henhaobio
Sign in

bioRxivEcologyPreprint

Reflectance spectra capture temporal variation in functional traits and their integration across leaf phenology

Cornelius Onyedikachi NichodemusiD, José Eduardo MeirelesiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Plant function depends not on single traits but on how traits covary, yet trait integration is typically estimated from a peak-season snapshot and treated as fixed. Currently, we do not understand if integration has a phenology and how it affects spectral trait models that depend on trait covariance. We measured traits and reflectance spectra weekly across a growing season in seven temperate species (7,515 spectra), compared trait and spectral correlation structure across three phenophases, and tested PLSR models trained on the full season or peak season alone, alongside a widely used published model. Traits followed species-specific trajectories, with spectral lability concentrated in the visible region. Correlation structure shifted with phenology: LMA--EWT, C--N, and N--QY held across phenophases, whereas LMA--N and VIS--NIR spectral correlations reversed sign. Full-season models tracked trait dynamics, whereas peak-season and published models failed outside their window, producing biologically impossible estimates. Traits and their integration have themselves a phenology, and a peak-season correlation matrix does not fully capture trait integration. Spectra recover trait dynamics only with phenologically representative training data, since models calibrated in one phenophase inherit its correlation structure -- a shift that coincides with, and may underlie, their failure outside it.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (10 Sept 2026) · CC BY 4.0

Weekly leaf spectra and traits in seven temperate species reveal phenophase-dependent trait correlations; peak-season PLSR models fail at senescence while full-season models track trait dynamics.

Trait integration and spectral trait models are usually calibrated on a single peak-season snapshot, implicitly assuming trait covariance is fixed even though leaf traits change at different rates through development and senescence. This work resolves correlation shifts within one leaf cohort across a season and tests whether partial least squares regression (PLSR) trait models transfer across phenophases, extending spatial transferability concerns to time.

Key results

  1. ShownPLSR models trained across the full season reproduced measured trait dynamics. The peak-season model and the published Kothari 2023 coefficients predicted poorly, and their regression slopes departed from the 1:1 line.

    All-season and week-as-covariate: LMA and EWT R^2 = 0.85-0.93 (%RMSE 4.48-4.77), N and QY R^2 = 0.6-0.66, C R^2 = 0.08-0.26. Peak-season and Kothari 2023: LMA R^2 = 0.03-0.25 (%RMSE 39.05-89.83), N R^2 = 0.11-0.14, C R^2 = 0.00, QY R^2 = 0.23. Only the full-season models had slope 95% CIs that overlapped the 1:1 line. · Assessing model performance across leaf…; Model performance across phenology

  2. ShownTrait and spectral correlation structure changed between young, mature and senescent leaves. The LMA-N trade-off and visible/near-infrared band correlations changed sign, while LMA-EWT, C-N and N-QY stayed positive and NIR-SWIR stayed negative.

    Contrasts at weeks 22, 32 and 44, tested at p < 0.05: LMA-N was negative in young and mature leaves and weakly positive in senescent leaves. LMA-QY and EWT-QY differed significantly between phenophases. LMA-EWT, C-N and N-QY stayed positive. VIS-NIR went from near zero to positive to negative and NIR-SWIR stayed negative; VIS-NIR and VIS-SWIR differences were significant. · Spectral and trait correlation; Spectral and trait correlation; The phenology of correlation structure

  3. ShownOutside their calibration window, models trained on one phenophase (including the published Kothari 2023 coefficients) gave impossible late-season values, such as negative concentrations and soluble fractions above 100%. They also implied a spurious rise in N during senescence.

    Share of Kothari 2023 estimates within bounds (concentrations ≥ 0, percentages < 100%): 98-100% at peak season vs 33-97% in late season. The late-season invalid share ranged from 3% for K to 67% for hemicellulose, and ⅓ of estimates were unreasonable for 6 of 9 traits. · Assessing model performance over time; Assessing model performance across leaf…; Model performance across phenology

  4. ShownSeasonal spectral change was concentrated in the pigment-driven visible region, with less change in near- and shortwave-infrared reflectance. Trait trajectories were species-specific; both evergreen Rhododendron species lost LMA from early to peak season and then regained it, unlike most other species.

    CV across 24 weekly mean spectra was highest at 400-700 nm and lower at 700-2400 nm. ANOVA: LMA, C and QY changed significantly in all species, N in all but R. maximum, and EWT in all but R. maximum and Q. rubra. In R. catawbiense and R. maximum, LMA fell sharply from early to peak season, then rose gradually. · Leaf spectra and trait measurement; Spectral and trait phenology; Spectral and trait phenology

  5. ShownAdding collection week as a covariate did not improve on spectra alone, implying the spectra carry the phenological signal. Only the full-season models matched measured LMA and EWT in every phenophase.

    Week-as-covariate predictions and metrics were statistically indistinguishable from the all-season model. ANOVA/Tukey tests found no significant predicted-vs-measured differences for LMA and EWT in any species or phenophase for full-season models. Peak-season and Kothari 2023 predictions did differ significantly, mainly in early and late season. · Assessing significant differences in traits; Assessing model performance across leaf…; Model performance across phenology

Methods, data and reproducibility

Leaves of seven species were sampled weekly for 24 weeks, pairing full-range reflectance spectra with LMA (leaf mass per area), EWT (equivalent water thickness) and photosystem II quantum yield, plus C and N at 5 weeks. Correlations were compared at three phenophase weeks, and PLSR models (all-season, week-as-covariate, peak-season) were compared with published Kothari 2023 coefficients.

  • Leaf-level reflectance spectra (350-2500 nm instrument, trimmed to 400-2400 nm), measured weekly May 31 to November 6, 2023 on the University of Maine Orono campus · Acer, Betula, Prunus, Quercus and Rhododendron spp. (7 species) · n = 7,515 spectra · 10.6084/m9.figshare.32060352
  • Weekly leaf traits: LMA and EWT from fresh and oven-dry mass and scanned leaf area, plus photosystem II quantum yield · Acer, Betula, Prunus, Quercus and Rhododendron spp. (7 species) · n = 27 leaves sampled per species per week, 24 weeks · 10.6084/m9.figshare.32060352
  • Leaf percent C and N by elemental analysis on 3 leaves per species at weeks 22, 26, 32, 39 and 44 · Acer, Betula, Prunus, Quercus and Rhododendron spp. (7 species) · n = 102 (C) and 104 (N) data points · 10.6084/m9.figshare.32060352
  • TRY database values for LMA, C and N, used to check measured trait ranges against global values
  • Canadian Airborne Biodiversity Observatory (CABO) trait data, used for a visual range comparison that included EWT
Methods & tools
SVC HR-1024i full-range field spectroradiometer with leaf clip, spectrolab R package (spectral processing), pls R package (PLSR, oscorespls fitting, onsigma component selection), caret R package (createDataPartition), magick R package (leaf area from scans), ggplot2, Fluorpen 1.1.0 (photosystem II quantum yield), Kothari 2023 published PLSR trait coefficients, Variable influence on projection (VIP), ANOVA with Tukey HSD post hoc tests, C/N elemental analysis at Central Appalachians Stable Isotope Facility (CASIF), TRY and CABO trait databases
Software & environment
R with spectrolab, pls, caret, magick and ggplot2 (package versions not stated); Fluorpen 1.1.0 for quantum yield.
Reproducibility gaps
The dataset is given only as a Figshare DOI, with no URL in the text. R package versions and session details are not reported.
Limitations (6)
  • Sampling covers one site, one growing season and seven species. Whether the trait trajectories, correlation shifts and peak-versus-late spectral divergence hold in other biomes, growth forms or climates is unknown. · stated by the authors
  • Carbon and nitrogen were measured at only 5 sampling weeks (102 and 104 data points) versus 24 weeks for LMA, EWT and QY. This weakens the C and N models and the precision of their correlations within each phenophase. · stated by the authors
  • Correlation structure was estimated at three representative weeks (22, 32, 44), so its phenology is a contrast among phenophases rather than a continuous trajectory. · stated by the authors
  • The design does not test whether correlation shifts cause late-season model failure. Extrapolation beyond the trained trait range and higher measurement uncertainty in senescing tissue remain untested alternatives. · stated by the authors
  • Peak-season models used 10 of 24 weeks of data, so their poorer accuracy mixes the effect of phenological coverage with smaller training sets and narrower trait ranges. No size-matched full-season control was fitted. · noted in this summary
  • Validation used a random 75%/25% split balanced by branch within one site and year, so the same individuals and branches feed both calibration and validation. Accuracy on new individuals, sites or years is not tested. · noted in this summary

Read this if you build or apply leaf spectral PLSR trait models outside peak season, or study how trait covariation and the leaf economic spectrum change through leaf development and senescence.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKothari '22Kothari '23Kothari '21Casaverde '23Cui '23Shiklomanov '18
70% semantic matchEcology
S. Kothari, R. Beauchamp-Rioux, F. Blanchard+12 more·bioRxiv·Jul 1, 2022