Using Machine Learning to Reveal the Relationships Between Plant Functional Traits and Flux Regimes at Eddy-Covariance Towers

The current-generation of land surface models (LSMs) are powerful tools used in predictions of the future global climate and carbon cycle. Many of these LSMs are parametrised using plant functional types (PFTs), often of a coarse nature with only relatively few possible groups. In turn, extensive use of eddy-covariance data is utilised when calibrating these LSMs, with the model PFT matched to the classification reported by the site owners. Importantly, the PFT group is one of the few site characteristics that is consistently supplied across FLUXNET sites. However, there are issues with this method of LSM calibration. It is well-known that many PFT classification schemes cannot be predicted from climate, and that traits may vary more within species or sites than between them.

Here we present our results assessing the suitability of PFTs for capturing site flux regimes using a suite of machine learning techniques. We explore natural groupings of sites based on the measurements used for LSM calibration and identify potential site characteristics and traits that might allow these natural groupings to be predicted. Our results identify driving characteristics of site flux regime differences, and can be used to direct LSM development and highlight priority locations for future eddy-covariance flux towers.