Robust estimation of key leaf traits from reflectance spectroscopy of herbarium specimens
Aaron K. Lee, Lauren Vander Esch, Jin Oong, Brett Fredericksen, Juan Ramirez‐Lerma, J. Antonio Guzmán, Dawson White, Tim Whitfeld, Jose Eduardo Meireles, Jeannine Cavender‐Bares, Ya YangAbstract
Community‐wide efforts to digitize herbarium specimens have facilitated novel uses of specimen data across scales. However, the need for destructive sampling has prevented large‐scale examination of foliar functional traits. We demonstrate that nondestructive reflectance spectra are an effective tool for estimating leaf traits from herbarium collections.
Using partial least squares regression (PLSR), we developed and evaluated trait estimation models for leaf mass per area (LMA), nitrogen content (N) and carbon content (C) from nondestructive leaf‐level spectra and destructive trait measurements from herbarium specimens. Our modelling dataset consisted of temperate species collected in the Minneapolis‐St. Paul area (Minnesota, USA) between 1876 and 2022. We also trained biased models excluding sample subsets corresponding to specimen age, habit, growth form and species categories to assess the transferability of PLSR trait estimation models. Finally, we evaluated whether leaf traits estimated from spectra recover the temporal trends observed in destructively measured traits.
Our PLSR models accurately estimate LMA and N, and reasonably estimate C. They are robust to sample greenness, specimen age and species, but sensitive to mismatches in functional group categories between the training and target data. Furthermore, spectrally estimated LMA and N reproduce temporal trends observed in measured traits over the 146‐year time series; estimated C reproduces most temporal trends despite poorer model performance than the other traits.
Given the limitations of broad destructive sampling from herbarium collections, we show that PLSR models trained on younger specimens can reliably estimate leaf traits from older specimens. While we caution that trait estimation models are only as good as their training data, we are optimistic that spectroscopy can capture the extensive functional data preserved in herbarium collections in a largely nondestructive manner.