Effects of soil composition and preparation on the prediction of particle size distribution using mid-infrared spectroscopy and partial least-squares regression.
Soil particle size distribution (PSD) is one of the most useful soil properties provided as a routine soil analysis service. It can be used to determine soil texture, and thus key soil processes affecting soil water dynamics and air diffusion, plant nutrition, microbial activity and nutrient and pollutant mobility. Clay content, in particular, has a large effect on soil structure by promoting the stability of soil aggregates and soil swell-shrink properties (Boivin et at. 2004). However, the standard laboratory method for PSD can be prohibitively expensive for large numbers of soil samples, and even more so for calcareous soils because of the need to remove carbonate.
A more inexpensive but still accurate surrogate method, without the use of reagents, would be a considerable benefit. Infrared spectroscopy, in combination with multivariate regression methods such as the commonly used partial least-squares regression (PLSR), offers such an alternative. In particular, the prediction of PSD using mid-infrared (MIR) spectrometry is based on the sensitivity of the MIR region to soil chemistry, the clay mineralogy associated with clay content and to the larger sand-sized particles dominated by quartz grains (Janik et al. 1998). The spectral response to the silt component is less certain, probably because its composition can range from less well-defined fine to coarse material. The diffuse reflectance Fourier transform (DRIFT)-PLSR technique (Janik et al. 1998, 2007) is rapid and inexpensive, apart from the cost of calibration development, is non-destructive and allows for the simultaneous prediction of multiple analytes. Apart from the advantage of quantitative prediction, interpretation of the spectral peaks responsible for the model allows some understanding of the most important soil properties affecting PSD prediction.
Most of the reported studies involving infrared devices for the prediction of PSD have used the visible-near infrared (vis-NIR) range (Waiser et al. 2007; Viscarra Rosscl et al. 2009; Bricklemyer and Brown 2010; Stenberg et al. 2010; Soriano-Disla et al. 2014). In particular, the study by Stenberg et at. (2010) reported on the effect of rewetting soil samples on particle size and organic carbon prediction. However, DRIFT spectroscopy using the MIR spectral range was also shown to be an effective method to predict PSD (Janik et al. 1998, 2009; Viscarra Rossel and Webster 2012; Soriano-Disla et al. 2014).
An important advantage of MIR spectroscopy is that it provides an interpretable 'fingerprint' of the spectral band shapes and frequencies for the soil. This is usually more difficult in the N1R spectral range because of the less well resolved bands. The soil spectra arc sensitive to bands for soil components that affect the accuracy of laboratory PSD analysis, such as carbonates (e.g. calcitc and dolomite) and the amount and type of soil organic matter (SOM) in the sample (Janik et al. 1998).
The presence of high carbonate content (usually > 10%), such as in Mediterranean and arid region soils (Isbell 2002), can have a major effect on the accuracy of laboratory determinations used to derive the multivariate regression models, and thus the calibration accuracy (Kunze and Dixon 1986; Arroyo et al. 2005). Soil carbonate can form a major soil component and its distribution can occur in unknown proportions throughout the clay, silt and sand fractions. It can also affect the PSD by coating the soil particles (Bowman and Hutka 2002), thus binding smaller aggregates into larger-sized particles. This problem can be alleviated by removing the carbonate, before laboratory analysis, with acid (McKenzie et al 2002). Because of the sensitivity of the MIR for carbonate (Nguyen et al. 1991), and thus the ability of the MIR-DRIFT technique to accurately measure carbonate (Janik et al. 1998), we can hypothesise that it should be possible for the DRIFT-PLSR method to predict PSD in the presence of carbonate, without the need for removing the carbonate before PSD analysis.
Apart from concerns about variations in soil composition and type, there are issues related to sample heterogeneity and particle size. In the first case (intcrparticular heterogeneity), the heterogeneity is due to a spatial variation in particle composition with regard to the sampling area of the infrared beam, exacerbating the variation in replicate spectral scans of the same sample (Stumpe et al. 2011). In other instances of heterogeneity (intraparticle heterogeneity), there can be a change in micro-aggregate composition from the surface to the interior of the sample (Tisdall and Oades 1982). There is also spectral scattering, resulting in the distortion of spectral peaks caused by refractive index effects, especially for diffuse reflectance of larger mineral (silicate) particles (Nguyen et al. 1991). Because of these issues, samples being scanned may need to be fine ground to reduce aggregate sizes and expose the interior of the sample matrix. Thus, there is a need to understand the relationship between the effects of fine grinding and spectral changes in samples, and how this may affect the DRIFT-PLSR models.
As a step towards implementing a global Australian calibration for PSD, individual soils sets from a range of geographical regions, constituting 'site-specific' calibrations, were used to build a single large, combined, soil set. Large-scale libraries tend to span over a wider range and a higher variability of the soil property under investigation, which actually appears to be the dominating factor affecting prediction errors (Stenberg et al. 2010). The lack of accuracy shown by spectroscopic models built with large-scale spectral libraries is also due to the complexity of the relationship between soil properties and spectra for heterogeneous soil samples.
Although Guerrero et al. (2010) showed that we do not necessarily need large databases to predict local samples, the most important thing is to have the right samples (right variability) in the calibration; in this case, we are building a global model (global library) for the prediction of particle size to account for the different soils across Australia. The accuracy for the prediction of local samples can then be improved by following different approaches as described by Guerrero et al. (2010). However, in many situations, such local site-specific data arc unavailable.
The aims of the present study were to assess the feasibility of deriving DRIFT-PLSR models from large archival calibration sets, to predict PSD and to examine the effects of sample grinding, soil composition and carbonate content on prediction accuracy.
Materials and methods
Four sets of soils and data (their locations are shown in Fig. 1) were used in the present study: 500 samples from throughout Queensland (Qld; north-eastern Australia), 674 samples from New South Wales (NSW; central eastern Australia) and 135 samples, provided by the Analytical Chemistry Unit (ACU) at CSIRO (Waite Campus, Urrbrae), from throughout southern Australia (the soil set, calcareous and non-calcareous soils from South Australia, is denoted as ACU). A set of 50 calcareous soils was sampled from the Eyre Peninsula (EP; south-western South Australia). Soils varied in classification and were derived mostly from the following Food and Agriculture Organization of the United Nations (FAO) soil units: Acrisols, Calcisols, Cambisols, Fcrralsols, Luvisols, Podzols, Solonchaks, Solonetz and Vertosols.
Soil laboratory analysis
Particle size distribution data for Qld and NSW were provided by the respective departments of natural resource management laboratories according to their standard methods (Qld Department of Environment and Resource Management Laboratory Method 2Z2; NSW Office of Environment and Heritage Laboratory Method P7B). The ACU and EP soils were oven dried at 40[degrees]C, sieved to <2 mm and then dispersed with a reciprocating shaker (in water containing sodium hexametaphosphate) before being pretreated with hydrogen peroxide to remove soil organic matter, followed by sedimentation. Subsamples of the <2-mm soils were scaled in a desiccator before infrared analysis.
The ACU and EP samples were sonicated to ensure any aggregates were dispersed and pretreated with hydrogen peroxide to remove soil organic matter. Carbonate concentrations were determined by using a pressure 'calcimeter' (Sherrod et al. 2002) and, if carbonate was detected, the soil was treated with glacial acetic acid to remove the carbonate, washed to remove any soluble salts and dispersed before being transferred to sedimentation cylinders for PSD determinations. Determinations of clay and silt contents for the ACU and EP soils used the pipette method (USDA Natural Resources Conservation Service National Soil Survey Center 1996; McKenzie et al. 2002). Sand fractions were determined gravimetrically, with the coarse sand passed through a 0.20-mm sieve after dispersion. The pretreatment for carbonate content was performed according to McKenzie et al. (2002), but substituting glacial acetic acid for hydrochloric acid to reduce the likelihood of clay decomposition by the stronger acid.
The actual amount of each particle size fraction accounting for carbonate concentration was expressed as is. For example, if there was 80% carbonate present in the sample, the clay, silt and sand fractions only totalled 20% and the rest of the material was deemed to be carbonate. For the Qld and NSW samples, the original PSD data were provided without accounting for the carbonate content (assumed to be negligible). In the few NSW and Qld soils where carbonate was identified from the subsequent infrared spectra, carbonate was predicted using archival MIR-PLSR calibration models (Janik et al. 1998) and PSD normalising to give a sum of mass fraction of soil plus carbonate equal to 100%. A summary of the results of the laboratory determination of PSD for all soil sets is presented in Table 1. Fine grinding to <0.1 mm for infrared analysis was performed with a vibrating steel ring-mill (LabTechnics LM1-P; Analytical Equipment) equipped with a 45-mm diameter, 440-g steel puck for 60 s.
All samples were scanned for 60 s with a bcnchtop Spectrum One (PE) spectrometer (Perkin Elmer), equipped with an AutoDiff diffuse reflectance accessory (Perkin-Elmer), in the frequency range 7800^150 [cm.sup.-1] and at a resolution of 8 [cm.sup.-1]. Approximately 70 mg of each sample was required for DRIFT scanning. The scans were arranged in sets: Qld, NSW, ACU and EP, with an additional set made up from a composite of the four sets. The sets were further subdivided into <2- and <0.1-mm fine-ground material. The archival samples (Qld, NSW and ACU) were scanned in duplicate and the EP samples were scanned in triplicate and averaged for multivariate analysis. A silicon carbide reference disc (Perkin Elmer) was used as a background (assumed to have a reflectance, [R.sub.0] of 1). The spectra were expressed in pseudo absorbance (A) units, calculated from the reflectance spectra of the sample (7?s), where A - [log.sub.10] ([R.sub.0]/[R.sub.s]). Spectra were imported into
Unscrambler-X VI 0.3 (Canto) and baseline corrected using the Unscramblcr 'De-trend' preprocessing application by fitting a first-order polynomial to each individual spectrum and then removing this from the spectrum.
Multivariate modelling Principal components analysis
Principal component analysis (PCA), trained by crossvalidation of 20 randomly selected sample segments, was performed on a composite of all soil sets. Individual PCA score plots for each of the datasets were then projected onto the composite PCA scores to assess the extent of overlap between datasets. The PCA loadings (measured as absorbance values, arbitrary units) and score distributions were used to discern the spectral features contributing to the spectral variance. For the EP soils, the Kennard-Stone algorithm was used to select the 30 most representative samples EP field samples from the 50 samples collected for laboratory analysis (Kennard and Stone 1969; Soriano-Disla et al. 2014).
Partial least-squares analysis
Before carrying out PLSR analysis, one-third of soil samples from each of the Qld, NSW, ACU and EP sets were randomly selected as 'independent' validation sets. The remaining two-thirds of the samples were used for calibration to train the PLSR models (by cross-validation). The same calibration and validation samples were also used in the combined composite set composed of all samples.
Calibrations for PLSR modelling for the individual datasets and the composite set were derived from the DRIFT spectra (X-predictor variables) and experimental PSD data (Y-dependent variables). The regression statistics for infrared predictions were reported in terms of the coefficient of determination ([R.sup.2]), non-bias corrected root mean square error (RMSE) and the ratio of the standard deviation of the reference values to the prediction error (RPD; Williams 1987). The quality of PLSR for prediction purposes can be roughly ascertained from the RPD: values <1.5 are considered poor, those in the range 1.5-1.9 suggest indicator quality, values 2.0-2.9 suggest good quality and values [less than or equal to] 3.0 indicate analytical quality (Sudduth and Hummel 1996; Janik et al. 1998).
Results and discussion
The soils varied in composition according to their geographic origin, as shown by the PCA loading plots in Fig. 2. Samples from tropical, subtropical and temperate regions (Qld and NSW) were mostly characterised by mixed kaolinitc, smectite, illite clays (peaks between 3695 and 3620 [cm.sup.-1] for kaolinite, 3620 [cm.sup.-1] for smectite and illite and a broad water band at 3600 and 3350 [cm.sup.-1] for smectite) and quartz (peaks between 2000-1820 [cm.sup.-1] a strong peak inversion near 1100-1000 [cm.sup.-1] and two sharp peaks at 800 and 700 [cm.sup.-1] Nguyen et al. 1991; Janik et al. 1998). The Qld and NSW soils were broadly similar in that there were no discernible peaks in the loadings due to carbonate, consistent with the very few calcareous soils in these sets.
In contrast, loadings for the calcareous soils (ACU and EP) were notably different to those of the Qld and NSW soils, particularly in the spectral region characterised by carbonate (calcitc, 2520 and 1810 [cm.sup.-1]) and other frequencies below 2000 [cm.sup.-1]. Although the first two loadings for the NSW and Qld sets were similar, the loading for principal component (PC)-3 in the NSW soils was dominated by kaolinitc and quartz, in contrast with the Qld set where smectite was the dominant clay, as well as weak negative soil organic matter peak near 2930 and 2850 [cm.sup.-1] and a stronger negative peak near 1730 [cm.sup.-1] due to the -COOH functional group. It was found that the PCA loadings for the ACU and EP soils were completely different to those of the Qld and NSW sets in that they were characterised by soil carbonate plus quartz. There was also a contribution from kaolinitc clay. The loading for PC-3 appeared to be strongly influenced by smectite clay.
Soil variability can also be conveniently illustrated by examination of plots of the PCA scores, which represent the variability between soils based on their spectral data. Fig. 3 shows the score plots of the first three PCs for the separate soil sets projected from the PCs of the combined (composite) set.
The PCA scores for the Qld samples (closed circles in Fig. 3a, b) showed a high degree of overlap with those of the combined set. However, there was a slight negative skew along the PC-1 axis, which, according to the first PCA loading plot in Fig. 2, suggested a bias towards lower sand and higher clay contents compared with the combined set. The PC-2 scores were centred onto the composite set, which was characterised by kaolinite clay. The scores for PC-3 were shifted positively with respect to the combined set, again being characterised by kaolinite clay and quartz, but also negatively characterised by carbonate.
There was a tighter grouping of scores for samples along the PC-1 axis in the NSW soil set (Fig. 3c, d), although there were several samples with higher PC-1 scores, suggesting higher sand contents in these samples. The projected scores for PC-3 were slightly skewed towards negative PC-3, suggesting slightly lower kaolinite contents. The scores for the ACU soils were characterised by a strong shift towards positive PC-2 (low kaolinite) and negative PC-3 (high carbonate low kaolinite; Fig. 3e, f). The EP soil scores were mainly located at negative PC-2 (high kaolinite) and PC-3 (high kaolinite) centred around zero, although there were two samples highly negative in PC-3 (Fig. 3g, h), corresponding to carbonate and low kaolinite.
Partial least-squares analysis
Soils < 2 mm sieved
As a step towards implementing an Australian 'global' calibration for PSD (such as using the composite set), the individual soils sets were first analysed separately to demonstrate the effect of varying soil composition or types on modelling accuracy. The PLSR cross-validation and prediction statistics are given in Table 2 for the <2-mm sieved soils.
The cross-validation calibration RPD values for the Qld soils were 2.9 (good accuracy) for clay, 1.6 (indicator accuracy) for silt and 2.6 (good accuracy) for sand. Corresponding [R.sup.2] values were 0.88 for clay, 0.61 for silt and 0.85 for sand. Validation of the Qld calibrations, given in Table 2, resulted in prediction accuracies similar to those of the cross-validation results. The predictions for clay and sand, although lower than for the calibrations, were still considered as good quality for prediction purposes, with [R.sup.2] values of 0.79 and 0.78 respectively, an RMSE of 9% and 12% respectively and an RPD of 2.2 and 2.1 respectively.
The PLSR loading weights (measured as absorbance values, arbitrary units) for factor-1 in the Qld PSD calibration cross-validations (Fig. 4) were essentially controlled by the strong correlation of sand with quartz and for clay by kaolinite. The sharp peaks near 3695-3630 [cm.sup.-1] for the Qld clay calibration were assigned to kaolinite clay and the very broad band from 3500 to 3100 [cm.sup.-1] was assigned to water in clay structures and goethite (near 3100 [cm.sup.-1]; Nguyen et al. 1991). For sand calibration, the two groups of bands near 2000-1840 and 1200-1000 [cm.sup.-1] were attributed to quartz.
A series of strong negative peaks for the Qld clay calibration were observed in the factor-2 loading weights, one near 3520-2450 [cm.sup.-1] due to gibbsite (Nguyen et al. 1991), and positive gibbsite and kaolinite peaks for the sand calibration. The loading weights for silt calibration were less defined, correlating mostly with illite (negative peak near 3620 [cm.sup.-1]; Nguyen et al. 1991) and organic matter alkyl peaks at 2930 and 2850 c[cm.sup.-1].
The PLSR modelling for NSW was essentially controlled by the strong correlation with sand (as quartz) and kaolinite clay. According to Table 2, the regression cross-validation accuracies for clay and sand contents in the NSW set were considered to be of good accuracy, with [R.sup.2] = 0.82, RMSE = 8% and RPD = 2.3 for clay, and [R.sup.2] = 0.81, RMSE= 10% and RPD = 2.3 for sand. The validation prediction accuracies for clay and sand were very similar to those of the calibration cross-validation (see Table 2). This confirmed that there was a sufficient number of samples in the calibration set to overcome any negative effects of any outliers on the performance of the calibration models. The calibration and validation accuracies for silt content were similar, with [R.sup.2] = 0.69-0.67 and RMSE = 6%, but the RPD of 1.8-1.7 suggested indicator accuracy.
The loading weights for factor-1 in the NSW clay calibration (Fig. 4a) showed a positive peak for kaolinite clay (sharp peak near 3695 cm ') and negative quartz peaks (Nguyen et al. 1991; Reeves 2010). The second loading weight (Fig. 4d) was almost entirely due to negative peaks for SOM, with peaks at 3400 [cm.sup.-1] (hydroxyl in SOM and water), weak-[CH.sub.2] peaks near 2930-2850 [cm.sup.-1] and peaks near 1650 [cm.sup.-1] due to the amide group. The first two PLSR loading weights for silt cross-validation (Fig. 4b, e) were less well defined, but were characterised almost entirely by negative quartz peaks, although the broad peak at 250-1100 [cm.sup.-1] was somewhat ambiguous in that it was atypical of quartz. There may have been some contribution to this peak resulting from specular reflectance (Reststrahlen bands) from clay mineral Si-0 vibrations. The first two loading weights for sand (Fig. 4c, f) suggested that the dominant soil composition associated with sand content was mostly due to quartz (first loading weight) and organic matter (second loading weight).
Cross-validation of carbonate in the < 2-mm ACU soil set resulted in high accuracy, with [R.sup.2] = 0.98, RMSE = 24% and RPD = 6.6 (data for carbonate not shown). Validation prediction accuracy for carbonate was of similar accuracy, with [R.sup.2] -0.93, RMSE=19% and RPD = 3.6. The cross-validation accuracy for clay content was also high, with [R.sup.2] = 0.83, RMSE = 7% and RPD = 2.5, but validation accuracy was only of indicator quality, with [R.sup.2] = 0.73, RMSE = 9% and RPD=1.8. One sample (#549) was identified as an outlier and removed from validation. The reduced validation accuracy was unexpectedly low, possibly due to the presence of carbonate as a major interfering constituent in the validation set. Cross-validation and validation prediction accuracies for silt content were poor, with [R.sup.2] = 0.52 and 0.51, RMSE = 7% and RPD =1.4 and 1.2 respectively. Sand content regression was considered good quality, with [R.sup.2], RMSE and RPD cross-validation and validation statistics of 0.87 and 0.77, 28% and 21% and 2.8 and 2.1 respectively.
The loading weights for the ACU soils (Fig. 4) were dominated by peaks due to kaolinite, quartz and carbonate, suggesting that the PSD data were driven predominantly by these minerals. The factor-1 loading weights for clay showed positive peaks for kaolinite clay (sharp peaks at 3695 and 3630 [cm.sup.-1]) and negative quartz peaks (2000-1800 [cm.sup.-1]). The factor-2 loading weight for clay content was due, almost entirely, to negative carbonate contribution. The cross-validation loading weights for silt were also mostly due to kaolinite and carbonate. As expected, the first sand loading weight was dominated by quartz peaks, although there was a significant negative correlation with carbonate, whereas the second loading weight was dominated by kaolinite and carbonate.
Because of the very small size of the EP dataset (only 30 samples with analytical data and, of these, 10 samples allocated for validation), a high degree of caution should be exercised in the interpretation of these PLSR results. Models determined from such small datasets arc extremely susceptible to outlier samples with high PLSR leverage, and can result in either too high or too low regression accuracy. Three such outlier calibration samples were identified: Samples #11, #5 and #13. These samples were closely monitored or omitted where required during PLSR modelling.
The cross-validation accuracy for carbonate in the EP soils was high, with [R.sup.2] = 0.92, RMSE = 28% and RPD = 3.5 (data not shown). However, validation prediction accuracy for carbonate was only indicator quality, with [R.sup.2] = 0.81, RMSE=I4% and RPD= 1.9, suggesting either that the model lacked the required detail (range variability) to model the validation samples or that there were spectral outliers in the validation set.
Cross-validation accuracy for clay was of good quality, with [R.sup.2] = 0.70, RMSE= 10% and RPD = 2.4, and indicator accuracy for validation with [R.sup.2] = 0.65, RMSE = 8% and RPD=1.5. Again, the relatively poor validation RPD may be due to the presence of outliers in the validation set.
Cross-validation for silt resulted in [R.sup.2] = 0.79, RMSE = 4%, and RPD = 2.2. This relatively high accuracy can be considered atypical, usually being less accurate, and may possibly be the result of a fortuitous effect of some unidentified high leverage outlier sample apart from the ones identified above. Also important here is the fact that the silt range was very narrow (only 0-12% silt) and so the RMSE of 4% accounted for 25% of the error. Part of this error would be, of course, be attributed to laboratory error because silt is usually determined by the difference between clay and sand contents. This may mean that the EP soils were composed mostly of clay minerals and sandsized quartz, but without having many intermediate siltsized particles. Validation prediction accuracy was poor, with [R.sup.2] = 0.39, RMSE = 3% and RPD= 1.3. Sand content regression accuracy was considered 'good', with cross-validation and validation [R.sup.2], RMSE and RPD statistics for cross-validation and validation of 0.82 and 0.91, 25% and 25%, and 2.4 and 2.8 respectively. In the PLSR analyses of the EP soils, it appears that 30 samples was simply not a large enough sample set to provide unambiguous validation models for PSD predictions.
The EP loading weights for clay, silt and sand were dominated by carbonate, kaolinite and quartz. The factor-1 loading weights in Fig. 4a, d for clay showed a positive peak for kaolinite clay (sharp peak near 3695 [cm.sup.-1]) and negative carbonate peaks (Nguyen et al. 1991; Janik et al. 1998; Reeves 2010). The factor-2 loading weights were due almost entirely to carbonate and negative quartz peaks.
Clearly, carbonate content in the EP soils played a pivotal role in the prediction of particle size determinations from MIR spectra of whole soils. The factor-1 loading weights for silt calibration were similar to those for clay content: mostly kaolinite and negative carbonate; however, the factor-2 loading weight had a very high negative contribution from -OH vibrations in organic matter, as indicated by the strong, broad negative band in the 3600-3000 [cm.sup.-1] region. As expected, the first sand loading weight was dominated by quartz peaks, with a contribution from kaolinite and negative carbonate, whereas the factor-2 loading weights were more characteristic of kaolinite and negative carbonate and organic matter or smectite clay.
As stated above, the purpose of building the composite set was to allow the prediction of PSD over a large range of soil types. However, it was necessary to first check to see whether a single combined calibration model would be possible to predict samples from the constituent specific soil sets. Widely different loading weights were required to model the separate Qld and NSW calibrations compared with the calcareous ACU and EP soils, and thus a combined sample set calibration may not succeed. The composite calibration set, constructed from a combination of the <2-mm Qld, NSW, ACU and EP samples, was tested with the validation samples, either as a composite validation or one for each of the specific sample set segments.
Table 3 describes the performance of the composite calibration model for the prediction of the individual sample sets as segments of the composite set. The cross-validation result for clay content calibration ([R.sup.2] = 0.82, RPD = 2.4) presented in Table 3 showed that the composite model (based on <2-mm samples) provided cross-validation accuracies similar to those of the separate NSW, Qld and ACU soil set cross-validations given in Table 2. The accuracy for the EP segment was lower ([R.sup.2] = 0.67, RPD= 1.0) compared with that of the separate EP set cross-validation ([R.sup.2] = 0.70, RPD= 1.9). This lower accuracy for the EP samples may have been due to the occurrence of carbonate, not well represented by the non calcarcous NSW and Qld datasets, with the model relying instead only on similarities with the calcareous ACU composite samples.
Validation for the clay fractions, for each of the individual dataset segments, was very similar ([R.sup.2] ~ 0.80, RPD ~ 2), except for the ACU samples ([R.sup.2] = 0.70) and EP (RPD= 1.1). As shown above, this compared well with the clay validation of the models built specifically for the individual Qld, NSW, ACU and EP sets.
The calibration cross-validation and validation accuracies for silt in the composite set were lower than for the individual set cross-validations. The reason for poor predictions for silt are probably the same as for the separate sets and due, in part, to laboratory analytical error in deriving silt content by subtracting the measured sand and clay contents from total soil weight and, in part, to the ill-defined composition of silt-sized particles. For example, there is a well-defined chemical or mineralogical compositional basis for clay and sand, clay being characterised primarily by high 2 : 1 clay mineralogy, high organic matter and low sand, and high sand content characterised by high quartz content (Janik et al. 1998). In contrast, silt has a wide range of PSD (ranging from 2 to ~50[micro]m), so there is the matter of the more ill-defined and less easily related silt particle size to soil chemistry or mineralogy.
Cross-validation of sand content for the composite <2-mm set ([R.sup.2] = 0.80, RPD = 2.3) compared well with the contributions from the Qld, NSW, ACU and EP segments, with respective [R.sup.2] values of 0.82, 0.78, 0.84 and 0.86 and RPD values of 2.3, 2.1, 2.2 and 2.6. This agreed well with the accuracies for individual cross-validations of the Qld, NSW, ACU and EP samples. Prediction of sand for the composite validation set resulted in an [R.sup.2] of 0.78 and an RPD of 1.6 (i.e. an indicator accuracy and far worse than for the cross-validation). This reduced accuracy could be explained by examining the results for the individual Qld, NSW, ACU and EP segments ([R.sup.2] = 0.78, 0.81, 0.68 and 0.80 respectively; RPD=1.60, 2.1, 2.3, 1.7 and 2.2 respectively). In general, the prediction accuracies were rated as good, but only indicator quality for the ACU set, which reduced the accuracy for the composite set. Presumably the reduced accuracies of the global set for sand content again reflected its inability to deal adequately with the calcareous ACU soils. Although composite models for clay and sand were similar to those obtained in the individual calibration sets, silt predictions remained poor. This may be indicative of the variation in modelling mechanisms responsible for the prediction of this analyte, a situation favoured when specific calibrations arc developed.
Finally, it seems that sand is better predicted than clay by MIR spectroscopy, probably indicating that coarse grain chemistry is based primarily on the presence of sand-sized quartz and thus better defined in comparison with clay mineral chemistry. In the NIR, physical soil properties, such as particle size distribution, may be measured through light scattering or reflection rather than chemistry (Ben Dor et al. 1999; Nocita et al. 2015). So, MIR has the advantage that sand (via quartz) has both a well-defined and specific signature, as well as characteristic spectral distortion and scattering effects (Nguyen et al. 1991).
In summary, the results in Table 3 showed that, for <2-mm sieved soils, there is a good prospect of being able to build a large composite or global dataset for a large range of soil types ranging from tropical acidic Kandosols and Fcrrosols and alkaline Vertosols to more neutral soil types in temperate regions rich in kaolinite and illite, through to the high-pH calcareous and sandy soils of southern Australia.
Fine grinding has been shown to improve prediction accuracy by reducing inter- and intraparticle heterogeneity (Soriano-Disla et al. 2014). Cross-validation results for the separate Qld, NSW and EP sets, after fine grinding, arc presented in Table 4 (the ACU samples were not included because sufficient samples were not available for fine grinding). Improvements in the composite cross-validation accuracies for clay, silt and sand were obtained by fine grinding ([R.sup.2] = 0.88, 0.62 and 0.87; RPD = 2.9, 1.6 and 2.8 for clay, silt and sand respectively). Results for the separate sample set segments confirmed this improvement (Table 4).
With regard to the separate datasets, the loading weights were found to differ considerably between the Qld <2-mm sieved and fine-ground samples (Fig. 5). There was a very strong increase in the loading weight for factor-1 for gibbsitc and organic matter (both negative) and a decrease in (negative) quartz for clay content (Fig. 5a, b) on fine grinding. There was also a very strong increase in kaolinite intensity in loading weight-2 for the fine ground material compared with the <2-mm sieved material.
The factor-1 loading weights for silt (Fig. 5c, d) were strongly affected by fine grinding, resulting in a strong reduction in intensities for quartz and increases in organic matter and (negative) kaolinite. The increased effect of fine grinding was apparent in the factor-1 loading weight for sand content (Fig. 5e, f), showing a strong reduction in quartz and an increase in kaolinite and gibbsite. The loading weight for factor-2 was notably more intense in organic matter (negative).
Loading weights for the NSW soil clay content calibration, shown in Fig. 6, were characterised by an increased (negative) peak due to SOM and reduced negative bands due to quartz. The second loading weight resulted in a strong kaolinite peak with fine grinding. With regard to silt, there were reduced negative peaks for quartz and increased negative peaks for smectite, particularly in the second loading weight. The very strong reduction in quartz and increase in organic matter was evident for the fine-ground soils.
The first clay loading weight for the <2-mm EP samples (Fig. 7) was due to kaolinite and (negative) carbonate, whereas the loading weights for fine-ground soils appeared to be due to an increase in carbonate and (negative) quartz. The silt loading weights for <2-mm sieved soils were dominated by kaolinite and (negative) carbonate. Carbonate and (negative) smectite dominated the factor-2 loading weights for the fine ground material. Fine grinding removed the contributions for quartz, kaolinite and carbonate for sand content calibrations.
There were important spectral differences between the <2-mm sieved and fine-ground samples (Fig. 8). To our knowledge, this is the first time that a formal link between the effects of fine grinding and sample composition, through examination of PLSR loading weights, has been proposed and demonstrated. For clay content in the composite dataset, there was stronger correlation with kaolinite in the fine-ground material compared with the <2-mm sieved material. This was particularly evident in factor-2 (Fig. 8b), where the factor-2 loading weights of the fine-ground samples were very strong in highly crystalline kaolinite (sharp peaks between 3695 and 3620 [cm.sup.-1]). The loading weights for silt were also affected by fine grinding, resulting in an increased negative correlation with carbonate and kaolinite in factor-1. The increased effect of fine grinding was also shown in the loading weights for sand content. Kaolinite and carbonate peaks in the loading weight for PC-1 increased in negative intensity by fine grinding, as did quartz and organic matter in factor-2.
This improvement was possibly due to better dispersion of the soil composition throughout the sample and reduced inter- and intraparticulate heterogeneity. Intra-aggregate heterogeneity, from the viewpoint of DRIFT spectroscopy, is effectively due to a change in chemistry from the outside surface of the sample to the interior within a few microns of the surface. This means that what is 'seen' by the infrared beam in <2-mm particles does not necessarily represent the true composition of the sample because the internal compositions of the aggregates are hidden from view. An example of this would be due to the presence of surface coatings, often composed of fine clay, iron oxide, calcrete and organic matter. Fine clay domains may also be encapsulated internally by organic and iron oxide layers, adding further complexity to the micro-aggregates.
Apart from simply homogenising the soil particles by grinding, the loading weights suggest that there may have also been a removal of surface layer material from the <2-mm sieved micro-aggregates and exposure of the interior matrix to the infrared beam. This conclusion is supported by the reduction in carbonate and goethite, probably as surface coatings, and a strong increase in high crystallinity kaolinite, quartz and organic matter.
The effects of the fine grinding soils to <0.1 mm on the accuracy of PSD determinations resulted in an overall improvement of-0.07 [R.sup.2] units and 1.7% RMSE units averaged across all soils and analytes. However, the improvement was highly dependent on the clay, sand and silt contents, and partly dependent on the particular soil set (number of samples and soil composition) and selection of the calibration and validation samples. For example, the highest improvement in the calibration occurred for the NSW set (0.12 [R.sup.2] units averaged across all analytes), although the maximum improvement in RMSE occurred for the Qld soils (RMSE = 3% PSD units averaged across all analytes). The least improvement in [R.sup.2] occurred for the EP set (0.03 [R.sup.2] units averaged across all analytes). Furthermore, the effect of grinding on validation accuracy was quite different to that of calibration cross-validation, possibly due to the manner of selecting the samples for calibration and validation sets, resulting in the inclusion of outlier samples in the validation sets. In the present study, there was no improvement for clay prediction for the NSW set, but a strong improvement in the EP sand validation (0.18 [R.sup.2] units). Although it has been shown that fine grinding generally improved prediction accuracy, the effects were highly variable and difficult to predict. From a practical viewpoint, the increases in RMSE and model accuracy from fine grinding arc small: management of soil with 8% or 11% clay would be similar.
As a step towards the development of a global PSD calibration for Australian soils, we have shown that spectra and data from several region-specific sample sets characteristic of tropical, subtropical, temperate and Mediterranean climate soils can be combined into a single regression model, and that this model can be used for reliable predictions of PSD in soils from specific regions. In some cases, reduced accuracies of the composite set calibrations for sand content reflected its inability to deal adequately with the less well-represented calcareous Mediterranean (ACU and EP) soils. MIR-PLSR has been shown to be able to provide a simple, rapid and relatively accurate prediction of PSD across different soil types and compositions in the presence of varying soil interferences (i.e. different levels of SOM and carbonate). The effect of fine grinding soils to <0.1 mm on the accuracy of PSD determinations was an overall improvement of -0.07 [R.sup.2] units and 1.7% RMSE units averaged across all soils and analytes. The strength of the PLSR regressions was confirmed by the fact that approximately 96% of the <0.1-mm samples had independently predicted sums of clay, silt, sand and carbonate predictions in the range 85-100%. However, the improvement in prediction accuracy was highly dependent on the clay, sand and silt contents, and partly on the soil set (number of samples and soil composition) and selection of the calibration and validation samples. In some soils, the improved accuracy due to fine grinding may be of limited importance in terms of soil management. It is suggested that soil calibrations can be developed using <2-mm material for routine application, but fine grinding may be advantageous for more rigorous applications. In addition, and for the first time, we have demonstrated that the spectral effects in the MIR of fine grinding are manifested through changes in the PLSR loading weights for the fine-ground versus <2-mm samples by allowing better access of the infrared beam to the soil matrix. After grinding, we observed a stronger contribution to the PLSR loading weights from kaolinite, quartz and organic matter and weaker contributions from carbonates and goethite, suggesting that surface coatings on the soil micro-aggregates, such as SOM, iron, aluminium oxides and carbonate, are removed and/or diluted with the soil matrix substrate as a result of grinding.
The authors acknowledge the support of the CSIRO Technology Accelerator fund in carrying out this research and the provision of analytical data by Drs Daniel Brough and Ben Harms from the Queensland Department of Science, Information Technology, Innovation and the Arts, Soil and Land Resources Science Division, Dr Andrew Rawson from the NSW Office of Environment and Heritage, and the CSIRO Land and Water Flagship, Urrbrae, Analytical Chemistry Unit.
Arroyo LJ, Li H, Teppen BJ, Boyd SA (2005) A simple method for partial purification of reference clays. Clays and Clay Minerals 53, 511-519. doi: 10.1346/CCMN.2005.0530507
Ben Dor F., Irons JR, Epema JF (1999) Soil reflectance. In 'Manual of remote sensing: remote sensing for the earth sciences. Vol. 3'. pp. 111-188. (John Wiley & Sons, New York, NY)
Boivin P, Garnier P, Tessier D (2004) Relationship between clay content, clay type, and shrinkage properties of soil samples. Soil Science Society of America Journal 68, 1145 1153. doi: 10.2136/sssaj2004.1145
Bowman G, Hutka J (2002) Particle size analysis. In 'Soil physical measurement and interpretation for land evaluation'. (Eds N McKenzie, KCoughlan, H Cresswell) pp. 224 239. (CSIRO Publishing: Melbourne, Vic.)
Bricklemyer RS, Brown DJ (2010) On-the-go VisNIR: Potential and limitations for mapping soil clay and organic carbon. Computers and Electronics in Agriculture 70, 209 216. doi:10.1016/j.compag.2009.10. 006
Guerrero C, Zomoza R, Gomez I, Mataix-Beneyto J (2010) Spiking of NIR regional models using samples from target sites: Effect of model size on prediction accuracy. Geoderma 158, 66-77.
Isbell R (2002) 'The Australian soil classification.' (CSIRO Publishing: Melbourne, Vic.)
Janik LJ, Merry RH, Skjemstad JO (1998) Can mid infrared diffuse reflectance analysis replace soil extractions? Australian Journal of Experimental Agriculture 38, 681 696. doi:10.1071/EA97144
Janik LJ, Merry RH, Forrester ST, Lanyon DM, Rawson A (2007) Rapid prediction of soil water retention using mid infrared spectroscopy. Soil Science Society of America Journal 71,507-514. doi : 10.2136/sssaj2005. 0391
Janik LJ. Forrester ST, Rawson A (2009) The prediction of soil chemical and physical properties from mid-infrared spectroscopy and combined partial least-squares regression and neural networks (PLS-NN) analysis. Chemometrics and Intelligent Laboratory Systems 97, 179 188. doi: 10. 1016/j.chemolab.2009.04.005
Kennard RW, Stone LA (1969) Computer aided design of experiments. Technometrics 11, 137-148. doi: 10.1080/00401706.1969.10490666
Kunze GW, Dixon JB (1986) Pretreatment for mincralogical analysis. In 'Methods of soil analysis, part I. Physical and mineralogica! methods'. 2nd edn. (Ed. A Klute) pp. 91 100. (American Society of Agronomy, Inc. and the Soil Science Society of America, Inc.: Madison, WI)
McKenzie N, Coughlan K, Cresswell H (2002) Particle size analysis. In 'Soil physical measurement and interpretation for land evaluations', pp. 224-239. (CSIRO Publishing: Melbourne)
Nguyen TT, Janik LJ, Raupach M (1991) Diffuse reflectance infrared Fourier transform (DRIFT) spectroscopy in soil studies. Australian Journal of Soil Research 29, 49-67. doi: 10.1071/SR9910049
Nocita M. Stevens A, van Wesemael B, Aitkenhead M, Bernard B, Barthes M, Ben Dor E, Brown DJ, Clairotte M, Csorba A, Dardcnnc P, Dematte JAM, Gcnot V, Guerrero C, Knadel M, Montanarella L, Noon C, Ramirez-Lopez L, Robertson J, Sakai H (2015) Soil spectroscopy: an alternative to wet chemistry for soil monitoring. Advances in Agronomy 132. 139 159. doi: 10.1016/bs.agron.2015.02.002
Reeves JB III (2010) Near-versus mid-infrared diffusive reflectance spectroscopy for soil analysis emphasizing carbon and laboratory versus on-site analysis: where are we and what needs to be done? Geoderma 158, 3 14. doi: 10.1016/j.gcoderma.2009.04.005
Sherrod LA, Dunn G, Peterson GA, Kolberg RL (2002) Inorganic carbon analysis by modified pressure calcimeter method. Soil Science Society of America Journal 66, 299-305. doi: 10.2136/sssaj2002.2990
Soriano-Disla JM, Janik L, Viscarra Rossel RA, McDonald LM, McLaughlin MJ (2014) The performance of visible, near and mid-infrared spectroscopy for prediction of soil physical, chemical and biological properties. Applied Spectroscopy Reviews 49, 139-186. doi : 10.1080/05704928.2013.811081
Stenberg B, Viscarra-Rossel RA, Mouazen AM, Wetterling J (2010) Visible and near-infrared spectroscopy in soil science. Advances in Agronomy 107, 163-215. doi: 10.1016/S0065-2113(10)07005-7
Stumpe B. Weihermuller L, Marschncr B (2011) Sample preparation and selection for qualitative and quantitative analyses of soil organic carbon with mid-infrared reflectance spectroscopy. European Journal of Soil Science 62, 849-862. doi. 10.1111/j.1365-2389.2011.01401.x
Sudduth KA, Hummel JW (1996) Geographic operating range evaluation of a NIR soil sensor. Transactions of the American Society of Agricultural Engineers 39, 1599-1604. doi: 10.13031/2013.27674
Tisdall JM, Oades JM (1982) Organic matter and water-stable aggregates in soils. Journal of Soil Science 33, 141 163. doi: 10.1111/j.1365-2389. 1982.tb01755.x
USDA Natural Resources Conservation Service, National Soil Survey Center (1996) Particle-size analysis. In 'Soil survey laboratory methods manual'. Soil Survey Investigation Report No. 42, Version 3.0. (USDA: Washington DC)
Viscarra Rossel RA, Webster R (2012) Predicting soil properties from the Australian soil visible-near infrared spectroscopic database. European Journal of Sod Science 63, 848 860. doi: 10.1111/j.1365-2389.2012. 01495.x
Viscarra Rossel RA, Cattle SR. Ortega A, Fouad Y (2009) In situ measurements of soil colour, mineral composition and clay content by vis-NIR spectroscopy. Geoderma 150, 253-266. doi: 10.1016/j.gcod erma.2009.01.025
Waiser TH, Morgan CLS, Brown DJ, Hallmark CT (2007) In situ characterization of soil clay content with visible near-infrared diffuse reflectance spectroscopy. Soil Science Society of America Journal 71, 389-396. doi : 10.2136/sssaj2006.0211
Williams PC (1987) Variables affecting near-infrared reflectance spectroscopy. In 'Near-infrared technology in the agricultural and food industries'. (Eds PC Williams, KH Norris) pp. 143 167. (American Association of Cereal Chemists Inc.: St Paul, MN)
Leslie J. Janik (A,B,C), Jose M. Soriano-Disla (A,B), Sean T. Forrester (A), and Michael J. McLaughlin (A,B)
(A) CSIRO Environmental Contaminant Mitigation and Technologies Program, CSIRO Land and Water, Waite Campus, Waite Road, Urrbrae, SA 5064, Australia.
(B) School of Agriculture, Food and Wine, University of Adelaide, Waite Campus, Waite Road, Urrbrae, SA 5064, Australia.
(C) Corresponding author. Email: email@example.com
Table 1. Soils used for particle size distribution (PSD) calibration and analytical data for soil carbonate and PSD (clay silt and sand) Particle size data are shown for actual soil fractions after carbonate has been removed. Laboratory carbonate data for Qld and NSW were unavailable, so these data were predicted using mid-infrared partial least-squares regression calibrations. Samples provided from Queensland (Qld), New South Wales (NSW), CSIRO (ACU) and Eyre Peninsula (HP) Soil set Statistic Carbonate Clay Silt Sand Qld No. samples 500 500 500 500 Minimum (%) 0.0 1 1 3 Maximum (%) 28 89 53 97 Median (%) 0.8 36 12 53 s.d. 0.6 20 9 24 NSW No. samples 674 674 674 674 Minimum (%) 0 0 1 1 Maximum (%) 14 84 60 100 Median (%) 0 22 18 50 s.d. 1.9 19 11 23 ACU No. samples 135 135 135 135 Minimum (%) 0 1 0 4 Maximum (%) 84 72 32 98 Median (%) 1 22 7 49 s.d. 24 18 7 28 EP No. samples 30 30 30 30 Minimum (%) 0 1 0 3 Maximum (%) 96 39 12 94 Median (%) 6 14 3 66 s.d. 24 10 4 22 Table 2. Calibration cross-validation and validation statistics for clay (Clay%), silt (Silt%) and sand (Sand%) for the <2-mm samples RMSK, root mean square error; [R.sup.2], coefficient of determination; RPD, ratio of the s.d. to the RMSE; Qld, Queensland; NSW, New South Wales; ACU, CSIRO; EP, Eyre Peninsula Sample set Analyte No. Median (A) Range RMSE samples (%) (%) (%) used Calibration cross-validation Qld Clay% 332 30 1-89 7 Silt% 332 12 1-52 6 Sand% 329 53 5-97 9 NSW Clay% 453 24 0-74 8 Silt% 453 19 0-60 6 Sand% 453 49 3-100 10 ACU Clay% 84 22 1-72 7 Silt% 85 7 0-32 5 Sand% 85 49 4-98 10 EP Clay% 18 14 0-39 10 Silt% 18 1 0-11 2 Sand% 19 61 0-81 10 Validation Qld Clay% 168 36 1-89 9 Silt% 168 11 1-53 6 Sand% 168 52 3-95 12 NSW Clay% 221 23 0-84 8 Silt% 221 17 1-51 6 Sand% 221 53 1-98 10 ACU Clay% (#549-OL) 49 28 2-68 9 Silt% 50 12 1-27 6 Sand% 50 48 15-97 10 EP Clay% 10 12 4-32 5 Silt% 10 2 0-12 3 Sand% 10 52 25-94 9 Sample set Analyte [R.sup.2] s.d. RPD Factors (B) (%) Calibration cross-validation Qld Clay% 0.88 20 2.9 14 Silt% 0.61 9 1.6 10 Sand% 0.85 24 2.6 13 NSW Clay% 0.82 18 2.3 19 Silt% 0.69 11 1.8 19 Sand% 0.81 22 2.3 19 ACU Clay% 0.83 18 2.5 6 Silt% 0.52 7 1.4 7 Sand% 0.87 28 2.8 5 EP Clay% 0.70 10 1.9 3 Silt% 0.79 4 2.2 5 Sand% 0.82 25 2.4 5 Validation Qld Clay% 0.79 21 2.2 14 Silt% 0.62 9 1.6 10 Sand% 0.78 25 2.1 13 NSW Clay% 0.83 20 2.4 19 Silt% 0.67 11 1.7 19 Sand% 0.82 24 2.3 19 ACU Clay% (#549-OL) 0.73 17 1.8 6 Silt% 0.51 7 1.2 7 Sand% 0.77 21 2.1 5 EP Clay% 0.65 8 1.5 3 Silt% 0.39 3 1.3 5 Sand% 0.91 25 2.8 5 (A) Median analytical value. (B) Number of partial least-squares regression factors used in the determinations. Table 3. Calibration cross-validation and validation statistics for clay (Clay%), silt (Silt%) and sand (Sand%) for the composite model (based on scanning of <2-mm sieved samples) and the statistics for the portions (segments) attributed to samples from Queensland (Qld), New South Wales (NSW), South Australia (SA) and the Eyre Peninsula (EP) RMSE, root mean square error; [R.sup.2], coefficient of determination; RPD, ratio of the s.d. to the RMSE; ACU, CSIRO Sample set Analyte No. samples Median (A) Range RMSE used (%) <%) (%) Calibration cross-validation Composite Clay% 1313 29 0-89 9 Silt% 1313 14 0-65 7 Sand% 1313 49 1-100 11 Qld (segment) Clay% 755 35 1-89 8 Silt% 756 12 1-52 6 Sand% 756 49 2-98 11 NSW (segment) Clay% 452 23 0-77 8 Silt% 452 21 0-65 9 Sand% 452 49 1-100 11 ACU (segment) Clay% 85 22 1-72 9 Silt% 85 7 0-32 7 Sand% 85 49 4-98 13 EP (segment) Clay% 20 14 1-42 11 Silt% 20 2 0-13 8 Sand% 20 65 3-91 9 Validation Composite Clay% 664 27 1-89 9 Silt% 664 49 3-98 12 Sand% 664 14 0-67 7 Qld (segment) Clay% 378 27 1-89 9 Silt% 378 14 1-53 6 Sand% 378 49 3-98 12 NSW (segment) Clay% 226 26 1-81 8 Silt% 226 20 1-67 9 Sand% 226 47 8-96 11 ACU (segment) Clay% 50 26 2-68 9 Silt% 50 12 1-27 7 Sand% 50 53 15-97 12 EP (segment) Clay% 10 12 4-34 8 Silt% 10 13 0-13 5 Sand% 10 95 39-95 9 Sample set Analyte [R.sup.2] s.d. (%) RPD Factors (B) Calibration cross-validation Composite Clay% 0.82 20 2.4 18 Silt% 0.59 11 1.6 18 Sand% 0.80 25 2.3 20 Qld (segment) Clay% 0.84 21 2.5 18 Silt% 0.57 9 1.5 18 Sand% 0.82 25 2.3 20 NSW (segment) Clay% 0.78 18 2.1 18 Silt% 0.51 13 1.4 18 Sand% 0.78 24 2.1 20 ACU (segment) Clay% 0.78 18 2.1 18 Silt% 0.32 7 1.0 18 Sand% 0.84 28 2.2 20 EP (segment) Clay% 0.67 10 1.0 18 Silt% 0.02 4 0.5 18 Sand% 0.86 23 2.6 20 Validation Composite Clay% 0.80 20 2.2 18 Silt% 0.78 25 2.1 20 Sand% 0.63 11 1.6 18 Qld (segment) Clay% 0.80 21 2.2 18 Silt% 0.61 9 1.6 18 Sand% 0.78 26 2.1 20 NSW (segment) Clay% 0.80 17 2.2 18 Silt% 0.59 13 1.5 18 Sand% 0.81 24 2.3 20 ACU (segment) Clay% 0.70 17 1.8 18 Silt% 0.29 7 1.0 18 Sand% 0.68 21 1.7 20 EP (segment) Clay% 0.78 9 1.1 18 Silt% 0.17 4 0.8 18 Sand% 0.80 20 2.2 20 (A) Median analytical value. (B) Number of partial least-squares regression factors used in the determinations. Table 4. Calibration cross-validation and validation statistics for clay (Clay%), silt (Silt%) and sand (Sand%) for the <0.1-nnn composite Queensland (Qld), New South Wales (NSW) and Eyre Peninsula (EP) samples RMSE, root mean square error; RPD. ratio of the s.d. to the RMSE Sample set Analyte No. Median (A) Range RMSE samples (%) (%) (%) used Calibration Separate sets Qld Clay% 330 31 1-89 5 Silt% 331 12 1-52 5 Sand% 330 53 5-97 7 NSW Clay% 450 24 0-74 6 Silt% 450 19 0-60 6 Sand% 450 49 3-100 8 EP Clay% 18 15 0-39 4 Silt% 19 2 0-11 2 Sand% 20 62 0-90 7 Composite set Qld + NSW + EP Clay% 792 24 0-89 7 Silt% 797 13 0-55 6 Sand% 793 49 1-100 9 Qld segment Clay% 755 35 1-89 8 Silt% 330 12 1-51 5 Sand% 329 66 3-97 7 NSW segment Clay% 444 21 0-81 7 Silt% 446 15 0-55 7 Sand% 443 47 1-100 9 EP segment Clay% 21 15 3-34 5 Silt% 21 3 0-13 6 Sand% 21 66 11-95 8 Validation Separate sets Qld Clay% 168 35 1-89 6 Silt% 168 11 1-53 7 Sand% 168 52 3-95 9 NSW Clay% 221 23 0-85 11 Silt% 221 17 1-51 6 Sand% 221 53 1-98 9 EP Clay% 10 12 4-32 4 Silt% 10 2 0-12 3 Sand% 10 52 25-94 5 Composite set Qld f NSW + EP Clay% 406 24 0-83 7 Silt% 330 12 1-51 5 Sand% 408 56 5-98 9 Qld segment Clay% 169 23 1-83 6 Silt% 169 15 1-53 6 Sand% 169 58 6-95 9 NSW segment Clay% 229 27 0-77 7 Silt% 231 17 1-54 7 Sand% 233 45 5-98 10 EP segment Clay% 8 11 6-42 4 Silt% 8 2 0-13 6 Sand% 8 68 34-91 8 Sample set Analyte [R.sup.2] s.d. RPD Factors (B) (%) Calibration Separate sets Qld Clay% 0.93 20 3.7 12 Silt% 0.64 9 1.7 15 Sand% 0.90 24 3.2 15 NSW Clay% 0.88 18 2.9 14 Silt% 0.74 11 2.0 21 Sand% 0.86 22 2.7 18 EP Clay% 0.88 10 2.9 5 Silt% 0.83 4 2.5 6 Sand% 0.93 26 3.7 5 Composite set Qld + NSW + EP Clay% 0.88 19 2.9 18 Silt% 0.62 10 1.6 17 Sand% 0.87 23 2.8 18 Qld segment Clay% 0.84 21 2.5 18 Silt% 0.59 8 1.5 17 Sand% 0.91 25 3.4 18 NSW segment Clay% 0.83 18 2.4 18 Silt% 0.59 10 1.6 17 Sand% 0.83 22 2.4 18 EP segment Clay% 0.71 9 1.7 18 Silt% 0.29 4 0.6 17 Sand% 0.86 20 2.6 18 Validation Separate sets Qld Clay% 0.93 21 3.7 12 Silt% 0.52 9 1.4 15 Sand% 0.87 25 2.8 15 NSW Clay% 0.87 20 1.8 14 Silt% 0.69 11 1.7 21 Sand% 0.88 24 2.9 18 EP Clay% 0.86 8 2.2 5 Silt% 0.42 3 1.3 6 Sand% 0.97 25 4.9 5 Composite set Qld f NSW + EP Clay% 0.88 19 2.9 18 Silt% 0.59 8 1.5 17 Sand% 0.84 23 2.5 18 Qld segment Clay% 0.89 18 2.9 18 Silt% 0.67 10 1.7 17 Sand% 0.85 23 2.6 18 NSW segment Clay% 0.87 19 2.8 18 Silt% 0.65 11 1.7 17 Sand% 0.84 24 2.4 18 EP segment Clay% 0.91 12 3.2 18 Silt% 0.36 5 0.8 17 Sand% 0.84 20 2.5 18 (A) Median analytical value. (B) Number of partial least-squares regression factors used in the determinations.
Please note: Some tables or figures were omitted from this article.
|Printer friendly Cite/link Email Feedback|
|Author:||Janik, Leslie J.; Soriano-Disla, Jose M.; Forrester, Sean T.; McLaughlin, Michael J.|
|Date:||Nov 1, 2016|
|Previous Article:||The influence of surfactant and organic matter content on wetting pattern of different non-water repellent soils.|
|Next Article:||Spatial and temporal variations of soil function in a Mediterranean serpentine ecosystem.|