European Journal of Soil Science, February 2010, 61, 1–13 doi: 10.1111/j.1365-2389.2009.01199.x
Predicting pasture root density from soil spectralreflectance: field measurement
B . H . K U S U M Oa,c , M . J . H E D L E Ya , C . B . H E D L E Ya,b , G . C . A R N O L D†,a,b
& M . P . T U O H Ya
aSoil and Earth Sciences, Institute of Natural Resources, College of Sciences, Massey University, Private Bag 11-222, Palmerston North,New Zealand, bLandcare Research, Private Bag 11-052, Palmerston North, New Zealand, and cDepartment of Soil Science, Faculty ofAgriculture, University of Mataram, Jl. Majapahit 62, Mataram 83125, Lombok, Indonesia
Summary
This paper reports the development and evaluation of a field technique for in situ measurement of rootdensity using a portable spectroradiometer. The technique was evaluated at two sites in permanent pasture oncontrasting soils (an Allophanic and a Fluvial Recent soil) in the Manawatu region, New Zealand. Using amodified soil probe, reflectance spectra (350–2500 nm) were acquired from horizontal surfaces at three depths(15, 30 and 60 mm) of an 80-mm diameter soil core, totalling 108 samples for both soils. After scanning,3-mm soil slices were taken at each depth for root density measurement and soil carbon (C) and nitrogen(N) analysis. The two soils exhibited a wide range of root densities from 1.53 to 37.03 mg dry root g−1 soil.The average root density in the Fluvial soil (13.21 mg g−1) was twice that in the Allophanic soil (6.88 mg g−1).Calibration models, developed using partial least squares regression (PLSR) of the first derivative spectra andreference data, were able to predict root density on unknown samples using a leave-one-out cross-validationprocedure. The root density predictions were more accurate when the samples from the two soil types wereseparated (rather than grouped) to give sub-populations (n = 54) of spectral data with more similar attributes.A better prediction of root density was achieved in the Allophanic soil (r2 = 0.83, ratio prediction to deviation(RPD) = 2.44, root mean square error of cross-validation (RMSECV) = 1.96 mg g−1) than in the Fluvial soil(r2 = 0.75, RPD = 1.98, RMSECV = 5.11 mg g−1). It is concluded that pasture root density can be predictedfrom soil reflectance spectra acquired from field soil cores. Improved PLSR models for predicting field rootdensity can be produced by selecting calibration data from field data sources with similar spectral attributes tothe validation set. Root density and soil C content can be predicted independently, which could be particularlyuseful in studies examining potential rates of soil organic matter change.
Introduction
Roots contribute to soil organic matter (SOM) content via growthand decomposition (Trujillo et al., 2006). The amount of SOM inpastoral soil is closely related to net productivity of roots, which isdetermined by previous soil and pasture management history (Nieet al., 1997; Cullen et al., 2006) and type and stage of growth ofthe pasture (Fisher et al., 2007). Land slope and aspect may influ-ence availability of water and soil fertility, which in turn influencepasture species and root density (Saggar et al., 1999) and SOMcontent.
Received 13 January 2009, revised version accepted 11 September 2009
Visible-near infrared reflectance spectroscopy (Vis-NIR) hasbeen used to quantify root density in glasshouse studies (Kusumoet al., 2009) and soil carbon (C) and nitrogen (N) content in lab-oratory (Chang & Laird, 2002; Moron & Cozzolino, 2002) andfield studies (e.g. Mouazen et al., 2007; Kusumo et al., 2008). Ifthis technique can be used to quantify root density independentlyfrom soil C and N content, it will be particularly useful for mon-itoring root-induced SOM dynamics. Vis-NIR spectroscopy doesnot provide a direct independent measure of root density, or ofany soil property. Acquired spectral properties must be calibratedagainst standard reference measurements (e.g. soil C content, orin our case, root density measured by separating roots from soilby wet sieving). Other researchers have applied calibration mod-els developed from soil samples studied in controlled conditionsin the laboratory to predict soil constituents of interest in fieldsamples (Mouazen et al., 2007) but have not obtained satisfactory
prediction of total soil C and organic C content by using the labo-ratory calibration models. Often, differences in soil moisture andstructural condition confound such an application.
The potential to predict root density from the Vis-NIRSspectra of soil was demonstrated by Kusumo et al. (2009) usingglasshouse-grown ryegrass plants. In the present paper, the methodis modified to acquire reflectance spectra from soil cores taken inthe field. This modified method is evaluated as a rapid techniquefor predicting root density in field soils.
In our previous glasshouse study (Kusumo et al., 2009), soilchromophores other than roots, such as SOM, iron oxides, clayminerals and aggregate surfaces, were homogenized by usingsieved and well-mixed soils in pots. Under field conditions, thesechromophores change significantly with distance and soil depth(Dematte et al., 2004). Even in the same soil type, the dynamicnature of soil chromophores will influence the range of informa-tion acquired in the reflectance spectra and possibly influence theprediction accuracy for root density. If field root density can bepredicted using reflectance spectroscopy, the technique would beextremely valuable in support of studies measuring the potentialfor C sequestration from fine root turnover, or, studies evaluatingnew deep-rooting cultivars being developed for more efficient Nand water use (Dunbabin et al., 2003; Crush et al., 2007).
The first objective of the present study was to test whether acalibration model developed using soil spectral and root densitydata from glasshouse-grown ryegrass plants (Kusumo et al., 2009)could be used to predict field root density in pastoral soils.The second objective was to test whether the prediction of fieldroot density in pastoral soils can be improved by developingcalibration models from field-acquired spectra and reference fieldroot densities.
Materials and methods
Site locations and soil property measurement
The root density assessment was evaluated at two permanentpasture sites (23.5 km apart): the first on Ramiha silt loam (Allo-phanic soil, Andic Dystrochrept derived from loess and andesiticash) and the second on Manawatu fine sandy loam (Fluvial Recentsoil, Dystric Fluventic Eutrochrept derived from greywacke allu-vium) in the Manawatu region, New Zealand (Hewitt, 1998). Thepermanent ryegrass (Lolium perenne L) and white clover (Tri-folium repens L) dominant pastures at both sites had been presentfor more than 20 years. A total of 18 soil cores, 10 m apartfrom each other, were collected from each site and cut into threedepths (15, 30 and 60 mm), providing 54 samples at each siteand a total of 108 soil samples. Malley & Martin (2003) statedthat at least 50 samples are required for developing initial NIRcalibration, although several other researchers have used fewerthan this; Shibusawa et al. (2001) used 15 samples for example.Soil reflectance spectra were acquired at three horizontal surfaces(15, 30 and 60 mm) cut through an 80-mm diameter soil core(Figure 1c–e). A 3-mm soil slice (slice A) was collected at eachsurface, put in a sealed plastic bag and stored as field moist at
4◦C for no more than 3 days before root densities were measured(Figure 1f) using the wet sieving method (Kusumo et al., 2009).Root density was expressed as mg dry root g−1 dry soil.
A further 3-mm slice (slice B) was collected and used fordetermining dry weight and moisture content (see Figure 1c).Water content of subsample slice B was determined by using thegravimetric method of drying at 105◦C in an oven until reachinga constant weight. Forty-two and 36 Ramiha and Manawatu soilsamples were analysed for total C and N on an air-dry subsampleof slice B with a Leco FP-2000 CNS analyser (LECO Corp.,St Joseph, MI, USA). Chemical properties, namely pHH2O (ratioof soil to water 1:2.5), P-retention, Olsen-P, cation exchangecapacity (CEC), total C and total N, and physical properties,namely bulk density and texture, of the two soils were determinedon a composite sample (six cores per sample) taken from the 0 to100 mm soil depth (Blakemore et al., 1987), with three replicatecomposite samples per site.
Principal component analysis
Prior to partial least squares regression (PLSR) analysis, aprincipal component analysis (PCA) was conducted on the firstderivative of the spectral data. A score plot of the first twocomponents (PC1 and PC2) of the PCA analysis was used toexplore the pattern of spectral differences between the Ramihaand Manawatu samples.
Developing a calibration model
Diffuse spectral reflectance of each freshly cut soil surface wasrecorded using a purpose-built soil probe (Figure 1a–e). Theprobe was developed from a commercially available plant contactprobe supplied by Analytical Spectral Devices (ASD), Colorado,USA. A greater intensity halogen lamp (parabolic reflector, 4.5 W)was fitted to the probe to replace the original lamp (elliptic reflec-tor, 4.5 W). A round casing was developed to avoid direct contactof the quartz window with soil and to provide a fixed distance(30.5 mm) between the soil surface and the sensor head. Duringmeasurement the soil core was rotated and the field of view ofthe sensor was 561 mm2 (Figure 1a,e).
The spectral data were pre-processed (for details see Kusumoet al., 2009) and first derivatives of 5-nm-spaced data calculatedusing SpectraProc V 1.1 software (Hueni & Tuohy, 2006). Thefirst derivative data were imported to Minitab 14 (MINITAB Inc.,2003) for PLSR analysis. Calibration models were developed byusing PLSR to fit the reference data (root density, soil C and N)to pre-processed spectral data. PLSR has been commonly used tobuild calibration models because it can deal with multicollinearity(inter-correlation) of hyperspectral data by reducing the dimen-sionality of the data into several components (latent variables)that are not correlated with each other (Miller & Miller, 2005).This multivariate analysis can also be used to build models fromthe whole Vis-NIR spectral data set (independent variables) as it isusually larger than the analytical data set (response variables). Theresulting regression models were then used to predict root density,
Figure 1 (a) Modified soil probe used (e) to collect reflectance spectra from (b and c) a soil core, and then (d) the soil slice (f) was washed to separateroot from soil using a wet-sieving method.
soil C and N in unknown samples. The accuracy of the models was
tested internally by using a leave-one-out cross-validation proce-
dure and externally using a separate validation sample set (only
on root density data). Separate calibration and validation sets were
created by ranking the soil samples from the smallest to the largest
root density, and odd and even ranked numbers were allocated
to calibration and validation set, respectively. This resulted in a
1:1 ratio of calibration and validation sets. The sample numbers
used for the separate calibration: validation sets were 53:53, 26:26
and 26:25 for amalgamated Ramiha and Manawatu soil, separate
Ramiha soil and separate Manawatu soil, respectively (Table 2).
The number of latent variables used to develop calibration modelswere those that produced the smallest predicted residual error sumof squares (PRESS) in the leave-one-out cross-validation proce-dure (Miller & Miller, 2005). During PLSR processing, samplesthat had a standardized residual >2.0 were removed as outliers(MINITAB Inc., 2003) from the calibration and validation sets.
Predicting field root density using the calibration modelconstructed from glasshouse data
Different root densities were created (Kusumo et al., 2009) bydifferential N and phosphorus (P) fertilizer addition to Italian
ryegrass (Lolium multiflorum Lam.) grown in pots of Ramiha andManawatu soil in a glasshouse; these were the same soil types asused in the field study. A PLSR calibration model was developedfrom the first derivative spectra and root density data from theglasshouse-grown ryegrass plants. This calibration model was thenused to predict root density in the present study. The root densityunit used in the glasshouse study was mg dry root cm−3 and usingpot bulk densities this was converted to mg dry root g−1 dry soilto make it compatible with the measurements made in the fieldstudy. The conversion equation is shown as follows:
Weight of dry root in mg
g soil= Weight of dry root in mg
cm3soil× 1
ρb
, (1)
where ρb is bulk density in each pot, which is presented as gramdry soil cm−3 soil.
Regression model accuracy. The ability of PLSR models topredict soil properties was assessed using the following statistics.Root mean square error (RMSE) is the standard deviation of thedifference between the measured and the predicted soil propertyvalues and that calculated from cross-validation is RMSECV, andthat from the validation data is RMSEP. The ratio of prediction todeviation (RPD) is the ratio of the standard deviation of measuredvalue of soil properties to the RMSE. The ratio error range (RER)is the ratio of the range of measured values of soil properties tothe RMSE. The best prediction model is shown by the largestRPD, RER and r2 and the smallest RMSECV or RMSEP (formore detailed explanation see Kusumo et al., 2009).
Results and discussion
Summary of pasture root density, and soil chemical andphysical data
The range of measured pasture root densities and soil chemical andphysical properties of Ramiha silt loam and Manawatu fine sandyloam soils are presented in Table 1. The mean of root density ateach depth in Ramiha soil was less than in Manawatu soil, exceptat the 60-mm depth. Differences in root density may have resultedfrom the different characteristics of each soil. For example, theManawatu soil has a fine sandy loam texture with a large availableP status (for Olsen-P see Table 1), poor P retention and greaterconcentration of exchangeable cations than the silt-loam textured,P-retentive Ramiha soil. Mechanical impedance to root growthwas unlikely in both soils, which had bulk densities below thecritical level (>1.4 g cm−3) that restricts root growth.
Root density decreases significantly with depth, both in theRamiha and the Manawatu soil (Table 1), except at the 30-and 60-mm depths in the Ramiha soil, where the root densitieswere not significantly different. If the average amount of rootmass in the 0–60 mm depth is expressed per hectare, the totalroot dry mass was 3642 and 9353 kg DM ha−1 in the Ramiha(Andic Dystrochrept) and the Manawatu soils (Dystric FluventicEutrochrept), respectively, consistent with the result reported byNie et al. (1997).
Table 1 Root density, and chemical and physical properties of Ramihaand Manawatu soils
Properties Ramiha Manawatu
Root density / mg g−1 (54 samples of eachof Ramiha and Manawatu soils)
Minimum 2.27 1.53Maximum 23.10 37.03Mean 6.82 13.21Standard deviation 4.86 9.95Coefficient of variation (CV) 0.71 0.75Mean root density (mg g−1 each depth)at 15-mm depth 12.39 22.51at 30-mm depth 4.85 14.49at 60-mm depth 3.20 2.63Total C / % (42 Ramiha and 36 Manawatu
samples)Minimum 6.19 2.13Maximum 10.19 4.47Mean 8.41 3.17Standard deviation 0.88 0.59Coefficient of variation (CV) 0.10 0.19Total N / % (42 Ramiha and 36 Manawatu
samples)Minimum 0.51 0.24Maximum 0.73 0.43Mean 0.61 0.32Standard deviation 0.05 0.05Coefficient of variation (CV) 0.08 0.14Water content / % (54 samples of each
Ramiha and Manawatu soils)Minimum 50.70 24.30Maximum 70.04 38.60Mean 57.11 31.24Standard deviation 5.31 3.86Coefficient of variation (CV) 0.09 0.12Soil (0–100 mm)pH 4.7 4.9P-Olsen / μgP g−1 18.9 71.1P retention / % 73 14CEC / molc kg−1 29.5 15.0Total C / % 7.38 3.18Total N / % 0.59 0.31Bulk density / g cm−3 0.89 1.18Texture Silt loam Fine sandy loam
Decreasing amounts of soil C with depth were also found inboth soils (Table 1). Interestingly, Ramiha soil with smaller rootdensities contained larger amounts of soil C. This is probablybecause SOM decomposition was inhibited by complexing withthe amorphous clay mineral allophane (Boudot et al., 1988), whichis abundant in this soil (Theng et al., 1986). The finer textureof the Ramiha soil also presents more surface area for clay-organo complexes to be formed (Theng et al., 1986) and protectsthe organic matter from decomposition in smaller pores or soilaggregates (Six et al., 2000; Bossuyt et al., 2002).
Figure 2 Spectra acquired from soils with different root densities from the glasshouse trial (Kusumo et al., 2009) and from soil cores collected at differentdepths (15, 30 and 60 mm) in the field; (a) Ramiha and (b) Manawatu soil. RD and WC are root density and water content, respectively.
Spectral reflectance of Ramiha and Manawatu soil
The reflectance spectra collected from the glasshouse trial(Kusumo et al., 2009) and their pattern, which changed with soildepth for each soil, are presented in Figure 2. Spectra were influ-enced not only by roots but also by other soil chromophores (e.g.
parent material, clay and non-clay minerals, iron oxides, waterand SOM) (Baumgardner et al., 1985; Ben-Dor, 2002). In theRamiha soil (Figure 2a), the spectra collected in the glasshouse atsmall root density depicted a much sharper angle at approximately780 nm because of increased reflectance near the yellow-red bandregion (600–750 nm), probably caused by reflected colour from
iron oxides. The effect of iron oxides can also be seen at theband near 950 nm; the spectrum was slightly concave as the ironoxides absorb more NIR light in this band region (Dematte &Garcia, 1999; Dematte et al., 2004). In the field soil spectra, thegreater root content masks the effect of iron oxides, and the anglenear 780 nm becomes more obtuse (Figure 2a). No obvious fea-tures of iron oxides were found in the Manawatu soil because ofits naturally small iron oxide content (Figure 2b). For the Man-awatu soil, the shapes of glasshouse-soil spectra and field-soilspectra (low root density) were similar. The effect of roots onthe spectral reflectance can be seen clearly from the surface sam-ples (15-mm depth), which had a more obtuse angle and tended tohave greater reflectance in the near infrared region 780–2500 nm,which is more pronounced in the Manawatu (Figure 2b) than inthe Ramiha soil (Figure 2a). Spectra from mixtures of dry soiland ground root (Kusumo et al., 2009) also had greater reflectancewith increasing root content. Air voids contained in the root andair-root cell interfaces probably caused the increased reflectancein the NIR region (Baumgardner et al., 1985). In addition, spectralreflectance from the 15-mm depth sample tended to resemble thespectral reflectance of senescent leaves (Gausman et al., 1975).They also resembled minimally decomposed soil material (fibric)reported by Stoner & Baumgardner (1981).
Principal component analysis
A principal component analysis was used to transform the firstderivative spectral data set into a smaller set of linear combina-tions, or principal components, that represent as much of the vari-ance of the data set as possible (Dillon & Goldstein, 1984; Miller& Miller, 2005). The first two principal components (PC) repre-sented the largest variance of the spectral data. The score plot ofthe first two PCs (Figure 3a) calculated from spectra acquired fromsamples of the two soils clearly differentiated between the spec-tral properties of the Ramiha and Manawatu soils. Chromophores,other than roots, such as clay type, metal oxides, moisture andSOM, may contribute to this obvious discrimination. PC1, whichcontains the largest spectral variance (33.7%), was not correlatedwith changes in root density (Figure 3b; r = −0.13) when soilsamples from both soils were considered, but samples with simi-lar root densities tended to have a vertical distribution in the scoreplot, showing that PC2 is correlated with density (Figure 3a). Byseparating the soils and thus removing the variance from soil type,root density becomes a source of variance that was better corre-lated with the first principal component (PC1) (Figure 3b).
If the soil property of interest causes the major variance in thespectral data, then the PCA of spectral data followed by a PC scoreplot analysis can be useful for choosing the sample population tobe used in the PLSR model calibration. However, such preliminaryanalysis is of little use if the soil property of interest causesonly small variance in the spectral data. PLSR analysis, however,selects the component that is most strongly correlated to a soilproperty, providing useful predictions even if that component isnot a large part of total soil variance (Kusumo et al., 2009). In
the case of the data presented in Figure 3a, the PCA analysis andthe subsequent correlation (Figure 3b) of principal componentsagainst root densities indicate that variance from root density isallocated to PC1 if the two soils are separated, and allocated toPC2 if the two soils are grouped.
Can the glasshouse calibration model be used to predictfield root density?
The best calibration model that was developed from the glasshousestudy (Kusumo et al., 2009) was used to predict root densityfrom acquired spectral data in the present study. Despite the soilmaterials used in the glasshouse study being sampled from thesame sites and depths as those used in the field study, field rootdensities were poorly predicted, which is shown by a very smallcoefficient of determination (r2 = 0.35), large error (RMSEP =10.32 mg g−1) and very poor RPD (0.82) (see Figure 4). Theglasshouse calibration model tended to under-estimate field rootdensity values, which is shown by poorer root density prediction(Figure 4). The poor prediction of the glasshouse calibrationmay be explained partly because the range and variance of theglasshouse root densities (range 0.60–5.46 mg g−1, variance 1.78)did not cover the field observations (range 1.53–37.03 mg g−1,variance 71.08). Glasshouse data have a smaller range andvariance compared with field root density data. It is well knownthat the range of the calibration set should cover the rangeof the validation set (Williams, 2001; Malley & Martin, 2003;Westerhaus et al., 2004) to get the best calibration.
The PC1 and PC2 score plot of the PCA of the first derivativespectra acquired from the field and glasshouse data (Figure 5)showed that the spectral properties of the glasshouse calibrationset (filled and open circles) do not embrace the spectral varianceof the validation set, the field spectral data (filled and opendiamonds). These spectral differences are caused by a combinationof different water contents of the glasshouse soils (Ramiha soil,52.0–69.5%; Manawatu soil, 38.3–52.4%) and field soils (Ramihasoil, 50.7–70.0%; Manawatu soil, 24.3–38.6%) and the non-homogenous nature and different structure and colour of the fieldsoils. Ideally, the scattered distribution of the validation set shouldalso be within the distribution of the calibration set (Miller &Miller, 2005). Better prediction may be obtained if the distributionpatterns between calibration and validation sets on the score plotwere more similar (Esbensen et al., 2006), and both sets shouldembrace the same variance dimensions (Williams, 2001).
Can pasture root density in the field be predicted from modelsestablished with field-acquired spectral reflectance?
Figure 6 illustrates the relationships between measured field rootdensity (from wet sieving) and the root density predicted fromthe field-acquired spectra. The sample populations and regressionstatistics are shown in Table 2. PLSR of the first derivative spec-tra and reference data produced good calibration models forthe separate soil types, which are shown by large coefficients
Figure 3 A score plot of the first and second components from the PCA of the first derivative of the reflectance spectra acquired from (a) combined Ramihaand Manawatu soils and (b) coefficients of correlation between root density and the principal components when the two soil data sets were combined andseparated.
of determination (r2 > 0.99) and small errors (RMSEC = 0.43and 0.93 mg g−1, Table 2). Weaker calibration relationships (r2 =0.86; RMSEC = 3.09 mg g−1) were found when samples fromboth soil types were amalgamated (Ramiha and Manawatu). Obvi-ous separation between Ramiha and Manawatu soils on the PCAscore plots (Figures 3a and 5) and better correlation between root
density and scores of PC1 when the two soils were separated
(Figure 3b), support development of the calibration models for
separate soils (Table 2). Firstly, because of the small number of
samples involved (less than 60 samples per soil), a leave-one-
out cross-validation procedure was used to test the two models
(Williams, 2001). More accurate cross-validation predictions were
Figure 4 Field root density predicted by theglasshouse PLSR calibration model com-pared with the measured field root densityfor Ramiha and Manawatu soils.
Figure 5 A score plot derived from the principlecomponent analysis of the spectral reflectancedata obtained from the glasshouse study and fieldstudy of Ramiha and Manawatu soils.
also found when samples from each soil type were separated rather
than grouped (Table 2). More accurate root density prediction
was found for Ramiha (RPD = 2.44; r2 cross-validation = 0.83)
than for Manawatu soil (RPD = 1.98; r2 cross-validation = 0.75).
Malley et al. (2004) categorized prediction accuracy into moder-
ately useful (in Manawatu soil) when RPD = 1.75−2.25 (with
RER = 8−10 and r2 = 0.70−0.80) and moderately successful
(in Ramiha soil) when RPD = 2.25−3 (with RER = 15−20 and
r2 = 0.80−0.90).
The poorer correlation between PC1 of the spectral data
and measured root density for the Manawatu soil (r = 0.748)
(Figure 3b) than for the Ramiha soil (r = −0.812), provides
evidence that the spectral data were a weaker predictor of root
density in the Manawatu soil. In PLSR analysis, components
that have good correlation with measured soil properties are
given extra weight and more predominantly determine the
prediction quality (Miller & Miller, 2005). The leave-one-out
cross-validation procedure fits n PLSR calibration models for each
sample of n data. The result is that the Manawatu PLSR leave-
one-out calibration models produce predicted root-density data
with a wider range and variance from the measured root density
(see Figure 6 and Table 2) than the Ramiha PLSR leave-one-out
calibration models. This is why leave-one-out cross-validation
produced a series of calibrations that together explain less of
Figure 6 Relationships between root density measured by wet sieving and predicted from soil spectral reflectance created by using PLSR calibration andleave-one-out cross-validation data (left), and separate PLSR calibration and validation data sets (right) of separated (above and middle) and combined(below) Ramiha and Manawatu soils data. The ratio of allocating data to the calibration and validation data sets was 1:1 (Table 2).
the variance in the root density than the non-validated single
calibration model for each soil (r2 both >0.99, Table 2).
When the Manawatu and Ramiha data sets were each divided
into separate calibration and validation sets (25–26 data points),
the calibration models produced ‘moderately’ accurate prediction
of root density (r2 = 0.8 and 0.92, respectively, Table 2 and
Figure 6), despite the small sample numbers in the external
Table 2 Calibration, cross-validation and validation of spectral and root-density data using PLSR when observations from Ramiha and Manawatu fieldpasture soils are combined and separated
Calibration values
Soil type Allocated samples Outliers n Comp r2 RMSECV
n = number of samples used; Comp = components (factors or latent variables); RMSEC, RMSECV and RMSEP in mg dry root g−1 soil; Bias = meandifference between measured and predicted root density; Slope = regression coefficient between measured and predicted root density.
Is root density predicted independently from soil C?
The interaction of NIR light with organic materials containingbonds such as N-H, C-H and O-H allows spectral reflectance to beused to predict organic materials including roots (Kusumo et al.,2009). Spectral reflectance collected from pasture soils is influ-enced by root attributes and decomposed SOM, both of whichdecreased with soil depth (Figure 2). Both root density (r2 = 0.83and 0.92, Table 2) and soil C (r2 = 0.86, Table 3) were predictedaccurately with the NIRS technique in the Ramiha soil (Figure 7b).Poor correlations existed, however, between measured root densityand total C in the Ramiha soil (r2 = 0.24, Figure 7a), indicatingthat root density and soil C were predicted independently by thePLSR of spectral data. Closer examination of the PLSR modelcoefficients shows that different wavebands were given differentlevels of importance in the regression models used to predict rootdensity and soil C (Figure 7c). Large PLSR coefficients for soil Cat wavelengths between 730 and 940 nm indicate that these wave-lengths were important for soil C prediction in the Ramiha soil.
Soil N content was also predicted with reasonable accuracy(r2 = 0.72, Table 3) in the Ramiha soil, despite a poor correlationbetween soil C and N (r2 = 0.38). Independent prediction betweensoil C and N has also been reported by Chang & Laird (2002). Inthe Manawatu soil, however, strong correlations existed betweenroot density and soil C (r2 = 0.73) and between soil C and N(r2 = 0.83). Because they are correlated in the Manawatu data,it is not possible to show that root density can be predictedindependently of change in SOM (Tables 2 and 3).
PLS regression coefficients, important bands for root density
Coefficients resulting from PLSR models for predicting rootdensity are presented graphically in Figure 8. The size of the
coefficient (negative or positive) in each model represents theimportance of each wavelength band in explaining the variation inroot density. Both calibration models had differences in selectionof wavelengths that varied with root density. These differencesbetween soils are to be expected because parts of the spectraare chromophores, other than roots, that are unique to each soil.Roots, an ‘intrusive’ feature of soils, will partially obscure theseunique chromophores. Other regions in the spectra are deriveddirectly from the root material. Interestingly, both models selectedthe visible band at 600 nm (yellow-orange) as an importantband, which is similar to the previous glasshouse study (Kusumoet al., 2009), where this band was associated with reflectancefrom ryegrass roots. Some other important bands in the nearinfrared region in describing ryegrass root density are presentedin Figure 8.
Conclusions
The inability of the glasshouse data to cover the same range ofmeasured root densities and spectral reflectance characteristicsimplies that the most robust regression models for predictingpasture root density in the field should be developed fromfield study rather than from glasshouse samples. An exceptioncould be when measured root densities and spectral reflectancecharacteristics for glasshouse and field samples occupy the samerange and variance dimensions.
Our results show that Vis-NIR spectroscopy can be used forprediction of pasture root density in the field. Accurate PLSRcalibration models can be developed from a small calibration setof field-acquired spectra from a soil slice with known referenceroot density data. The calibration models can predict root densitiesfrom a larger population of acquired spectral data. This approach
Figure 7 Relationships between (a) measured root density and soil C, (b) measured soil C and predicted soil C and (c) the coefficients of the PLSR modelsused to predict root density (RD) and % C of the Ramiha soil in the field.
Table 3 The partial least squares regression calibration and cross-validation of spectral and soil C and N data from Ramiha and Manawatu field pastoralsoils
Comp = components (factors or latent variables); RMSEC and RMSECV in %; Bias = mean difference between measured and predicted soil C (or N);Slope = regression coefficient between measured and predicted soil C (or N).
Figure 8 Coefficients of the PLSR models (leave-one-out cross-validation) for predicting root density in Ramiha and Manawatu pastoral soils.
can reduce the number of samples that must be measured for rootdensity, especially when dealing with a large sample set (e.g. formapping purposes). To improve the accuracy, separate calibrationmodels for different soil types should be explored.
In addition, root density can be predicted independently fromsoil C, which suggests that soil spectral reflectance will beparticularly useful in studies examining potential rates of SOMsynthesis from root production. Further research is required toimprove the accuracy of this technique. Improvements may bepossible by enlarging the field of view of the spectrometer and byensuring that the soil sample scanned by the sensor represents thesame sample used for the reference measurement of root densityor C and N concentration.
Acknowledgements
We thank Andreas Hueni for creating the software (SpectraProc)for the spectral data pre-processing.
References
Baumgardner, M.F., Silva, L.R.F., Biehl, L.L. & Stoner, E.R. 1985.Reflectance properties of soils. Advances in Agronomy , 38, 1–44.
Ben-Dor, E. 2002. Quantitative remote sensing of soil properties.Advances in Agronomy , 75, 173–243.
Blakemore, L.C., Searle, P.L. & Daly, B.K. 1987. Methods for Chemical
Analysis of Soils. New Zealand Soil Bureau Scientific Report: 80.
Department of Scientific and Industrial Research, Lower Hutt.
Bossuyt, H., Six, J. & Hendrix, P.F. 2002. Aggregate-protected carbon
in no-tillage and conventional tillage agroecosystem using carbon in
labelled plant residue. Soil Science Society of America Journal , 66,
1965–1973.
Boudot, J.P., Bel Hadi Brahim, A. & Chone, T. 1988. Dependence of
carbon and nitrogen mineralization rates upon amorphous metallic
constituents and allophanes in highland soils. Geoderma, 42, 245–260.
Arnold, G.C. 2008. The use of diffuse reflectance spectroscopy for insitu carbon and nitrogen analysis of pastoral soils. Australian Journalof Soil Research, 46, 623–635.
Kusumo, B.H., Hedley, M.J., Hedley, C.B., Hueni, A., Arnold, G.C. &Tuohy, M.P. 2009. The use of Vis-NIR spectral reflectance fordetermining root density: evaluation of ryegrass roots in a glasshousetrial. European Journal of Soil Science, 60, 22–32.
Malley, D.F. & Martin, P.D. 2003. The use of near-infrared spectroscopyfor soil analysis. In: Tools for Nutrient and Pollutant Management:
Application to Agriculture and Environmental Quality (eds L.D. Currie& J.A. Hanly), pp. 371–404. Fertilizer and Lime Research Centre,Massey University, Palmerston North.
Malley, D.F., Martin, P.D. & Ben-Dor, E. 2004. Application in analysisof soils. In: Near-infrared Spectroscopy in Agriculture, AgronomyMonograph, Vol. 44 (eds K.A. Barbarick, C.A. Roberts, W.A. Dick,J. Workman Jr, J.B. Reeves III & L. Al-Amoodi), pp. 729–784.American Society of Agronomy, Crop Science Society of America, SoilScience Society of America, Madison, WI.
Miller, J.N. & Miller, J.C. 2005. Statistics and Chemometrics for Analyt-ical Chemistry . Pearson Prentice Hall, Harlow.
MINITAB Inc. 2003. MINITAB Statistical Software. Minitab Inc., StateCollege, PA.
Moron, A. & Cozzolino, D. 2002. Application of near infrared reflectancespectroscopy for the analysis of organic C, total N and pH in soils ofUruguay. Journal of Near Infrared Spectroscopy , 10, 215–221.
Mouazen, A.M., Maleki, M.R., De Baerdemaeker, J. & Ramon, H. 2007.On-line measurement of selected soil properties using a VIS-NIR sensor.Soil & Tillage Research, 93, 13–27.
Nie, Z.N., Mackay, A.D., Valentine, I., Barker, D.J. & Hodgson, J. 1997.Influence of pastoral fallow on plant root growth and soil physical andchemical characteristics in a hill pasture. Plant & Soil , 197, 201–208.
Saggar, S., Mackay, A.D. & Hedley, C.B. 1999. Hill slope effects onthe vertical fluxes of photosynthetically fixed 14C in a grazed pasture.Australian Journal of Soil Research, 37, 655–666.
Shibusawa, S., Made Anom, I., Sato, S.W., Sasao, A. & Hirako, S. 2001.Soil mapping using the real-time soil spectrophotometer. In: Third
European Conference on Precision Agriculture (ECPA) (eds G. Granier& S. Blackmore), pp. 497–508. Agro, Montpellier.
Six, J., Paustian, K., Elliot, E.T. & Combrink, C. 2000. Soil structureand soil organic matter. I. Distribution of aggregate size classes andaggregate-associated carbon. Soil Science Society of America Journal ,64, 681–689.
Stoner, E.R. & Baumgardner, M.F. 1981. Characteristic variation ofreflectance of surface soils. Soil Science Society of America Journal ,45, 1161–1165.
Theng, B.K.G., Churchman, G.J. & Newman, R.H. 1986. The occurrenceof interlayer clay-organic complexes in two New Zealand soils. Soil
Science, 142, 262–266.Trujillo, W., Fisher, M.J. & Lal, R. 2006. Root dynamics of native
savanna and introduced pastures in the Eastern Plains of Colombia.Soil & Tillage Research, 87, 28–38.
Westerhaus, M., Workman J. Jr, Reeves III, J.B. & Mark, H. 2004.Quantitative analysis. In: Near-infrared Spectroscopy in Agriculture,Agronomy Monograph, Vol. 44 (eds K.A. Barbarick, C.A. Roberts,W.A. Dick, J. Workman Jr, J.B. Reeves III & L. Al-Amoodi),pp. 133–174. American Society of Agronomy, Crop Science Societyof America, Soil Science Society of America, Madison, WI.
Williams, P.C. 2001. Implementation of near-infrared technology. In:Near-infrared Technology in the Agriculture and Food industries (edsP. Williams & K. Norris), pp. 145–169. American Association ofCereal Chemists, St Paul, MI.
