sea level (a.s.l.) representing a gradient in mean annual temperature between +10.6 and 1.6 1C, and in mean annual precipitation from 722 to 2327 mm. Undisturbed grassland

225

soils were divided into horizons of 0–5 cm, 5–10 cm, 10–20 cm, 0–10 cm or 0–20 cm. Disturbed soils from arable sites were taken as one horizon from 0 to 20 cm. A detailed description of the sampling technique, together with site characteristics such as soil properties, climate, geology and land-use, are given in Desaules and Studer (1993). A second set with 70 soil samples were obtained from a biodiversity study in temperate grasslands on the Swiss Central Plateau. These samples originated from three different regions at altitudes between 420 and 670 m a.s.l. with an annual precipitation between 964 and 1333 mm. Each soil sample was a composite of 20 soil cores taken from the top 20 cm. Management intensities of these grassland sites varied considerably, as described in detail by Buholzer et al. (2005). Since soil samples included in a PLS prediction model should have similar spectral properties, only carbonate-free soils with an organic matter (OM) content of less than 15% were used. Both, carbonate and OM cause distinct peaks in infrared spectra, which may interfere with other peaks. Bulk soil samples were dried at 40 1C, crushed, and particles 42 mm were removed. Silt and clay contents were determined by the pipette method and sand content calculated by difference (FAC, 1989). Carbon and nitrogen contents of all archived bulk soil samples were measured newly before fractionating the soil samples after dry combustion with an elemental analyzer (Vario EL, Elementar). 2.2. Fractionation procedure Soil samples were separated by means of physical and chemical procedures, as shown in Fig. 1. Thirty grams of soil material (o2 mm) were dissolved in 150 ml water and dispersed using a calibrated ultrasonic probe-type (Sonopuls, Bandelin) with an output-energy of 22 J ml1. This energy application breaks up labile aggregates but does not disrupt sand-sized plant fragments (Amelung and Zech, 1999). The dispersed suspension was then wet-sieved with a 63-mm sieve until the rinsing water was clear. Particles remaining on the sieve consisted of sand and stable aggregates (S+A) as well as non-protected POM. POM was separated by stirring all particles 463 mm with sodium polytungstate with a density of 1.8 g cm3 (Sohi et al., 2001). After centrifugation at 1000g for 15 min and decanting the light POM fraction, the POM and the S+A fractions were washed with deionized water to remove all sodium polytungstate, and then dried at 40 1C. Silt and clay particles (s+c) were obtained by filtering the suspension o63 mm through a 0.45 mm nylon mesh. An aliquot of the filtrate o0.45 mm was frozen and used to measure the amount of dissolved organic carbon (DOC). The s+c fraction was dried at 40 1C and weighted. Out of this fraction, we isolated a chemically resistant carbon fraction (rSOC) by NaOCl oxidation after a method of Kaiser and Guggenberger (2003). One gram of s+c was oxidized during 18 h at 25 1C with 50 ml of 6% NaOCl

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

226

suspension < 0.45 µm

DOC labile fraction

bulk soil < 2 mm dispersion with 22 J ml-1, wet sieving to 63 µm

light fraction

POM

heavy fraction

S+A

fraction > 63 µm, afterwards densityseparation at 1.8 g cm-3

stabilized fraction fraction < 63 µm

soil fractions

s+c

s+c-rSOC

NaOCl resistant fraction

rSOC

carbon fractions

resistant fraction

Fig. 1. Procedure to obtain the following organic carbon fractions: dissolved organic carbon (DOC), particulate organic matter (POM), sand and stable aggregates (S+A), silt and clay particles (s+c), and oxidation-resistant carbon (rSOC). Adding POM+DOC yields a combined labile fraction, and S+A+(s+crSOC) a combined stabilized fraction.

(wt/wt) adjusted to pH 8 with concentrated HCl. Residues were centrifuged at 1000g for 15 min, decanted, washed with deionized water, and then centrifuged again. This oxidation step was repeated twice. Carbon and nitrogen contents in all solid fractions were measured by combustion with an elemental analyzer (Vario EL, Elementar), and in liquid samples by thermal oxidation with a liquid analyser (Dimatoc 2000, Dimatec). With reference to the stabilizing mechanisms described above, OC contents of the fractions DOC and POM were combined to yield one labile fraction, and OC contents of S+A and s+c minus rSOC were combined to obtain one stabilized fraction. 2.3. DRIFT spectroscopy Infrared spectra were recorded in reflectance mode with a Perkin Elmer Spectrum One spectrometer with Diffuse Reflectance Infrared Fourier Transformation (DRIFT) inlet. Bulk soil samples were diluted with 97% KBr to reduce scatter light intensity, homogenized for 30 s in a ball mill, and then scanned 16 times in the range from 4000 to 400 cm1 (mid-infrared region) with a resolution of 4 cm1. KBr background spectra were subtracted from the measured spectra. DRIFT spectra were corrected against atmospheric CO2 and water vapour and automatically adjusted to an internal CH4 cell. In the following, infrared regions are ascribed to functional groups according to Baes and Bloom (1989).

2.4. PLS quantification For PLS, spectral information is arranged as n  m matrix which consists of n spectra with absorbance values for m wavelengths, and the calibration data is expressed as a single vector with the measured values for these spectra (Kramer, 1998). The PLS1 algorithm decomposes the mdimensional spectra space into few factors termed latent variables (LV), which represent the best projections of the calibration vector onto the n  m matrix. One of the advantages of PLS compared to other chemometric methods like principal component analysis is the possibility to interpret the first few LV, because they show the correlations between the property values and the spectral features (Kramer, 1998). Furthermore, PLS takes as well variations of the absorbance as variations of the calibration data into account. A separate prediction model has to be evaluated for every compound that will be estimated by PLS. The exact decomposition process conducted by the PLS1 algorithm is described in detail in Martens and Naes (1989). We divided all DRIFT spectra for each PLS prediction model into two separate sets for calibration and validation. All soil samples (n ¼ 111) were arranged for each case according to measured property values (i.e. clay content, total OC and nitrogen, or amount of OC in the different fractions). The DRIFT spectra of every 10th soil sample, beginning with number 5, was allocated to the validation subset (n ¼ 11). Individual prediction models for every

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

where xm are measured values, xp PLS-predicted values, and n is the number of samples used for validation. The prediction model with the lowest SEP and a slope closest to 1 was considered as best fit (Fig. 2). To compare SEP values of prediction models for different data ranges, relative SEP (rel SEP) was calculated as rel SEP ¼

SEP  100 x¯

(2)

with x¯ as the mean of the validation data range. The appropriateness of a chemometric prediction method was evaluated by the residual prediction deviation (RPD), which is the ratio between the standard deviation (s.d.) of the validation sample set and the SEP (Willimas and Norris, 1987): RPD ¼

s:d: . SEP

(3)

In agricultural applications, RPD values 43 are considered satisfactory, whereas prediction models with

5.0

1.10

4.5

best fit with 12 LV

1.08

y = 1.021x SEP = 2.79

3.5 1.04

slope (

)

1.06

)

4.0

SEP (

property were then computed with the software Spectrum Quant+ (Perkin Elmer, Version 4.51.02) using the data from the 100 calibration spectra. Spectra were imported into the software in absorbance units and scaled to a mean of zero. The data range was reduced to between 4000 and 600 cm1 in order to remove signal noise in the region between 600 and 400 cm1. The software was forced to decompose the spectra matrix into 30 LV, and to validate this decomposition process internally by full cross-validation. During full cross-validation, each spectrum was in turn excluded from the calibration sample set and was predicted by the PLS model calculated with the remaining spectra. To detect outliers, leverage values were consulted, which can be conceived as the distance between a single spectrum and the centre of the entire data matrix. Spectra with leverage values higher than twice the mean leverage value of all spectra were excluded from the calibration. By decomposing the spectra into 30 LV, it was assumed that the prediction model would be over-fitted because signal noise of the spectral measurements could also be correlated with the property vector (Martens and Naes, 1989). The optimal numbers of LV were found by testing 1–30 LV in the prediction model and comparing the predicted values of each PLS model with the measured ones. For this, the prediction models calibrated on 100 DRIFT spectra were applied to the 11 DRIFT spectra of the validation set, and the PLS predicted and the laboratorial quantified values compared. To decide which prediction model fitted best, standard errors of prediction (SEP) and slopes of the regression equations through the origin between predicted and measured values for 1–30 LV were compared. Standard error of prediction (SEP) values were calculated as sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P ðxm  xp Þ2 SEP ¼ , (1) n1

227

3.0 1.02

2.5

2.0

1.00 6

8

10 12 14 number of LV

16

18

Fig. 2. Determination of the optimal number of latent variables (LV): relationship between number of number LV and standard error of prediction (SEP) (m) and slope of regression equation through zero (J) for measured vs. PLS-predicted data for total organic carbon.

RPD values o3 should only be used as screening methods (Malley et al., 2000). 3. Results and discussion 3.1. Carbon distribution among fractions Means and distributions of results for the measured contents of OC, nitrogen and clay in bulk soils, and for OC contents of different fractions are summarized in Fig. 3. OC content (7s.d.) of bulk soil averaged for all samples was 27.5 (712.1) mg C g1 soil. Bulk soils from arable sites had the lowest OC content with 18.1 (76.7) mg C g1 soil, while OC contents in samples from permanent grasslands were 27.3 (711.1) mg C g1 soil and from alpine grasslands 43.4 (714.2) mg C g1 soil, in agreement with results from earlier surveys (Leifeld et al., 2005). POM as well as rSOC contained on average 9% of the total OC. Only 2% of total OC was dissolved by water during sieving. About 29% of OC in bulk soil was contained in the S+A fraction, and 51% in the s+c-rSOC fraction. The distribution of OC in the latter two fractions varied considerably among land-use classes. In samples from arable sites, the fraction of OC stabilized in S+A averaged 13%, in temperate grasslands 29%, and in alpine grasslands 55%. Accordingly, the relative amount of OC in the fraction of s+c minus rSOC was 29% in alpine grasslands, 51% in lowland grasslands, and 67% in soils from arable sites. The differences in the distribution of OC between the S+A and s+c-rSOC fractions between sites probably

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

6

50

5

mg g-1soil

%

30 25 20 15

40 30

4 3

10

20

2

5

10

1

0

0

0

40

8

30

6

Cs+c-rSOC

CrSOC

30 25

CDOC

10

10

8

8

6

6

20

20

4

15

mg g-1soil

60

10

mg g-1soil

7

35

50

mg g-1soil

8

70

mg g-1soil

80

40

CPOM

CS+A

Ntot

mg g-1soil

Ctot

clay 45

mg g-1soil

228

4

4

10 10

2

0

0

5 0

2

2

0

0

Fig. 3. Clay content, total organic carbon and nitrogen contents, and OC concentrations in the fractions (S+A ¼ sand and stable aggregates, POM ¼ particulate organic matter, s+crSOC ¼ silt and clay minus resistant soil organic carbon, rSOC ¼ resistant soil organic carbon, and DOC ¼ dissolved organic carbon) for all samples (n ¼ 111) expressed as box plots with median (point), 25% and 75% percentiles (box), and minimum and maximum values (whiskers).

40

20 15 10 5 0

soil)

50

N PLS predicted (mg

25

y = 1.021x rel SEP = 10% 12 LV

g-1

30

C PLS predicted (mg g-1soil)

clay PLS predicted (%)

6

60 y = 1.013x rel SEP = 10% 4 LV

35

40 30 20 10 0

0

5

10 15 20 25 30 35 40

clay measured (%)

y = 0.976x rel SEP = 9% 7 LV

5 4 3 2 1 0

0

10

20

30

40

50

C measured (mg g-1soil)

60

0

1

2

3

4

5

6

N measured (mg g-1soil)

Fig. 4. Scatter plots of measured vs. PLS-predicted values for total clay, SOC and N, with corresponding regression equations with zero-intercept, relative standard errors of prediction (rel SEP) and number of latent variables (LV) used.

resulted from the stronger aggregation under the less intensive soil management (Bossuyt et al., 2002). Relative contents of OC in the POM, rSOC and DOC varied only slightly among land-use classes.

3.2. Quantifying bulk soil properties PLS predictions for clay, OC and nitrogen of bulk soil samples agreed well with measured data (Fig. 4 and Table 1). Pearson’s correlation coefficients (r) for measured vs. PLS-predicted values for the validation sample sets were 0.97 for clay and OC, and 0.96 for nitrogen, and the data were very close to the 1:1 line. Slopes for the regressions were 1.013 for clay, 1.021 for OC, and 0.976 for nitrogen. Also, rel SEP of 10% for clay and OC, and 9% for nitrogen were in an acceptable range. Residual prediction deviation values of 3.5 for clay, 4.1 for OC and 4.0 for nitrogen confirmed the successful prediction. Correlation coefficients (r) within the calibration sample sets after full crossvalidation were 0.96 for clay, 0.99 for OC and 0.96 for

nitrogen, and thus in the same range for both subsets of data. The correlations between measured and PLS-predicted values were slightly better than those of Janik and Skjemstad (1995) who obtained R2-values of 0.80 for clay, 0.92 for OC and 0.82 for nitrogen. These authors used 298 mid-infrared spectra of non-diluted Australian soil samples and computed PLS models with 6 LV for clay, 9 LV for OC and 8 LV for nitrogen. In comparison, we used 4 LV for the prediction of clay, 12 LV for OC and 7 LV for nitrogen. McCarty et al. (2002) computed PLS prediction models from mid-infrared and near-infrared spectra of 177 agricultural soil samples without sample dilution with KBr from diverse locations in the central United States to estimate OC, CaCO3 and pH. They used 17 LV in the prediction model for OC and received a value of 0.94 for R2 between measured and predicted results for mid-infrared spectra. As prediction models based on near-infrared spectra were less accurate in their study (R2 ¼ 0:82), these authors concluded that mid-infrared spectra are better suited than near-infrared spectra to predict OC of soils.

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

229

Table 1 Correlations between measured and PLS-predicted values for the calibration and validation (n ¼ 11) sample sets with number of latent variables (LV) used for calibration, Pearson’s correlation coefficients (r), regression equations through zero, relative standard errors of predictions (rel SEP), and residual prediction deviations (RPD) Property

Clay Ctot Ntot CDOC CPOM CS+A Cs+crSOC CrSOC Clabile fractions Cstabile fractions

Calibration

Validation

LV

r

n

Regression equation

r

rel SEP %

RPD

4 12 7 5 9 6 7 8 12 11

0.96 0.99 0.96 0.90 0.91 0.93 0.89 0.85 0.96 0.99

91 90 93 90 92 94 94 93 92 92

y ¼ 1:013x y ¼ 1:021x y ¼ 0:976x y ¼ 0:738x y ¼ 0:993x y ¼ 1:025x y ¼ 0:975x y ¼ 0:991x y ¼ 0:965x y ¼ 0:954x

0.97 0.97 0.96 0.90 0.97 0.93 0.94 0.89 0.98 0.98

10 10 9 51 22 36 12 16 16 11

3.5 4.1 4.0 2.0 3.4 2.6 2.8 2.0 4.1 4.1

Predicted values were abbreviated as Ctot, total organic carbon; Ntot, total nitrogen; CDOC, dissolved organic carbon; CPOM, particulate organic carbon; CS+A, carbon in sand and stable aggregates; Cs+crSOC, carbon in silt and clay fraction without resistant soil organic carbon (rSOC); Clabile fractions, carbon in CDOC plus C POM; and Cstabile fractions, carbon in CS+A plus Cs+crSOC.

3.3. Quantifying OC of soil fractions The PLS prediction accuracy for OC differed considerably between fractions (Table 1). Correlation coefficient for measured vs. predicted OC in the DOC fraction was 0.90, but the RPD value was 2.0. As discussed by Malley et al. (2000), prediction models with such low RPD values should only be used for screening, and should not be accepted as measurements. The rel SEP was 51% and the slope of the regression was 0.738, which indicates that the PLS model predicted DOC inaccurately. This result could have been caused by the fractionation procedure used in this study. To obtain a soil-specific amount of DOC, it would be necessary to rinse the soil samples with a constant ratio of water to soil. However, as we aimed to determine the amount of OC lost during sieving and not the amount of effectively water-soluble OC, we measured DOC in the rinsing water and not in a constant amount of water. Another reason for the inaccuracy could be the small amount of DOC compared to the total mass; on average, DOC accounted for only 0.04% of the total soil mass. Conversely, PLS prediction for OC in the POM fraction was excellent (r ¼ 0.97, slope of the regression ¼ 0.993), and the RPD value of 3.4 also indicated a high reliability of the prediction model, although rel SEP was 22%. The prediction model for OC in the S+A fraction was less accurate. The PLS model did not overestimate the amount of OC in S+A (slope ¼ 1.025) and r was 0.93, but rel SEP of 36% was high, and the RPD of 2.6 was low. The same situation applied to the prediction of OC in the s+c without rSOC. For this fraction, the RPD value of 2.8 was also low, but with a rel SEP of 12%, a regression slope of 0.975 and r ¼ 0.94, the PLS prediction model still yielded acceptable results for the quantification of OC stabilized in this fraction. The correlation coefficient between measured and predicted OC in the fraction rSOC was 0.89 and the rel SEP 16%. However, the RPD value of 2.0 was very low.

But with a slope of the regression of 0.991, the PLS prediction did not lead to systematic over- or underestimations. Chang et al. (2001) recommended to accept prediction models with RPD of o2.0 in cases where regression slopes are not significantly different from 1, and r2 values are at least 0.80. Overall, the results of these statistical analyses suggest that OC of all fractions except DOC could be estimated with PLS using their bulk soil infrared spectra, but that rel SEP and RPD were not always convincing. One reason for the latter could be the low OC concentrations of the different fractions. 3.4. Prediction of carbon in combined soil fractions To improve prediction models, we tested the predictive capability to quantify OC contents of combined labile and combined stabilized fractions (see Fig. 1). PLS models for these combined fractions were computed using the same methods as before. By relating measured to PLS-predicted OC values, we obtained an r-value of 0.98 for both fractions. Also RPD values for both prediction models were excellent (4.1). The PLS model for the stabilized fraction had a rel SEP of 11% and for the labile fraction of 16%. Both prediction models only slightly underestimated OC contents. The slope of the regression for the combined labile fractions was 0.965 and for stabilized fractions it was 0.954. The improvement in the predictions for these combined fractions relative to the results for the separate fractions could be explained by the higher amount of soil components that is correlated to the spectra. Cozzolino and Moron (2006) obtained RPD values between 2.0 and 3.1 for the estimation of OC in particle-size fractions based on near-infrared spectra. Attained rel SEP were comparable to standard errors of OC in repeated fractionation steps. Leifeld and Ko¨gel-Knabner (2005) obtained standard errors of OC in POM and sand-sized fractions of

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

230

stabilized OC

labile OC

resistant OC

(i)

(ii)

(iii)

(iv)

4000

2000

1000 600

4000

2000

cm-1

1000 600

cm-1

4000

2000

1000 600

cm-1

Fig. 5. First four (i–iv) latent variables (LV) of PLS models for OC content in labile, stabilized and resistant fractions. Dotted lines are zero lines.

9–21%, and Six et al. (2000) obtained standard errors of OC in size classes and density fractions in macroaggregates of 11%. To investigate which infrared regions contributed most to the prediction of the labile, stabilized and rSOC fractions, we compared the first four LV of the prediction models (Fig. 5). These first four LV contributed 96.9% of the totally explained spectral variance of the PLS model for the labile and stabilized fractions, and 95.8% for rSOC. Therefore, the uppermost portion of spectral information used for computing the prediction models was within these four first LV. The combined labile fraction consisted of DOC and POM, and therefore contained almost no mineral compounds. This is also evident from the shape of the LV. Positive peaks in the LV correlate with the measured data and negative peaks with interfering components (Janik and Skjemstad, 1995). Positive peaks occurred in the region from 3000 to 2800 cm1 (C–H stretches), from 1340 to 1065 cm1 (C–O–H stretches), and from 830 to 730 cm1 (aromatic C–H bends). Negative peaks were in the range 3700–3000 cm1 (O–H of mineral compounds and H2O) and 1100–1000 cm1 (Si–O of quartz). The first LV of the prediction models for the stabilized and the rSOC fraction were similar. Positive peaks occurred in the region from 3700 to 2800 cm1, which can be ascribed to O–H of mineral compounds and alkyl C–H stretches, and negative interfering peaks in the region from 1100 to 1000 cm1 (Si–O peaks). The similarity of these two LV is plausible,

because the rSOC fraction was obtained from the s+c fraction. But after the first LV, additional LVs differed considerably between PLS models for stabilized and resistant fractions. In contrast to the PLS model for the stabilized fraction, the model for OC of rSOC had almost no positive peaks between 3000 and 2800 cm1 (alkyl C–H). This reflects the effect of the chemical treatment used to obtain the resistant fraction. NaOCl oxidizes all methane, methylene and methyl groups (Mikutta et al., 2005). Consequently, rSOC is free of alkyl C–H, and the very small peaks between 3000 and 2800 cm1 in the LV for rSOC compared to the LV for stabilized OC revealed that alkyl C was not taken into consideration by PLS for the prediction model of rSOC. This analysis suggests that the successful application of the method to quantify labile, stabilized and resistant fractions is not only supported by statistical parameters, but also by the plausibility of the relationship between spectral properties and functional groups of the LV. 4. Conclusions Prediction models for bulk soil properties estimated the laboratorial measured data very well and were of comparable accuracy as for previously published studies. Moreover, statistical parameters between measured and PLSpredicted OC contents of differently stabilized soil fractions suggest that mid-infrared spectroscopy in combination with PLS is appropriate to measure OC contents of

ARTICLE IN PRESS M. Zimmermann et al. / Soil Biology & Biochemistry 39 (2007) 224–231

labile, stabilized and resistant OC fractions out of bulk soil spectra. Infrared regions considered by PLS to compute the prediction models emphasize the reasonable relationship between spectral properties and functional OC groups of the fractions. Furthermore, the quantification of soil carbon fractions from bulk soil DRIFT spectra, which takes just some minutes, is much faster than fractionating soil samples by chemical and physical methods and measuring the carbon contents of the fractions afterwards. Thus, the proposed method can be very helpful in analyzing large numbers of samples for effects of agricultural management and land-use or land-use changes on total OC content, and on differently stabilized OC fractions. Acknowledgements This study was financed by the Swiss Federal Office for the Environment. We thank the Swiss National Soil Survey (NABO) and H.-R. Oberholzer for providing the soil samples. References Amelung, W., Zech, W., 1999. Minimisation of organic matter disruption during particle-size fractionation of grassland epipedons. Geoderma 92, 73–85. Baes, A.U., Bloom, P.R., 1989. Diffuse reflectance and transmission Fourier-transform infrared (Drift) spectroscopy of humic- and fulvicacids. Soil Science Society of America Journal 53, 695–700. Bossuyt, H., Six, J., Hendrix, P.F., 2002. Aggregate-protected carbon in no-tillage and conventional tillage agroecosystems using carbon-14 labelled plant residue. Soil Science Society of America Journal 66, 1965–1973. Buholzer, S., Jeanneret, P., Bigler, F., 2005. Evaluation der O¨komassnahmen-Bereich Biodiversita¨t. Schriftenreihe der FAL 56, Zurich, Switzerland. Chang, C.W., Laird, D.A., Mausbach, M.J., Hurburgh, C.R., 2001. Nearinfrared reflectance spectroscopy—principal components regression analyses of soil properties. Soil Science Society of America Journal 65, 480–490. Cozzolino, D., Moron, A., 2006. Potential of near-infrared reflectance spectroscopy and chemometric to predict soil organic carbon fractions. Soil & Tillage Research 85, 76–85. DeGryze, S., Six, J., Paustian, K., Morris, S.J., Paul, E.A., Merckx, R., 2004. Soil organic carbon pool changes following land-use conversions. Global Change Biology 10, 1120–1132. Desaules, A., Studer K., 1993. Nationales Bodenbeobachtungsnetz. Messresultate 1985–1991. Schriftreihe fu¨r Umwelt (Nr. 200). Bundesamt fu¨r Umwelt, Wald und Landschaft, Bern, Switzerland. FAC, 1989. Methoden fu¨r Bodenuntersuchungen. Schriftenreihe der FAC, Nr. 5. Eidgeno¨ssische Forschungsanstalt fu¨r Agrikulturchemie und Umwelthygiene, Liebefeld, Bern, Switzerland. Franzluebbers, A.J., Stuedemann, J.A., 2002. Particulate and nonparticulate fractions of soil organic carbon under pastures in the Southern Piedmont USA. Environmental Pollution 116, 53–62. Haaland, D.M., Thomas, E.V., 1988. Partial least-squares methods for spectral analyses. 1. Relation to other quantitative calibration methods and the extraction of qualitative information. Analytical Chemistry 60, 1193–1202. Hassink, J.A., Whitmore, P., Kubat, J., 1997. Size and density fractionation of soil organic matter and the physical capacity of soils to protect organic matter. European Journal of Agronomy 7, 189–199.

231

IPCC, 2000. Land Use, Land-use Change, and Forestry. Intergovernmental Panel on Climate Change. Cambridge University Press, UK. Janik, L.J., Skjemstad, J.O., 1995. Characterization and analysis of soils using midinfrared partial least-squares. 2. Correlations with some laboratory data. Australian Journal of Soil Research 33, 637–650. John, B., Yamashita, T., Ludwig, B., Flessa, H., 2005. Storage of organic carbon in aggregate and density fractions of silty soils under different types of land use. Geoderma 128, 63–79. Kramer, R., 1998. Chemometric Techniques for Quantitative Analysis. Marcel Dekker, New York. Kaiser, K., Guggenberger, G., 2003. Mineral surfaces and soil organic matter. European Journal of Soil Science 54, 219–236. Krull, E.S., Skjemstad, J.O., 2003. Delta C-13 and delta N-15 profiles in C-14-dated Oxisol and Vertisols as a function of soil chemistry and mineralogy. Geoderma 112, 1–29. Lal, R., 2004. Soil carbon sequestration impacts on global climate change and food security. Science 304, 1623–1627. Leifeld, J., Bassin, S., Fuhrer, J., 2005. Carbon stocks in Swiss agricultural soils predicted by land-use, soil characteristics, and altitude. Agriculture, Ecosystems & Environment 105, 255–266. Leifeld, J., Ko¨gel-Knabner, I., 2005. Soil organic matter fractions as early indicators for carbon stock changes under different land-use? Geoderma 124, 143–155. Malley, D.F., Lockhart, L., Wilkinson, P., Hauser, B., 2000. Determination of carbon, nitrogen, and phosphorus in freshwater sediments by nearinfrared reflectance spectroscopy: rapid analysis and a check on conventional analytical methods. Journal of Paleolimnology 24, 415–425. Martens, H., Naes, T., 1989. Multivariate Calibration. Wiley, Chichester. McCarty, G.W., Reeves, J.B., Reeves, V.B., Follett, R.F., Kimble, J.M., 2002. Mid-infrared and near-infrared diffuse reflectance spectroscopy for soil carbon measurement. Soil Science Society of America Journal 66, 640–646. Mikutta, R., Kleber, M., Kaiser, K., Jahn, R., 2005. Review: organic matter removal from soils using hydrogen peroxide, sodium hypochlorite, and disodium peroxodisulfate. Soil Science Society of America Journal 69, 120–135. Paul, E.A., Follett, R.F., Leavitt, S.W., Halvorson, A., Peterson, G.A., Lyon, D.J., 1997. Radiocarbon dating for determination of soil organic matter pool sizes and dynamics. Science Society of America Journal 61, 1058–1067. Siregar, A., Kleber, M., Mikutta, R., Jahn, R., 2005. Sodium hypochlorite oxidation reduces soil organic matter concentrations without affecting inorganic soil constituents. European Journal of Soil Science 56, 481–490. Six, J., Merckx, R., Kimpe, K., Paustian, K., Elliott, E.T., 2000. A reevaluation of the enriched labile soil organic matter fraction. European Journal of Soil Science 51, 283–293. Six, J., Guggenberger, G., Paustian, K., Haumaier, L., Elliott, E.T., Zech, W., 2001. Sources and composition of soil organic matter fractions between and within soil aggregates. European Journal of Soil Science 52, 607–618. Six, J., Conant, R.T., Paul, E.A., Paustian, K., 2002a. Stabilization mechanisms of soil organic matter: implications for C-saturation of soils. Plant and Soil 241, 155–176. Six, J., Callewaert, P., Lenders, S., De Gryze, S., Morris, S.J., Gregorich, E.G., Paul, E.A., Paustian, K., 2002b. Measuring and understanding carbon storage in afforested soils by physical fractionation. Soil Science Society of America Journal 66, 1981–1987. Sohi, S.P., Mahieu, N., Arah, J.R.M., Powlsen, D.S., Madari, B., Gaunt, J.L., 2001. A procedure for isolating soil organic matter fractions suitable for modeling. Soil Science Society of America Journal 65, 1121–1128. Viscarra-Rossel, R.A., Walvoort, D.J.J., McBratney, A.B., Janik, L.J., Skjemstad, J.O., 2006. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131, 59–75. Willimas, P., Norris, K., 1987. Near-Infrared Technology in the Agricultural and Food Industries. American Association of Cereal Chemists, St. Paul, MN.