Plant functional traits serve as tools to understand how plants as primary producers contribute – and react to changing environmental conditions. They also provide means to further our understanding of ecosystem functioning and the relationships between environmental conditions and vegetation patterns on different scales. In order to enable multidimensional analyses, efforts into establishing large databases of functional traits have increased with one database – the TRY project - being unique in its coverage and scale within the plant trait community. However, despite its unprecedented coverage - about 3 million records - complete information on all traits is still lacking for the majority of species within TRY. Another problem is the sparse distribution of trait measurements, especially in remote and tropical regions, which further limits spatially explicit analyses. Here we present two novel machine-learning approaches, which overcome these limitations: hierarchical and advanced hierarchical probabilistic matrix factorization (HPMF and aHPMF respectively).
Results/Conclusions
We show HPMFs’ high accuracy in filling the sparse TRY matrix, compared to using species mean (MEAN), a simple intuitive method frequently used in ecology to fill gaps in plant trait matrixes. HPMF-predicted traits have significantly lower root mean square errors on all taxonomic hierarchies, as well as across 13 structural and physiological traits than MEAN. Trait-trait correlations are better preserved using HPMF compared to MEAN, especially with respect to extreme values. Another advantage of HPMF is that a Gibbs sampler generates measures of uncertainty for each HPMF-generated trait prediction. We illustrate the potential of aHPMF to extrapolate from point measurements to larger spatial scales, using the example of foliar nitrogen concentration and specific leaf area predictions across the species range of Acer saccharum and Pinus sylvestris.
We discuss how our methods may serve as tools to fill sparse trait matrices and extrapolate trait measurements on different spatial scales. By providing measures of uncertainty for each predicted trait value, they indicate where to concentrate sampling efforts in order to decrease these uncertainties. This may aid in improving our understanding of multidimensional trait constellations as well as possible vegetation changes with future environmental change.