Foliar functional traits refer to a range of biochemical and physiognomic characteristics of vegetation that control photosynthesis, nutrient and water cycling, and can be used to describe the functional diversity of ecosystems. In this study, we utilize NASA AVIRIS - Next Generation imaging spectroscopy data to map 15 foliar functional traits in a grassland experiment at the Cedar Creek Ecosystem Science Reserve, a Long-Term Ecological Research site in Central Minnesota, across three years. To estimate traits, we compared the widely used partial least squares regression (PLSR) with Gaussian processes regression (GPR) to assess differences in model performance and uncertainty estimates. PLSR is attractive for its straightforward implementation and interpretation, but requires bootstrapping-based methods to estimate and map prediction uncertainties. On the other hand, GPR is more complex to implement and interpret, but provides explicit estimates of uncertainties. Our results indicated that foliar functional traits can be retrieved in these grasslands with moderate to high accuracies using either method. Highest validation accuracies were obtained for leaf mass per area (LMA), soluble cell contents, hemicellulose and cellulose (all with R2 > 0.8), and lower accuracies for lignin, nitrogen, and some pigments with both techniques. The estimations for the mass-based pigments were more accurate than the area-based pigments (at least 5% improvement in normalized RMSE). Overall, GPR and PLSR performed comparably with respect to both skill of predictions and the selection of most informative spectral regions. Maps of uncertainties corresponded well between the two models, with highest uncertainties related to low vegetation cover, high diversity levels, or under irrigation and nitrogen treatments (not represented in the field sampling). The maps showed that trait values in each plot were relatively stable across three years of managed species richness. Our results provide a template for mapping foliar traits and their uncertainties in grasslands, and point to the need for extensive ground data across time to properly evaluate performance of trait mapping algorithms.
Bibliographical noteFunding Information:
This work was supported by grant DEB-1342778 to P.A.T. and DEB-1342872 grant to J.C.B. and S.E.H through the NSF-NASA Dimensions of Biodiversity program, and by the Cedar Creek NSF Long-Term Ecological Research program ( DEB-1234162 ). We are grateful for the help of Brett Fredericksen, Ian Carriere, Cathleen Nguyen, Shan Kothari, Clayton Kingdon, Robi Phetteplace, Aidan Mazur, Melanie Sinnen with field data collection and chemical analyses, and for the support of Kally Worm and Troy Mielke and the entire CCESR staff. We thank the anonymous reviewers for their helpful comments.
- Gaussian processes regression
- Imaging spectroscopy
- Trait map