Combustion and Spectral Methods for Quantifying Carbon and Nitrogen Concentrations in Pacific Northwest Douglas-Fir Forest Soils

Traditional combustion methods for assessing soil carbon (C) and nitrogen (N) stocks are time consuming and expensive; visible and near-infrared (VNIR) methods offer a quick and inexpensive alternative for establishing soil C and N concentrations. We compared combustion and spectral methods for quantifying soil carbon and nitrogen concentrations. We sampled organic and mineral soil horizons in managed and old-growth Douglas-fir (Pseudotsuga menziesii) forests in western Oregon. We applied combustion methods to determine total soil carbon and nitrogen concentrations of these samples. We then applied VNIR methods to derive a reference spectral library for analyzing Pacific Northwest (PNW) andesitic forest soils. Our spectral analysis confirmed that visible range spectra (especially in the 650-750 nm range) are the most useful for distinguishing differences in sample soil organic matter content. Our results provide a robust calibration model for applying spectral analysis combined with Partial Least Square Regression (PLSR) to quantify carbon and nitrogen stocks in PNW Douglas-fir forest soils. Model validation resulted in R values ranging from 0.92 to 0.95 for C and from 0.73 to 0.84 for N.


Introduction
Awareness of the rate at which anthropogenic activities alter global carbon (C) (Marland, Boden, Andres, Brenkert, & Johnston, 2002) and nitrogen (N) (Vitousek et al., 1997;Galloway & Cowling, 2002) dynamics has led to widespread interest in modeling and monitoring global carbon (e.g.Solomon et al., 2007;IPCC, 2014) and nitrogen (e.g.International Nitrogen Initiative, http://initrogen.org,founded 2003) cycles.In particular, policy initiatives aimed at reducing net greenhouse gas emissions, such as California Assembly Bill 32, the Regional Greenhouse Gas Initiative (RGGI) and the Kyoto Protocol, promote interest in quantifying C stored in managed landscapes.Due to significant soil organic matter (SOM) stocks in forest soils and evidence of a seasonal carbon sink in the northern hemisphere (Fung, Tucker, & Prentice, 1987;Pacala et al., 2001;Goodale et al., 2002;Fahey et al., 2010), temperate and boreal forest ecosystems have received particular attention from carbon researchers.Tools for quantifying SOM stocks, including C and N content, can support policy initiatives aimed at increasing C storage, can provide data for calibrating and validating models of C and N cycling, and can facilitate the monitoring of soil fertility in managed forest ecosystems.
Forests of the Pacific Northwest (PNW) demonstrate significant soil C storage potential.Sun et al. (2004) reported mineral soil C stocks ranging from 8-36 kg C m -2 (to a depth of 100 cm) and forest floor C of 0.5-1.2kg C m -2 across 13 study sites in the PNW.Similarly, Hudiburg et al. (2009) quantified the significant potential for C accumulation in forests of the PNW and Hutyra, Yoon, and Alberti (2011) report even urban C stocks in the PNW exceed average U.S. forest carbon stocks.Franklin, & Spies, 1992, 1993), and the change in ecosystem properties along stand edges (Griffiths & Swanson, 2001).This mosaic of different-aged forest stands varies in its potential to sequester carbon, as demonstrated by the range of mineral soil C, forest floor C and soil respiration observed across different aged stands by Sun et al. (2004).In a landscape of diverse forest management, accurate quantification of soil carbon stocks for monitoring policy initiatives will combine observational and model analysis.
Because soil properties are extremely variable in time and space (e.g.Burrough, 1983), accurately quantifying variation in soil element distribution requires intensive sampling.Soil sampling incurs significant time and monetary costs and is destructive in nature.In order to verify carbon accounting policy objectives, better tools are needed to quantify the spatial distribution of carbon in forest ecosystems (Fahey et al., 2010).
Creating spectral reference libraries is an important first step toward developing simple techniques for rapid soil assessment.Once reference libraries are established, spectral analysis is a relatively quick and inexpensive method for soil analysis.Tools that facilitate our ability to monitor soil properties have the potential to significantly improve our understanding of the effect of management practices on soil formation and soil C storage.We examined the potential for VNIR methods to serve as such a tool in PNW Douglas-fir ecosystems.In this paper we report a comparison of combustion and spectral methods for quantifying element stocks in andesitic soils.First, soils were analyzed for total soil C and N using combustion methods.Subsequently, we applied VNIR analysis to this spatially distributed soil data set to create a spectral database for quantifying soil C and N stocks.Our analysis included rigorous model validation using independent data not used in model development and calibration.In addition, we assessed whether VNIR methods would be robust if limited soil samples are available, by testing the accuracy of models developed from aggregated soil horizon spectral data and by testing the accuracy of models when applied to different aged stands.

Site Characteristics
Old-growth and harvested Douglas-fir-dominated stands were studied at the H. J. Andrews Experimental Forest in the western Cascades of Oregon, United States.Study site soils are Andisols classified as Andic Dystrochrepts (Dyrness, Norgren, Sollins, & Vance-Borland, 1996).Andisols are highly productive soils formed from volcanic ash parent material and with a mineral profile that forms strong bonds with organic matter.The sampled watersheds are predominantly from the Carpenter soil series, with a surface horizon of dark brown gravelly sandy loam, lower horizons of dark brown gravelly sandy loam to loam, and virtually unlimited effective rooting depth due to deep, unconsolidated parent material (Dyrness & Hawk, 1972).The mean monthly temperatures range from 0.6 o C in January to 17.8 o C in July; mean annual precipitation is 2257 mm with 70% of rainfall occurring between November and March (Bierlmaier & McKee, 1989).The selectively harvested stand (watershed 7 or W7) is a 15.4 ha drainage and the old-growth stand (watershed 8 or W8) is a 21.5 ha drainage.The harvested stand drains at an elevation of 940 m and the old-growth stand at 995 m above sea level.An east-west transect was established at 1060 m elevation in the old-growth stand and at 1040 m elevation in the harvested stand.The watersheds include slopes with a 15-45% grade, and south and southeast aspects.The stands share a border, with the harvested stand lying to the west of the old-growth stand.The selectively harvested stand was cut in 1974, broadcast burned in 1975, and planted with Douglas-fir seedlings in 1976.Samples were collected in the summer of 1997, when the harvested stand was a 21-year-aged stand.Analysis of fire history concluded that the old-growth stand contains individuals up to 450 years old (Heyerdahl, Swanson, Berry, & Agee, 1994).

Field Sampling
Organic and mineral soil samples were collected at 2 m intervals along 256 m linear transects.In both stands, location '0 m' corresponds to the plot along the edge shared by the old-growth and harvested stands.As a result, higher plot location values indicate farther distance in meters from the stand boundary.The old-growth stand plot numbers increase from west to east, while the harvested stand plot numbers increase from east to west.
At each sampling location the litter layer was removed, an organic soil core was taken across the entire depth of the organic soil horizon, and a 15 cm deep mineral soil sample was extracted.The upper organic soil was sampled with a 15 cm deep bulb planter and deeper soils where sampled with a 30 cm deep, 2 cm diameter corer.In the field, soil samples were stored in Ziploc® plastic bags in a cooler with ice.Soil samples were air dried until they attained a constant weight and subsequently passed through a 2 mm sieve.Fine roots that filtered through the sieve were removed from each sample using tweezers.

Element Analysis
Soil samples were ground in glass scintillation vials using stainless steel rods; mineral and organic soil replicates were ground separately.Organic and mineral soil samples were combusted to analyze total carbon and nitrogen concentration on a Perkin-Elmer 2400 Series II CHNS/O Analyzer (Waltham, MA).The weight of sample tins ranged from 17-28 mg for organic soil samples and from 30-50 mg for mineral soil samples.Error analysis was conducted for each sample type by taking subsamples from a single soil sample.Analytic organic, pine needle, or peach leaf standards were run every 10 samples to confirm autoanalyzer precision.Replicates were run every 5 samples to test data repeatability.

Statistical Analysis
We calculated summary statistics for total soil carbon and nitrogen concentrations determined using combustion methods.We calculated sample mean, standard deviation, and standard error and compared summary statistics across harvested and old-growth stands using t-test analysis.The relationship between sample total carbon and nitrogen concentration was assessed using correlation analysis.
The spatial distribution of total carbon and nitrogen concentrations was analyzed using autocorrelation and cross-correlation analysis (Wei, 1990).We tested the data series for stationarity using a power transformation test; the smallest residual sums of squares occurred for λ = 1, confirming that the data are stationary and that a variance stabilizing transformation is not warranted (Wei, 1990).Patterns in organic and mineral soil element concentrations were compared within a data series using the autocorrelation function (ACF, Appendix 2 Eq.A2.1) and across two different data series using the cross-correlation function (CCF, Appendix 2 Eq.A2.2) analyses (Wei, 1990).The ACF equation tests for patterns of correlation among adjacent sample points within a single data set.An extension of the ACF equation, the CCF equation tests for such patterns across two separate data sets.Autocorrelation significance at a 95% confidence level is defined as values of the autocorrelation or cross-correlation coefficient (Equation A2.1, A2.2) that fall outside the region defined by ±2/√n (Wei, 1990).The value of an autocorrelation analysis depends on dataset length, with longer transects allowing for autocorrelation analysis of samples separated by longer distances, or lag, intervals.In this study, sample collection occurred at 2 m intervals along 256 m transects, allowing for the comparison of autocorrelation for sampling locations separated by 2-40 m.For the 256 m transect studied, a lagged spatial analysis included 100 data pairs at the largest lag considered (40 m).All programs used for this analysis were written in Matlab (The MathWorks, Inc., Natick, MA).

Spectral Analysis
Air-dried soils sampled for the combustion analysis were scanned using a FieldSpec Pro hyperspectral sensor (Analytical Spectral Devices, Inc., Boulder, CO: ASD, 1997).Absolute reflectance (decimal percent) of samples was recorded over the 350 to 2500 nm range at a 1-nm spectral resolution, yielding a total of 2150 data points per spectrum.Soil samples were placed into 4 cm diameter optical quality petri dishes and illuminated with a quartz tungsten halogen lamp light source held inside a Muglight sensor attachment.Reflectance was recorded through the glass bottom of each dish with a constant angle (55 o from horizontal) at a distance of 4 cm from the sensor.To avoid possible spectral differences originating from particle size variation within a soil sample, after five consecutive readings the sample was rotated 90 • and five additional readings were collected.Sensor calibration was performed using three standards: Spectralon (white spectra), kaolinite, and soil with known spectral characteristics.The white reference sample was run every 20-30 samples to check for machine accuracy; if spectral drift was noticed the lens was cleaned and a new baseline was established.The kaolinite soil reference samples were re-sampled at the end of each sampling day.Analysis was conducted making sure no light escaped through the soil dish.
Conversion of spectra to ASCII was conducted using ViewSpec Pro (Analytical Spectral Devices, Inc., Boulder, CO).We used S-PLUS (TIBCO Software Inc., Palo Alto, CA) to create a master data file with one representative reflectance spectrum per soil sample.Representative reflectance spectra were created for each spectral band by averaging the ten readings obtained at two different positions per sample.In order to identify patterns in the spectral results, we applied a principal components analysis (PCA) to the raw reflectance results using all VNIR spectral bands prior to model calibration.We analyzed each data set separately to allow for a detailed characterization of the spectra.Using PCA allows us to transform the data set into linearly uncorrelated variables-the principal components (PCs).The transformation is defined such that the first PC accounts for the largest variability in the data set.The PC with the highest coefficients or loadings contributes more information toward characterizing the data set.The goal of applying PCA is to reduce the number of relevant variables for characterizing the data.

Statistical Model Calibration and Validation
The accuracy of spectral methods is affected by factors such as sample range, calibration model quality, and calibration technique.We calibrated soil reflectance to soil properties using partial least squares regression (PLSR) analysis.This method quantifies the relationship between two data matrices, X and Y, by a linear multivariate model and is widely used in VNIR reflectance spectroscopy (Daniel, Tripathi, Honda, & Apisit, 2004).In this study X was the soil reflectance and Y was the soil property established with combustion methods.Unlike multiple linear regression (MLR), PLSR can handle data with strong collinearity in the independent (X) variables, which may occur with higher frequency relative to the observations (Y).Vasques et al. (2008) compared the accuracy of different calibration techniques (stepwise MLR, PCA, PLSR, classification and regression tree (CART)) as well as different pre-processing techniques for the estimation of soil C and found that all approaches produced similar results with very successful estimation accuracy.However, the best result was obtained using the PLSR method.PLSR models a linear relationship between variable and reflectance spectra, while models such as multivariate adaptive regression splines (MARS) or CART are used to model a nonlinear relationship.The linear relationship in our soil C and N spectral analysis suggests PLSR is the best model for our system (Vasques et al., 2008).Taking the first derivative of spectra did not improve our estimations (data not shown), rather the raw spectra provided the best results.
In PLSR applications, selection of the number of latent variables is critical to prevent over-or under-fitting the data, which would create models with poor predictive capability.We robustly calibrated the PLSR model using leave-one-out cross-validation (LOOCV).In this method the data is divided into subsets.The calibration model is constructed using all but one subset, and the model is validated against the removed subset.This process is repeated with a different subset removed for model testing in each iteration.We applied LOOCV calibration and cross-validation techniques using only 70% of the data set.In addition to cross-validation, model accuracy was assessed through validation using the remaining 30% of the data set.Random numbers were generated in Excel for this analysis.The PLSR analysis was performed on the ASCII-formatted spectral data using the Unscrambler V.8.0.5 (CAMO, Oslo, Norway).
In addition, we tested the predictive ability of models developed using data from aggregated soil horizons by developing calibration models using spectral outcomes from both the organic and mineral horizons in each stand.Lastly, we tested whether a model calibrated for a particular stand management had predictive power across different stands.We used a model calibrated in the harvested stand to predict element distribution in the old-growth stand as well as the converse scenario.

Prediction Accuracy
In our validation analysis, model accuracy was evaluated using: i) the coefficient of determination (R 2 ) between measured and predicted sample values, ii) the root mean square error of prediction ( ) based on the difference between measured and predicted values, and iii) the ratio of standard deviation to RMSEP ( ).An R 2 > 0.75 and RPD > 2.5 suggests the model is accurate and reliable; spectral results with RPD > 2 are categorized as reliable, while a range of 1.5 < RPD < 2 suggests the spectral analysis is promising, but with inaccuracies (Chang et al., 2001;Dunn, Beecher, Batten, & Ciavarella, 2002;Shepherd & Walsh, 2002).

Combustion Analysis and Summary Statistics
As expected, total carbon and nitrogen content was abundant in organic soil and low in mineral soil (Table 1).There was insufficient organic soil present at some sampling locations to gather a sample, with significant gaps in the presence of a distinct organic soil horizon in the harvested stand.Total soil C and N concentrations were normally distributed.T-test analyses comparing C and N concentration between harvested and old-growth stands showed differences were statistically significant at the α = 0.05 level (p < 0.001 for all comparisons except mineral soil %N with p = 0.026).
A correlation analysis of within sample C versus N concentration confirmed the notion of a constant C:N ratio commonly applied when modeling surface soil horizons.Correlation coefficient results demonstrated strong positive correlation between sample C and N concentration in organic soil of the harvested stand (r = 0.74, n = 89), mineral soil of the harvested stand (r = 0.75, n = 129), organic soil of the old-growth stand (r = 0.84, n = 125), and mineral soil of the old-growth stand (r = 0.7, n = 129).There was no correlation between C and N concentration across organic and mineral soil horizons.
The distribution of total soil carbon and nitrogen concentrations varied widely around the mean over small spatial scales (Figures 1 and 2).This high-frequency change in C and N concentration suggests that there was variation at scales smaller than the scale of measurement.Given a standard error of the replicates that was one to three orders of magnitude smaller than the standard deviation in sample C and N concentrations, the observed meter-scale variation in carbon and nitrogen concentration exceeded the variation introduced by error in the combustion analysis methods (Table 1).
Soil C and N concentration did not exhibit a clear spatial pattern.In the harvested stand, mineral and organic soil C and N concentration showed less high-frequency variation, with a trend toward moderate, clumped troughs and isolated, elevated peaks (Figure 1).In the old-growth stand, organic and mineral soil C and N concentration varied drastically in adjacent plots, with values fluctuating above and below the mean across the entire stand (Figure 2).In both stands, the pattern of variation in C and N concentration was strongly correlated.
In mineral soil samples from the harvested stand, positive autocorrelation in N concentration occurred at 2 and 4 m intervals, demonstrated by ACF values that exceeded the significance criteria, r k = ±2/√n.No significant autocorrelation was observed in the distribution of soil C and N concentration in the old-growth stand.Cross-correlation analysis did not uncover compelling patterns.When the old-growth model was applied to predict harvested stand carbon content, the validation showed an R 2 = 0.72 for mineral soils, but R 2 = 0.40 for organic soils.Table 3. Statistical assessment of model cross-validation and independent validation.Results show model validation for the harvested (W7) and old-growth (W8) stands.Within each stand category, model calibration was implemented for 1) combined organic and mineral soil samples, 2) organic samples only, or 3) mineral samples only Note.†: Validation using samples independent of the calibration data set; ‡: Root Mean Square Error of Prediction; §: Standard Error of Calibration; α: Ratio of Deviation to Prediction (SD/RMSEP).

Carbon and Nitrogen Distribution
Mean carbon and nitrogen content observed in this study was similar to Dyrness et al. (1996) H. J. Andrews experimental forest soil survey observations which resulted in mean element concentrations of 9.2% C and 0.18% N for data averaged across stand ages.Our element analysis also overlapped with observations from a 70-100-year-aged stand that naturally regenerated following fire in the Siskiyou National Forest in western Oregon.This Siskiyou stand was sampled to observe soil variability, with resulting O-horizon element concentrations ranging from 13.1-51.6%C and 0.25-1.6%N and the 0-15 cm mineral soil element concentration ranging from 2.5-13.5% C and 0.09-0.34%N (Homann, Bormann, & Boyle, 2001).
Relative to the old-growth stand, the harvested stand demonstrated a 20% reduction in carbon concentration and a 10% reduction in nitrogen concentration of the organic soil layer.A 14% reduction in mineral soil carbon concentration in the harvested stand relative to the old-growth stand was also observed.Various mechanisms could have caused the reduced soil C and N observed in the harvested stand, relative to the old-growth stand.
Reduced soil C and N could result from the disturbance caused by harvesting, including forest biomass removal during the harvest resulting in reduced litter inputs, post-harvest broadcast burning of organic matter, or soil erosion following the harvest.In addition, reduced soil C and N in the harvested relative to the old-growth stand could have resulted from vegetation and environmental differences during stand succession.The open canopy of a regenerating stand creates a different microclimate (Chen et al., 1992(Chen et al., , 1993)), exposing the surface soils to different post-harvest decomposition conditions.The open canopy of a regenerating stand also supports different understory vegetation, a property that influences watershed C and N cycling.Post-harvest changes in soil organic matter have been documented in other temperate forest ecosystems.Klopatek (2002) found that C and N were lost from harvested Douglas-fir stands 20 and 40 years after harvest, while C and N were gained in old-growth stands.Johnson (1995) found carbon was preferentially lost from SOM relative to nitrogen, causing lower soil C:N ratios relative to unharvested stands 8 years after a whole-tree clear-cut in northeastern U.S. hardwood forests.
The observed variation lacking a distinct spatial pattern at the scale of sampling is consistent with Griffiths and Swanson (2001) work examining soil properties along the edge of different aged Douglas-fir stands.The fine-scale variation observed in this study is consistent with the high variation in soil C and N concentration observed in 70-and 100-year-old Douglas-fir forests of western Oregon, with coefficients of variation ranging from 20-30% reported for C and N concentrations of the organic and mineral soil (Homann et al., 2001).The rapid change from soil C and N concentration peaks to troughs at the meter scale suggests that there is variation at finer spatial resolutions.Consistent with Burrough's (1983) notion that processes that act in a nested and spatially autocorrelated manner at large map scales may appear as a white noise (or random) process at other scales, the data hint at patterns over 50-100 meter lags.However a data series with larger spatial extent is necessary to address pattern formation at this scale.

Soil Reflectance
The maximum overall reflectance magnitude (albedo) of our sampled soils was 45%, which is consistent with previous studies (e.g.Daniel et al., 2004;Summers, Lewis, Ostendorf, & Chittleborough, 2011).As soil water content and organic matter increases, the overall reflectance decreases.Greater absorption of light was observed in organic soil samples relative to mineral soil samples (Figure 3); organic samples exhibit lower reflectance and higher absorbance in the visible and short NIR wavelength range (up to 900 nm).Galvão and Vitorello (1998) as well as Jensen (2000) reported a greater amount of OM is associated with a greater absorption of incident energy, and therefore with a lower spectral reflectance.Similarly, Daniel et al. (2004) reported that organic matter reduced the reflectance in the 550 and 700 nm range.Our spectral trends are consistent with the organic matter spectral dynamics observed in the literature.
The spectral characteristics and PCA results from the visible spectral range demonstrated a consistent relationship to measured C and N concentrations.PC2 had significant negative correlation with measured carbon and nitrogen (Table 2).All wavelengths in the visible range had positive loadings with PC2, while PC2 had negative correlation with measured C and N.Because high organic matter increases visible-range spectral absorption and reduces reflectivity, it is consistent for PC2, which is a grouping of visible range spectra, to have negative correlation with soil C and N content.
Our PC2 trends were consistent with the effect of soil organic carbon (SOC) trends observed in the literature.The highest positive loading was observed in the 650 to 750 nm wavelength (Figure 4d).This is consistent with results from Bartholomeus et al. (2008) that showed the highest correlation coefficients between SOC and 600 and 700 nm range wavebands.Vasques et al. (2008) reported wavelengths in the visible range, such as 430, 460, and 590 nm, to be the most significant bands with the highest stepwise multiple linear regression coefficients for the estimation of soil carbon content.

Application of Spectral Models
The predictive capacity of VNIR methods is primarily dependent on the extent of the relationship between the soil property of interest and the spectral reflectance.Variables such as soil organic matter, carbon, soil texture, and iron are directly related to soil spectral characteristics and specific absorption bands within the VNIR wavelength range.Because SOM is directly correlated to spectral reflectance, our model has good predictive capability.Some soil properties are only indirectly related to soil spectra, such as cation-exchange capacity, and the success of VNIR in estimation of these variables is related to the degree of correlation of the variable with soil parameters that are directly related to soil spectra (Bilgili et al., 2010).
Our analysis showed that VNIR was well-suited to characterize the carbon and nitrogen content of PNW forests on andesitic soils.Organic and mineral soil carbon and nitrogen were successfully estimated using cross-validation and independent validation methods in both harvested and old-growth stands (Table 3).Based on the RPD metric, individual soil horizon calibration models were most accurate for estimating SOC in organic soils of the old-growth site (W8-organic), followed by organic soils of the harvested stand (W7-organic), mineral soils of the harvested stand (W7-mineral) and lastly in mineral soils of the old-growth site (W8-mineral).Model accuracy for nitrogen followed the same sequence as carbon, except model application to mineral soils of the harvested stand (W7-mineral) was slightly more accurate than that in organic soil of the harvested stand (W7-organic).A model calibrated using aggregated soil horizon spectra and observations (W7-combined, W8-combined) likewise resulted in good validation outcomes.Model validation showed reliable outcomes when applied to predict soils with a similar management history; however, model application to predict soil element concentration of different aged stands showed moderate predictive ability.Overall, our calibration model successfully estimates soil carbon and nitrogen based on the RPD metric for all sites without the application of spectral pre-treatment methods.Validation outcomes were similar across methods.However validation using independent samples is a more robust test of the predictive capacity of the calibration model for estimating soil C and N.
The estimation of carbon in all four data sets was consistently more successful than the estimation of nitrogen.
The high accuracy of SOC estimation results from the strong absorption of OM in the visible-range.Some soil N can be estimated due to functional groups existing in the spectra that are directly related to N (Kusumo, Hedley, Hedley, & Tuohy, 2011).In our analysis, the estimation of nitrogen appears to be largely indirect, and mostly due to the relationship of nitrogen with soil carbon.The relationship between measured soil nitrogen and carbon and the resulting R 2 validation metrics clarify the mechanism of nitrogen estimation.Correlation coefficient (r) values between nitrogen and carbon were 0.84, 0.75, 0.74, and 0.70 for organic soil of the old-growth stand (W8-organic), mineral soil of the harvested stand (W7-mineral), organic soil of the harvested stand (W7-organic), and mineral soil of the old-growth stand (W8-mineral), respectively.The highest R 2 value for nitrogen estimation was obtained for the data set of W8-organic (R 2 = 0.84), followed by W7-mineral (R 2 = 0.81), W7-organic (R 2 = 0.75) and W8-mineral (R 2 = 0.73).The validation R 2 values and correlation coefficients paralleled each other, indicating that the estimation of nitrogen was indirect and due to its relationship to carbon.
Earlier studies using the PLSR methodology have likewise obtained successful estimation (R 2 > 0.75) of soil carbon and nitrogen, though these studies used larger data sets for model calibration relative to our analysis (McCarty, Reeves, Reeves, Follett, & Kimble, 2002;Morón & Cossolino, 2002;Dunn et al., 2002).In our study, soil C and N were successfully estimated using fewer than 100 calibration samples, an outcome similar to more recent studies (Vasques et al., 2008;Kusumo et al., 2011;Bartholomeus et al., 2008).

Conclusions
Our analysis applied outcomes from combustion methods for assessing total soil C and N concentrations to develop spectral libraries of soil C and N using VNIR analysis and the PLSR methodology for model development.Our analysis demonstrated that VNIR methodologies are reliable for estimating soil C and N concentrations in soils from managed and old-growth PNW Douglas-fir forest stands.We found that calibration models using separate soil horizon spectra, as well as models using combined horizon spectra, successfully predicted total soil C and N concentrations.While the PLSR models were most accurate when applied to soils of a similar stand age, the models demonstrated moderate reliability in predicting total soil C and N when applied to different aged stands.Given interest in land management that supports increased SOC sequestration, monitoring carbon storage in forest ecosystems will require the integration of observations and modeling tools.Overall, our spectral library can be a robust and inexpensive method for estimating soil C and N concentration for application to management of PNW forests.

Table 1 .
Mean and standard deviation (SD) of carbon (%C) and nitrogen (%N) concentrations for harvested and old-growth stands.Standard error (SE) of the replicates and sample mean C:N are reported.*Mean only includes plots for which there was a sample