Speaker
Description
Relationships between plant functional traits and environmental variables are important to understand the reasons underlying patterns of vegetation characteristics and quantify potential changes due to climate change. However, plant functional traits have been predicted from environmental data only with moderate success and the reasons for this remain partly unclear.
Here, we test the prediction of 18 key plant functional traits from climate and soil variables using machine learning and examine differences between datasets and aggregation approaches and the role of plant functional types (PFTs). We use trait data directly from the global plant trait database TRY and from sPlotOpen. We found that for TRY and sPlotOpen, community weighted mean (CWM) traits could be predicted with R2 of about 0.4-0.7 and relative RMSE (rRMSE) of 0.05-1.5. Adding PFT information only increased R2 moderately overall but by more than 0.2 for individual PFTs and some traits. Adding PFT information had notably larger impacts on trait prediction when using PFT-level aggregated traits either at plot-level (sPlotOpen) or at grid-cell level (50 km) for both TRY and sPlotOpen data. The strongest predictability in terms or R2 and rRMSE was achieved when calculating grid-cell-level CWM traits with fractional PFT cover as weighting factors. This led to 0.9 > R2 > 0.6 and 0.05 < rRMSE < 0.4 for all traits when using PFT weighted mean sPlotOpen data and slightly worse performance when using TRY data. Our findings have important implications for upscaling in-situ plant traits to generate large-scale maps and indicate the relevance of PFTs to refine environment-based trait predictions.
| Status Group | Postdoctoral Researcher |
|---|
