The study was performed in the alpine parts of the northern, subarctic zone of Norway (Finnmark). Plant species were sampled from eight river valleys, extending the temperature gradient of the region. The sampled plant species (Table 1) are typical for the riparian zone, dominated by willows, forbs and grasses, and for the surrounding dry heath tundra, dominated by dwarf shrubs. Each sample consisted of several plant individuals from the same location, and habitat, with a maximum distance of 25 m from each other. Plant individuals were always sampled by harvesting the green non-woody plant parts; the whole plant in the case of grasses, forbs and horsetails, and the leaves in the case of shrubs (Table 1). Plant material was cut some centimeters above the soil level and sampled plant material was always visually controlled to be free from soil particles. Samples of 29 plant species, representing different growth forms, were collected making a total of 442 plant samples (Table 1). In order to cover plant and Si structural variations over several plant development stages, we sampled plant species at the beginning (June), during (July) and at the end of the growing season (August).

Plant samples were oven dried for 3 days at 75°C. We chose not to wash the plant samples as sampling routines excluded the presence of soil particles and the ecological context (high precipitation, dense vegetation, long distance to roads) restricted the possibilities for aeolean dust/soil transport. Subsequently, samples were pulverized using a ball mill (Mixer Mill, MM301; Retsch GmbH and Co. Haan, Germany) in order to get very fine and homogeneous plant material. We applied a chemical digestion followed by a colorimetric analysis, as the gravimetric approach has been proven to show a poor reproducibility (Herbauts et al., 1994). Plant bSi was extracted by the wet alkaline (0.1 M Na₂CO₃) extraction, a technique recently confirmed to be suitable for bSi analysis of plant material (Meunier et al., 2013). About 30 mg of pulverized plant material was incubated for 4 h in 0.1 M Na₂CO₃ at 80°C (DeMaster, 1991) and then 10 ml of the extract was filtered (Chromafil® A-45/25, pore size of 0.45 μm). After extraction, samples were stored in a dark room at 3°C and were analyzed colorimetrically for extracted dissolved Si (DSi) within a maximum of 2 weeks, using a Continuous Flow Analyser (CFA; SKALAR SA 1500, Smith and Milne, 1981). Blank extractions were subtracted to account for Si release from recipients and chemicals. The error of the entire analysis procedure was measured as the (reproducibility) standard deviation of the analysis of 12 subsamples of a single pulverized plant sample (Deschampsia cespitosa).

Remaining pulverized plant samples were made into tablets (Ø16 mm, >1 mm thick) by applying 6 tons of pressure with a hydraulic press, in order to obtain a homogeneous flat surface. A homogeneous flat surface reduces random light scatter of loose powder, thus reducing random variation in spectral signatures. Before NIRS analyses, the plant tablets were oven dried at 50°C for 2 h in order to remove the remaining water film on the hydrophilic plant material. The dry samples were then cooled in a desiccator until scanning by NIRS. The NIRS-spectra, measured as the logarithm of the inversed reflectance (log(1/R)), were recorded with monochromatic radiation at 1.4 nm intervals from 350 to 1050 nm, and at 2 nm intervals from 1000 to 2500 nm, using a FieldSpec 3 (ASD Inc., Boulder, Colorado, USA) and were registered at 1 nm interpolated intervals. Each sample spectrum was recorded as the average of three replicate scans. NIRS scanning was performed at room temperature during one single winter day, which most likely resulted in a stable room temperature as there was no sunlight to warm up the room.

Three nested calibration sets were created; one on the total sample size, representing all studied plant species (n = 442); one for a subset containing all graminoids (n = 264), which are, as active Si accumulators, most often studied in plant-Si studies, and finally, in order to test the performance of NIRS when applied to one single plant species, one calibration for a subset containing all samples of the grass species Deschampsia cespitosa (n = 75). All data sets were treated similarly: in order to get a similar spectral diversity within the calibration and validation dataset, calibration and validation samples were selected using the Kennard-Stone algorithm (Kennard and Stone, 1969) included in the “prospectr” package 0.1.3 (Stevens and Ramirez-Lopez, 2014) in R 3.0.1 (R Development Core Team, 2013). Spectral outliers were identified by means of Mahalanobis distances (75% quantile as cutoff point; Mahalanobis, 1936; De Maesschalck et al., 2000) and removed in a single operation. Before starting the calibration, the wavelength regions where the NIRS sensors overlapped (350–380 nm, 760–840 nm, 1700–1800 nm and 2450–2500 nm) were removed to avoid false correlations from instrumental noise.

The statistical procedure for the calibration, using partial least-squares regression (Martens and Næs, 1993), in the pls-package 2.4–3 (Mevik et al., 2013) in R 3.1.0. Different transformations on the spectral data set were considered for the calibrations: scaling and/or centering, smoothing before applying derivatives, and the use of first (1D) and second order derivatives (2D) by using the earlier mentioned “prospectr” package (Stevens and Ramirez-Lopez, 2014) in R.

The most parsimonious calibration model was chosen based on evaluation of two calibration statistics: a high coefficient of multiple determination (R²), and the lowest root mean square error of calibration (RMSEC) for a given number of selected model components k. Overfitting was avoided by selecting the combination of model parameters k showing the lowest standard error of cross-validation (SECV; Mark and Workman, 1991). Finally, the chosen calibration model was used to predict the Si concentrations of the validation sample sets. The R², the RMSEP (“root mean square error of prediction,” i.e., the mean error rate between predicted and real values), the bias (i.e., the systematic error of the linear regression), and the intercept and slope of the linear fit of the predictions were used to assess the applicability of the developed NIRS calibration models. We tested the quality of all calibration models with and without the visible region (380–720 nm). Chlorophyll has a strong signature in the visible light spectra, and seasonal variation in chlorophyll content could consequently be confounded with variation in bSi content. As the models with visible light were consistently of lower quality, we only present models without the visible light.

We found plant Si concentrations to range from the detection limit (<0.024%Si) to 10.0% of Si by dried plant weight. The average error of the wet chemical analysis method, measured as the reproducibility standard deviation of the repeated measurement of a Deschampsia cespitosa sample with average concentration of 0.34%Si, was 0.022% of Si by dry weight, which corresponds to an average relative error of 2.6%. Across all species, plant Si concentration showed large variation with a coefficient of variation of 69%. However, the concentrations of Si were low (<0.1%Si) for almost all shrubs, dwarf shrubs and forbs, while most graminoids and horsetails showed average Si concentrations higher than 0.5% (Table 1). The selection of the calibration and validation datasets based on the spectral characteristics of the samples resulted in comparable Si concentration frequency distributions of the calibration and validation datasets for the three separate models (Figure 1).

The best calibration model for Si across all plant species was based on the second order derivatives (2D) of the smoothed (17 wavelength intervals) NIR spectra with 23 components, which showed highest loading values at 1050, 1900, and 2300 nm (Figure 2). The best calibration model for graminoids was based on the multiplicative-scatter corrected spectra, and 33 components. Highest loading values were found for wavelengths around 1900 and 2300 nm (Figure 3). Finally, the best calibration model for the grass species Deschampsia cespitosa was based on the centered, second order derivative spectra with 19 components, with highest loading values for the same wavelengths as the calibration model including all plant species (Figure 4).

Our calibration model including 442 samples from 29 different plant species showed a RMSEP of 0.24% of Si by dry weight (Table 2, Figure 5), corresponding to a relative error of 28% for the average measured Si concentration. Our separate calibration model for graminoids showed a RMSEP of 0.10% of Si by dry weight, corresponding to an average relative error of only 8.3% (Table 2, Figure 6). Narrowing the graminoid dataset to the most abundant grass species, Deschampsia cespitosa, increased the calibration model accuracy; a RMSEP of 0.023% of Si by dry weight, corresponding to a relative error of 1.3% (Table 2, Figure 7). The bias was also highest for the model including all plant species, whereas it was lowest and similar for the graminoids and Deschampsia cespitosa (Table 2). For the models including all plant species and the graminoids, the slope and intercept of the validation regression lines were close to 0 and 1 respectively, whereas the Deschampsia cespitosa model showed a slope close to one, and an intercept which differed more from 0 (Table 2). The R² values of the validation regression were all equal to or higher than 0.90 and showed a similar pattern as the bias, with the highest value for the graminoids model (0.95), the lowest for the model including all plant species (0.90) and an intermediate value for the Deschampsia cespitosa model (0.93; Table 2, Figures 5–7).

Our results show that the Si content of plants can successfully be measured by NIRS across a large range of plant species. All three calibration models showed a high coefficient of determination (R² ≥ 0.9). The calibration model for the grass Deschampsia cespitosa showed the best overall performance with the lowest root mean square error in the validation (RMSEP), and a slope closest to 1 (Table 2). Taking into account that the average error of the wet chemical extraction method followed by a colorimetric analysis is 0.022% of Si by dry weight, using NIRS for graminoids increases the average error with 0.078% of Si by dry weight, which is acceptable when working with plants accumulating Si (Si > 0.5%Si by dry weight; Raven, 2003). For the Deschampsia cespitosa model, the average error became slightly lower compared to the reference chemical analysis method (0.015%Si vs. 0.022%Si). Also, a further specification of the calibration model from all plant species to the plant group level and finally to a single plant species increased the accuracy. Prudence is called for when using the calibration model across all plant species in studies that require very high accuracy. Although the model containing all plant species showed a high R², and an intercept and slope close to 0 and 1, respectively, it was characterized by a RMSEP of 0.24% of Si by dry weight, which is rather high, even for plant species actively accumulating Si.

Our three calibration models showed remarkably better validation statistics compared to earlier studies on the application of NIRS of plant material (Halgerson et al., 2004; Jin and Chen, 2007; Ostatek-Boczynski et al., 2013), and is probably a result of (1) the use of a well-proven and widely used reference method for plant Si analysis (Meunier et al., 2013), and (2) a calibration dataset which consists of plant species representing almost the entire range of natural plant Si concentrations (Hodson et al., 2005). We also hypothesize that the specific pre-treatment of the plant samples before NIRS scanning, where we pressed ground plant powder into a tablet, may have increased the quality of the calibration. The ball-mill produced very fine powder, resulting in a flat tablet surface which may have reduced both light scattering from powder and original structural differences between plant species. However, as we were not able to measure the fineness, particle size distribution and homogeneity of the ground plant samples, this issue remains subject for further research and the quality of the NIRS calibration could potentially be increased by using a cyclotech mill instead of the ball mill procedure. The cyclotech mill grinding method ensures a narrow particle size distribution, which has been proven to strongly influence the NIR spectrum (Casler and Shenk, 1985).

Our success in measuring Si with NIRS across several plant lineages, representing different plant phenological stages and a range of ecological contexts, may also be due to the single type of Si structure in plants. Whereas the often transient and species-specific character of associations between minerals and organic or hydrated molecules (Clark et al., 1987) may limit the applicability of NIRS for mineral analysis to a well-defined population (plant species, sampling time; Foley et al., 1998), Si in plants is only known to occur as a hydrated condensate of orthosilicic acid (Si(OH)₄). However, narrowing the calibration dataset to graminoids and finally to Deschampsia cespitosa resulted in a clear increase in accuracy (lower RMSEP) and a regression slope close to one in the case of Deschampsia cespitosa. We suggest that this increase in calibration model accuracy is related to plant group/species dependent locations of bSi formation. Cell-silicification shows clear phylogenetic differences. In monocotyledons, silica is more often deposited in the cell lumen of “silica cells,” and as “silica bodies” on cells, while dicotyledons more often show silica deposits within or underneath the cell wall (Piperno, 1988). In addition, plant lineages differ in the type of tissues which are silicified; in grasses, Si is preferentially deposited as a layer beneath the cuticle (Prychid et al., 2004), the Pteridaceae (ferns) and Equisetaceae (horsetails) mostly show silicification within and at epidermal cells (Kaufman et al., 1971; Sundue, 2009), while orchids (Orchidaceae) are characterized by silica bodies in the sclerenchyma (Møller and Rasmussen, 1984). Moreover, depending on the condensation environment (e.g., Si(OH)₄ concentration, temperature, pH), the silica condensate can differ significantly in density and composition (Perry et al., 2003). Si also shows strong interactions with plant biomolecules such as cellulose (Perry and Lu, 1992), phenol- or lignin-carbohydrate complexes (Inanaga and Okasaka, 1995; Inanaga et al., 1995), callose (Law and Exley, 2011), and proteins (Perry and Keeling-Tucker, 2003). Although little is known about these chemical interactions, Si possibly links to these complex plant biomolecules through dihydrogen bonds (Zhang et al., 2013). As different condensation locations are characterized by different plant biomolecules, this strong interaction with the organic plant environment may result in an additional chemical and spectral difference in silica between plant lineages, which may explain the better calibration model for graminoids separately. Plant lineages show a large variation in the absolute bSi concentration (Hodson et al., 2005), possibly also influencing the characteristics of the individual Si condensates and its interactions with the organic environment. Finally, the three calibration models may also be confounded by other plant compounds as lignin, cellulose, and phenolics, which are, in some plant families, closely related to plant silica (Schoelynck et al., 2010; Cooke and Leishman, 2012). However, our dataset was not large enough to compare NIRS calibration model characteristics for different plant groups and further research is needed to assess the importance of plant group/species specific calibrations.

Using NIRS for plant Si analysis significantly reduces the operational and time costs. The time for sample drying is dependent on the amount of samples dried at the same time, but in all cases it is small and identical for both methods, and is therefore omitted from this estimation. In both methods, it took about 5 min to grind each sample using a ball-mill. In the reference method, about 30 mg of each plant sample was weighed, bSi was extracted in an alkaline solution and the extract was analyzed spectrophotometrically. Because this reference method was already highly optimized, e.g., by using an automated spectrophotometric device (SKALAR, 332 samples per run), each of these three steps only took about 3 min. When plant Si was analyzed by using NIRS scanning, it took 3 min per sample for the production of a tablet and 1.5 min per sample for the NIRS scan itself. As such, our calibration shows that the use of NIRS may reduce the analysis time by at least 30% compared to the chosen reference method for plant Si analysis, which is comparable to that reported for XRF spectroscopy (Reidinger et al., 2012). However, most time was still spent on the pre-treatment (drying, grinding, making tablet) of the plant sample. We believe that this pre-treatment may explain our strong calibration across largely differing plant species. However, omitting the grinding and rather extending the NIRS method to dried unground plant samples may be worthwhile as it results in a considerable progress in time efficiency, reducing the analysis time to 1.5 min after drying (90% reduction). The application to unground samples would introduce the structure of the plant leaf as an important calibration factor. Leaf structure strongly differs between plant species. Although a scatter correction of the NIR spectra may reduce the influence of structure (e.g., Gislum et al., 2004), the earlier described plant family specific location of silica deposition probably requires separate calibrations for each plant family or even plant species. Application of NIRS on fresh plant material, which would exclude all sample pre-treatment and allow for in situ non-destructive field monitoring of plant silicon, shows some larger challenges as all wavelength zones with high influence on our calibration model were situated in zones typical for O–H bonds of water, which should be avoided when working with fresh plant samples (e.g., Foley et al., 1998; Morón et al., 2007). Finally, we recommend other studies to carefully consider whether washing of the plant samples is needed in order to remove dust/soil particles. This will extend the analysis time, but is a common factor for both the chemical and the spectral analysis method.

An even stronger gain in efficiency and reduction in operational costs is obtained when applying NIRS for simultaneous analysis of different plant compounds. As a result of its central role in the functioning of plants, Si biochemistry in plants is closely linked to that of numerous other chemical compounds (Cooke and Leishman, 2011). Studies that quantify Si in plants often also quantify plant structural compounds such as lignin and cellulose (Schoelynck et al., 2010), as well as compounds related to abiotic and biotic plant stress (toxic metals, secondary metabolites) and plant macronutrients (e.g., N and P). Numerous studies have shown the applicability of NIRS for the analysis of many of these Si-related plant compounds (e.g., Joffre et al., 1992; Gillon et al., 1999; Stolter et al., 2006; González-Martín et al., 2007). Especially in studies where focus is on different plant compounds, the advantages of using NIRS as a method are unprecedented both in efficiency and cost.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.