Predicting leaf traits across functional groups using reflectance spectroscopy

Plant ecologists use functional traits to describe how plants respond to and influence their environment. Reflectance spectroscopy can provide rapid, non-destructive estimates of leaf traits, but it remains unclear whether general trait-spectra models can yield accurate estimates across functional groups and ecosystems. We measured leaf spectra and 22 structural and chemical traits for nearly 2000 samples from 104 species. These samples span a large share of known trait variation and represent several functional groups and ecosystems. We used partial least-squares regression (PLSR) to build empirical models for estimating traits from spectra. Within the dataset, our PLSR models predicted traits like leaf mass per area (LMA) and leaf dry matter content (LDMC) with high accuracy (R2>0.85; %RMSE<10). Models for most chemical traits, including pigments, carbon fractions, and major nutrients, showed intermediate accuracy (R2=0.55-0.85; %RMSE=12.7-19.1). Micronutrients such as Cu and Fe showed the poorest accuracy. In validation on external datasets, models for traits like LMA and LDMC performed relatively well, while carbon fractions showed steep declines in accuracy. We provide models that produce fast, reliable estimates of several widely used functional traits from leaf reflectance spectra. Our results reinforce the potential uses of spectroscopy in monitoring plant function around the world.

 We provide models that produce fast, reliable estimates of several widely used functional traits 31 from leaf reflectance spectra. Our results reinforce the potential uses of spectroscopy in 32 monitoring plant function around the world. 33

Introduction 37
Plant functional ecology relies on the measurement of traits that influence how plants interact with their 38 abiotic and biotic environment. Such traits can be used to interpret or predict outcomes at the community-39 or ecosystem-scale (Funk et al. 2017), ranging from species sorting across environmental gradients 40 (Kunstler et al. 2016) to rates of carbon and nutrient cycling (Cornwell et al. 2008;Ollinger et al. 2008). The resources needed to carry out intensive trait measurement campaigns using conventional methods has 46 motivated the search for other approaches that could yield fast, reliable trait estimates. 47 48 One such approach is to predict plant traits from reflectance spectra, which report the fraction of light 49 reflected from a surface across a range of wavelengths. For vegetation, the range most often studied is 50 around 350-2500 nm, which includes >97% of solar radiation (American Society for Testing and 51 Materials 2020). Because the structure and chemistry of plant tissues determine their optical properties, 52 many traits can be estimated from reflectance spectra measured at leaf or canopy scales (Jacquemoud & 53 Ustin 2019). At the leaf level, linking reflectance spectra and functional traits provides a fast way to 54 estimate large numbers of leaf traits from individual plants. At the canopy level, it enables researchers to 55 map traits using remote sensing, which may enable monitoring of plant community structure (Durán et al. The performance of statistical models is often evaluated by splitting a dataset at random into calibration 118 and validation subsets. This procedure represents a best-case scenario, since the data are all collected with 119 the same protocols and, based on the mathematics of random sampling, the two subsets must have similar 120 distributions of predictor and response variables. Therefore, we validated our models on both a random 121 subset of our dataset (internal validation) and on independent datasets (external validation)-the latter 122 perhaps representing a more realistic test of how they would perform in practical applications. We assembled a database (n = 1971; species = 104) of leaf spectra and functional traits measured as part 143 of the Canadian Airborne Biodiversity Observatory (CABO; www.caboscience.org). We aggregated data 144 from ten individual projects carried out during 2018 and 2019 at sites across temperate Canada and one 145 site in southwestern Australia (Table 1). The sites span a wide range of edaphic properties, from nutrient-146 poor bogs and phosphorus-limited woodlands to highly fertile post-agricultural and wetland sites. Each 147 sample comprised a large, homogeneous group of sunlit leaves, which we divided up for spectral and trait 148 measurements. We classified samples into seven functional groups: broadleaf trees, coniferous trees, 149 ferns, forbs, graminoids, shrubs, and vines (Table 1)

Leaf spectral and trait measurements 153
We measured directional-hemispherical reflectance and transmittance spectra (350-2500 nm) using an 154 HR-1024i spectroradiometer equipped with a DC-R/T integrating sphere from Spectra Vista Corporation 155 (Poughkeepsie, NY, USA). For each sample, we measured spectra from the adaxial surface of six leaves 156 or (for small or narrow leaves) leaf arrays. We resampled spectra to 1 nm resolution, averaged spectra few projects that account for about one-third (n = 678) of the total samples. Further details on trait 167 measurement protocols can be found in the Supplementary Materials. We compared our trait distributions 168 to the TRY trait database to evaluate how well they span the global range of trait variation. Although it 169 has known geographic and taxonomic biases, TRY is the most comprehensive database of its kind (Kattge 170 et al. 2020). We also performed a principal components analysis (PCA) on the scaled and centered trait 171 data to visualize patterns of trait variation among our samples. 172 173 Within a dataset, chemical traits may vary more in proportion to the mass or the area of the leaf (Osnas et 174 al. 2013). When a trait is distributed mainly proportional to leaf area, mass-normalization induces a 175 negative correlation with LMA. Since interspecific variation in pigments tends to be area-proportional, 176 Kattenborn et al. (2019) used this principle, among others, to argue that pigment content (per unit area) is 177 a better target than concentration (per unit mass) for remote sensing-based estimation. We applied our 178 PLSR modeling approach to both mass-and area-based traits, but we focus on mass-based models since 179 conventional protocols measure most chemical traits on a mass basis. Chemical traits should be assumed 180 here to be mass-based when left unspecified. We also calculated the mass-proportionality of each 181 chemical trait in our dataset as the ordinary least squares slope between the log-transformed area-182 normalized trait and log-transformed LMA (Osnas et al. 2013). 183

Trait estimation 186
Due to the high dimensionality and multicollinearity of spectral data, we used a PLSR modeling approach 187 to build trait estimation models. PLSR projects the high-dimensional input dataset (here, spectra) onto a 188 smaller number of latent components constructed to maximize their covariance with the variable(s) to be 189 predicted (traits). We built separate models using reflectance, transmittance, and absorptance as 190 predictors, in each case using the whole trimmed spectrum (400-2400 nm). We did not transform the 191 distributions of any trait to reduce skewness, as is sometimes recommended (Burnett et al. 2021); in preliminary tests, such transformations did not yield any consistent improvement in model performance. 193 We performed all analyses in R version 3.6.3 (R Core Team 2020). 194

195
We calibrated and validated our main set of models ('primary models') largely following the example of 196 divided the full dataset into 75% calibration and 25% validation sets at random, stratified by functional 198 group (see Table 1). We fit a preliminary PLSR model for each trait on the calibration subset, using 10-199 fold cross-validation to select the number of components to use in our final models. We chose the 200 smallest number of components that brought the cross-validation root mean squared error of prediction 201 (RMSEP) within one standard deviation of the global minimum-a distinct number (4-27) for each trait 202 and type of spectrum. We calculated the variable influence on projection (VIP) metric (Wold et al. 2001) 203 to determine which spectral regions were most important for predicting each trait. 204 205 Next, we fit our final models through a jackknife resampling procedure in which we further divided the 206 75% calibration dataset 100 times into random 70% testing and 30% training subsets. For each iteration, 207 we built a model from the 70% testing dataset using the previously chosen number of components and 208 applied it to the 30% training dataset. This procedure allowed us to get a distribution of summary 209 statistics that shows how model performance varies based on random changes in the training and testing 210 datasets. The primary model comprises the 100 sets of coefficients for each trait-one from each 211 resampling iteration-which we applied to the 25% validation dataset to get a distribution of estimates for 212 each sample. We compared the measured values to the mean estimates for each trait to produce summary 213 statistics (R 2 , RMSE, %RMSE) for internal validations. In all cases, we report RMSE relative to the 1:1 214 line. For robustness to outliers, we defined %RMSE as the RMSE divided by the range of the inner 95% 215 other elements were only available for certain projects and we could not guarantee adequate coverage of 245 all functional groups. We quantified performance using R 2 , RMSE, and %RMSE-calculating %RMSE 246 using the inner 95% trait range of the full dataset in the denominator rather than the inner 95% trait range 247 of only the validation functional group. This metric, which we call %RMSEfull for clarity, avoids apparent 248 reductions in model performance due to a narrow trait range in the validation functional group. 249

250
We conjectured that for a given trait, model performance would decline as the calibration and validation 251 datasets became more dissimilar in their trait distributions (Hypothesis 1). We quantified dissimilarity 252 using the Hellinger distance between the calibration and validation datasets' distributions for each trait, as 253 (Poncet et al. 2019). Here we focused on %RMSEfull because it is 254 normalized to the scale of the data (unlike RMSE) and incorporates bias (unlike R 2 ). We tested the 255 influence of trait dissimilarity on %RMSEfull with analysis of covariance (ANCOVA) using Hellinger 256 distance (continuous), the trait identity (categorical), and their interaction as predictors. 257 258

259
Trait and spectral variability 260 The ranges for most traits in our dataset spanned most of the trait ranges in TRY. Many traits, including 261 LMA, EWT, and all elements other than C and N, had distributions with strong positive skew in both 262 TRY and CABO data, resulting in a long upper tail of high values (Fig. 1  Our samples showed the typical shape of leaf spectra, including a sharp increase in reflectance and 283 transmittance ('red edge') between the visible and near-infrared (NIR; 700-1000 nm) ranges and 284 prominent water absorption bands in the short-wave infrared (SWIR; >1000 nm) range ( Fig. 2; Fig. S2). 285 Compared to other functional groups, needleleaf conifers had low reflectance and transmittance and high 286 absorptance across the spectrum. Forbs also tended to have lower transmittance and higher absorptance in 287 the SWIR. In absolute terms, absorptance and transmittance spectra had greater variation than reflectance 288 spectra, especially in the SWIR. 289

PLSR modeling 291
Internal calibration and validation 292 The accuracy of trait estimates from reflectance spectra varied widely by trait (Table 2 represented upper tail of the measured range, while for C the models overestimated values at the sparse 301 lower tail (Fig. 3). Trait models with lower accuracy on the internal validation data also had greater 302 variability in accuracy across iterations in the resampling analysis-even when leaving out elements other 303 than C and N, which are only available for a few projects (Fig. S10-12). 304

305
The VIP metric revealed broad similarities in the regions of the spectrum that were important for 306 predicting different traits from reflectance spectra ( Fig. S13-15). In particular, major features of the 307 visible region, including the green hump (centered at 550-560 nm) and especially the inflection point of 308 the red edge (710-720 nm) were important for predicting most traits besides EWT. These visible-range 309 importance peaks (especially at the red edge) were often higher for traits like Al, Zn, Mn, and 310 hemicellulose than for pigments, which are the dominant drivers of visible reflectance (Jacquemoud & 311 Ustin 2019). In general, VIP declined into the NIR but had several peaks in the SWIR, including 312 relatively distinct peaks for multiple traits at 1390 nm and 1880 nm, and weaker ones at 1720 nm, 2020 313 nm, 2150 nm, and 2290 nm. For EWT, LMA, Cu, Fe, and Na, VIP remained high throughout a larger 314 portion of the SWIR; uniquely, EWT had its global maximum VIP at 1390 nm. 315

316
In the dataset, most chemical traits were distributed mainly proportional to mass (Table S1). Exceptions 317 included all pigments and several macro-and micronutrients, which showed mass-proportionality less 318 than 0.5. We produced estimates of area-based chemical traits both by building models to estimate them 319 directly and by multiplying estimated mass-based traits by estimated LMA (Table S1; Fig. S16-19). For 320 most traits, estimating area-based traits directly yielded more accurate estimates; the difference was 321 greatest for pigments, Ca, Cu, N, P, which all have area-proportionality less than 0.6. However, multiplying mass-based estimates by LMA estimates was slightly more accurate for many highly mass-323 proportional traits, including Al, C, Na, and cellulose. 324 325 For most traits, transmittance and absorptance spectra yielded slight improvements in performance 326 relative to reflectance (Table S2; Fig. S3-12). Excluding traits that showed very poor performance in 327 general (average R 2 < 0.35), the greatest improvements between reflectance and either transmittance or 328 absorptance spectra (average Δ%RMSE > 1) occurred for LMA, LDMC, EWT, pigments, Ca, K, N, and 329 P. Using transmittance or absorptance spectra never worsened performance considerably (average 330 Δ%RMSE < -1). 331 332

External validation 333
We applied our reflectance-based primary models for each trait to three external validation datasets whose 334 spectral and trait measurement protocols differed from the CABO dataset. In general, models performed 335 less well on these independent datasets than the internal validation (Table 2  also had the highest R 2 in the external validation. As a result, structural and water-related traits (LMA, 338 LDMC, and EWT) had moderate-to-high R 2 in the external validation (0.589-0.842 averaged across 339 datasets). The RMSE for LDMC predictions was only slightly higher than in the internal validation (40.6-340 46.4 vs. 32.9 mg g -1 ). In contrast, both LMA and EWT were greatly overestimated for most samples, so 341 even though the R 2 was moderate-to-high, RMSE was also high (LMA: 0.0855-0.137 vs 0.0175 kg m -2 ; 342 EWT: 0.0395-0.0705 vs 0.0334 mm). Pigments, N, and particularly C and carbon fractions had 343 correspondingly lower prediction accuracy and varying amounts of bias. Many other macro-and 344 micronutrients (besides K) had very low prediction accuracy (R 2 < 0.15).
For each trait, models generally performed worse when applied to a functional group left out from the 348 calibration dataset than when (as in our primary models) calibrated and applied to random selections of 349 samples (Fig. 5-6; Table S3). In particular, models calibrated using the other functional groups tended to 350 show poor performance when applied to graminoids or forbs. Carbon fractions tended to be estimated 351 with relatively low accuracy (high RMSE, low R 2 ); because these carbon fractions were predicted well We built PLSR models to predict widely used foliar traits from reflectance spectra across nearly 2000 369 samples of 104 species, including several functional groups and ecosystems. Our findings underscore that 370 leaf spectra integrate many aspects of leaf phenotypes in a single measurement (Cavender-Bares et al.

Alternate PLSR models 418
We focused on estimating chemical traits on a mass basis, but the remote sensing literature contains 419 powerful arguments for estimating certain constituents on an area basis. In particular, the optical 420 properties of leaves are driven more by the absolute quantity of a constituent than its quantity relative to 421 other constituents. This consideration should matter most for traits distributed in proportion to area 422 (Kattenborn et al. 2019). In our dataset, pigments and several nutrients were mostly area-proportional 423 (Table S1). As Kattenborn et al. (2019) suggested, estimating area-based traits directly was usually more 424 accurate than multiplying mass-based estimates by LMA estimates, especially for traits distributed in proportion to area (Table S1). This consideration is important given that many research questions may 426 require traits expressed on an area basis. 427 428 Using transmittance or absorptance spectra instead of reflectance spectra increases accuracy for most 429 traits, but only slightly (Table S2). It may be relevant that the traits that show the greatest improvement 430 include most of the highly area-proportional traits. Given that transmittance or absorptance spectra 431 Aside from our internal validation, we also considered model performance under less ideal scenarios: we 437 tested the primary models on external validation datasets, and built another set of models to be tested on a 438 functional group left out during their construction. When transferring models to left-out functional 439 groups, performance was usually worse (lower R 2 , higher RMSEfull) than when models were evaluated on 440 a random internal validation dataset (Fig. 6; Table S3). This finding is consistent with previous studies 441 transferring models among species (Helsen et al. 2021) or sites (Yan et al. 2021). In general, trait models 442 performed worse when the trait distributions in the calibration and validation datasets were less similar, 443 but the strength and even the direction of the trend varied among traits (Fig. 6B). Thus, support for 444 Hypothesis 1 was equivocal. Predictions of LMA among conifers and carbon fractions among graminoids 445 were strongly biased and failed to capture the functional distinctiveness of these groups. These results 446 underscore that models must be built using samples whose diversity encompasses the models' expected The external validation was a more severe test of model generality because the data were collected on 450 different species with different instruments using different trait measurement protocols. This kind of scenario may better represent a realistic use case for such models. All traits showed declines in 452 performance from internal to external validation (Table 2) Our models draw from samples representing several functional groups and a wide range of temperate 486 ecosystems, including wetlands, grasslands, and closed-canopy forests (Table 1). However, these samples 487 represent just a portion of Earth's vast plant diversity-omitting, for example, tundra, drylands, and 488 tropical biomes. These biomes would surely contain combinations of optically important traits missing in 489 our dataset, which would need to be represented in a universal model that could be used in mapping 490 functional trait variation worldwide. We hope plant ecologists can use our data and models not only as a 491 tool to address ecological questions, but also as a foundation for building yet more general models. 492

493
Our results imply that some traits can consistently be estimated better than others. We can compare our disparate subsets of the world's plant diversity. Both found that models for LMA and water content were 497 most accurate (R 2 > 0.85), followed by, in varying orders, C and N, pigments, and carbon fractions (R 2 = 498 0.55-0.85). In both studies, models for other elements were typically worse (R 2 < 0.65), but better for Ca 499 and K than many other elements. Both studies also found that transmittance spectra yielded slightly better 500 performance for most traits. The stability in rankings suggests a general hierarchy of traits that are the 501 most practical targets for reliable model-based estimation. 502 The emerging body of research showing that spectral models can yield fast, reliable estimates of many 504 leaf traits across functional groups and ecosystems could expand our ability to map and monitor plant 505 function around the world, bridging between conventional trait measurements and remote sensing. Our 506 findings clarify both the opportunities and challenges of using spectral trait estimates. Research has not 507 yet progressed to the point where plant ecologists can use existing empirical models to estimate traits in 508 their study systems without further validation. Aside from the issues of taxonomic and functional 509 representativeness, our external validation underscores the challenges of transferring models among 510 instruments, which might require new methodological developments to overcome. Even under ideal 511 conditions, some aspects of leaf function (e.g. elemental composition) may be hard to measure accurately 512 based on spectroscopy alone. We hope that the promise of rapid, large-scale monitoring of plant function 513 inspires the coordinated effort needed to understand and (when possible) overcome these challenges. 514

515
In sum, we built models to predict traits from reflectance spectra using nearly 2000 leaf samples from 104 516 species. Traits varied a great deal how well they could be estimated using spectral models. Our models for 517 structural and water-related traits were very accurate in internal validation and showed great promise in 518 external validation; in contrast, our models for many micronutrients performed very poorly. We suggest 519 that these patterns reflect a general hierarchy that emerges mainly from the varying degrees of influence 520 these traits have on the spectrum. Leaf reflectance spectra thus bear the imprint of many aspects of leaf 521 function-but not all to the same degree. Most importantly, we show that many traits can be estimated 522 accurately from spectra using general models that span multiple functional groups and ecosystems. This 523 finding represents a significant advance in leveraging spectra towards major goals in functional ecology.

Acknowledgements 525
We conducted this research at institutions and field sites located on the largely unceded ancestral and 526 contemporary land of many First Nations and Aboriginal people. We thank Jocelyne Ayotte, Zachary All fresh-leaf spectral data are available through the CABO data portal (https://data.caboscience.org/leaf).
Ecological Spectral Information System (EcoSIS, https://ecosis.org/), and upload models to the 551 Ecological Spectral Model Library (EcoSML, https://ecosml.org/). At that stage, we will update this 552 section accordingly. Analysis scripts are available as a repository on GitHub 553 (https://github.com/ShanKothari/CABO-trait-models).        We collected spectra and traits for all specimens using consistent protocols. For each sample, we 852 collected sunlit leaves (>3 h per day of sun exposure) that were relatively homogeneous-for trees and 853 shrubs, at about the same canopy position in the same individual. When possible, we aimed to collect 854 enough leaves in each sample to carry out all destructive trait measurements while retaining several grams 855 of leaf tissue in long-term storage. We avoided leaves that showed clear signs of senescence, and 856 (particularly for spectral measurements) aimed to choose healthy-looking leaves without visible herbivore 857 or pathogen damage. On a subset of leaves, we measured reflectance and transmittance spectra and 858 punched leaf disks for pigment analyses. We also selected random subsets of visually similar leaves to 859 measure other structural or chemical traits. To prepare each sample for chemical trait measurements, we 860 dried several grams of leaves at 65 °C for three days and ground them in a 2 mm cyclone mill. All trait 861 measurements exclude petioles but include raches of compound leaves, since they are functionally 862 equivalent to the midrib of a simple leaf. and ferns (n = 7; species = 1). In total, the database includes samples from 31 families. 876 877

Spectral measurements and processing 878
We measured directional-hemispherical reflectance and transmittance spectra (350-2500 nm) using an 879 HR-1024i spectroradiometer equipped with a DC-R/T integrating sphere from Spectra Vista Corporation 880 (Poughkeepsie, NY, USA). The instrument's nominal bandwidth varies among sensors: ≤ 1.5 nm through 881 the range of the silicon detector array (~350-1000 nm), ≤ 3.8 nm through the first InGaAs array (~1000-882 1890 nm), and ≤ 2.5 nm through the second InGaAs array (~1890-2500 nm). The spectral resolution 883 likewise ranges from 3.3 to 9.5 nm, depending on the spectral region. For each sample, we measured 884 spectra from the adaxial surface of six leaves or leaf arrays (Laliberté & Soffer 2018a). For leaves that 885 were too small or narrow to cover the port of the integrating sphere, we followed a protocol adjusted from 886 We processed reflectance and transmittance spectra following Schweiger & Laliberté (2020). We 892 resampled the spectra linearly to 1 nm resolution and interpolated linearly over the sensor overlap regions. 893 We then averaged spectra from leaves of the same sample and reduced noise by applying a Savitzky-894 Golay filter with varying order and length: order 3 and length 21 from 350-715 nm, order 3 and length 35 895 from 715-1390 nm, order 3 and length 75 from 1390-1880 nm, and order 5 and length 175 from 1880-896 2500 nm. We trimmed spectra to 400-2400 nm to remove the particularly noisy ends of the spectrum. 897 Finally, we calculated absorptance at each wavelength by subtracting the sum of reflectance and 898 transmittance from 1. We performed all spectral processing using spectrolab v. 0.0.10 (Meireles et al.

Trait measurements 916
Leaf mass and area 917 We measured the fresh mass of a subset of leaves from each sample in the field shortly after collection. 918 We rehydrated the leaves in sealed plastic bags with damp paper towels and stored them in a refrigerator 919 for at least 12 h. We then measured their rehydrated mass, scanned them, and measured their area using 920 WinFOLIA (Regent Instruments, Québec, QC, CA). Finally, we dried them at 65 °C for at least 72 h 921 before measuring their dry mass. We used total masses and areas, summed across all leaves from a 922 sample, in trait calculations. We calculated leaf mass per area (LMA) as the dry mass divided by leaf 923 area; leaf dry matter content (LDMC) as dry mass divided by fresh mass; and equivalent water thickness 924 (EWT) as fresh mass minus dry mass, all divided by leaf area (Laliberté 2018). 925 We stored leaf disks in a cooler with ice in the field and transferred them first to a -20 °C freezer in the 953 lab, then to a -80 °C freezer for long-term storage. When we needed to ship them (e.g. the Warren 954 project), we freeze-dried them first. We extracted pigments from frozen leaf disks (or cut pieces, for dataset, we did not measure initial fresh mass of leaves and instead used the rehydrated mass. This 967 difference could cause EWT to be overestimated and LDMC to be underestimated. However, relative 968 water content in the remaining projects was usually close to 100% (median: 87.7%, 2.5-97.5 th percentile: 969 68.3-97.6%). We also removed solubles and hemicellulose values from the Dessain dataset for the same 970 species whose values we removed from the main CABO dataset. Finally, we used a conservative outlier 971 detection procedure-the same as we used for the main CABO dataset-to remove eight data points from 972 Dessain ICP-OES data. In LOPEX, there were two estimates of both cellulose and lignin for each sample, 973 so we took the average for each one. 974

975
Besides applying the main set of models described in the main text, we also tested whether certain 976 transformations of the spectral data could improve performance on the external validation data-as, for used: (1) brightness normalization, which normalizes each spectrum to a unit vector while preserving its 979 shape (Feilhauer et al. 2010), and (2) continuum removal, which performs an albedo normalization by 980 reporting the difference between the actual spectrum and a linear interpolation of its convex hull (Clark & 981 Roush 1984). We first transformed the spectra in the CABO dataset using either brightness normalization 982 or continuum removal and built models using the calibration and validation approaches described in the 983 main text (model performance for internal validation not shown). We then applied the models to the 984 external validation data transformed likewise. Neither transformation caused any systematic improvement 985 in trait estimates (Tables S4-5).            Al (mg g -1 ) 0.003 0.0742 30.6 Ca (mg g -1 ) 0.005 11.7 37 Cu (mg g -1 ) 0.001 0.0145 33.4 Fe (mg g -1 ) 0.046 0.0526 26.4 K (mg g -1 ) 0.468 9.02 28.6 Mg (mg g -1 ) 0.123 2.37 55.1 Mn (mg g -1 ) 0.004 0.503 54.5 Na (mg g -1 ) 0.084 0.907 23.3 P (mg g -1 ) 0.063 2.33 58 Zn (mg g -1 ) 0.037 0.107 31.7