Novel Data-driven Method for the Prediction of Remotely Sensed Leafing Phenology

Abstract:

Changes in plant phenology induced by climate change have important implications in conservation, agriculture, and public health and call for urgently needed improvements in the predictability of phenology. Previous models that assume linear relationships between the timing of events and “critical environmental cues” often fail to accurately predict phenology, as mechanisms of phenology are highly nonlinear. To better inform decision-making in a changing world, we forecast the nonlinear dynamics of leafing phenology with a state-of-the-art data-driven model, Gaussian process regression with spatio-temporal delay embedding (GP-SDM). Specifically, leafing phenology at each location was modeled as a function of past leafing phenology and environmental conditions in the neighborhood. We adopted a hierarchical model structure that allows this nonlinear function to be correlated across locations. After fitting the model with high-performance computing, we were able to accurately predict the continuous development of canopies characterized by Enhanced Vegetation Index from satellite remote sensing and that characterized by Green Chromatic Coordinate from near-surface digital imagery (PhenoCams) in the near term (one year). Evaluated with diagnostics including root mean squared prediction error and Pearson correlation coefficient, we show that the non-parametric GP-SDM outperformed traditional parametric methods (e.g., using growing degree-days) in predicting the leaf onset and senescence. In addition, the model allowed us to detect key environmental drivers of canopy development at different times and locations. Finally, we designed and implemented a sparse online GP-SDM algorithm for the iterative data assimilation. This project demonstrates a novel and promising approach to improve mechanistic understanding and the predictability of the Earth system.