During fall 2018, we collected senescent leaf litter from the Forests and Biodiversity experiment (FAB1) one- to two-year-old trees were planted in a 0.5 × 0.5 m grid within plots that varied in their species richness (1, 2, 5, or 12 species) and composition. The species pool included eight broadleaf angiosperms and four needleleaf conifers (Table 1). We left out one species (Juniperus virginiana) from sampling because its scaly needle morphology meant that individual trees seldom produced enough litter at once for spectroscopic or chemical analyses. To get enough samples of another species (Acer negundo), we also collected material from mature trees in a residential yard in Minneapolis, MN, USA 50.0 km from FAB1. We collected litter samples as close as we feasibly could to the time of senescence using both litterfall traps on the ground and hand-collection of senesced leaves from trees. We only took leaves from trees when the petiole had a complete abscission layer and the leaf could be detached with a gentle pull (Chapin & Moilanen 1991). Because tree species dropped their leaves at different times, we collected litter repeatedly from September to November 2018 to ensure that the litter we collected had recently 132 133 134 135 136 137 138 139 140 141 senesced. To avoid leaching we discarded litter samples from traps after rain, although nitrogen and carbon leaching tends to be low even under extreme wetting (Schreeg et al. 2013).

We collected 322 litter samples, seeking to represent the variation in leaf traits within each species (Table 1). While our sampling did not explicitly account for the species richness and composition treatments in FAB1, we collected leaves from each species across neighborhoods and crown positions to enhance intraspecific variation. Each sample included 1.5-3.0 g of tissue and could comprise one leaf or multiple similar leaves. When a single trap contained more than 2.5-3.0 g of leaves from a single species, we divided them into two or more samples by keeping together leaves that we judged by eye to be more visually similar. We dried each sample in a dark shed at 40 °C for at least three days, and then stored them in a cool, dark room until spectral and trait measurements. 143 144 145 146 147 listed are deciduous angiosperms, and all needleleaf species are evergreen conifers. We did not sample J. virginiana. Twelve of the A. negundo samples come from a residential yard rather than FAB1. One B. papyrifera sample included in the tally below was omitted from ground litter analyses due to lack of tissue.

Species

virginiana

ellipsoidalis

macrocarpa

Total Code ACNE ACRU BEPA JUVI PIBA PIRE PIST QUAL QUEL Functional group Broadleaf Broadleaf Broadleaf Needleleaf Needleleaf Needleleaf Needleleaf Broadleaf Broadleaf QUMA Broadleaf QURU TIAM Broadleaf Broadleaf # samples 27 28 36 n/a 20 31 26 37 33 34 33 17 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172

We measured reflectance spectra of each dried sample, both intact and ground, using a PSR+ 3500 fullused the Spectral Evolution leaf clip assembly. We used a built-in 99% Spectralon white reflectance standard to calibrate readings before each sample. Among broadleaf species, many of the leaves curled in on themselves during senescence and drying; when possible, we sought to measure spectra in relatively flat portions of the leaf to minimize specular reflection (Petibon et al. 2021). Among needleleaf species, we made mats by laying out needles in a single layer; because of their shape, there often remained small gaps or areas of overlap between needles. We discarded spectra with large discrepancies in sensor overlap regions, or (particularly for needleleaf species) with very high (>70%) or very low (<30%) peak reflectance, which often indicated that the needle mat was highly non-uniform. There remained at least three reflectance spectra for each sample, and more for most of the broadleaf samples that had multiple leaves.

After removing their petioles, we ground leaf samples by placing them in individual plastic vials with ball bearings and shaking them with a paint shaker for several hours. To measure ground-leaf spectra, we used the Spectral Evolution benchtop reflectance probe and glass-windowed sample trays. Before measuring each sample, we measured a white Spectralon panel placed in a sample tray for calibration. For sample measurements, we put at least 0.6 g of leaf powder into a sample tray. Preliminary tests showed that adding more material did not change the spectra, suggesting that transmittance was close to zero. The benchtop probe pressed the loose powder into an even pellet. We measured three spectra per sample4 turning the sample tray 120° between the first two measurements, and loosening and mixing the powder between the second and third measurements.

The spectrometer automatically interpolated the spectra to 1 nm resolution. We performed all subsequent processing using spectrolab v. 0.0.18 (Meireles et al. 2023) in R v. 3.6.1 (R Core Team, 2019). Because 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 the 350-400 nm and 2400-2500 nm regions could be noisy, we trimmed each spectrum to 400-2400 nm. For intact leaves only, we corrected discontinuities at the overlap region between the Si and first InGaAs detectors (970-1000 nm) using the match_sensors() function, and spline-smoothed only this region using the smooth_spline() function. Finally, we averaged all intact spectra and all ground spectra for a given sample. As a metric of measurement consistency, we calculated the mean spectral angle between measurements of the same sample (Kruse et al. 1993). Intact measurements (median across samples: 6.8°; 2.5th-97.5th percentile: 2.9°-20.4°) showed much less consistency than ground measurements (median: 0.6°; 2.5th-97.5th percentile: 0.2°-3.2°).

Researchers risk underestimating resorption efficiency when they fail to account for leaf mass loss during resorption. To avoid this risk, many studies either express nutrient content of non-senesced and senesced leaves on a per-area basis, or equivalently include a mass loss correction factor (van Heerwaarden et al. 2003). We measured LMA of each litter sample to enable conversion between a per-mass and per-area basis. We did not attempt to quantify or correct for the smaller bias sometimes caused by leaf shrinkage (van Heerwaarden et al. 2003).

For broadleaf species, we could not measure the area of whole leaves because many were curled and brittle. Instead, we used a hole punch to take four to six 0.3 cm2 disks from each dried leaf sample. We sought to capture the variation within the sample, including veins. We calculated LMA as the total mass divided by the total area of all disks. This method has generally been shown to provide accurate estimates of the LMA of whole leaves (Perez et al. 2020).

For needleleaf species, we measured the mass and area of five needles per sample. We estimated area as the length times the maximum width of the needle measured using digital calipers. We then measured the mass of each needle and calculated LMA as the total mass divided by the total area.

We measured leaf carbon and nitrogen concentration (Cmass and Nmass) on ground, oven-dried samples using dry combustion gas chromatography performed by an elemental analyzer (Costech ECS 4010 Analyzer, Valencia, California, USA). We converted Cmass and Nmass to an area basis (Carea and Narea) by multiplying them with LMA.

We measured carbon fractions on ground, oven-dried samples using an ANKOM 200 Fiber Analyzer (Ankom Technology, Macedon, New York, USA). The first digestion in a hot neutral detergent solution washes out plant solubles. The second digestion in a hot acidic detergent solution further washes out bound proteins and hemicellulose (henceforth 8hemicellulose9 for simplicity). The remaining fraction comprises cellulose and acid-unhydrolyzable residue (AUR) such as lignin (henceforth 8recalcitrant carbon9). We estimated each component using the changes in sample mass between steps. Because the paint shaker ground our samples into a very fine powder, we had to use small-pored ANKOM F58 fiber bags, which cannot be used in the final acid detergent lignin digestion to measure AUR concentration alone. We also did not determine the ash concentration of the samples.

We removed some trait values prior to statistical analyses because of clear issues during trait data collection4for example, the rupture of sample bags while measuring carbon fractions. Unlike some other studies (e.g., Gillon et al. 1999), we only removed samples due to known issues with trait measurements, not simply because they had large prediction residuals during statistical analyses.

We modeled the relationship between traits and spectra using partial least-squares regression (PLSR), which is well-suited to handle datasets that include many collinear predictors (Burnett et al. 2021). PLSR reduces the full matrix of spectra to a smaller number of orthogonal latent components that best explain 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 the variation in the variables to be predicted. We implemented our PLSR models in R package pls v. 2.7.1 (Mevik et al. 2019).

Previous studies using a PLSR modeling framework have often restricted the range of wavelengths used to predict certain traits in order to select bands that, based on prior research or known absorption features, may be most causally linked to the traits in question (e.g., Serbin et al. 2014). In our main set of analyses, we used the full spectrum (400-2400 nm) to predict each trait on the grounds that biochemical changes during senescence may alter the relationship between integrative functional traits like Nmass and specific optical features. However, we also explored two ways of restricting the spectral range for intact-leaf spectra only. First, we built models using only 400-1000 nm (8VIS/NIR models9) given that many lessexpensive spectrometers only measure this range. Second, we built models using only 1300-2400 nm (8SWIR models9). The pigment composition of leaves changes rapidly during senescence, including breakdown products (8brown pigments9) which may absorb up to 1300 nm (Fourty et al. 1996). The SWIR models could perform better in cases where pigments are poor indicators of the overall composition of the leaf.

Researchers may transform reflectance of ground-leaf spectra to pseudoabsorbance (-log R; Blackburn 1998; Serbin et al. 2014) or its derivatives, which may somewhat linearize the relationship between traits and spectral features and make those features more prominent. We found in preliminary tests that these transformations did not substantially improve model performance. Likewise, brightness normalization of intact-leaf spectra (Feilhauer et al. 2010) did not have a strong influence on model performance.

We divided our samples into calibration (75%) and validation (25%) datasets, stratified such that the proportions of each species were identical in each dataset. We fit models for each trait on the complete calibration dataset, using 10-fold cross-validation to select the optimal number of model components to use in further analyses while avoiding overfitting. We selected the smallest number of components whose 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 cross-validation root mean squared error of prediction (RMSEP) was within one standard deviation of the very lowest RMSEP at any number of components. We used these models to calculate the variable importance in projection metric (VIP) for each trait (Wold et al. 1994) but we do not report their accuracy.

We developed our main set of models using a resampling procedure in which we divided the calibration data at random 200 times into 70% training and 30% testing subsets (Burnett et al. 2021). During each resampling iteration, we trained a model for each trait on the 70% and assessed its performance (RMSE and R2) on the remaining 30%, using the previously determined number of components for that trait. This procedure left us with an ensemble of 200 models for each trait, which could be used both to make trait predictions and to examine the variability of model performance under random variation in training data.

We used these ensembles of models to predict traits from the validation (25%) dataset. We quantified model performance for each trait by calculating R2 and root mean squared error (RMSE) between the predictions (averaged for each validation sample across the 200 estimates) and the measured values. We also report the %RMSE, calculated as the RMSE divided by the 2.5% trimmed range of the measured values in the validation data (Kothari et al. 2023a). set.

Litter trait complete trait range across the full dataset. %RMSE is calculated as RMSE divided by the 2.5% trimmed range of data within the validation data RMSE %RMSE

RMSE %RMSE 273 274 275 276

Fig. 1: Distributions of spectral reflectance and its coefficient of variation (CV) among intact (top) and ground (bottom) leaf litter samples. The black line represents the median reflectance at each wavelength, flanked by the 25-75 percentile region (dark blue), and the 2.5-97.5 percentile region (light blue). The solid red line shows the coefficient of variation among spectra at each wavelength. 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302

Results There was a high level of variation in each trait both across the entire dataset and within species. For example, Nmass varied nearly 11-fold; in contrast, Cmass only varied about 1.5-fold (Table 2). In spite of the high level of intraspecific variation, especially among broadleaf species, species identity alone could account for between 46.0% (Narea) and 93.7% (LMA) of the variation in each trait. In general, needleleaf species had much higher LMA, Cmass, recalcitrant carbon, and lower Nmass than broadleaf species (Table S1), consistent with trends in non-senescent leaves (Serbin et al. 2014; Kothari et al. 2023a).

Consequently, we see correlations among pairs of leaf economic traits across the whole dataset (for log(LMA) and log(Nmass), R2 = 0.592; p < 10-15), but a weaker correlation among just needleleaf samples (R2 = 0.255; p < 10-5) and none among just broadleaf samples (R2 = 0.011; p = 0.111).

Among both ground-leaf and intact-leaf spectra, the coefficient of variation (CV) was highest in the visible range and decreased into the near-infrared range (NIR; Fig. 1). For intact-leaf spectra only, the CV once again increased into the short-wave infrared range (SWIR). There was a global maximum CV around 440-490 nm for intact spectra, shifted to lower wavelengths for ground spectra; there were also local maxima centered at around 670-675 nm for both kinds of spectra. Both maxima lie near absorption peaks of chlorophylls, and reflectance at 670 nm is particularly sensitive to variation in chlorophyll at low levels, such as during late senescence (Gitelson & Merzlyak 1996).

Each of our traits could be predicted with moderate-to-high success from full intact-leaf and ground-leaf spectra (Figs. 2-4; Table 2). The models for Nmass and (especially for intact-leaf spectra) LMA achieved the highest calibration and validation accuracy, followed by recalcitrant carbon, solubles, Cmass, and finally hemicellulose (Table 2). There was no general tendency for estimation of chemical traits to be better using intact- or ground-leaf spectra, but estimation of LMA was better using intact-leaf spectra.

Because we used LMA to convert from a mass to an area basis, we built models predicting Narea and Carea (nitrogen and carbon on a per-area basis) only from intact-leaf spectra (Fig. 4). The Carea model (%RMSE 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 = 8.02) had better performance than the Cmass model (%RMSE = 14.0), but the Narea model (%RMSE = 13.2) had worse performance than the Nmass model (%RMSE = 8.58; Table 2). Jackknife analyses showed that traits for which the ensemble model performance was poorest4such as hemicellulose, Cmass, and Narea4also had the greatest variability in model performance (Figs. S1 and S2).

For every trait prediction model trained on the full spectrum, the VIP metric also showed a global maximum around 677 nm (intact) or 672 nm (ground; Fig. 5), next to one of the aforementioned chlorophyll absorption peaks. This pattern suggests that the amount of chlorophyll remaining at abscission may be closely related to many traits, especially Nmass. More generally, the visible range was most important for predicting all traits, and the NIR range was also important for LMA. There were further peaks in the SWIR around 1430-1450 nm (for intact leaves), 1910-1950 nm and 2130-2220 nm. In fresh tissue, variation in reflectance around 1910-1950 nm is often interpreted as a measure of absorption by water, but the trace amounts of water remaining in our dried samples likely have no direct biological relevance. In general, reflectance in the visible range and parts of the SWIR range seemed most informative in predicting nutrients and carbon fractions in litter.

Our VIS/NIR models for intact leaves had considerably worse performance than our full-range models for every trait, particularly Nmass, Narea, and solubles. By contrast, our SWIR models had very similar performance to the full-range models (Table S2). Both sets of restricted-range models had a similar rankordering of model performance across traits to each other and the full models. Moreover, the VIP metric for these models revealed that the regions that were most important for these predictions were mainly the same bands in their respective regions that were important for the full-range models; however, for the

VIS/NIR models, very low wavelengths close to 400 nm also had high importance. 328 329 330 331 332

Fig. 2: Validation results for predictions of solubles, hemicellulose, and recalcitrant carbon from intactleaf litter spectra (left) and ground-leaf litter spectra (right). In each panel, a separate OLS regression line is shown for each of the 11 species, overlaid on top of the thick dashed 1:1 line. The error bars for each data point are 95% intervals calculated from the distribution of predictions based on the model coefficients from the 200 jackknife iterations. See Table 1 for species codes. 334 335 336

Fig. 3: Validation results for predictions of Cmass, Nmass, and LMA from intact-leaf litter spectra (left) and ground-leaf litter spectra (right). Regression lines and error bars are as in Fig. 2. See Table 1 for species codes. 339 340 341

Fig. 4: Validation results for predictions of Carea and Narea from intact-leaf litter spectra. Regression lines and error bars are as in Fig. 2. See Table 1 for species codes. 343 344 345

Fig. 5: The variable importance of prediction (VIP) metric in intact (top) and ground (bottom) models.

The dashed black line represents a threshold of 0.8 suggested as a rough heuristic of importance by

Burnett et al. (2021). 424 425 fraction analysis takes about 25 minutes of operator time per sample. The time saved through spectroscopic trait estimation grows as the number of traits of interest increases.

Many studies that leverage spectroscopy to address ecological questions use models tailored to the particular species and tissues under study (e.g., McTiernan et al. 2003; Hobbie 2005; Fortunel et al. 2009). While this practice allows researchers more confidence in their trait estimates, it requires them to spend time and money doing the manual trait measurements needed for calibration, which they must repeat for any new study system they adopt. Spectroscopic trait estimation is unlikely to serve as a complete substitute for standard measurements unless we can create models that are demonstrably accurate across a wide range of species and conditions.

The good validation performance of our models does not guarantee accurate trait estimates from any set of litter samples. To take one example, consider the aforementioned pattern that the ground-leaf model for LMA often performs poorly within species, despite its strong performance across species. If it were used to study intraspecific variation in litter traits, the model9s estimates could be useless or even misleading .

Moreover, when models simply 8learn9 to exploit any features that happen to correspond with LMA among species within a given dataset4without respect to their causal relationship with LMA4they may also transfer poorly to new species (Kothari & Schweiger 2022). This case is closely linked to the more general problem in machine learning of distinguishing spurious and invariant correlations (Arjovsky et al. 2019) . As the volume of available spectral data increases, taking care to assess out-of-sample generalization and borrowing tools from machine learning could help us train better spectral models and make more realistic judgments about their performance (Kattenborn et al. 2022; Cherif et al. 2023).

Recent studies have attempted to tackle the challenge of building global trait-spectra models. For example, Serbin et al. (2019) show that a single model can accurately predict the LMA of fresh, nonsenesced leaves across 11 sites from the tropics to the Arctic. It remains to be seen whether any single 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 model could predict traits like Nmass or LMA from leaf tissue across many species at all stages of senescence and decomposition. The predictive success of global models may rely on the constrained patterns of trait covariation among non-senesced leaves4for example, along continua from sun to shade leaves or conservative to acquisitive leaves (Osnas et al. 2018)4which may help in estimating traits that are not themselves associated with strong absorption features (Nunes et al. 2017; Kothari & Schweiger 2022). However, the dramatic shifts in leaf chemistry and cellular structure during senescence (Keskitalo et al. 2005)4including pigment degradation and nutrient resorption4could weaken some of these relationships among traits, or between traits and their optical features. If so, it could be hard to build a global model for senesced leaves that achieves the predictive success of global models for non-senesced leaves.

Some evidence suggests that litter traits do in fact predictably covary with each other, and with nonsenesced leaf traits (Freschet et al. 2010; Freschet et al. 2012; Jackson et al. 2013) . But while most fresh green leaves look alike, each species9 leaves may senesce in its own way. For example, some of our species (A. rubrum, Q. alba, Q. ellipsoidalis, Q. rubra) tend to produce reddish anthocyanins during senescence, but the rest do not. The breakdown of certain metabolites during senescence also creates a complex mixture of brown pigments (Fourty et al. 1996), which are poorly described and may vary from species to species. Such variation could make the relationship between traits and optical features more contingent for senescent foliage.

Even setting aside these concerns, model generality may also be complicated by differences in sample preparation (e.g., grind size) and instruments, which may affect the spectrum in subtle but important ways (Foley et al. 1998; Petibon et al. 2021). Overcoming these challenges may require investing in protocol development, or in trying algorithms that are more flexible than PLSR in 8learning9 the complex forms of biological variation in heterogeneous datasets. In the meantime, it would be prudent for researchers who 475 476 plan to use existing spectral models in new systems to take some conventional trait measurements in order to validate model performance.

Finally, we note that plants produce and shed many kinds of tissue besides leaves, and these other litter sources also contribute to nutrient cycling through resorption and decomposition (Lü et al. 2012).

Reflectance spectroscopy may help measure nutrient-related traits in other tissues. For example, fine root decomposition is an important but little-understood part of nutrient cycling; fine roots may contribute nearly half of annual litter inputs in forests (Freschet et al. 2013) and vary considerably in resorption efficiency (Freschet et al. 2010). Elle et al. (2019) showed that PLSR models built using near-infrared reflectance spectroscopy can predict fine root lignin, which increases recalcitrance (See et al. 2019; but see, in contrast, Sun et al. 2018). Continuing to develop models for other plant tissues could make it easier to study how plants alter nutrient cycling in a holistic way. Likewise, nutrients other than nitrogen often limit or co-limit growth and decomposition, and tools to estimate them quickly and cheaply could advance the study of plant nutrient economies. However, estimates of some micronutrients from reflectance spectra may be unreliable (Nunes et al. 2017; Kothari et al. 2023a), and other methods may prove more useful (e.g., X-ray fluorescence spectroscopy; van der Ent et al. 2018).

Ultimately, whether accurate global models for predicting litter traits from spectra can be built is an empirical question that can only be settled by amassing a wide variety of data. We take an initial step towards this long-term goal by showing that reflectance spectroscopy can provide accurate litter trait estimates across many temperate tree species that vary widely in their litter traits. This line of research shows great promise in relieving some of the major limitations toward understanding plant functional ecology in an ecosystem context. 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519

Cotrozzi, L., Couture, J. J., Cavender-Bares, J., Kingdon, C. C., Fallon, B., Pilz, G., Pellegrini, E., Nali, C., & Townsend, P. A. (2017). Using foliar spectral properties to assess the effects of drought on plant water potential. Tree Physiology, 37(11), 158231591.

Coûteaux, M.-M., Sarmiento, L., Hervé, D., & Acevedo, D. (2005). Determination of water-soluble and total extractable polyphenolics in biomass, necromass and decomposing plant material using near-infrared reflectance spectroscopy (NIRS). Soil Biology and Biochemistry, 37(4), 7953799.

Couture, J. J., Singh, A., Rubert‐Nason, K. F., Serbin, S. P., Lindroth, R. L., & Townsend, P. A. (2016). Spectroscopic determination of ecologically relevant plant secondary metabolites. Methods in Ecology and Evolution, 7(11), 140231412.

Dechant, B., Cuntz, M., Vohland, M., Schulz, E., & Doktor, D. (2017). Estimation of photosynthesis traits from leaf reflectance spectra: Correlation to nitrogen content as the dominant mechanism. Remote

Diehl, P., Mazzarino, M. J., Funes, F., Fontenla, S., Gobbi, M., & Ferrari, J. (2003). Nutrient conservation strategies in native Andean-Patagonian forests. Journal of Vegetation Science, 14(1), 63370. El Zein, R., Bréda, N., Gérant, D., Zeller, B., & Maillard, P. (2011). Nitrogen sources for current-year shoot growth in 50-year-old sessile oak trees: An in situ 15N labeling approach. Tree Physiology, 31(12), near-infrared spectroscopy. Scientific Reports, 9(1), 6396.

Feilhauer, H., Asner, G. P., Martin, R. E., & Schmidtlein, S. (2010). Brightness-normalized Partial Least Squares Regression for hyperspectral data. Journal of Quantitative Spectroscopy and Radiative Transfer, 111(12), 194731957.

Foley, W. J., McIlwee, A., Lawler, I., Aragones, L., Woolnough, A. P., & Berding, N. (1998). Ecological applications of near infrared reflectance spectroscopy 3 a tool for rapid, cost-effective prediction of the composition of plant and animal tissues and aspects of animal performance. Oecologia, 116(3), 2933305. Fortunel, C., Garnier, E., Joffre, R., Kazakou, E., Quested, H., Grigulis, K., Lavorel, S., Ansquer, P., Castro, H., Cruz, P., Dole}al, J., Eriksson, O., Freitas, H., Golodets, C., Jouany, C., Kigel, J., Kleyer, M., Lehsten, V., Lepa, J., … Zarovali, M. (2009). Leaf traits capture the effects of land use changes and climate on litter decomposability of grasslands across Europe. Ecology, 90(3), 5983611. Fourty, Th., Baret, F., Jacquemoud, S., Schmuck, G., & Verdebout, J. (1996). Leaf optical properties with explicit description of its biochemical composition: Direct and inverse problems. Remote Sensing of Environment, 56(2), 1043117.

Freschet, G. T., Aerts, R., & Cornelissen, J. H. C. (2012). A plant economics spectrum of litter decomposability. Functional Ecology, 26(1), 56365. 621 622 623 resorption from leaves, stems and roots in a subarctic flora: What is the link with other resource economics traits? New Phytologist, 186(4), 8793889.

Freschet, G. T., Cornwell, W. K., Wardle, D. A., Elumeeva, T. G., Liu, W., Jackson, B. G., Onipchenko, V. G., Soudzilovskaia, N. A., Tao, J., & Cornelissen, J. H. C. (2013). Linking litter decomposition of above- and below-ground organs to plant3soil feedbacks worldwide. Journal of Ecology, 101(4), 9433

Gill, A. L., Schilling, J., & Hobbie, S. E. (2021). Experimental nitrogen fertilisation globally accelerates, then slows decomposition of leaf litter. Ecology Letters, 24(4), 8023811.

Gillon, D., Joffre, R., & Ibrahima, A. (1999). Can Litter Decomposability Be Predicted by Near Infrared Reflectance Spectroscopy? Ecology, 80(1), 1753186.

Gitelson, A. A., & Merzlyak, M. N. (1996). Signature Analysis of Leaf Reflectance Spectra: Algorithm Development for Remote Sensing of Chlorophyll. Journal of Plant Physiology, 148(3), 4943500. Grossman, J. J., Cavender‐ Bares, J., & Hobbie, S. E. (2020). Functional diversity of leaf litter mixtures slows decomposition of labile but not recalcitrant carbon over two years. Ecological Monographs, 90(3), e01407.

Grossman, J. J., Cavender-Bares, J., Hobbie, S. E., Reich, P. B., & Montgomery, R. A. (2017). Species richness and traits predict overyielding in stem growth in an early-successional tree diversity experiment.

Ecology, 98(10), 260132614. 647 648 649 resorption efficiency in plants of contrasting nutrient-acquisition strategies along a 2-million-year dune chronosequence. Journal of Ecology, 102(2), 3963410.

Heerwaarden, L. M. V., Toet, S., & Aerts, R. (2003). Current measures of nutrient resorption efficiency lead to a substantial underestimation of real resorption efficiency: Facts and solutions. Oikos, 101(3),

Hobbie, S. E. (2005). Contrasting Effects of Substrate and Fertilizer Nitrogen on the Early Stages of Litter Decomposition. Ecosystems, 8(6), 6443656.

Hobbie, S. E. (2015). Plant species effects on nutrient cycling: Revisiting litter feedbacks. Trends in Ecology & Evolution, 30(6), 3573363.

Jackson, B. G., Peltzer, D. A., & Wardle, D. A. (2013). Are functional traits and litter decomposability coordinated across leaves, twigs and wood? A test using temperate rainforest tree species. Oikos, 122(8), 113131142. 488.

Joffre, R., Gillon, D., Dardenne, P., Agneessens, R., & Biston, R. (1992). The use of near-infrared reflectance spectroscopy in litter decomposition studies. Annales des Sciences Forestières, 49(5), 4813 Kattenborn, T., Schiefer, F., Frey, J., Feilhauer, H., Mahecha, M. D., & Dormann, C. F. (2022). Spatially autocorrelated training and validation samples inflate performance assessment of convolutional neural networks. ISPRS Open Journal of Photogrammetry and Remote Sensing, 5, 100018. 673 674 675 decomposition controlled by manganese redox cycling. Proceedings of the National Academy of Sciences, 112(38), E52533E5260.

Keskitalo, J., Bergquist, G., Gardeström, P., & Jansson, S. (2005). A Cellular Timetable of Autumn Senescence. Plant Physiology, 139(4), 163531648.

Kobe, R. K., Lepczyk, C. A., & Iyer, M. (2005). Resorption Efficiency Decreases with Increasing Green Leaf Nutrients in a Global Data Set. Ecology, 86(10), 278032792.

Kothari, S., Beauchamp-Rioux, R., Blanchard, F., Crofts, A. L., Girard, A., Guilbeault-Mayers, X., Hacker, P. W., Pardo, J., Schweiger, A. K., Demers-Thibeault, S., Bruneau, A., Coops, N. C., Kalacska, M., Vellend, M., & Laliberté, E. (2023a). Predicting leaf traits across functional groups using reflectance spectroscopy. New Phytologist, 238(2), 5493566.

Kothari, S., Beauchamp-Rioux, R., Laliberté, E., & Cavender-Bares, J. (2023b). Reflectance spectroscopy allows rapid, accurate and non-destructive estimates of functional traits from pressed leaves. Methods in

Kothari, S., & Schweiger, A. K. (2022). Plant spectra as integrative measures of plant phenotypes. Journal of Ecology, 110(11), 253632554.

Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., & Goetz, A. F. H. (1993). The spectral image processing system (SIPS)4Interactive visualization and analysis of imaging spectrometer data. Remote Sensing of Environment, 44(2), 1453163. 699 700 701 Synergistic effects: A common theme in mixed-species litter decomposition. New Phytologist, 227(3), 7573765. resorption proficiency potentially reinforces plant3soil feedbacks and microscale heterogeneity in a semiarid grassland. Journal of Ecology, 100(1), 1443150.

Madritch, M., Cavender-Bares, J., Hobbie, S. E., & Townsend, P. (2020). Linking foliar traits to belowground processes. In J. Cavender-Bares, J. Gamon, & P. Townsend (Eds.), Remote Sensing of Plant Biodiversity (pp. 1733198). Springer, New York.

McTiernan, K. B., Coûteaux, M.-M., Berg, B., Berg, M. P., Calvo de Anta, R., Gallardo, A., Kratz, W., Piussi, P., Remacle, J., & Virzo De Santo, A. (2003). Changes in chemical composition of Pinus sylvestris needle litter during decomposition along a European coniferous forest climatic transect. Soil

Meireles, J. E., Schweiger, A. K., & Cavender-Bares, J. (2023). spectrolab: Class and Methods for Spectral Data in R. R package v.0.0.18. [WWW document] URL https:// CRAN.RMevik, B.-H., Wehrens, R., & Liland, K. H. (2019). PLS: partial least squares and principal component regression. R package v.2.7-1. [WWW document] URL https:// CRAN.R-project.org/package=pls Norris, K. H., Barnes, R. F., Moore, J. E., & Shenk, J. S. (1976). Predicting Forage Quality by Infrared measure the impact of soil type on leaf traits. Biogeosciences, 14(13), 337133385.

Osnas, J. L. D., Katabuchi, M., Kitajima, K., Wright, S. J., Reich, P. B., Bael, S. A. V., Kraft, N. J. B., Samaniego, M. J., Pacala, S. W., & Lichstein, J. W. (2018). Divergent drivers of leaf trait variation within species, among species, and among functional groups. Proceedings of the National Academy of Sciences, 115(21), 548035485.

Perez, T. M., Rodriguez, J., & Heberling, J. M. (2020). Herbarium-based measurements reliably estimate three functional traits. American Journal of Botany, 107(10), 145731464.

Petibon, F., Czyż, E. A., Ghielmetti, G., Hueni, A., Kneubühler, M., Schaepman, M. E., & Schuman, M. C. (2021). Uncertainties in measurements of leaf optical properties are small compared to the biological variation within and between individuals of European beech. Remote Sensing of Environment, 264, 112601.

Petit Bon, M., Böhner, H., Kaino, S., Moe, T., & Bråthen, K. A. (2020). One leaf for all: Chemical traits of single leaves measured at the leaf surface using near-infrared reflectance spectroscopy. Methods in Ecology and Evolution, 11(9), 106131071.

Prescott, C. E. (2010). Litter decomposition: What controls it and how can we alter it to sequester more carbon in forest soils? Biogeochemistry, 101(1), 1333149. 749 750 751

R Core Team. 2020. R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. [WWW document] URL https://www.R-project.org/ [accessed 9 March 2020].

Rea, A. M., Mason, C. M., & Donovan, L. A. (2018). Evolution of nutrient resorption across the herbaceous genus Helianthus. Plant Ecology, 219(8), 8873899.

Richardson, A. D., & Reeves III, J. B. (2005). Quantitative reflectance spectroscopy as an alternative to traditional wet lab analysis of foliar chemistry: Near-infrared and mid-infrared calibrations compared.

Schreeg, L. A., Mack, M. C., & Turner, B. L. (2013). Nutrient-specific solubility patterns of leaf litter across 41 lowland tropical woody species. Ecology, 94(1), 943105.

Schweitzer, J. A., Bailey, J. K., Rehill, B. J., Martinsen, G. D., Hart, S. C., Lindroth, R. L., Keim, P., & Whitham, T. G. (2004). Genetically based trait in a dominant tree affects ecosystem processes. Ecology Letters, 7(2), 1273134.

See, C. R., McCormack, M. L., Hobbie, S. E., Flores‐Moreno, H., Silver, W. L., & Kennedy, P. G. (2019). Global patterns in fine root decomposition: Climate, chemistry, mycorrhizal association and woodiness. Ecology Letters, 22(6), 9463953.

Serbin, S. P., Singh, A., McNeil, B. E., Kingdon, C. C., & Townsend, P. A. (2014). Spectroscopic determination of leaf morphological and biochemical traits for northern temperate and boreal tree species.

Ecological Applications, 24(7), 165131669. 775 776 777 Wang, Z., & Rogers, A. (2019). From the Arctic to the tropics: Multibiome prediction of leaf mass per area using leaf reflectance. New Phytologist, 224(4), 155731568. Contrasting dynamics and trait controls in first-order root compared with leaf litter decomposition.

Talbot, J. M., & Treseder, K. K. (2012). Interactions among lignin, cellulose, and nitrogen drive litter chemistry3decay relationships. Ecology, 93(2), 3453354. van der Ent, A., Przybyłowicz, W. J., de Jonge, M. D., Harris, H. H., Ryan, C. G ., Tylko, G., Paterson, D. J., Barnabas, A. D., Kopittke, P. M., & Mesjasz-Przybyłowicz, J. (2018). X -ray elemental mapping techniques for elucidating the ecophysiology of hyperaccumulator plants. New Phytologist, 218(2), 4323 452.

Vergutz, L., Manzoni, S., Porporato, A., Novais, R. F., & Jackson, R. B. (2012). Global resorption efficiencies and concentrations of carbon and nutrients in leaves of terrestrial plants. Ecological Monographs, 82(2), 2053220.

Wold, S. (1994). PLS for multivariate linear modeling. In H. van de Waterbeemd, editor. Chemometric methods in molecular design, methods and principles in medicinal chemistry, Verlag-Chemie (Weinheim, Germany), 1953218.

Wright, I. J., & Westoby, M. (2003). Nutrient concentration, resorption and lifespan: Leaf traits of Australian sclerophyll species. Functional Ecology, 17(1), 10319.

Ecology, 96(2), 3733380. eight leaf litter traits under consideration. See Table 1 in main text for species codes.

RMSE 11.3 16.6 comps

R2 0.939

Table S2: Summary statistics of PLSR model validation based on intact- and ground-leaf spectra restricted to the VIS/NIR range (400-1000 nm) or to part of the SWIR range (1300-2400 nm). %RMSE is calculated as RMSE divided by the 2.5% trimmed range of data within the validation data set. Trait LMA (g m-1) Recalcitrant carbon (%)