A novel model to accurately predict continental-scale timing of forest green-up

Background/Question/Methods

: The yearly cycles in vegetation greenness are among the most important drivers of ecosystem processes. Predictive models for the timing of vegetation greenup and senescence are crucial for understanding how biological communities respond to global change. Greenup timing is closely tied to climate and also tracks yearly variability in temperature, and the strength of this relationship varies spatio-temporally. Local studies have been useful in understanding underlying mechanisms but they are insufficient in explaining larger scale variabilities. Large-scale studies using remotely-sensed data have the potential to harness regional dynamics, even if underlying mechanisms remain unknown, Yet predictive power using these approaches is low. Here, we predict vegetation phenology across Eastern North America via a novel class of Bayesian regression model. Our modeling framework provides continental-level peak greenup time predictions with high accuracy using satellite observations from the MODerate resolution Imaging Spectroradiometer (MODIS). In addition to taking into account temporal structure at individual sites, our models make use of information from the entire study extent regardless of their spatial proximity.

Results/Conclusions

: Models were built from 2000 to 2016 and showed high prediction accuracy (R2 > 95%). Out-of-sample predictions for the years 2017 and 2018 showed accuracy within days of the predicted peaks, even though yearly greenup timing can vary by up to 30 days across the study region. Performance was remarkably high across deciduous and mixed forest types. Our method is generalizable to temperate forests across the globe and provides a basis for backcasting and forecasting forest greenup for any time periods where daily temperatures, whether directly measured or modeled, are available.