Spectroscopic Remote Sensing of Non-Structural Carbohydrates in Forest Canopies

Non-structural carbohydrates (NSC) are products of photosynthesis, and leaf NSC concentration may be a prognostic indicator of climate-change tolerance in woody plants. However, measurement of leaf NSC is prohibitively labor intensive, especially in tropical forests, where foliage is difficult to access and where NSC concentrations vary enormously by species and across environments. Imaging spectroscopy may allow quantitative mapping of leaf NSC, but this possibility remains unproven. We tested the accuracy of NSC remote sensing at leaf, canopy and stand levels using visible-to-shortwave infrared (VSWIR) spectroscopy with partial least squares regression (PLSR) techniques. Leaf-level analyses demonstrated the high precision (R2 = 0.69–0.73) and accuracy (%RMSE = 13%–14%) of NSC estimates in 6136 live samples taken from 4222 forest canopy species worldwide. The leaf spectral data were combined with a radiative transfer model to simulate the role of canopy structural variability, which led to a reduction in the precision and accuracy of leaf NSC estimation (R2 = 0.56; %RMSE = 16%). Application of the approach to 79 one-hectare plots in Amazonia using the Carnegie Airborne Observatory VSWIR spectrometer indicated the good precision and accuracy of leaf NSC estimates at the forest stand level (R2 = 0.49; %RMSE = 9.1%). Spectral analyses indicated strong contributions of the shortwave-IR (1300–2500 nm) region to leaf NSC determination at all scales. We conclude that leaf NSC can be remotely sensed, opening doors to monitoring forest canopy physiological responses to environmental stress and climate change.


Introduction
Non-structural carbohydrates (NSC) are the mobile portion of a plant's carbon stock, comprised primarily of sugars, starch and pectin [1,2].Also known as non-structural carbon or soluble carbon, plant NSC are produced and stored in leaves and can be transported to and stored in stems and roots.Plant NSC stocks reflect a balance between carbon fixation via photosynthesis and demand for longer-lasting compounds, such as cellulose and lignin.As such, NSC measurements provide biochemically-based insight into physiological performance (carbon "source") relative to whole plant growth (carbon "sink") [3,4].
NSC are important in trees and other perennial plants, because they provide chemical energy at times of reduced resource availability, such as during dry periods or leaf-off periods in deciduous plants [5].Recent work reveals that NSC are an important determinant of tree survival during drought [6].In that study, tropical tree species with inherently low NSC stocks in foliage, stems, and roots were about twice as likely to die in drought compared to those with naturally high NSC stocks.O'Brien et al. [6] also found that foliar NSC concentrations mirror those in stems and roots, despite the fact that NSC stocks may be higher in woody tissues compared to foliage [4].This suggests that leaf NSC can be used as a general diagnostic for plant NSC, at least in a spatial or geographic context.Given the observed and predicted increases in drought frequency and severity in tropical forest regions [7,8], there is a need for studies to determine foliar and whole-plant NSC concentrations as potentially powerful predictors of drought tolerance [9].This is particularly true in tropical forests, where leaf NSC concentrations display strong phylogenetic and regional variation [10].
The foliage of forest canopies is notoriously difficult to measure at any scale.Tall canopies and complex architectures preclude access and limit field collections to a few leaves or branches per crown.The data resulting from such approaches are challenging to interpret or to scale up and are not often random or systematic samples of an individual tree, a canopy of trees or a landscape [11].Remote measurement of foliar traits is the only way to accurately sample at stand and ecosystem levels, but remote sensing approaches for NSC estimation have not been broadly developed.Two scales of analysis are needed: (i) leaf-level studies to determine the maximum potential expression of NSC to spectral reflectance and transmittance measurements; and (ii) canopy-level studies to understand if leaf NSC-spectral relationships scale up amidst a background of varying structure (e.g., leaf area index, leaf angle distribution, canopy gaps).
Here, we present a multi-scale analysis establishing the potential for remote measurement of foliar NSC concentrations in forest canopies.We put emphasis on humid tropical and sub-tropical forest canopies, because they exhibit widely varying NSC concentrations [10].From a remote sensing perspective, we focus on NSC estimation using optical spectroscopy, which provides measurements of leaf reflectance and transmittance, and canopy reflectance, in the 400-2500 nm wavelength range.Spectroscopy has an established history in the detection and analysis of leaf and canopy chemical traits [12,13], but its use for NSC estimation is unproven.From both a theoretical and a laboratory spectroscopy standpoint, remote sensing of foliar NSC should be possible based on the specific wavelength sensitivities of sugars, starch and pectin in the shortwave region of the spectrum [14,15].Spectral features centered at 1450, 1490, 1510-1580, 1780, 1900, 1960, 2080-2100, and 2270-2280 nm are particularly promising [14], but none or few wavelengths are likely to provide the information needed to predict NSC concentrations accurately.Instead a combination of spectral features, available through full-spectrum chemometric approaches, is likely the best candidate for operational NSC determination at leaf and canopy scales.

General Approach
We took a three-tier approach to our assessment of NSC from optical spectroscopy (Figure 1).First, we collected leaf reflectance and transmittance spectra (400-2500 nm) in fresh foliage taken from thousands of forest canopies in the field and paired these spectra with laboratory assays of leaf NSC concentration.Second, we combined the leaf spectra with a canopy radiative transfer model that simulates vegetation structural variability and sensor noise and used the simulated canopy reflectance spectra to assess the potential limits of foliar NSC determination at the canopy scale.Finally, we used airborne imaging spectroscopy to assess foliar NSC concentrations at the stand level in forest plots distributed throughout the Amazon basin.For consistency, our leaf and airborne spectral measurements and field-based leaf collections were carried out in the early dry season for each study forest.At each of these scales, the chemometric method, called partial least squares regression (PLSR) analysis, as described below, was used to estimate NSC at leaf, modeled-canopy and actual forest stand scales.We developed PLSR-ready datasets for leaves, modeled canopies and real tropical forest field plots using visible-to-shortwave infrared (VSWIR) (400-2500 nm) spectroscopy.These three datasets provided a means to assess the precision and accuracy of NSC determinations from the simplest leaf-level case to the most complex tropical forest stand case.

Leaf Spectral Properties
A total of 6136 leaf samples representing 4222 different canopy species were collected from 61 sites distributed among tropical and sub-tropical forests in the Amazon Basin, Australia, Borneo, the Caribbean region, Central America, the Hawaiian Islands and Madagascar (Figure 2, Table 1).The global distribution of the samples, and their breakdown into plant families, genera and species, were described in Asner et al. [16].Briefly, the dataset is comprised of the most common plant habits found in tropical forest canopies, including tree (n = 5233), liana (648), palm (74), hemi-epiphyte (109) and vine (51).Across all sites, mean annual temperature (MAT) ranges from 8 to 27 °C, mean annual precipitation (MAP) ranges from 1200 to 6100 mm•yr −1 and ground elevation varies from 0 to 3660 m.Soil type varies strongly across sites, from nutrient-poor Oxisols (clays) and Entisols (white sands) to nutrient-rich Inceptisols.Hemispherical reflectance and transmittance in the 400-2500-nm wavelength range were measured on six randomly selected fresh leaves immediately after acquiring them from each forest canopy in the field (n = 6136 × 6 = 18,408 reflectance and 18,408 transmittance spectra).The spectral measurements were taken at or close to the mid-point between the main vein and the leaf edge and approximately half way from petiole to leaf tip.The spectra were collected with a field spectrometer (FS-3 with custom detectors and exit slit configurations to maximize signal-to-noise performance; Analytical Spectra Devices, Inc., Boulder, CO USA), an integrating sphere designed for high-resolution spectroscopic measurements and a custom illumination collimator [17].Twenty-five spectra per sample were averaged and then referenced to a calibration block within the integrating sphere (Spectralon, Labsphere Inc., Durham, NH).An integrating sphere and collimated light source are required to obtain directional-hemispherical reflectance and transmittance measurements, which are subsequently required for use in scaling up to the canopy level with radiative transfer models [18][19][20].The high-fidelity measurement capability of our field instruments resulted in leaf spectra that did not require smoothing or other filters commonly used in leaf optical studies.

Leaf NSC Assays
The method for the chemical determination of leaf NSC was reported by Asner and Martin [21], and the protocol is provided on the Spectranomics website (http://spectranomics.ciw.edu).Briefly, NSC concentration was determined in 0.5 g dry ground leaf tissue through using a neutral detergent fiber (NDF) solution in a fiber analyzer (Model 200/220, Ankom Technology, Macedon, NY, USA).The leaf samples were placed in 1800-1900 mL NDF and agitated for 75 min, then rinsed for 5 min in deionized water.This procedure was repeated three times to ensure maximum removal of NSC.The change in dry mass of the sample, before and after extraction, was used to calculate NSC concentration.Leaf standards were used as references with each digestion to ensure consistency across assays.

Canopy Reflectance Modeling
We projected the mean leaf reflectance and transmittance spectra (n = 6136 pairs) to the canopy level using the 4SAIL2 (4-Stream Scattering by Arbitrarily Inclined Leaves) radiative transfer model [20].This model simulates top-of-canopy spectral reflectance based on the measured leaf hemispherical-directional reflectance and transmittance spectra, along with the variation of leaf area index (LAI), leaf angle distribution and other crown geometric-optical properties.For each canopy simulation, a random combination of canopy structural parameters was selected from a very wide range of potential values based on growth-form (e.g., tree, liana), which was combined with the measured leaf spectra, to generate a canopy reflectance signature.Value ranges for each canopy structural parameter are listed in Table 2.
Table 2. Ranges of canopy structural parameters randomly selected during canopy radiative transfer model simulations.LAI = leaf area index; LAD = leaf angle distribution; Cv = crown covering the ground at nadir; Zeta = ratio of crown diameter to tree height.The modeled LAI variation is considered extreme for tropical canopies [22].Note: 1 Leaf angle distribution (LAD) is described by a two-parameter model, with the first parameter controlling average leaf inclination and the second parameter controlling the bimodality of the leaf angle distribution [20].The ranges shown here are extremely wide for tropical vegetation canopies.
These ranges are extreme in most cases, particularly with respect to LAI and leaf angle, both of which do not vary as widely as tested here at spatial resolutions typical of airborne spectroscopy (>1 m).Here, we selected extreme ranges compiled from the literature by Asner et al. [23] and Asner and Martin [17] as a means to create possible worst-case scenarios of canopy structure overpowering leaf spectral-chemical variation.Our modeling also included a treatment of both random and systematic spectrometer sensor noise using the technique described by Asner et al. [16].Together, this provided a method to estimate the effects of extreme canopy structural variability and sensor noise on leaf NSC estimation using canopy reflectance in the 400-2500-nm wavelength range.A useful aspect of this approach is that it likely represents the chemical and structural variability of broadleaf forests worldwide.

Airborne NSC Study
We tested our ability to quantitatively estimate leaf NSC concentrations in actual tropical forest canopies.Airborne imaging spectrometer data were acquired over 79 one-hectare field plots distributed throughout the Peruvian Amazon using the Carnegie Airborne Observatory (CAO) (Figure 3), which includes a high-fidelity VSWIR imaging spectrometer [24].In each field plot, multiple full-sunlight canopies were selected for leaf collections (n = 3-38 per plot) as described in detail by Asner et al. [11].Foliage from these canopies was assayed for NSC using the same technique described earlier for the global leaf collection.The field-collected NSC data were averaged to the plot level.The VSWIR data were collected over each field plot from an altitude of 2000 m above ground level (a.g.l.), an average flight speed of 55-60 m s −1 and a mapping swath of 1200 m.The VSWIR spectrometer measures spectral radiance in 480 channels spanning the 252-2648 nm wavelength range in 5-nm increments (full-width at half-maximum).The spectrometer has a 34° field-of-view and an instantaneous field-of-view of 1 mrad.At 2000 m a.g.l., the VSWIR data collection provided a 2.0-m ground sampling distance, or pixel size, throughout each study landscape.The VSWIR data were radiometrically corrected from raw DN values to radiance (W sr −1 m −2 ) using a flat-field correction, radiometric calibration coefficients and spectral calibration data collected in the laboratory.The image data were atmospherically corrected from radiance to apparent surface reflectance using the ACORN-5 (Atmospheric Correction Now) model (Imspec LLC, Glendale, CA USA).The reflectance imagery was corrected for cross-track brightness gradients using a bidirectional reflectance distribution function (BRDF) modeling approach described by Colgan et al. [25].The sunlit portions of canopies in each 1-ha plot were filtered and averaged based on the method described by Asner et al. (2015).This method removes pixels unsuitable for sunlit canopy spectroscopic measurement, including non-canopy surfaces (bare ground), shaded canopy pixels and pixels with low foliar content.To achieve this, we use a LiDAR (light detection and ranging) scanner flown with the spectrometer to estimate the height of the vegetation at a resolution of 8 laser shots per VSWIR spectral pixel and to model inter-canopy shade (Asner et al., 2015).To ensure that remaining candidate spectral pixels have sufficient foliar content, we apply a minimum NDVI threshold of 0.8.Spectral pixels that pass this combination of filters are considered suitable for NSC analysis at a mapping resolution of 1 ha.The resulting average reflectance spectrum of each 1-ha field plot was trimmed at the far ends (<400 nm, >2450 nm) of the measured wavelength range, as well as in regions dominated by atmospheric water vapor (1350-1480, 1780-2032 nm).Water vapor features preclude the use of these wavelength regions in canopy chemical analyses.

Chemometric Analyses
We used PLSR analysis [26] to characterize the strength of NSC expression in the spectroscopy of leaves, modeled canopies and actual canopies at the stand level.Leaf reflectance and transmittance were each tested using the field-based measurements followed by re-sampling to the 10-nm full-width half-maximum bandwidth spanning the 400-2500 nm spectral range and averaging of six individual leaf measurements (n = 6136 leaf samples).Canopy reflectance (n = 6136 simulated canopies) was tested in a similar configuration, but with the 1330-1430-nm and 1760-2030-nm atmospheric water vapor regions removed from the simulated VSWIR data [27].Actual forest stand data (n = 79 field plots) from the Amazon were treated the same way.
For all PLSR analyses, we minimized statistical overfitting by utilizing the prediction residual error sum of squares (PRESS) statistic [28].The PRESS statistic was calculated through a cross-validation prediction for each PLSR model.This cross-validation procedure iteratively generates regression models while reserving a portion of the samples (10% for input datasets with >100 samples; leave-one-out for <100 input samples) from the input dataset until the root mean square error (RMSE) for the PRESS statistic is minimized.The leaf and modeled canopy datasets were each randomly divided into 40 subsets to generate 40 unique PLSR models containing approximately 150 samples each.For the measured stand-level reflectance data, 70% were randomly selected on an iterative basis to generate 1000 PLSR models.For each subset (leaf and modeled canopy) or iteration (stand level) in this evaluation, PLSR weighting coefficients derived for each spectral band were multiplied by the spectral values to generate predictions of NSC concentration.This provided a way to calculate the variation in the calibration predictions of NSC concentrations from each dataset.Statistics were computed for the PLSR equations resulting in robust models (R 2 > the mean of the 40 or 1000 iterations at the leaf and modeled canopy or actual canopy, respectively) to assess the precision and accuracy of NSC remote sensing with spectroscopy.The coefficient of determination (R 2 ) was used as the metric of PLSR model precision and the RMSE as an assessment of the model accuracy.The mean PLSR weighting coefficients are provided in Appendix Table A1.

Leaf-Level Performance
The global compilation of leaf reflectance and transmittance spectra indicated very wide-ranging values at nearly all wavelengths (Figure 4).Maximum reflectance variation occurred in the near-infrared (IR) and shortwave-IR between 800 and 1800 nm.For leaf transmittance, maximum variation occurred in the shortwave-IR from 1500-2450 nm.The variation expressed in these spectra meets or exceeds the variation reported for leaf optical properties in other biomes [29][30][31][32][33][34] and matches the reported variability achieved in models of leaf optical properties [19,35].This suggests that our leaf optical database is globally relevant.Chemical analyses of the 6136 leaf samples indicated extremely wide-ranging concentrations of NSC, from 16.5% to 84.5% of total leaf mass (Table 3).Parallel to the spectra, this NSC range approximately meets the range reported for live, fresh foliar material worldwide [1,5,9].Based on this range of chemical values, the PLSR analyses indicated that NSC can be estimated at the leaf level from reflectance with high precision (mean R 2 = 0.73 ± 0.06) and accuracy (average %RMSE = 12.9%) (Table 3).Transmittance-based results showed similar performance, with mean R 2 = 0.69 ± 0.03 and average %RMSE = 13.9%.Analysis of sub-regional leaf datasets, such as from the Amazon basin (Table 1), indicated similar performance levels.Spectral weightings from the PLSR models with the largest deviations from zero (positive or negative) indicate that the shortwave-IR (2000-2450 nm) and, to a slightly lesser extent, the 1300-1800 nm region, were most critical to the determination of NSC concentrations from leaf reflectance and transmittance (Figure 5a,b).PLSR spectral weights are derived from matrix multiplication that simultaneously takes into account spectral variation relative to changing chemical concentration.In Figure 5, departures from zero indicate the increasing importance of spectral features in determining a chemical concentration [36].The most important features align with known centers of NSC expression in the spectrum, particularly from 1450 to 2270 nm, as detailed by Curran [14].Table 3. Calibration performances for non-structural carbohydrates (NSC) derived from optical spectroscopy in the 400-2450-nm wavelength range at leaf, simulated crown and forest stand levels.Mean ± standard deviation values are presented with minimum and maximum values in parentheses.RMSE = root mean squared error; %RMSE is standardized to the mean value.Vectors indicate the number of spectral dimensions selected by the partial least squares regression (PLSR) analyses to achieve the reported precision (R 2 ) and accuracy (RMSE) results.The range of vector numbers indicates variation among repeated PLSR model runs.For leaf reflectance and transmittance, as well as canopy reflectance, total n = 6136; these samples were randomly partitioned into 40 sub-samples with nsub = 150 for iterative PLSR analyses.For stand-level reflectance, n = 79; these samples were randomly partitioned into 1000 sub-samples with nsub = 55 for iterative PLSR analyses.

Canopy-Level Performance
Canopy simulations using the leaf reflectance and transmittance dataset (Figure 4) integrated with variable canopy structure and sensor noise resulted in an extremely wide range of canopy reflectance (Figure 6a), relative to actual stand-level canopy observations (Figure 6b).Maximum reflectance variation was observed in the near-IR (800-1300 nm) and secondarily in the shortwave-IR between 1500 and 1800 nm.The extreme structural variation incorporated into the modeled spectra reduced the precision (mean R 2 = 0.56 ± 0.09) and accuracy (avg.%RMSE = 16.1%) of foliar NSC estimates, as compared to leaf-level results (Table 3).Spectral PLSR weightings also shifted at the canopy scale, with the maximum signal for NSC estimates expressed in the 1200-2300 nm wavelength ranges (Figure 5c).

Airborne Imaging Spectroscopy
Airborne imaging spectroscopy of the 79 one-hectare plots showed much less spectral variability compared to the simulated canopies (Figure 6b).This was expected given the approach of automated averaging of spectra pre-filtered and selected for sunlit, highly foliated portions of canopies in each field plot [11].Moreover, actual stand-level spectral data were of somewhat lower overall reflectance in the near-IR as compared to the simulated canopy-level spectral data.This is caused by within-canopy shading, which is present in the real data and absent in the simulations.Despite these differences, similar to the canopy simulations, we found that maximum stand-level reflectance variation was expressed in the near-IR (800-1300 nm) and secondarily in the shortwave-IR from 1500-1800 nm.PLSR analyses of the filtered airborne spectra against leaf NSC samples taken from the field plots indicated good precision (mean R 2 = 0.49 ± 0.14) and excellent accuracy (average %RMSE = 9.1%) (Table 3).Whereas precision declined slightly in comparison to leaf-level NSC retrievals, the error (RMSE, %RMSE) decreased relative to the canopy-scale simulations.This results from the averaging of spectra and leaf samples within each field plot, compared to a very liberal use of canopy structural variation to the modeled spectra (Table 2).Spectral PLSR weightings indicated the importance of the 700-to 720-nm range ("red-edge") and the shortwave-IR (1500-2300 nm) in determining foliar NSC concentrations at this ecological scale (Figure 5d).

Discussion
Using a three-tier strategy, we determined that leaf NSC concentrations can be estimated using high-fidelity spectroscopy at leaf, canopy and whole-stand scales with demonstrably high precision and accuracy compared to laboratory-based chemical assays.Laboratory wet-chemical assays of NSC typically result in relative errors of about 9% on a per sample basis [37], and our lab sampling protocols result in a similar error level.Leaf-level spectroscopy yielded relative errors ranging from just 13%-14% (Table 3).When averaging samples and spectral signatures at the forest-stand level, relative uncertainties were reduced to about 9%.From either ecological or chemometric standpoints, the spectral-and wet chemical-based approaches are nearly indistinguishable.
Our model-based results should be considered quite liberal in the context of canopy structural variation, serving as a contributor to canopy reflectance.We varied LAI from 1.0 to 8.0 units (Table 2), which not only incorporates a global range of LAI values [23], but is also far more variable than what is typically encountered in closed-canopy tropical forests [38].Our leaf angle distributions and other model parameters were also structurally very wide ranging [18,20].Surprisingly, we found that leaf NSC estimates from modeled canopy spectroscopy remained demonstrably good, although precision and accuracy did suffer under such extreme canopy structural variability compared to the leaf-level analyses.This is important because, although past studies highlight the potential use of spectroscopy for canopy chemical trait analysis [13,14,39], other studies have emphasized the potentially dominant role of canopy structure over chemistry in determining canopy spectral reflectance [40].Here, we found insufficient evidence to suggest that structural variability in forest canopies will critically impair NSC estimation from imaging spectroscopy.
An additional finding was that the portions of the reflectance spectrum responsible for NSC determinations shifted in wavelength when going from leaf to canopy and stand levels (Figure 5).Whereas the shortwave-IR between 1300 and 2500 nm was almost uniformly important to NSC estimation at the leaf level, canopy-scale estimates were shifted to shorter wavelengths (closer to 1200 nm).This is caused by increased absorption at the canopy level, which is related to increased leaf area index [17,19].The leftward shift of the PLSR weightings was more pronounced at the stand level, where the spectral red-edge (~750 nm) was a co-predictor of NSC, in addition to spectral features in the shortwave-IR.This too is caused by canopy-scale absorption.Finally, we note that most of the PLSR weighting "peaks" (departures from zero to negative or positive values; Figure 5) are in line with many of the NSC-related features first suggested by Curran [14].This suggests consistency in the expression of foliar NSC in leaf and canopy reflectance spectra.
The total carbon content of a leaf is approximately equal to the sum of NSC, cellulose, hemi-cellulose and lignin [41].One might therefore predict that remote sensing of NSC is somehow inversely related to remote estimation of the other carbon molecules, and indeed, lignin and cellulose have long been a focus of imaging spectroscopy [12,39].In a previous study, we found that the spectral reflectance features associated with NSC were anti-correlated with spectral features expressed by lignin (r = −0.64)and cellulose (r = −0.77;p < 0.01) [11].This internal consistency in the diversity of spectral expression among major leaf carbon constituents suggests that all three sets of compounds can be estimated from VSWIR imaging spectroscopy.If so, this will open new doors to large-scale ecological studies of multiple carbon fractions in the foliage of forest canopies.The resulting data could, in turn, be used to estimate the relative strength of carbon sources (photosynthesis) and sinks (growth; tissue allocation) at ecological scales unobtainable using field and laboratory sampling alone.The dynamics of these environmentally-responsive carbon pools will be a focus of forthcoming studies with airborne high-fidelity imaging spectroscopy from the Carnegie Airborne Observatory.
Remote sensing of foliar NSC will allow for measurement and monitoring of the direct products of photosynthesis in forest canopies at an ecological scale never before achieved.This will facilitate large-scale assessments of forest canopy responses to changes in multiple environmental factors, including climate.Drought and temperature tolerance are critically important applications of NSC remote sensing with spectroscopy.With the future promise of global imaging spectroscopy via Germany's EnMap (Environmental Mapping and Analysis Program) and NASA's HyspIRI (Hyperspectral Infrared Imager) missions, climate change effects on forest canopy NSC concentrations will, for the first time, be mapped and placed in an Earth system context.Until then, airborne imaging spectrometers will be the best way forward to understand NSC dynamics at landscape to regional scales, as well as at the scale of individual canopies and species.The spectroscopy of leaf NSC is a gateway to developing a chemical basis for remote sensing of forest physiology at the macroscale.

Conclusions
Remote sensing of leaf non-structural carbohydrates (NSC) may provide a spatially-explicit understanding of forest canopy exposure or susceptibility to increasing temperatures, decreasing precipitation and other climate perturbations.Using a global dataset of leaf chemical and spectral properties in tropical forests, we found that leaf NSC concentrations in 6136 plants can be estimated with high precision (R 2 = 0.69-0.73)and accuracy (%RMSE = 13%-14%).Incorporating leaf spectral data into a radiative transfer model resulted in a reduction in the precision and accuracy of leaf NSC estimation at the canopy level (R 2 = 0.56; %RMSE = 16%).However, application of imaging spectroscopy to 79 one-hectare plots in Amazonia indicated the good precision and accuracy of leaf NSC estimates at the forest stand level (R 2 = 0.49; %RMSE = 9.1%).Spectral analyses indicated strong contributions of the shortwave-IR (1300-2500 nm) region to leaf NSC determination at all scales.These remotely-sensed estimates of NSC are indistinguishable, in terms of precision and accuracy, from those made via laboratory assay and can now be carried out at ecological scales otherwise intractable via ground-and laboratory-based studies.Future airborne and satellite-based spectrometers should be designed to deliver high-quality spectra, such as those used in this study, in order to advance NSC mapping for studies of biospheric responses to climate change.

Figure 1 .
Figure 1.Overview of our three-tier approach to estimate non-structural carbohydrates at (a) leaf, (b) simulated canopy and (c) forest stand scales.

Figure 2 .
Figure 2. Map of global sampling locations for foliar spectral and non-structural carbohydrates (NSC) in tropical and sub-tropical forest tree canopies.Dots indicate the sites listed in Table 1.Site names BCI, PNM, and PNSL are Barro Colorado Island, Parque Natural Metropolitano, and Parque National San Lorenzo, respectively.

Figure 3 .
Figure 3.The regional distribution of one-hectare field plots in the Amazon basin and Andean forests of Peru.The map is a digital elevation model (DEM) of Peru derived from NASA Shuttle Radar Topography Mission (SRTM) data.Circles contain clusters of field plots as indicated and are arrayed geographically to maximize environmental and floristic variation as reported by Asner et al. [11].

Figure 4 .
Figure 4. Leaf-level (a) reflectance and (b) transmittance spectra of fresh (live) samples collected from 6136 forest canopy growth-forms in sub-tropical and tropical forests worldwide.The locations of the sample collections are shown in Figure 2, and the general site information is provided in Table 1.The thick dashed line indicates mean values; dark grey areas indicate one standard deviation; light grey areas indicate the total range of values.

Figure 5 .
Figure 5. PLSR spectral weighting factors for NSC estimation in foliage at three scales: (a) leaf reflectance and (b) transmittance; (c) modeled canopy reflectance; and (d) actual forest stand-level reflectance.In each panel, the black line indicates the mean values, and the grey shading indicates one standard deviation among PLSR models run iteratively on random subsets of samples from each leaf, canopy and stand-level database.

Figure 6 .
Figure 6.(a) Simulated canopy reflectance calculated using all leaf spectra (Figure 4) with a canopy radiative transfer model and a random selection of canopy structural properties (Table 2).(b) Actual forest stand-level reflectance spectra from 79 one-hectare field plots in the Amazon basin.The thick dashed line indicates mean values; dark grey areas indicate one standard deviation; light grey areas indicate the total range of values.

Table 1 .
. Site names BCI, PNM, and PNSL are Barro Colorado Island, Parque Natural Metropolitano, and Parque National San Lorenzo, respectively.Collection sites and taxonomic partitioning of the leaf spectral and NSC dataset.